Merge pull request #6090 from neondatabase/releases/2023-12-11

Release 2023-12-11

.github/workflows/build_and_test.yml

@@ -199,6 +199,10 @@ jobs:
           #
           git config --global --add safe.directory ${{ github.workspace }}
           git config --global --add safe.directory ${GITHUB_WORKSPACE}
+          for r in 14 15 16; do
+            git config --global --add safe.directory "${{ github.workspace }}/vendor/postgres-v$r"
+            git config --global --add safe.directory "${GITHUB_WORKSPACE}/vendor/postgres-v$r"
+          done
 
       - name: Checkout
         uses: actions/checkout@v3
@@ -404,7 +408,7 @@ jobs:
         uses: ./.github/actions/save-coverage-data
 
   regress-tests:
-    needs: [ check-permissions, build-neon ]
+    needs: [ check-permissions, build-neon, tag ]
     runs-on: [ self-hosted, gen3, large ]
     container:
       image: 369495373322.dkr.ecr.eu-central-1.amazonaws.com/rust:pinned
@@ -436,6 +440,7 @@ jobs:
         env:
           TEST_RESULT_CONNSTR: ${{ secrets.REGRESS_TEST_RESULT_CONNSTR_NEW }}
           CHECK_ONDISK_DATA_COMPATIBILITY: nonempty
+          BUILD_TAG: ${{ needs.tag.outputs.build-tag }}
 
       - name: Merge and upload coverage data
         if: matrix.build_type == 'debug' && matrix.pg_version == 'v14'
@@ -1096,6 +1101,10 @@ jobs:
           #
           git config --global --add safe.directory ${{ github.workspace }}
           git config --global --add safe.directory ${GITHUB_WORKSPACE}
+          for r in 14 15 16; do
+            git config --global --add safe.directory "${{ github.workspace }}/vendor/postgres-v$r"
+            git config --global --add safe.directory "${GITHUB_WORKSPACE}/vendor/postgres-v$r"
+          done
 
       - name: Checkout
         uses: actions/checkout@v3

.github/workflows/neon_extra_builds.yml

@@ -142,6 +142,10 @@ jobs:
           #
           git config --global --add safe.directory ${{ github.workspace }}
           git config --global --add safe.directory ${GITHUB_WORKSPACE}
+          for r in 14 15 16; do
+            git config --global --add safe.directory "${{ github.workspace }}/vendor/postgres-v$r"
+            git config --global --add safe.directory "${GITHUB_WORKSPACE}/vendor/postgres-v$r"
+          done
 
       - name: Checkout
         uses: actions/checkout@v4
@@ -238,6 +242,20 @@ jobs:
       options: --init
 
     steps:
+      - name: Fix git ownership
+        run: |
+          # Workaround for `fatal: detected dubious ownership in repository at ...`
+          #
+          # Use both ${{ github.workspace }} and ${GITHUB_WORKSPACE} because they're different on host and in containers
+          #   Ref https://github.com/actions/checkout/issues/785
+          #
+          git config --global --add safe.directory ${{ github.workspace }}
+          git config --global --add safe.directory ${GITHUB_WORKSPACE}
+          for r in 14 15 16; do
+            git config --global --add safe.directory "${{ github.workspace }}/vendor/postgres-v$r"
+            git config --global --add safe.directory "${GITHUB_WORKSPACE}/vendor/postgres-v$r"
+          done
+
       - name: Checkout
         uses: actions/checkout@v4
         with:

.gitignore

@@ -18,3 +18,6 @@ test_output/
 *.o
 *.so
 *.Po
+
+# pgindent typedef lists
+*.list

Cargo.toml

@@ -45,12 +45,11 @@ azure_storage_blobs = "0.16"
 flate2 = "1.0.26"
 async-stream = "0.3"
 async-trait = "0.1"
-aws-config = { version = "0.56", default-features = false, features=["rustls"] }
-aws-sdk-s3 = "0.29"
-aws-smithy-http = "0.56"
-aws-smithy-async = { version = "0.56", default-features = false, features=["rt-tokio"] }
-aws-credential-types = "0.56"
-aws-types = "0.56"
+aws-config = { version = "1.0", default-features = false, features=["rustls"] }
+aws-sdk-s3 = "1.0"
+aws-smithy-async = { version = "1.0", default-features = false, features=["rt-tokio"] }
+aws-smithy-types = "1.0"
+aws-credential-types = "1.0"
 axum = { version = "0.6.20", features = ["ws"] }
 base64 = "0.13.0"
 bincode = "1.3"
@@ -89,6 +88,7 @@ humantime-serde = "1.1.1"
 hyper = "0.14"
 hyper-tungstenite = "0.11"
 inotify = "0.10.2"
+ipnet = "2.9.0"
 itertools = "0.10"
 jsonwebtoken = "8"
 libc = "0.2"
@@ -109,7 +109,7 @@ pin-project-lite = "0.2"
 prometheus = {version = "0.13", default_features=false, features = ["process"]} # removes protobuf dependency
 prost = "0.11"
 rand = "0.8"
-regex = "1.4"
+regex = "1.10.2"
 reqwest = { version = "0.11", default-features = false, features = ["rustls-tls"] }
 reqwest-tracing = { version = "0.4.0", features = ["opentelemetry_0_19"] }
 reqwest-middleware = "0.2.0"
@@ -122,14 +122,17 @@ rustls-pemfile = "1"
 rustls-split = "0.3"
 scopeguard = "1.1"
 sysinfo = "0.29.2"
+sd-notify = "0.4.1"
 sentry = { version = "0.31", default-features = false, features = ["backtrace", "contexts", "panic", "rustls", "reqwest" ] }
 serde = { version = "1.0", features = ["derive"] }
 serde_json = "1"
+serde_path_to_error = "0.1"
 serde_with = "2.0"
 serde_assert = "0.5.0"
 sha2 = "0.10.2"
 signal-hook = "0.3"
 smallvec = "1.11"
+smol_str = { version = "0.2.0", features = ["serde"] }
 socket2 = "0.5"
 strum = "0.24"
 strum_macros = "0.24"
@@ -146,7 +149,7 @@ tokio-postgres-rustls = "0.10.0"
 tokio-rustls = "0.24"
 tokio-stream = "0.1"
 tokio-tar = "0.3"
-tokio-util = { version = "0.7", features = ["io"] }
+tokio-util = { version = "0.7.10", features = ["io", "rt"] }
 toml = "0.7"
 toml_edit = "0.19"
 tonic = {version = "0.9", features = ["tls", "tls-roots"]}

Dockerfile.compute-node

@@ -393,7 +393,9 @@ RUN case "${PG_VERSION}" in \
         export TIMESCALEDB_CHECKSUM=6fca72a6ed0f6d32d2b3523951ede73dc5f9b0077b38450a029a5f411fdb8c73 \
         ;; \
       *) \
-        echo "TimescaleDB not supported on this PostgreSQL version. See https://github.com/timescale/timescaledb/issues/5752" && exit 0;; \
+        export TIMESCALEDB_VERSION=2.13.0 \
+        export TIMESCALEDB_CHECKSUM=584a351c7775f0e067eaa0e7277ea88cab9077cc4c455cbbf09a5d9723dce95d \
+        ;; \
     esac && \
     apt-get update && \
     apt-get install -y cmake && \
@@ -714,6 +716,23 @@ RUN wget https://github.com/pksunkara/pgx_ulid/archive/refs/tags/v0.1.3.tar.gz -
     cargo pgrx install --release && \
     echo "trusted = true" >> /usr/local/pgsql/share/extension/ulid.control
 
+#########################################################################################
+#
+# Layer "wal2json-build"
+# Compile "wal2json" extension
+#
+#########################################################################################
+
+FROM build-deps AS wal2json-pg-build
+COPY --from=pg-build /usr/local/pgsql/ /usr/local/pgsql/
+
+ENV PATH "/usr/local/pgsql/bin/:$PATH"
+RUN wget https://github.com/eulerto/wal2json/archive/refs/tags/wal2json_2_5.tar.gz && \
+    echo "b516653575541cf221b99cf3f8be9b6821f6dbcfc125675c85f35090f824f00e wal2json_2_5.tar.gz" | sha256sum --check && \
+    mkdir wal2json-src && cd wal2json-src && tar xvzf ../wal2json_2_5.tar.gz --strip-components=1 -C . && \
+    make -j $(getconf _NPROCESSORS_ONLN) && \
+    make -j $(getconf _NPROCESSORS_ONLN) install
+
 #########################################################################################
 #
 # Layer "neon-pg-ext-build"
@@ -750,6 +769,7 @@ COPY --from=rdkit-pg-build /usr/local/pgsql/ /usr/local/pgsql/
 COPY --from=pg-uuidv7-pg-build /usr/local/pgsql/ /usr/local/pgsql/
 COPY --from=pg-roaringbitmap-pg-build /usr/local/pgsql/ /usr/local/pgsql/
 COPY --from=pg-embedding-pg-build /usr/local/pgsql/ /usr/local/pgsql/
+COPY --from=wal2json-pg-build /usr/local/pgsql /usr/local/pgsql
 COPY pgxn/ pgxn/
 
 RUN make -j $(getconf _NPROCESSORS_ONLN) \

Makefile

@@ -260,6 +260,44 @@ distclean:
 fmt:
 	./pre-commit.py --fix-inplace
 
+postgres-%-pg-bsd-indent: postgres-%
+	+@echo "Compiling pg_bsd_indent"
+	$(MAKE) -C $(POSTGRES_INSTALL_DIR)/build/$*/src/tools/pg_bsd_indent/
+
+# Create typedef list for the core. Note that generally it should be combined with
+# buildfarm one to cover platform specific stuff.
+# https://wiki.postgresql.org/wiki/Running_pgindent_on_non-core_code_or_development_code
+postgres-%-typedefs.list: postgres-%
+	$(ROOT_PROJECT_DIR)/vendor/postgres-$*/src/tools/find_typedef $(POSTGRES_INSTALL_DIR)/$*/bin > $@
+
+# Indent postgres. See src/tools/pgindent/README for details.
+.PHONY: postgres-%-pgindent
+postgres-%-pgindent: postgres-%-pg-bsd-indent postgres-%-typedefs.list
+	+@echo merge with buildfarm typedef to cover all platforms
+	+@echo note: I first tried to download from pgbuildfarm.org, but for unclear reason e.g. \
+		REL_16_STABLE list misses PGSemaphoreData
+	# wget -q -O - "http://www.pgbuildfarm.org/cgi-bin/typedefs.pl?branch=REL_16_STABLE" |\
+	# cat - postgres-$*-typedefs.list | sort | uniq > postgres-$*-typedefs-full.list
+	cat $(ROOT_PROJECT_DIR)/vendor/postgres-$*/src/tools/pgindent/typedefs.list |\
+		cat - postgres-$*-typedefs.list | sort | uniq > postgres-$*-typedefs-full.list
+	+@echo note: you might want to run it on selected files/dirs instead.
+	INDENT=$(POSTGRES_INSTALL_DIR)/build/$*/src/tools/pg_bsd_indent/pg_bsd_indent \
+		$(ROOT_PROJECT_DIR)/vendor/postgres-$*/src/tools/pgindent/pgindent --typedefs postgres-$*-typedefs-full.list \
+		$(ROOT_PROJECT_DIR)/vendor/postgres-$*/src/ \
+		--excludes $(ROOT_PROJECT_DIR)/vendor/postgres-$*/src/tools/pgindent/exclude_file_patterns
+	rm -f pg*.BAK
+
+# Indent pxgn/neon.
+.PHONY: pgindent
+neon-pgindent: postgres-v16-pg-bsd-indent neon-pg-ext-v16
+	$(MAKE) PG_CONFIG=$(POSTGRES_INSTALL_DIR)/v16/bin/pg_config CFLAGS='$(PG_CFLAGS) $(COPT)' \
+		FIND_TYPEDEF=$(ROOT_PROJECT_DIR)/vendor/postgres-v16/src/tools/find_typedef \
+		INDENT=$(POSTGRES_INSTALL_DIR)/build/v16/src/tools/pg_bsd_indent/pg_bsd_indent \
+		PGINDENT_SCRIPT=$(ROOT_PROJECT_DIR)/vendor/postgres-v16/src/tools/pgindent/pgindent \
+		-C $(POSTGRES_INSTALL_DIR)/build/neon-v16 \
+		-f $(ROOT_PROJECT_DIR)/pgxn/neon/Makefile pgindent
+
+
 .PHONY: setup-pre-commit-hook
 setup-pre-commit-hook:
 	ln -s -f $(ROOT_PROJECT_DIR)/pre-commit.py .git/hooks/pre-commit

README.md

@@ -149,6 +149,9 @@ tenant 9ef87a5bf0d92544f6fafeeb3239695c successfully created on the pageserver
 Created an initial timeline 'de200bd42b49cc1814412c7e592dd6e9' at Lsn 0/16B5A50 for tenant: 9ef87a5bf0d92544f6fafeeb3239695c
 Setting tenant 9ef87a5bf0d92544f6fafeeb3239695c as a default one
 
+# create postgres compute node
+> cargo neon endpoint create main
+
 # start postgres compute node
 > cargo neon endpoint start main
 Starting new endpoint main (PostgreSQL v14) on timeline de200bd42b49cc1814412c7e592dd6e9 ...
@@ -185,8 +188,11 @@ Created timeline 'b3b863fa45fa9e57e615f9f2d944e601' at Lsn 0/16F9A00 for tenant:
 (L) main [de200bd42b49cc1814412c7e592dd6e9]
 (L) ┗━ @0/16F9A00: migration_check [b3b863fa45fa9e57e615f9f2d944e601]
 
+# create postgres on that branch
+> cargo neon endpoint create migration_check --branch-name migration_check
+
 # start postgres on that branch
-> cargo neon endpoint start migration_check --branch-name migration_check
+> cargo neon endpoint start migration_check
 Starting new endpoint migration_check (PostgreSQL v14) on timeline b3b863fa45fa9e57e615f9f2d944e601 ...
 Starting postgres at 'postgresql://cloud_admin@127.0.0.1:55434/postgres'
 

compute_tools/Cargo.toml

@@ -38,3 +38,4 @@ toml_edit.workspace = true
 remote_storage = { version = "0.1", path = "../libs/remote_storage/" }
 vm_monitor = { version = "0.1", path = "../libs/vm_monitor/" }
 zstd = "0.12.4"
+bytes = "1.0"

compute_tools/src/bin/compute_ctl.rs

@@ -31,7 +31,7 @@
 //!             -C 'postgresql://cloud_admin@localhost/postgres' \
 //!             -S /var/db/postgres/specs/current.json \
 //!             -b /usr/local/bin/postgres \
-//!             -r {"bucket": "neon-dev-extensions-eu-central-1", "region": "eu-central-1"}
+//!             -r http://pg-ext-s3-gateway
 //! ```
 //!
 use std::collections::HashMap;
@@ -51,7 +51,7 @@ use compute_api::responses::ComputeStatus;
 
 use compute_tools::compute::{ComputeNode, ComputeState, ParsedSpec};
 use compute_tools::configurator::launch_configurator;
-use compute_tools::extension_server::{get_pg_version, init_remote_storage};
+use compute_tools::extension_server::get_pg_version;
 use compute_tools::http::api::launch_http_server;
 use compute_tools::logger::*;
 use compute_tools::monitor::launch_monitor;
@@ -60,7 +60,7 @@ use compute_tools::spec::*;
 
 // this is an arbitrary build tag. Fine as a default / for testing purposes
 // in-case of not-set environment var
-const BUILD_TAG_DEFAULT: &str = "5670669815";
+const BUILD_TAG_DEFAULT: &str = "latest";
 
 fn main() -> Result<()> {
     init_tracing_and_logging(DEFAULT_LOG_LEVEL)?;
@@ -74,10 +74,18 @@ fn main() -> Result<()> {
     let pgbin_default = String::from("postgres");
     let pgbin = matches.get_one::<String>("pgbin").unwrap_or(&pgbin_default);
 
-    let remote_ext_config = matches.get_one::<String>("remote-ext-config");
-    let ext_remote_storage = remote_ext_config.map(|x| {
-        init_remote_storage(x).expect("cannot initialize remote extension storage from config")
-    });
+    let ext_remote_storage = matches
+        .get_one::<String>("remote-ext-config")
+        // Compatibility hack: if the control plane specified any remote-ext-config
+        // use the default value for extension storage proxy gateway.
+        // Remove this once the control plane is updated to pass the gateway URL
+        .map(|conf| {
+            if conf.starts_with("http") {
+                conf.trim_end_matches('/')
+            } else {
+                "http://pg-ext-s3-gateway"
+            }
+        });
 
     let http_port = *matches
         .get_one::<u16>("http-port")
@@ -198,7 +206,7 @@ fn main() -> Result<()> {
         live_config_allowed,
         state: Mutex::new(new_state),
         state_changed: Condvar::new(),
-        ext_remote_storage,
+        ext_remote_storage: ext_remote_storage.map(|s| s.to_string()),
         ext_download_progress: RwLock::new(HashMap::new()),
         build_tag,
     };
@@ -266,7 +274,13 @@ fn main() -> Result<()> {
             let mut state = compute.state.lock().unwrap();
             state.error = Some(format!("{:?}", err));
             state.status = ComputeStatus::Failed;
-            drop(state);
+            // Notify others that Postgres failed to start. In case of configuring the
+            // empty compute, it's likely that API handler is still waiting for compute
+            // state change. With this we will notify it that compute is in Failed state,
+            // so control plane will know about it earlier and record proper error instead
+            // of timeout.
+            compute.state_changed.notify_all();
+            drop(state); // unlock
             delay_exit = true;
             None
         }

compute_tools/src/compute.rs

@@ -22,10 +22,10 @@ use utils::id::{TenantId, TimelineId};
 use utils::lsn::Lsn;
 
 use compute_api::responses::{ComputeMetrics, ComputeStatus};
-use compute_api::spec::{ComputeMode, ComputeSpec};
+use compute_api::spec::{ComputeFeature, ComputeMode, ComputeSpec};
 use utils::measured_stream::MeasuredReader;
 
-use remote_storage::{DownloadError, GenericRemoteStorage, RemotePath};
+use remote_storage::{DownloadError, RemotePath};
 
 use crate::checker::create_availability_check_data;
 use crate::pg_helpers::*;
@@ -59,8 +59,8 @@ pub struct ComputeNode {
     pub state: Mutex<ComputeState>,
     /// `Condvar` to allow notifying waiters about state changes.
     pub state_changed: Condvar,
-    ///  the S3 bucket that we search for extensions in
-    pub ext_remote_storage: Option<GenericRemoteStorage>,
+    /// the address of extension storage proxy gateway
+    pub ext_remote_storage: Option<String>,
     // key: ext_archive_name, value: started download time, download_completed?
     pub ext_download_progress: RwLock<HashMap<String, (DateTime<Utc>, bool)>>,
     pub build_tag: String,
@@ -252,7 +252,7 @@ fn create_neon_superuser(spec: &ComputeSpec, client: &mut Client) -> Result<()>
                     IF NOT EXISTS (
                         SELECT FROM pg_catalog.pg_roles WHERE rolname = 'neon_superuser')
                     THEN
-                        CREATE ROLE neon_superuser CREATEDB CREATEROLE NOLOGIN REPLICATION IN ROLE pg_read_all_data, pg_write_all_data;
+                        CREATE ROLE neon_superuser CREATEDB CREATEROLE NOLOGIN REPLICATION BYPASSRLS IN ROLE pg_read_all_data, pg_write_all_data;
                         IF array_length(roles, 1) IS NOT NULL THEN
                             EXECUTE format('GRANT neon_superuser TO %s',
                                            array_to_string(ARRAY(SELECT quote_ident(x) FROM unnest(roles) as x), ', '));
@@ -277,6 +277,17 @@ fn create_neon_superuser(spec: &ComputeSpec, client: &mut Client) -> Result<()>
 }
 
 impl ComputeNode {
+    /// Check that compute node has corresponding feature enabled.
+    pub fn has_feature(&self, feature: ComputeFeature) -> bool {
+        let state = self.state.lock().unwrap();
+
+        if let Some(s) = state.pspec.as_ref() {
+            s.spec.features.contains(&feature)
+        } else {
+            false
+        }
+    }
+
     pub fn set_status(&self, status: ComputeStatus) {
         let mut state = self.state.lock().unwrap();
         state.status = status;
@@ -728,7 +739,12 @@ impl ComputeNode {
 
         // Write new config
         let pgdata_path = Path::new(&self.pgdata);
-        config::write_postgres_conf(&pgdata_path.join("postgresql.conf"), &spec, None)?;
+        let postgresql_conf_path = pgdata_path.join("postgresql.conf");
+        config::write_postgres_conf(&postgresql_conf_path, &spec, None)?;
+        // temporarily reset max_cluster_size in config
+        // to avoid the possibility of hitting the limit, while we are reconfiguring:
+        // creating new extensions, roles, etc...
+        config::compute_ctl_temp_override_create(pgdata_path, "neon.max_cluster_size=-1")?;
         self.pg_reload_conf()?;
 
         let mut client = Client::connect(self.connstr.as_str(), NoTls)?;
@@ -749,6 +765,10 @@ impl ComputeNode {
         // 'Close' connection
         drop(client);
 
+        // reset max_cluster_size in config back to original value and reload config
+        config::compute_ctl_temp_override_remove(pgdata_path)?;
+        self.pg_reload_conf()?;
+
         let unknown_op = "unknown".to_string();
         let op_id = spec.operation_uuid.as_ref().unwrap_or(&unknown_op);
         info!(
@@ -809,7 +829,17 @@ impl ComputeNode {
 
         let config_time = Utc::now();
         if pspec.spec.mode == ComputeMode::Primary && !pspec.spec.skip_pg_catalog_updates {
+            let pgdata_path = Path::new(&self.pgdata);
+            // temporarily reset max_cluster_size in config
+            // to avoid the possibility of hitting the limit, while we are applying config:
+            // creating new extensions, roles, etc...
+            config::compute_ctl_temp_override_create(pgdata_path, "neon.max_cluster_size=-1")?;
+            self.pg_reload_conf()?;
+
             self.apply_config(&compute_state)?;
+
+            config::compute_ctl_temp_override_remove(pgdata_path)?;
+            self.pg_reload_conf()?;
         }
 
         let startup_end_time = Utc::now();
@@ -957,12 +987,12 @@ LIMIT 100",
         real_ext_name: String,
         ext_path: RemotePath,
     ) -> Result<u64, DownloadError> {
-        let remote_storage = self
-            .ext_remote_storage
-            .as_ref()
-            .ok_or(DownloadError::BadInput(anyhow::anyhow!(
-                "Remote extensions storage is not configured",
-            )))?;
+        let ext_remote_storage =
+            self.ext_remote_storage
+                .as_ref()
+                .ok_or(DownloadError::BadInput(anyhow::anyhow!(
+                    "Remote extensions storage is not configured",
+                )))?;
 
         let ext_archive_name = ext_path.object_name().expect("bad path");
 
@@ -1018,7 +1048,7 @@ LIMIT 100",
         let download_size = extension_server::download_extension(
             &real_ext_name,
             &ext_path,
-            remote_storage,
+            ext_remote_storage,
             &self.pgbin,
         )
         .await

compute_tools/src/config.rs

@@ -93,5 +93,25 @@ pub fn write_postgres_conf(
         writeln!(file, "neon.extension_server_port={}", port)?;
     }
 
+    // This is essential to keep this line at the end of the file,
+    // because it is intended to override any settings above.
+    writeln!(file, "include_if_exists = 'compute_ctl_temp_override.conf'")?;
+
+    Ok(())
+}
+
+/// create file compute_ctl_temp_override.conf in pgdata_dir
+/// add provided options to this file
+pub fn compute_ctl_temp_override_create(pgdata_path: &Path, options: &str) -> Result<()> {
+    let path = pgdata_path.join("compute_ctl_temp_override.conf");
+    let mut file = File::create(path)?;
+    write!(file, "{}", options)?;
+    Ok(())
+}
+
+/// remove file compute_ctl_temp_override.conf in pgdata_dir
+pub fn compute_ctl_temp_override_remove(pgdata_path: &Path) -> Result<()> {
+    let path = pgdata_path.join("compute_ctl_temp_override.conf");
+    std::fs::remove_file(path)?;
     Ok(())
 }

compute_tools/src/extension_server.rs

@@ -71,18 +71,16 @@ More specifically, here is an example ext_index.json
     }
 }
 */
-use anyhow::Context;
 use anyhow::{self, Result};
+use anyhow::{bail, Context};
+use bytes::Bytes;
 use compute_api::spec::RemoteExtSpec;
 use regex::Regex;
 use remote_storage::*;
-use serde_json;
-use std::io::Read;
-use std::num::NonZeroUsize;
+use reqwest::StatusCode;
 use std::path::Path;
 use std::str;
 use tar::Archive;
-use tokio::io::AsyncReadExt;
 use tracing::info;
 use tracing::log::warn;
 use zstd::stream::read::Decoder;
@@ -138,23 +136,31 @@ fn parse_pg_version(human_version: &str) -> &str {
 pub async fn download_extension(
     ext_name: &str,
     ext_path: &RemotePath,
-    remote_storage: &GenericRemoteStorage,
+    ext_remote_storage: &str,
     pgbin: &str,
 ) -> Result<u64> {
     info!("Download extension {:?} from {:?}", ext_name, ext_path);
-    let mut download = remote_storage.download(ext_path).await?;
-    let mut download_buffer = Vec::new();
-    download
-        .download_stream
-        .read_to_end(&mut download_buffer)
-        .await?;
+
+    // TODO add retry logic
+    let download_buffer =
+        match download_extension_tar(ext_remote_storage, &ext_path.to_string()).await {
+            Ok(buffer) => buffer,
+            Err(error_message) => {
+                return Err(anyhow::anyhow!(
+                    "error downloading extension {:?}: {:?}",
+                    ext_name,
+                    error_message
+                ));
+            }
+        };
+
     let download_size = download_buffer.len() as u64;
+    info!("Download size {:?}", download_size);
     // it's unclear whether it is more performant to decompress into memory or not
     // TODO: decompressing into memory can be avoided
-    let mut decoder = Decoder::new(download_buffer.as_slice())?;
-    let mut decompress_buffer = Vec::new();
-    decoder.read_to_end(&mut decompress_buffer)?;
-    let mut archive = Archive::new(decompress_buffer.as_slice());
+    let decoder = Decoder::new(download_buffer.as_ref())?;
+    let mut archive = Archive::new(decoder);
+
     let unzip_dest = pgbin
         .strip_suffix("/bin/postgres")
         .expect("bad pgbin")
@@ -222,29 +228,32 @@ pub fn create_control_files(remote_extensions: &RemoteExtSpec, pgbin: &str) {
     }
 }
 
-// This function initializes the necessary structs to use remote storage
-pub fn init_remote_storage(remote_ext_config: &str) -> anyhow::Result<GenericRemoteStorage> {
-    #[derive(Debug, serde::Deserialize)]
-    struct RemoteExtJson {
-        bucket: String,
-        region: String,
-        endpoint: Option<String>,
-        prefix: Option<String>,
-    }
-    let remote_ext_json = serde_json::from_str::<RemoteExtJson>(remote_ext_config)?;
+// Do request to extension storage proxy, i.e.
+// curl http://pg-ext-s3-gateway/latest/v15/extensions/anon.tar.zst
+// using HHTP GET
+// and return the response body as bytes
+//
+async fn download_extension_tar(ext_remote_storage: &str, ext_path: &str) -> Result<Bytes> {
+    let uri = format!("{}/{}", ext_remote_storage, ext_path);
 
-    let config = S3Config {
-        bucket_name: remote_ext_json.bucket,
-        bucket_region: remote_ext_json.region,
-        prefix_in_bucket: remote_ext_json.prefix,
-        endpoint: remote_ext_json.endpoint,
-        concurrency_limit: NonZeroUsize::new(100).expect("100 != 0"),
-        max_keys_per_list_response: None,
-    };
-    let config = RemoteStorageConfig {
-        storage: RemoteStorageKind::AwsS3(config),
-    };
-    GenericRemoteStorage::from_config(&config)
+    info!("Download extension {:?} from uri {:?}", ext_path, uri);
+
+    let resp = reqwest::get(uri).await?;
+
+    match resp.status() {
+        StatusCode::OK => match resp.bytes().await {
+            Ok(resp) => {
+                info!("Download extension {:?} completed successfully", ext_path);
+                Ok(resp)
+            }
+            Err(e) => bail!("could not deserialize remote extension response: {}", e),
+        },
+        StatusCode::SERVICE_UNAVAILABLE => bail!("remote extension is temporarily unavailable"),
+        _ => bail!(
+            "unexpected remote extension response status code: {}",
+            resp.status()
+        ),
+    }
 }
 
 #[cfg(test)]

compute_tools/src/http/api.rs

@@ -123,7 +123,7 @@ async fn routes(req: Request<Body>, compute: &Arc<ComputeNode>) -> Response<Body
             }
         }
 
-        // download extension files from S3 on demand
+        // download extension files from remote extension storage on demand
         (&Method::POST, route) if route.starts_with("/extension_server/") => {
             info!("serving {:?} POST request", route);
             info!("req.uri {:?}", req.uri());
@@ -227,7 +227,7 @@ async fn handle_configure_request(
 
         let parsed_spec = match ParsedSpec::try_from(spec) {
             Ok(ps) => ps,
-            Err(msg) => return Err((msg, StatusCode::PRECONDITION_FAILED)),
+            Err(msg) => return Err((msg, StatusCode::BAD_REQUEST)),
         };
 
         // XXX: wrap state update under lock in code blocks. Otherwise,

compute_tools/src/http/openapi_spec.yaml

@@ -156,17 +156,17 @@ paths:
                 description: Error text or 'OK' if download succeeded.
                 example: "OK"
         400:
-        description: Request is invalid.
-        content:
-          application/json:
-            schema:
-              $ref: "#/components/schemas/GenericError"
+          description: Request is invalid.
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/GenericError"
         500:
-        description: Extension download request failed.
-        content:
-          application/json:
-            schema:
-              $ref: "#/components/schemas/GenericError"
+          description: Extension download request failed.
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/GenericError"
 
 components:
   securitySchemes:

compute_tools/src/pg_helpers.rs

@@ -193,16 +193,11 @@ impl Escaping for PgIdent {
 /// Build a list of existing Postgres roles
 pub fn get_existing_roles(xact: &mut Transaction<'_>) -> Result<Vec<Role>> {
     let postgres_roles = xact
-        .query(
-            "SELECT rolname, rolpassword, rolreplication, rolbypassrls FROM pg_catalog.pg_authid",
-            &[],
-        )?
+        .query("SELECT rolname, rolpassword FROM pg_catalog.pg_authid", &[])?
         .iter()
         .map(|row| Role {
             name: row.get("rolname"),
             encrypted_password: row.get("rolpassword"),
-            replication: Some(row.get("rolreplication")),
-            bypassrls: Some(row.get("rolbypassrls")),
             options: None,
         })
         .collect();

compute_tools/src/spec.rs

@@ -118,19 +118,6 @@ pub fn get_spec_from_control_plane(
     spec
 }
 
-/// It takes cluster specification and does the following:
-/// - Serialize cluster config and put it into `postgresql.conf` completely rewriting the file.
-/// - Update `pg_hba.conf` to allow external connections.
-pub fn handle_configuration(spec: &ComputeSpec, pgdata_path: &Path) -> Result<()> {
-    // File `postgresql.conf` is no longer included into `basebackup`, so just
-    // always write all config into it creating new file.
-    config::write_postgres_conf(&pgdata_path.join("postgresql.conf"), spec, None)?;
-
-    update_pg_hba(pgdata_path)?;
-
-    Ok(())
-}
-
 /// Check `pg_hba.conf` and update if needed to allow external connections.
 pub fn update_pg_hba(pgdata_path: &Path) -> Result<()> {
     // XXX: consider making it a part of spec.json
@@ -265,8 +252,6 @@ pub fn handle_roles(spec: &ComputeSpec, client: &mut Client) -> Result<()> {
         let action = if let Some(r) = pg_role {
             if (r.encrypted_password.is_none() && role.encrypted_password.is_some())
                 || (r.encrypted_password.is_some() && role.encrypted_password.is_none())
-                || !r.bypassrls.unwrap_or(false)
-                || !r.replication.unwrap_or(false)
             {
                 RoleAction::Update
             } else if let Some(pg_pwd) = &r.encrypted_password {
@@ -298,14 +283,22 @@ pub fn handle_roles(spec: &ComputeSpec, client: &mut Client) -> Result<()> {
         match action {
             RoleAction::None => {}
             RoleAction::Update => {
-                let mut query: String =
-                    format!("ALTER ROLE {} BYPASSRLS REPLICATION", name.pg_quote());
+                // This can be run on /every/ role! Not just ones created through the console.
+                // This means that if you add some funny ALTER here that adds a permission,
+                // this will get run even on user-created roles! This will result in different
+                // behavior before and after a spec gets reapplied. The below ALTER as it stands
+                // now only grants LOGIN and changes the password. Please do not allow this branch
+                // to do anything silly.
+                let mut query: String = format!("ALTER ROLE {} ", name.pg_quote());
                 query.push_str(&role.to_pg_options());
                 xact.execute(query.as_str(), &[])?;
             }
             RoleAction::Create => {
+                // This branch only runs when roles are created through the console, so it is
+                // safe to add more permissions here. BYPASSRLS and REPLICATION are inherited
+                // from neon_superuser.
                 let mut query: String = format!(
-                    "CREATE ROLE {} CREATEROLE CREATEDB BYPASSRLS REPLICATION IN ROLE neon_superuser",
+                    "CREATE ROLE {} INHERIT CREATEROLE CREATEDB IN ROLE neon_superuser",
                     name.pg_quote()
                 );
                 info!("role create query: '{}'", &query);
@@ -687,6 +680,9 @@ pub fn handle_extension_neon(client: &mut Client) -> Result<()> {
     info!("create neon extension with query: {}", query);
     client.simple_query(query)?;
 
+    query = "UPDATE pg_extension SET extrelocatable = true WHERE extname = 'neon'";
+    client.simple_query(query)?;
+
     query = "ALTER EXTENSION neon SET SCHEMA neon";
     info!("alter neon extension schema with query: {}", query);
     client.simple_query(query)?;

control_plane/src/bin/attachment_service.rs

@@ -9,6 +9,7 @@ use clap::Parser;
 use hex::FromHex;
 use hyper::StatusCode;
 use hyper::{Body, Request, Response};
+use pageserver_api::shard::TenantShardId;
 use serde::{Deserialize, Serialize};
 use std::path::{Path, PathBuf};
 use std::{collections::HashMap, sync::Arc};
@@ -173,7 +174,8 @@ async fn handle_re_attach(mut req: Request<Body>) -> Result<Response<Body>, ApiE
         if state.pageserver == Some(reattach_req.node_id) {
             state.generation += 1;
             response.tenants.push(ReAttachResponseTenant {
-                id: *t,
+                // TODO(sharding): make this shard-aware
+                id: TenantShardId::unsharded(*t),
                 gen: state.generation,
             });
         }
@@ -196,8 +198,15 @@ async fn handle_validate(mut req: Request<Body>) -> Result<Response<Body>, ApiEr
     };
 
     for req_tenant in validate_req.tenants {
-        if let Some(tenant_state) = locked.tenants.get(&req_tenant.id) {
+        // TODO(sharding): make this shard-aware
+        if let Some(tenant_state) = locked.tenants.get(&req_tenant.id.tenant_id) {
             let valid = tenant_state.generation == req_tenant.gen;
+            tracing::info!(
+                "handle_validate: {}(gen {}): valid={valid} (latest {})",
+                req_tenant.id,
+                req_tenant.gen,
+                tenant_state.generation
+            );
             response.tenants.push(ValidateResponseTenant {
                 id: req_tenant.id,
                 valid,
@@ -247,6 +256,13 @@ async fn handle_attach_hook(mut req: Request<Body>) -> Result<Response<Body>, Ap
     tenant_state.pageserver = attach_req.node_id;
     let generation = tenant_state.generation;
 
+    tracing::info!(
+        "handle_attach_hook: tenant {} set generation {}, pageserver {}",
+        attach_req.tenant_id,
+        tenant_state.generation,
+        attach_req.node_id.unwrap_or(utils::id::NodeId(0xfffffff))
+    );
+
     locked.save().await.map_err(ApiError::InternalServerError)?;
 
     json_response(

control_plane/src/bin/neon_local.rs

@@ -415,6 +415,7 @@ fn handle_tenant(tenant_match: &ArgMatches, env: &mut local_env::LocalEnv) -> an
                 None,
                 None,
                 Some(pg_version),
+                None,
             )?;
             let new_timeline_id = timeline_info.timeline_id;
             let last_record_lsn = timeline_info.last_record_lsn;
@@ -495,6 +496,7 @@ fn handle_timeline(timeline_match: &ArgMatches, env: &mut local_env::LocalEnv) -
                 None,
                 None,
                 Some(pg_version),
+                None,
             )?;
             let new_timeline_id = timeline_info.timeline_id;
 
@@ -582,6 +584,7 @@ fn handle_timeline(timeline_match: &ArgMatches, env: &mut local_env::LocalEnv) -
                 start_lsn,
                 Some(ancestor_timeline_id),
                 None,
+                None,
             )?;
             let new_timeline_id = timeline_info.timeline_id;
 
@@ -608,11 +611,9 @@ fn handle_endpoint(ep_match: &ArgMatches, env: &local_env::LocalEnv) -> Result<(
     };
     let mut cplane = ComputeControlPlane::load(env.clone())?;
 
-    // All subcommands take an optional --tenant-id option
-    let tenant_id = get_tenant_id(sub_args, env)?;
-
     match sub_name {
         "list" => {
+            let tenant_id = get_tenant_id(sub_args, env)?;
             let timeline_infos = get_timeline_infos(env, &tenant_id).unwrap_or_else(|e| {
                 eprintln!("Failed to load timeline info: {}", e);
                 HashMap::new()
@@ -672,6 +673,7 @@ fn handle_endpoint(ep_match: &ArgMatches, env: &local_env::LocalEnv) -> Result<(
             println!("{table}");
         }
         "create" => {
+            let tenant_id = get_tenant_id(sub_args, env)?;
             let branch_name = sub_args
                 .get_one::<String>("branch-name")
                 .map(|s| s.as_str())
@@ -716,6 +718,18 @@ fn handle_endpoint(ep_match: &ArgMatches, env: &local_env::LocalEnv) -> Result<(
                 (Some(_), true) => anyhow::bail!("cannot specify both lsn and hot-standby"),
             };
 
+            match (mode, hot_standby) {
+                (ComputeMode::Static(_), true) => {
+                    bail!("Cannot start a node in hot standby mode when it is already configured as a static replica")
+                }
+                (ComputeMode::Primary, true) => {
+                    bail!("Cannot start a node as a hot standby replica, it is already configured as primary node")
+                }
+                _ => {}
+            }
+
+            cplane.check_conflicting_endpoints(mode, tenant_id, timeline_id)?;
+
             cplane.new_endpoint(
                 &endpoint_id,
                 tenant_id,
@@ -728,8 +742,6 @@ fn handle_endpoint(ep_match: &ArgMatches, env: &local_env::LocalEnv) -> Result<(
             )?;
         }
         "start" => {
-            let pg_port: Option<u16> = sub_args.get_one::<u16>("pg-port").copied();
-            let http_port: Option<u16> = sub_args.get_one::<u16>("http-port").copied();
             let endpoint_id = sub_args
                 .get_one::<String>("endpoint_id")
                 .ok_or_else(|| anyhow!("No endpoint ID was provided to start"))?;
@@ -758,80 +770,28 @@ fn handle_endpoint(ep_match: &ArgMatches, env: &local_env::LocalEnv) -> Result<(
                     env.safekeepers.iter().map(|sk| sk.id).collect()
                 };
 
-            let endpoint = cplane.endpoints.get(endpoint_id.as_str());
+            let endpoint = cplane
+                .endpoints
+                .get(endpoint_id.as_str())
+                .ok_or_else(|| anyhow::anyhow!("endpoint {endpoint_id} not found"))?;
+
+            cplane.check_conflicting_endpoints(
+                endpoint.mode,
+                endpoint.tenant_id,
+                endpoint.timeline_id,
+            )?;
 
             let ps_conf = env.get_pageserver_conf(pageserver_id)?;
             let auth_token = if matches!(ps_conf.pg_auth_type, AuthType::NeonJWT) {
-                let claims = Claims::new(Some(tenant_id), Scope::Tenant);
+                let claims = Claims::new(Some(endpoint.tenant_id), Scope::Tenant);
 
                 Some(env.generate_auth_token(&claims)?)
             } else {
                 None
             };
 
-            let hot_standby = sub_args
-                .get_one::<bool>("hot-standby")
-                .copied()
-                .unwrap_or(false);
-
-            if let Some(endpoint) = endpoint {
-                match (&endpoint.mode, hot_standby) {
-                    (ComputeMode::Static(_), true) => {
-                        bail!("Cannot start a node in hot standby mode when it is already configured as a static replica")
-                    }
-                    (ComputeMode::Primary, true) => {
-                        bail!("Cannot start a node as a hot standby replica, it is already configured as primary node")
-                    }
-                    _ => {}
-                }
-                println!("Starting existing endpoint {endpoint_id}...");
-                endpoint.start(&auth_token, safekeepers, remote_ext_config)?;
-            } else {
-                let branch_name = sub_args
-                    .get_one::<String>("branch-name")
-                    .map(|s| s.as_str())
-                    .unwrap_or(DEFAULT_BRANCH_NAME);
-                let timeline_id = env
-                    .get_branch_timeline_id(branch_name, tenant_id)
-                    .ok_or_else(|| {
-                        anyhow!("Found no timeline id for branch name '{branch_name}'")
-                    })?;
-                let lsn = sub_args
-                    .get_one::<String>("lsn")
-                    .map(|lsn_str| Lsn::from_str(lsn_str))
-                    .transpose()
-                    .context("Failed to parse Lsn from the request")?;
-                let pg_version = sub_args
-                    .get_one::<u32>("pg-version")
-                    .copied()
-                    .context("Failed to `pg-version` from the argument string")?;
-
-                let mode = match (lsn, hot_standby) {
-                    (Some(lsn), false) => ComputeMode::Static(lsn),
-                    (None, true) => ComputeMode::Replica,
-                    (None, false) => ComputeMode::Primary,
-                    (Some(_), true) => anyhow::bail!("cannot specify both lsn and hot-standby"),
-                };
-
-                // when used with custom port this results in non obvious behaviour
-                // port is remembered from first start command, i e
-                // start --port X
-                // stop
-                // start <-- will also use port X even without explicit port argument
-                println!("Starting new endpoint {endpoint_id} (PostgreSQL v{pg_version}) on timeline {timeline_id} ...");
-
-                let ep = cplane.new_endpoint(
-                    endpoint_id,
-                    tenant_id,
-                    timeline_id,
-                    pg_port,
-                    http_port,
-                    pg_version,
-                    mode,
-                    pageserver_id,
-                )?;
-                ep.start(&auth_token, safekeepers, remote_ext_config)?;
-            }
+            println!("Starting existing endpoint {endpoint_id}...");
+            endpoint.start(&auth_token, safekeepers, remote_ext_config)?;
         }
         "reconfigure" => {
             let endpoint_id = sub_args
@@ -1252,7 +1212,7 @@ fn cli() -> Command {
     let remote_ext_config_args = Arg::new("remote-ext-config")
         .long("remote-ext-config")
         .num_args(1)
-        .help("Configure the S3 bucket that we search for extensions in.")
+        .help("Configure the remote extensions storage proxy gateway to request for extensions.")
         .required(false);
 
     let lsn_arg = Arg::new("lsn")
@@ -1437,15 +1397,7 @@ fn cli() -> Command {
                 .subcommand(Command::new("start")
                     .about("Start postgres.\n If the endpoint doesn't exist yet, it is created.")
                     .arg(endpoint_id_arg.clone())
-                    .arg(tenant_id_arg.clone())
-                    .arg(branch_name_arg.clone())
-                    .arg(timeline_id_arg.clone())
-                    .arg(lsn_arg)
-                    .arg(pg_port_arg)
-                    .arg(http_port_arg)
                     .arg(endpoint_pageserver_id_arg.clone())
-                    .arg(pg_version_arg)
-                    .arg(hot_standby_arg)
                     .arg(safekeepers_arg)
                     .arg(remote_ext_config_args)
                 )
@@ -1458,7 +1410,6 @@ fn cli() -> Command {
                 .subcommand(
                     Command::new("stop")
                     .arg(endpoint_id_arg)
-                    .arg(tenant_id_arg.clone())
                     .arg(
                         Arg::new("destroy")
                             .help("Also delete data directory (now optional, should be default in future)")

control_plane/src/endpoint.rs

@@ -45,6 +45,7 @@ use std::sync::Arc;
 use std::time::Duration;
 
 use anyhow::{anyhow, bail, Context, Result};
+use compute_api::spec::RemoteExtSpec;
 use serde::{Deserialize, Serialize};
 use utils::id::{NodeId, TenantId, TimelineId};
 
@@ -124,6 +125,7 @@ impl ComputeControlPlane {
         let http_port = http_port.unwrap_or_else(|| self.get_port() + 1);
         let pageserver =
             PageServerNode::from_env(&self.env, self.env.get_pageserver_conf(pageserver_id)?);
+
         let ep = Arc::new(Endpoint {
             endpoint_id: endpoint_id.to_owned(),
             pg_address: SocketAddr::new("127.0.0.1".parse().unwrap(), pg_port),
@@ -168,6 +170,30 @@ impl ComputeControlPlane {
 
         Ok(ep)
     }
+
+    pub fn check_conflicting_endpoints(
+        &self,
+        mode: ComputeMode,
+        tenant_id: TenantId,
+        timeline_id: TimelineId,
+    ) -> Result<()> {
+        if matches!(mode, ComputeMode::Primary) {
+            // this check is not complete, as you could have a concurrent attempt at
+            // creating another primary, both reading the state before checking it here,
+            // but it's better than nothing.
+            let mut duplicates = self.endpoints.iter().filter(|(_k, v)| {
+                v.tenant_id == tenant_id
+                    && v.timeline_id == timeline_id
+                    && v.mode == mode
+                    && v.status() != "stopped"
+            });
+
+            if let Some((key, _)) = duplicates.next() {
+                bail!("attempting to create a duplicate primary endpoint on tenant {tenant_id}, timeline {timeline_id}: endpoint {key:?} exists already. please don't do this, it is not supported.");
+            }
+        }
+        Ok(())
+    }
 }
 
 ///////////////////////////////////////////////////////////////////////////////
@@ -476,11 +502,24 @@ impl Endpoint {
             }
         }
 
+        // check for file remote_extensions_spec.json
+        // if it is present, read it and pass to compute_ctl
+        let remote_extensions_spec_path = self.endpoint_path().join("remote_extensions_spec.json");
+        let remote_extensions_spec = std::fs::File::open(remote_extensions_spec_path);
+        let remote_extensions: Option<RemoteExtSpec>;
+
+        if let Ok(spec_file) = remote_extensions_spec {
+            remote_extensions = serde_json::from_reader(spec_file).ok();
+        } else {
+            remote_extensions = None;
+        };
+
         // Create spec file
         let spec = ComputeSpec {
             skip_pg_catalog_updates: self.skip_pg_catalog_updates,
             format_version: 1.0,
             operation_uuid: None,
+            features: vec![],
             cluster: Cluster {
                 cluster_id: None, // project ID: not used
                 name: None,       // project name: not used
@@ -497,7 +536,7 @@ impl Endpoint {
             pageserver_connstring: Some(pageserver_connstring),
             safekeeper_connstrings,
             storage_auth_token: auth_token.clone(),
-            remote_extensions: None,
+            remote_extensions,
         };
         let spec_path = self.endpoint_path().join("spec.json");
         std::fs::write(spec_path, serde_json::to_string_pretty(&spec)?)?;

control_plane/src/pageserver.rs

@@ -11,6 +11,7 @@ use std::io::{BufReader, Write};
 use std::num::NonZeroU64;
 use std::path::PathBuf;
 use std::process::{Child, Command};
+use std::time::Duration;
 use std::{io, result};
 
 use anyhow::{bail, Context};
@@ -522,19 +523,24 @@ impl PageServerNode {
         &self,
         tenant_id: TenantId,
         config: LocationConfig,
+        flush_ms: Option<Duration>,
     ) -> anyhow::Result<()> {
         let req_body = TenantLocationConfigRequest { tenant_id, config };
 
-        self.http_request(
-            Method::PUT,
-            format!(
-                "{}/tenant/{}/location_config",
-                self.http_base_url, tenant_id
-            ),
-        )?
-        .json(&req_body)
-        .send()?
-        .error_from_body()?;
+        let path = format!(
+            "{}/tenant/{}/location_config",
+            self.http_base_url, tenant_id
+        );
+        let path = if let Some(flush_ms) = flush_ms {
+            format!("{}?flush_ms={}", path, flush_ms.as_millis())
+        } else {
+            path
+        };
+
+        self.http_request(Method::PUT, path)?
+            .json(&req_body)
+            .send()?
+            .error_from_body()?;
 
         Ok(())
     }
@@ -559,6 +565,7 @@ impl PageServerNode {
         ancestor_start_lsn: Option<Lsn>,
         ancestor_timeline_id: Option<TimelineId>,
         pg_version: Option<u32>,
+        existing_initdb_timeline_id: Option<TimelineId>,
     ) -> anyhow::Result<TimelineInfo> {
         // If timeline ID was not specified, generate one
         let new_timeline_id = new_timeline_id.unwrap_or(TimelineId::generate());
@@ -572,6 +579,7 @@ impl PageServerNode {
             ancestor_start_lsn,
             ancestor_timeline_id,
             pg_version,
+            existing_initdb_timeline_id,
         })
         .send()?
         .error_from_body()?

control_plane/src/tenant_migration.rs

@@ -14,7 +14,6 @@ use pageserver_api::models::{
 use std::collections::HashMap;
 use std::time::Duration;
 use utils::{
-    generation::Generation,
     id::{TenantId, TimelineId},
     lsn::Lsn,
 };
@@ -93,6 +92,22 @@ pub fn migrate_tenant(
     // Get a new generation
     let attachment_service = AttachmentService::from_env(env);
 
+    fn build_location_config(
+        mode: LocationConfigMode,
+        generation: Option<u32>,
+        secondary_conf: Option<LocationConfigSecondary>,
+    ) -> LocationConfig {
+        LocationConfig {
+            mode,
+            generation,
+            secondary_conf,
+            tenant_conf: TenantConfig::default(),
+            shard_number: 0,
+            shard_count: 0,
+            shard_stripe_size: 0,
+        }
+    }
+
     let previous = attachment_service.inspect(tenant_id)?;
     let mut baseline_lsns = None;
     if let Some((generation, origin_ps_id)) = &previous {
@@ -101,40 +116,26 @@ pub fn migrate_tenant(
         if origin_ps_id == &dest_ps.conf.id {
             println!("🔁 Already attached to {origin_ps_id}, freshening...");
             let gen = attachment_service.attach_hook(tenant_id, dest_ps.conf.id)?;
-            let dest_conf = LocationConfig {
-                mode: LocationConfigMode::AttachedSingle,
-                generation: gen.map(Generation::new),
-                secondary_conf: None,
-                tenant_conf: TenantConfig::default(),
-            };
-            dest_ps.location_config(tenant_id, dest_conf)?;
+            let dest_conf = build_location_config(LocationConfigMode::AttachedSingle, gen, None);
+            dest_ps.location_config(tenant_id, dest_conf, None)?;
             println!("✅ Migration complete");
             return Ok(());
         }
 
         println!("🔁 Switching origin pageserver {origin_ps_id} to stale mode");
 
-        let stale_conf = LocationConfig {
-            mode: LocationConfigMode::AttachedStale,
-            generation: Some(Generation::new(*generation)),
-            secondary_conf: None,
-            tenant_conf: TenantConfig::default(),
-        };
-        origin_ps.location_config(tenant_id, stale_conf)?;
+        let stale_conf =
+            build_location_config(LocationConfigMode::AttachedStale, Some(*generation), None);
+        origin_ps.location_config(tenant_id, stale_conf, Some(Duration::from_secs(10)))?;
 
         baseline_lsns = Some(get_lsns(tenant_id, &origin_ps)?);
     }
 
     let gen = attachment_service.attach_hook(tenant_id, dest_ps.conf.id)?;
-    let dest_conf = LocationConfig {
-        mode: LocationConfigMode::AttachedMulti,
-        generation: gen.map(Generation::new),
-        secondary_conf: None,
-        tenant_conf: TenantConfig::default(),
-    };
+    let dest_conf = build_location_config(LocationConfigMode::AttachedMulti, gen, None);
 
     println!("🔁 Attaching to pageserver {}", dest_ps.conf.id);
-    dest_ps.location_config(tenant_id, dest_conf)?;
+    dest_ps.location_config(tenant_id, dest_conf, None)?;
 
     if let Some(baseline) = baseline_lsns {
         println!("🕑 Waiting for LSN to catch up...");
@@ -170,31 +171,25 @@ pub fn migrate_tenant(
         }
 
         // Downgrade to a secondary location
-        let secondary_conf = LocationConfig {
-            mode: LocationConfigMode::Secondary,
-            generation: None,
-            secondary_conf: Some(LocationConfigSecondary { warm: true }),
-            tenant_conf: TenantConfig::default(),
-        };
+        let secondary_conf = build_location_config(
+            LocationConfigMode::Secondary,
+            None,
+            Some(LocationConfigSecondary { warm: true }),
+        );
 
         println!(
             "💤 Switching to secondary mode on pageserver {}",
             other_ps.conf.id
         );
-        other_ps.location_config(tenant_id, secondary_conf)?;
+        other_ps.location_config(tenant_id, secondary_conf, None)?;
     }
 
     println!(
         "🔁 Switching to AttachedSingle mode on pageserver {}",
         dest_ps.conf.id
     );
-    let dest_conf = LocationConfig {
-        mode: LocationConfigMode::AttachedSingle,
-        generation: gen.map(Generation::new),
-        secondary_conf: None,
-        tenant_conf: TenantConfig::default(),
-    };
-    dest_ps.location_config(tenant_id, dest_conf)?;
+    let dest_conf = build_location_config(LocationConfigMode::AttachedSingle, gen, None);
+    dest_ps.location_config(tenant_id, dest_conf, None)?;
 
     println!("✅ Migration complete");
 

docs/rfcs/029-pageserver-wal-disaster-recovery.md

@@ -0,0 +1,205 @@
+# Name
+
+Created on: 2023-09-08
+Author: Arpad Müller
+
+## Summary
+
+Enable the pageserver to recover from data corruption events by implementing
+a feature to re-apply historic WAL records in parallel to the already occurring
+WAL replay.
+
+The feature is outside of the user-visible backup and history story, and only
+serves as a second-level backup for the case that there is a bug in the
+pageservers that corrupted the served pages.
+
+The RFC proposes the addition of two new features:
+* recover a broken branch from WAL (downtime is allowed)
+* a test recovery system to recover random branches to make sure recovery works
+
+## Motivation
+
+The historic WAL is currently stored in S3 even after it has been replayed by
+the pageserver and thus been integrated into the pageserver's storage system.
+This is done to defend from data corruption failures inside the pageservers.
+
+However, application of this WAL in the disaster recovery setting is currently
+very manual and we want to automate this to make it easier.
+
+### Use cases
+
+There are various use cases for this feature, like:
+
+* The main motivation is replaying in the instance of pageservers corrupting
+  data.
+* We might want to, beyond the user-visible history features, through our
+  support channels and upon customer request, in select instances, recover
+  historic versions beyond the range of history that we officially support.
+* Running the recovery process in the background for random tenant timelines
+  to figure out if there was a corruption of data (we would compare with what
+  the pageserver stores for the "official" timeline).
+* Using the WAL to arrive at historic pages we can then back up to S3 so that
+  WAL itself can be discarded, or at least not used for future replays.
+  Again, this sounds a lot like what the pageserver is already doing, but the
+  point is to provide a fallback to the service provided by the pageserver.
+
+## Design
+
+### Design constraints
+
+The main design constraint is that the feature needs to be *simple* enough that
+the number of bugs are as low, and reliability as high as possible: the main
+goal of this endeavour is to achieve higher correctness than the pageserver.
+
+For the background process, we cannot afford a downtime of the timeline that is
+being cloned, as we don't want to restrict ourselves to offline tenants only.
+In the scenario where we want to recover from disasters or roll back to a
+historic lsn through support staff, downtimes are more affordable, and
+inevitable if the original had been subject to the corruption. Ideally, the
+two code paths would share code, so the solution would be designed for not
+requiring downtimes.
+
+### API endpoint changes
+
+This RFC proposes two API endpoint changes in the safekeeper and the
+pageserver.
+
+Remember, the pageserver timeline API creation endpoint is to this URL:
+
+```
+/v1/tenant/{tenant_id}/timeline/
+```
+
+Where `{tenant_id}` is the ID of the tenant the timeline is created for,
+and specified as part of the URL. The timeline ID is passed via the POST
+request body as the only required parameter `new_timeline_id`.
+
+This proposal adds one optional parameter called
+`existing_initdb_timeline_id` to the request's json body. If the parameter
+is not specified, behaviour should be as existing, so the pageserver runs
+initdb.
+If the parameter is specified, it is expected to point to a timeline ID.
+In fact that ID might match `new_timeline_id`, what's important is that
+S3 storage contains a matching initdb under the URL matching the given
+tenant and timeline.
+
+Having both `ancestor_timeline_id` and `existing_initdb_timeline_id`
+specified is illegal and will yield in an HTTP error. This feature is
+only meant for the "main" branch that doesn't have any ancestors
+of its own, as only here initdb is relevant.
+
+For the safekeeper, we propose the addition of the following copy endpoint:
+
+```
+/v1/tenant/{tenant_id}/timeline/{source_timeline_id}/copy
+```
+it is meant for POST requests with json, and the two URL parameters
+`tenant_id` and `source_timeline_id`. The json request body contains
+the two required parameters `target_timeline_id` and `until_lsn`.
+
+After invoking, the copy endpoint starts a copy process of the WAL from
+the source ID to the target ID. The lsn is updated according to the
+progress of the API call.
+
+### Higher level features
+
+We want the API changes to support the following higher level features:
+
+* recovery-after-corruption DR of the main timeline of a tenant. This
+  feature allows for downtime.
+* test DR of the main timeline into a special copy timeline. this feature
+  is meant to run against selected production tenants in the background,
+  without the user noticing, so it does not allow for downtime.
+
+The recovery-after-corruption DR only needs the pageserver changes.
+It works as follows:
+
+* delete the timeline from the pageservers via timeline deletion API
+* re-create it via timeline creation API (same ID as before) and set
+  `existing_initdb_timeline_id` to the same timeline ID
+
+The test DR requires also the copy primitive and works as follows:
+
+* copy the WAL of the timeline to a new place
+* create a new timeline for the tenant
+
+## Non Goals
+
+At the danger of being repetitive, the main goal of this feature is to be a
+backup method, so reliability is very important. This implies that other
+aspects like performance or space reduction are less important.
+
+### Corrupt WAL
+
+The process suggested by this RFC assumes that the WAL is free of corruption.
+In some instances, corruption can make it into WAL, like for example when
+higher level components like postgres or the application first read corrupt
+data, and then execute a write with data derived from that earlier read. That
+written data might then contain the corruption.
+
+Common use cases can hit this quite easily. For example, an application reads
+some counter, increments it, and then writes the new counter value to the
+database.
+On a lower level, the compute might put FPIs (Full Page Images) into the WAL,
+which have corrupt data for rows unrelated to the write operation at hand.
+
+Separating corrupt writes from non-corrupt ones is a hard problem in general,
+and if the application was involved in making the corrupt write, a recovery
+would also involve the application. Therefore, corruption that has made it into
+the WAL is outside of the scope of this feature. However, the WAL replay can be
+issued to right before the point in time where the corruption occured. Then the
+data loss is isolated to post-corruption writes only.
+
+## Impacted components (e.g. pageserver, safekeeper, console, etc)
+
+Most changes would happen to the pageservers.
+For the higher level features, maybe other components like the console would
+be involved.
+
+We need to make sure that the shadow timelines are not subject to the usual
+limits and billing we apply to existing timelines.
+
+## Proposed implementation
+
+The first problem to keep in mind is the reproducability of `initdb`.
+So an initial step would be to upload `initdb` snapshots to S3.
+
+After that, we'd have the endpoint spawn a background process which
+performs the replay of the WAL to that new timeline. This process should
+follow the existing workflows as closely as possible, just using the
+WAL records of a different timeline.
+
+The timeline created will be in a special state that solely looks for WAL
+entries of the timeline it is trying to copy. Once the target LSN is reached,
+it turns into a normal timeline that also accepts writes to its own
+timeline ID.
+
+### Scalability
+
+For now we want to run this entire process on a single node, and as
+it is by nature linear, it's hard to parallelize. However, for the
+verification workloads, we can easily start the WAL replay in parallel
+for different points in time. This is valuable especially for tenants
+with large WAL records.
+
+Compare this with the tricks to make addition circuits execute with
+lower latency by making them perform the addition for both possible
+values of the carry bit, and then, in a second step, taking the
+result for the carry bit that was actually obtained.
+
+The other scalability dimension to consider is the WAL length, which
+is a growing question as tenants accumulate changes. There are
+possible approaches to this, including creating snapshots of the
+page files and uploading them to S3, but if we do this for every single
+branch, we lose the cheap branching property.
+
+### Implementation by component
+
+The proposed changes for the various components of the neon architecture
+are written up in this notion page:
+
+https://www.notion.so/neondatabase/Pageserver-disaster-recovery-one-pager-4ecfb5df16ce4f6bbfc3817ed1a6cbb2
+
+### Unresolved questions
+
+none known (outside of the mentioned ones).

libs/compute_api/src/spec.rs

@@ -26,6 +26,13 @@ pub struct ComputeSpec {
     // but we don't use it for anything. Serde will ignore missing fields when
     // deserializing it.
     pub operation_uuid: Option<String>,
+
+    /// Compute features to enable. These feature flags are provided, when we
+    /// know all the details about client's compute, so they cannot be used
+    /// to change `Empty` compute behavior.
+    #[serde(default)]
+    pub features: Vec<ComputeFeature>,
+
     /// Expected cluster state at the end of transition process.
     pub cluster: Cluster,
     pub delta_operations: Option<Vec<DeltaOp>>,
@@ -68,6 +75,19 @@ pub struct ComputeSpec {
     pub remote_extensions: Option<RemoteExtSpec>,
 }
 
+/// Feature flag to signal `compute_ctl` to enable certain experimental functionality.
+#[derive(Serialize, Clone, Copy, Debug, Deserialize, PartialEq, Eq)]
+#[serde(rename_all = "snake_case")]
+pub enum ComputeFeature {
+    // XXX: Add more feature flags here.
+
+    // This is a special feature flag that is used to represent unknown feature flags.
+    // Basically all unknown to enum flags are represented as this one. See unit test
+    // `parse_unknown_features()` for more details.
+    #[serde(other)]
+    UnknownFeature,
+}
+
 #[derive(Clone, Debug, Default, Deserialize, Serialize)]
 pub struct RemoteExtSpec {
     pub public_extensions: Option<Vec<String>>,
@@ -187,8 +207,6 @@ pub struct DeltaOp {
 pub struct Role {
     pub name: PgIdent,
     pub encrypted_password: Option<String>,
-    pub replication: Option<bool>,
-    pub bypassrls: Option<bool>,
     pub options: GenericOptions,
 }
 
@@ -229,7 +247,10 @@ mod tests {
     #[test]
     fn parse_spec_file() {
         let file = File::open("tests/cluster_spec.json").unwrap();
-        let _spec: ComputeSpec = serde_json::from_reader(file).unwrap();
+        let spec: ComputeSpec = serde_json::from_reader(file).unwrap();
+
+        // Features list defaults to empty vector.
+        assert!(spec.features.is_empty());
     }
 
     #[test]
@@ -241,4 +262,22 @@ mod tests {
         ob.insert("unknown_field_123123123".into(), "hello".into());
         let _spec: ComputeSpec = serde_json::from_value(json).unwrap();
     }
+
+    #[test]
+    fn parse_unknown_features() {
+        // Test that unknown feature flags do not cause any errors.
+        let file = File::open("tests/cluster_spec.json").unwrap();
+        let mut json: serde_json::Value = serde_json::from_reader(file).unwrap();
+        let ob = json.as_object_mut().unwrap();
+
+        // Add unknown feature flags.
+        let features = vec!["foo_bar_feature", "baz_feature"];
+        ob.insert("features".into(), features.into());
+
+        let spec: ComputeSpec = serde_json::from_value(json).unwrap();
+
+        assert!(spec.features.len() == 2);
+        assert!(spec.features.contains(&ComputeFeature::UnknownFeature));
+        assert_eq!(spec.features, vec![ComputeFeature::UnknownFeature; 2]);
+    }
 }

libs/pageserver_api/Cargo.toml

@@ -18,6 +18,7 @@ enum-map.workspace = true
 strum.workspace = true
 strum_macros.workspace = true
 hex.workspace = true
+thiserror.workspace = true
 
 workspace_hack.workspace = true
 

libs/pageserver_api/src/control_api.rs

@@ -4,7 +4,9 @@
 //! See docs/rfcs/025-generation-numbers.md
 
 use serde::{Deserialize, Serialize};
-use utils::id::{NodeId, TenantId};
+use utils::id::NodeId;
+
+use crate::shard::TenantShardId;
 
 #[derive(Serialize, Deserialize)]
 pub struct ReAttachRequest {
@@ -13,7 +15,7 @@ pub struct ReAttachRequest {
 
 #[derive(Serialize, Deserialize)]
 pub struct ReAttachResponseTenant {
-    pub id: TenantId,
+    pub id: TenantShardId,
     pub gen: u32,
 }
 
@@ -24,7 +26,7 @@ pub struct ReAttachResponse {
 
 #[derive(Serialize, Deserialize)]
 pub struct ValidateRequestTenant {
-    pub id: TenantId,
+    pub id: TenantShardId,
     pub gen: u32,
 }
 
@@ -40,6 +42,6 @@ pub struct ValidateResponse {
 
 #[derive(Serialize, Deserialize)]
 pub struct ValidateResponseTenant {
-    pub id: TenantId,
+    pub id: TenantShardId,
     pub valid: bool,
 }

libs/pageserver_api/src/key.rs

@@ -140,3 +140,7 @@ impl Key {
         })
     }
 }
+
+pub fn is_rel_block_key(key: &Key) -> bool {
+    key.field1 == 0x00 && key.field4 != 0
+}

libs/pageserver_api/src/models.rs

@@ -10,7 +10,6 @@ use serde_with::serde_as;
 use strum_macros;
 use utils::{
     completion,
-    generation::Generation,
     history_buffer::HistoryBufferWithDropCounter,
     id::{NodeId, TenantId, TimelineId},
     lsn::Lsn,
@@ -180,6 +179,8 @@ pub struct TimelineCreateRequest {
     #[serde(default)]
     pub ancestor_timeline_id: Option<TimelineId>,
     #[serde(default)]
+    pub existing_initdb_timeline_id: Option<TimelineId>,
+    #[serde(default)]
     pub ancestor_start_lsn: Option<Lsn>,
     pub pg_version: Option<u32>,
 }
@@ -262,10 +263,19 @@ pub struct LocationConfig {
     pub mode: LocationConfigMode,
     /// If attaching, in what generation?
     #[serde(default)]
-    pub generation: Option<Generation>,
+    pub generation: Option<u32>,
     #[serde(default)]
     pub secondary_conf: Option<LocationConfigSecondary>,
 
+    // Shard parameters: if shard_count is nonzero, then other shard_* fields
+    // must be set accurately.
+    #[serde(default)]
+    pub shard_number: u8,
+    #[serde(default)]
+    pub shard_count: u8,
+    #[serde(default)]
+    pub shard_stripe_size: u32,
+
     // If requesting mode `Secondary`, configuration for that.
     // Custom storage configuration for the tenant, if any
     pub tenant_conf: TenantConfig,
@@ -306,31 +316,14 @@ impl std::ops::Deref for TenantConfigRequest {
 
 impl TenantConfigRequest {
     pub fn new(tenant_id: TenantId) -> TenantConfigRequest {
-        let config = TenantConfig {
-            checkpoint_distance: None,
-            checkpoint_timeout: None,
-            compaction_target_size: None,
-            compaction_period: None,
-            compaction_threshold: None,
-            gc_horizon: None,
-            gc_period: None,
-            image_creation_threshold: None,
-            pitr_interval: None,
-            walreceiver_connect_timeout: None,
-            lagging_wal_timeout: None,
-            max_lsn_wal_lag: None,
-            trace_read_requests: None,
-            eviction_policy: None,
-            min_resident_size_override: None,
-            evictions_low_residence_duration_metric_threshold: None,
-            gc_feedback: None,
-        };
+        let config = TenantConfig::default();
         TenantConfigRequest { tenant_id, config }
     }
 }
 
 #[derive(Debug, Deserialize)]
 pub struct TenantAttachRequest {
+    #[serde(default)]
     pub config: TenantAttachConfig,
     #[serde(default)]
     pub generation: Option<u32>,
@@ -338,7 +331,7 @@ pub struct TenantAttachRequest {
 
 /// Newtype to enforce deny_unknown_fields on TenantConfig for
 /// its usage inside `TenantAttachRequest`.
-#[derive(Debug, Serialize, Deserialize)]
+#[derive(Debug, Serialize, Deserialize, Default)]
 #[serde(deny_unknown_fields)]
 pub struct TenantAttachConfig {
     #[serde(flatten)]
@@ -392,7 +385,9 @@ pub struct TimelineInfo {
     /// The LSN that we are advertizing to safekeepers
     pub remote_consistent_lsn_visible: Lsn,
 
-    pub current_logical_size: Option<u64>, // is None when timeline is Unloaded
+    pub current_logical_size: u64,
+    pub current_logical_size_is_accurate: bool,
+
     /// Sum of the size of all layer files.
     /// If a layer is present in both local FS and S3, it counts only once.
     pub current_physical_size: Option<u64>, // is None when timeline is Unloaded

libs/pageserver_api/src/shard.rs

@@ -1,13 +1,15 @@
 use std::{ops::RangeInclusive, str::FromStr};
 
+use crate::key::{is_rel_block_key, Key};
 use hex::FromHex;
 use serde::{Deserialize, Serialize};
+use thiserror;
 use utils::id::TenantId;
 
-#[derive(Ord, PartialOrd, Eq, PartialEq, Clone, Copy, Serialize, Deserialize, Debug)]
+#[derive(Ord, PartialOrd, Eq, PartialEq, Clone, Copy, Serialize, Deserialize, Debug, Hash)]
 pub struct ShardNumber(pub u8);
 
-#[derive(Ord, PartialOrd, Eq, PartialEq, Clone, Copy, Serialize, Deserialize, Debug)]
+#[derive(Ord, PartialOrd, Eq, PartialEq, Clone, Copy, Serialize, Deserialize, Debug, Hash)]
 pub struct ShardCount(pub u8);
 
 impl ShardCount {
@@ -38,7 +40,7 @@ impl ShardNumber {
 /// Note that the binary encoding is _not_ backward compatible, because
 /// at the time sharding is introduced, there are no existing binary structures
 /// containing TenantId that we need to handle.
-#[derive(Eq, PartialEq, PartialOrd, Ord, Clone, Copy)]
+#[derive(Eq, PartialEq, PartialOrd, Ord, Clone, Copy, Hash)]
 pub struct TenantShardId {
     pub tenant_id: TenantId,
     pub shard_number: ShardNumber,
@@ -71,19 +73,28 @@ impl TenantShardId {
         )
     }
 
-    pub fn shard_slug(&self) -> String {
-        format!("{:02x}{:02x}", self.shard_number.0, self.shard_count.0)
+    pub fn shard_slug(&self) -> impl std::fmt::Display + '_ {
+        ShardSlug(self)
+    }
+}
+
+/// Formatting helper
+struct ShardSlug<'a>(&'a TenantShardId);
+
+impl<'a> std::fmt::Display for ShardSlug<'a> {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        write!(
+            f,
+            "{:02x}{:02x}",
+            self.0.shard_number.0, self.0.shard_count.0
+        )
     }
 }
 
 impl std::fmt::Display for TenantShardId {
     fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
         if self.shard_count != ShardCount(0) {
-            write!(
-                f,
-                "{}-{:02x}{:02x}",
-                self.tenant_id, self.shard_number.0, self.shard_count.0
-            )
+            write!(f, "{}-{}", self.tenant_id, self.shard_slug())
         } else {
             // Legacy case (shard_count == 0) -- format as just the tenant id.  Note that this
             // is distinct from the normal single shard case (shard count == 1).
@@ -139,6 +150,89 @@ impl From<[u8; 18]> for TenantShardId {
     }
 }
 
+/// For use within the context of a particular tenant, when we need to know which
+/// shard we're dealing with, but do not need to know the full ShardIdentity (because
+/// we won't be doing any page->shard mapping), and do not need to know the fully qualified
+/// TenantShardId.
+#[derive(Eq, PartialEq, PartialOrd, Ord, Clone, Copy)]
+pub struct ShardIndex {
+    pub shard_number: ShardNumber,
+    pub shard_count: ShardCount,
+}
+
+impl ShardIndex {
+    pub fn new(number: ShardNumber, count: ShardCount) -> Self {
+        Self {
+            shard_number: number,
+            shard_count: count,
+        }
+    }
+    pub fn unsharded() -> Self {
+        Self {
+            shard_number: ShardNumber(0),
+            shard_count: ShardCount(0),
+        }
+    }
+
+    pub fn is_unsharded(&self) -> bool {
+        self.shard_number == ShardNumber(0) && self.shard_count == ShardCount(0)
+    }
+
+    /// For use in constructing remote storage paths: concatenate this with a TenantId
+    /// to get a fully qualified TenantShardId.
+    ///
+    /// Backward compat: this function returns an empty string if Self::is_unsharded, such
+    /// that the legacy pre-sharding remote key format is preserved.
+    pub fn get_suffix(&self) -> String {
+        if self.is_unsharded() {
+            "".to_string()
+        } else {
+            format!("-{:02x}{:02x}", self.shard_number.0, self.shard_count.0)
+        }
+    }
+}
+
+impl std::fmt::Display for ShardIndex {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        write!(f, "{:02x}{:02x}", self.shard_number.0, self.shard_count.0)
+    }
+}
+
+impl std::fmt::Debug for ShardIndex {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        // Debug is the same as Display: the compact hex representation
+        write!(f, "{}", self)
+    }
+}
+
+impl std::str::FromStr for ShardIndex {
+    type Err = hex::FromHexError;
+
+    fn from_str(s: &str) -> Result<Self, Self::Err> {
+        // Expect format: 1 byte shard number, 1 byte shard count
+        if s.len() == 4 {
+            let bytes = s.as_bytes();
+            let mut shard_parts: [u8; 2] = [0u8; 2];
+            hex::decode_to_slice(bytes, &mut shard_parts)?;
+            Ok(Self {
+                shard_number: ShardNumber(shard_parts[0]),
+                shard_count: ShardCount(shard_parts[1]),
+            })
+        } else {
+            Err(hex::FromHexError::InvalidStringLength)
+        }
+    }
+}
+
+impl From<[u8; 2]> for ShardIndex {
+    fn from(b: [u8; 2]) -> Self {
+        Self {
+            shard_number: ShardNumber(b[0]),
+            shard_count: ShardCount(b[1]),
+        }
+    }
+}
+
 impl Serialize for TenantShardId {
     fn serialize<S>(&self, serializer: S) -> Result<S::Ok, S::Error>
     where
@@ -209,6 +303,261 @@ impl<'de> Deserialize<'de> for TenantShardId {
     }
 }
 
+/// Stripe size in number of pages
+#[derive(Clone, Copy, Serialize, Deserialize, Eq, PartialEq, Debug)]
+pub struct ShardStripeSize(pub u32);
+
+/// Layout version: for future upgrades where we might change how the key->shard mapping works
+#[derive(Clone, Copy, Serialize, Deserialize, Eq, PartialEq, Debug)]
+pub struct ShardLayout(u8);
+
+const LAYOUT_V1: ShardLayout = ShardLayout(1);
+/// ShardIdentity uses a magic layout value to indicate if it is unusable
+const LAYOUT_BROKEN: ShardLayout = ShardLayout(255);
+
+/// Default stripe size in pages: 256MiB divided by 8kiB page size.
+const DEFAULT_STRIPE_SIZE: ShardStripeSize = ShardStripeSize(256 * 1024 / 8);
+
+/// The ShardIdentity contains the information needed for one member of map
+/// to resolve a key to a shard, and then check whether that shard is ==self.
+#[derive(Clone, Copy, Serialize, Deserialize, Eq, PartialEq, Debug)]
+pub struct ShardIdentity {
+    pub number: ShardNumber,
+    pub count: ShardCount,
+    stripe_size: ShardStripeSize,
+    layout: ShardLayout,
+}
+
+#[derive(thiserror::Error, Debug, PartialEq, Eq)]
+pub enum ShardConfigError {
+    #[error("Invalid shard count")]
+    InvalidCount,
+    #[error("Invalid shard number")]
+    InvalidNumber,
+    #[error("Invalid stripe size")]
+    InvalidStripeSize,
+}
+
+impl ShardIdentity {
+    /// An identity with number=0 count=0 is a "none" identity, which represents legacy
+    /// tenants.  Modern single-shard tenants should not use this: they should
+    /// have number=0 count=1.
+    pub fn unsharded() -> Self {
+        Self {
+            number: ShardNumber(0),
+            count: ShardCount(0),
+            layout: LAYOUT_V1,
+            stripe_size: DEFAULT_STRIPE_SIZE,
+        }
+    }
+
+    /// A broken instance of this type is only used for `TenantState::Broken` tenants,
+    /// which are constructed in code paths that don't have access to proper configuration.
+    ///
+    /// A ShardIdentity in this state may not be used for anything, and should not be persisted.
+    /// Enforcement is via assertions, to avoid making our interface fallible for this
+    /// edge case: it is the Tenant's responsibility to avoid trying to do any I/O when in a broken
+    /// state, and by extension to avoid trying to do any page->shard resolution.
+    pub fn broken(number: ShardNumber, count: ShardCount) -> Self {
+        Self {
+            number,
+            count,
+            layout: LAYOUT_BROKEN,
+            stripe_size: DEFAULT_STRIPE_SIZE,
+        }
+    }
+
+    pub fn is_unsharded(&self) -> bool {
+        self.number == ShardNumber(0) && self.count == ShardCount(0)
+    }
+
+    /// Count must be nonzero, and number must be < count. To construct
+    /// the legacy case (count==0), use Self::unsharded instead.
+    pub fn new(
+        number: ShardNumber,
+        count: ShardCount,
+        stripe_size: ShardStripeSize,
+    ) -> Result<Self, ShardConfigError> {
+        if count.0 == 0 {
+            Err(ShardConfigError::InvalidCount)
+        } else if number.0 > count.0 - 1 {
+            Err(ShardConfigError::InvalidNumber)
+        } else if stripe_size.0 == 0 {
+            Err(ShardConfigError::InvalidStripeSize)
+        } else {
+            Ok(Self {
+                number,
+                count,
+                layout: LAYOUT_V1,
+                stripe_size,
+            })
+        }
+    }
+
+    fn is_broken(&self) -> bool {
+        self.layout == LAYOUT_BROKEN
+    }
+
+    pub fn get_shard_number(&self, key: &Key) -> ShardNumber {
+        assert!(!self.is_broken());
+        key_to_shard_number(self.count, self.stripe_size, key)
+    }
+
+    /// Return true if the key should be ingested by this shard
+    pub fn is_key_local(&self, key: &Key) -> bool {
+        assert!(!self.is_broken());
+        if self.count < ShardCount(2) || (key_is_shard0(key) && self.number == ShardNumber(0)) {
+            true
+        } else {
+            key_to_shard_number(self.count, self.stripe_size, key) == self.number
+        }
+    }
+
+    pub fn shard_slug(&self) -> String {
+        if self.count > ShardCount(0) {
+            format!("-{:02x}{:02x}", self.number.0, self.count.0)
+        } else {
+            String::new()
+        }
+    }
+
+    /// Convenience for checking if this identity is the 0th shard in a tenant,
+    /// for special cases on shard 0 such as ingesting relation sizes.
+    pub fn is_zero(&self) -> bool {
+        self.number == ShardNumber(0)
+    }
+}
+
+impl Serialize for ShardIndex {
+    fn serialize<S>(&self, serializer: S) -> Result<S::Ok, S::Error>
+    where
+        S: serde::Serializer,
+    {
+        if serializer.is_human_readable() {
+            serializer.collect_str(self)
+        } else {
+            // Binary encoding is not used in index_part.json, but is included in anticipation of
+            // switching various structures (e.g. inter-process communication, remote metadata) to more
+            // compact binary encodings in future.
+            let mut packed: [u8; 2] = [0; 2];
+            packed[0] = self.shard_number.0;
+            packed[1] = self.shard_count.0;
+            packed.serialize(serializer)
+        }
+    }
+}
+
+impl<'de> Deserialize<'de> for ShardIndex {
+    fn deserialize<D>(deserializer: D) -> Result<Self, D::Error>
+    where
+        D: serde::Deserializer<'de>,
+    {
+        struct IdVisitor {
+            is_human_readable_deserializer: bool,
+        }
+
+        impl<'de> serde::de::Visitor<'de> for IdVisitor {
+            type Value = ShardIndex;
+
+            fn expecting(&self, formatter: &mut std::fmt::Formatter) -> std::fmt::Result {
+                if self.is_human_readable_deserializer {
+                    formatter.write_str("value in form of hex string")
+                } else {
+                    formatter.write_str("value in form of integer array([u8; 2])")
+                }
+            }
+
+            fn visit_seq<A>(self, seq: A) -> Result<Self::Value, A::Error>
+            where
+                A: serde::de::SeqAccess<'de>,
+            {
+                let s = serde::de::value::SeqAccessDeserializer::new(seq);
+                let id: [u8; 2] = Deserialize::deserialize(s)?;
+                Ok(ShardIndex::from(id))
+            }
+
+            fn visit_str<E>(self, v: &str) -> Result<Self::Value, E>
+            where
+                E: serde::de::Error,
+            {
+                ShardIndex::from_str(v).map_err(E::custom)
+            }
+        }
+
+        if deserializer.is_human_readable() {
+            deserializer.deserialize_str(IdVisitor {
+                is_human_readable_deserializer: true,
+            })
+        } else {
+            deserializer.deserialize_tuple(
+                2,
+                IdVisitor {
+                    is_human_readable_deserializer: false,
+                },
+            )
+        }
+    }
+}
+
+/// Whether this key is always held on shard 0 (e.g. shard 0 holds all SLRU keys
+/// in order to be able to serve basebackup requests without peer communication).
+fn key_is_shard0(key: &Key) -> bool {
+    // To decide what to shard out to shards >0, we apply a simple rule that only
+    // relation pages are distributed to shards other than shard zero. Everything else gets
+    // stored on shard 0.  This guarantees that shard 0 can independently serve basebackup
+    // requests, and any request other than those for particular blocks in relations.
+    //
+    // In this condition:
+    // - is_rel_block_key includes only relations, i.e. excludes SLRU data and
+    // all metadata.
+    // - field6 is set to -1 for relation size pages.
+    !(is_rel_block_key(key) && key.field6 != 0xffffffff)
+}
+
+/// Provide the same result as the function in postgres `hashfn.h` with the same name
+fn murmurhash32(mut h: u32) -> u32 {
+    h ^= h >> 16;
+    h = h.wrapping_mul(0x85ebca6b);
+    h ^= h >> 13;
+    h = h.wrapping_mul(0xc2b2ae35);
+    h ^= h >> 16;
+    h
+}
+
+/// Provide the same result as the function in postgres `hashfn.h` with the same name
+fn hash_combine(mut a: u32, mut b: u32) -> u32 {
+    b = b.wrapping_add(0x9e3779b9);
+    b = b.wrapping_add(a << 6);
+    b = b.wrapping_add(a >> 2);
+
+    a ^= b;
+    a
+}
+
+/// Where a Key is to be distributed across shards, select the shard.  This function
+/// does not account for keys that should be broadcast across shards.
+///
+/// The hashing in this function must exactly match what we do in postgres smgr
+/// code.  The resulting distribution of pages is intended to preserve locality within
+/// `stripe_size` ranges of contiguous block numbers in the same relation, while otherwise
+/// distributing data pseudo-randomly.
+///
+/// The mapping of key to shard is not stable across changes to ShardCount: this is intentional
+/// and will be handled at higher levels when shards are split.
+fn key_to_shard_number(count: ShardCount, stripe_size: ShardStripeSize, key: &Key) -> ShardNumber {
+    // Fast path for un-sharded tenants or broadcast keys
+    if count < ShardCount(2) || key_is_shard0(key) {
+        return ShardNumber(0);
+    }
+
+    // relNode
+    let mut hash = murmurhash32(key.field4);
+    // blockNum/stripe size
+    hash = hash_combine(hash, murmurhash32(key.field6 / stripe_size.0));
+
+    ShardNumber((hash % count.0 as u32) as u8)
+}
+
 #[cfg(test)]
 mod tests {
     use std::str::FromStr;
@@ -318,4 +667,91 @@ mod tests {
 
         Ok(())
     }
+
+    #[test]
+    fn shard_identity_validation() -> Result<(), ShardConfigError> {
+        // Happy cases
+        ShardIdentity::new(ShardNumber(0), ShardCount(1), DEFAULT_STRIPE_SIZE)?;
+        ShardIdentity::new(ShardNumber(0), ShardCount(1), ShardStripeSize(1))?;
+        ShardIdentity::new(ShardNumber(254), ShardCount(255), ShardStripeSize(1))?;
+
+        assert_eq!(
+            ShardIdentity::new(ShardNumber(0), ShardCount(0), DEFAULT_STRIPE_SIZE),
+            Err(ShardConfigError::InvalidCount)
+        );
+        assert_eq!(
+            ShardIdentity::new(ShardNumber(10), ShardCount(10), DEFAULT_STRIPE_SIZE),
+            Err(ShardConfigError::InvalidNumber)
+        );
+        assert_eq!(
+            ShardIdentity::new(ShardNumber(11), ShardCount(10), DEFAULT_STRIPE_SIZE),
+            Err(ShardConfigError::InvalidNumber)
+        );
+        assert_eq!(
+            ShardIdentity::new(ShardNumber(255), ShardCount(255), DEFAULT_STRIPE_SIZE),
+            Err(ShardConfigError::InvalidNumber)
+        );
+        assert_eq!(
+            ShardIdentity::new(ShardNumber(0), ShardCount(1), ShardStripeSize(0)),
+            Err(ShardConfigError::InvalidStripeSize)
+        );
+
+        Ok(())
+    }
+
+    #[test]
+    fn shard_index_human_encoding() -> Result<(), hex::FromHexError> {
+        let example = ShardIndex {
+            shard_number: ShardNumber(13),
+            shard_count: ShardCount(17),
+        };
+        let expected: String = "0d11".to_string();
+        let encoded = format!("{example}");
+        assert_eq!(&encoded, &expected);
+
+        let decoded = ShardIndex::from_str(&encoded)?;
+        assert_eq!(example, decoded);
+        Ok(())
+    }
+
+    #[test]
+    fn shard_index_binary_encoding() -> Result<(), hex::FromHexError> {
+        let example = ShardIndex {
+            shard_number: ShardNumber(13),
+            shard_count: ShardCount(17),
+        };
+        let expected: [u8; 2] = [0x0d, 0x11];
+
+        let encoded = bincode::serialize(&example).unwrap();
+        assert_eq!(Hex(&encoded), Hex(&expected));
+        let decoded = bincode::deserialize(&encoded).unwrap();
+        assert_eq!(example, decoded);
+
+        Ok(())
+    }
+
+    // These are only smoke tests to spot check that our implementation doesn't
+    // deviate from a few examples values: not aiming to validate the overall
+    // hashing algorithm.
+    #[test]
+    fn murmur_hash() {
+        assert_eq!(murmurhash32(0), 0);
+
+        assert_eq!(hash_combine(0xb1ff3b40, 0), 0xfb7923c9);
+    }
+
+    #[test]
+    fn shard_mapping() {
+        let key = Key {
+            field1: 0x00,
+            field2: 0x67f,
+            field3: 0x5,
+            field4: 0x400c,
+            field5: 0x00,
+            field6: 0x7d06,
+        };
+
+        let shard = key_to_shard_number(ShardCount(10), DEFAULT_STRIPE_SIZE, &key);
+        assert_eq!(shard, ShardNumber(8));
+    }
 }

libs/pq_proto/src/lib.rs

@@ -289,10 +289,10 @@ impl FeStartupPacket {
         // We shouldn't advance `buf` as probably full message is not there yet,
         // so can't directly use Bytes::get_u32 etc.
         let len = (&buf[0..4]).read_u32::<BigEndian>().unwrap() as usize;
-        // The proposed replacement is `!(4..=MAX_STARTUP_PACKET_LENGTH).contains(&len)`
+        // The proposed replacement is `!(8..=MAX_STARTUP_PACKET_LENGTH).contains(&len)`
         // which is less readable
         #[allow(clippy::manual_range_contains)]
-        if len < 4 || len > MAX_STARTUP_PACKET_LENGTH {
+        if len < 8 || len > MAX_STARTUP_PACKET_LENGTH {
             return Err(ProtocolError::Protocol(format!(
                 "invalid startup packet message length {}",
                 len
@@ -975,4 +975,10 @@ mod tests {
         let params = make_params("foo\\ bar \\ \\\\ baz\\  lol");
         assert_eq!(split_options(&params), ["foo bar", " \\", "baz ", "lol"]);
     }
+
+    #[test]
+    fn parse_fe_startup_packet_regression() {
+        let data = [0, 0, 0, 7, 0, 0, 0, 0];
+        FeStartupPacket::parse(&mut BytesMut::from_iter(data)).unwrap_err();
+    }
 }

libs/remote_storage/Cargo.toml

@@ -9,18 +9,18 @@ anyhow.workspace = true
 async-trait.workspace = true
 once_cell.workspace = true
 aws-smithy-async.workspace = true
-aws-smithy-http.workspace = true
-aws-types.workspace = true
+aws-smithy-types.workspace = true
 aws-config.workspace = true
 aws-sdk-s3.workspace = true
 aws-credential-types.workspace = true
 bytes.workspace = true
 camino.workspace = true
 hyper = { workspace = true, features = ["stream"] }
+futures.workspace = true
 serde.workspace = true
 serde_json.workspace = true
 tokio = { workspace = true, features = ["sync", "fs", "io-util"] }
-tokio-util.workspace = true
+tokio-util = { workspace = true, features = ["compat"] }
 toml_edit.workspace = true
 tracing.workspace = true
 scopeguard.workspace = true

libs/remote_storage/src/azure_blob.rs

@@ -1,21 +1,24 @@
 //! Azure Blob Storage wrapper
 
+use std::borrow::Cow;
 use std::collections::HashMap;
 use std::env;
 use std::num::NonZeroU32;
+use std::pin::Pin;
 use std::sync::Arc;
-use std::{borrow::Cow, io::Cursor};
 
 use super::REMOTE_STORAGE_PREFIX_SEPARATOR;
 use anyhow::Result;
 use azure_core::request_options::{MaxResults, Metadata, Range};
+use azure_core::RetryOptions;
 use azure_identity::DefaultAzureCredential;
 use azure_storage::StorageCredentials;
 use azure_storage_blobs::prelude::ClientBuilder;
 use azure_storage_blobs::{blob::operations::GetBlobBuilder, prelude::ContainerClient};
+use bytes::Bytes;
+use futures::stream::Stream;
 use futures_util::StreamExt;
 use http_types::StatusCode;
-use tokio::io::AsyncRead;
 use tracing::debug;
 
 use crate::s3_bucket::RequestKind;
@@ -49,7 +52,8 @@ impl AzureBlobStorage {
             StorageCredentials::token_credential(Arc::new(token_credential))
         };
 
-        let builder = ClientBuilder::new(account, credentials);
+        // we have an outer retry
+        let builder = ClientBuilder::new(account, credentials).retry(RetryOptions::none());
 
         let client = builder.container_client(azure_config.container_name.to_owned());
 
@@ -116,7 +120,8 @@ impl AzureBlobStorage {
         let mut metadata = HashMap::new();
         // TODO give proper streaming response instead of buffering into RAM
         // https://github.com/neondatabase/neon/issues/5563
-        let mut buf = Vec::new();
+
+        let mut bufs = Vec::new();
         while let Some(part) = response.next().await {
             let part = part.map_err(to_download_error)?;
             if let Some(blob_meta) = part.blob.metadata {
@@ -127,10 +132,10 @@ impl AzureBlobStorage {
                 .collect()
                 .await
                 .map_err(|e| DownloadError::Other(e.into()))?;
-            buf.extend_from_slice(&data.slice(..));
+            bufs.push(data);
         }
         Ok(Download {
-            download_stream: Box::pin(Cursor::new(buf)),
+            download_stream: Box::pin(futures::stream::iter(bufs.into_iter().map(Ok))),
             metadata: Some(StorageMetadata(metadata)),
         })
     }
@@ -217,9 +222,10 @@ impl RemoteStorage for AzureBlobStorage {
         }
         Ok(res)
     }
+
     async fn upload(
         &self,
-        mut from: impl AsyncRead + Unpin + Send + Sync + 'static,
+        from: impl Stream<Item = std::io::Result<Bytes>> + Send + Sync + 'static,
         data_size_bytes: usize,
         to: &RemotePath,
         metadata: Option<StorageMetadata>,
@@ -227,13 +233,12 @@ impl RemoteStorage for AzureBlobStorage {
         let _permit = self.permit(RequestKind::Put).await;
         let blob_client = self.client.blob_client(self.relative_path_to_name(to));
 
-        // TODO FIX THIS UGLY HACK and don't buffer the entire object
-        // into RAM here, but use the streaming interface. For that,
-        // we'd have to change the interface though...
-        // https://github.com/neondatabase/neon/issues/5563
-        let mut buf = Vec::with_capacity(data_size_bytes);
-        tokio::io::copy(&mut from, &mut buf).await?;
-        let body = azure_core::Body::Bytes(buf.into());
+        let from: Pin<Box<dyn Stream<Item = std::io::Result<Bytes>> + Send + Sync + 'static>> =
+            Box::pin(from);
+
+        let from = NonSeekableStream::new(from, data_size_bytes);
+
+        let body = azure_core::Body::SeekableStream(Box::new(from));
 
         let mut builder = blob_client.put_block_blob(body);
 
@@ -312,3 +317,153 @@ impl RemoteStorage for AzureBlobStorage {
         Ok(())
     }
 }
+
+pin_project_lite::pin_project! {
+    /// Hack to work around not being able to stream once with azure sdk.
+    ///
+    /// Azure sdk clones streams around with the assumption that they are like
+    /// `Arc<tokio::fs::File>` (except not supporting tokio), however our streams are not like
+    /// that. For example for an `index_part.json` we just have a single chunk of [`Bytes`]
+    /// representing the whole serialized vec. It could be trivially cloneable and "semi-trivially"
+    /// seekable, but we can also just re-try the request easier.
+    #[project = NonSeekableStreamProj]
+    enum NonSeekableStream<S> {
+        /// A stream wrappers initial form.
+        ///
+        /// Mutex exists to allow moving when cloning. If the sdk changes to do less than 1
+        /// clone before first request, then this must be changed.
+        Initial {
+            inner: std::sync::Mutex<Option<tokio_util::compat::Compat<tokio_util::io::StreamReader<S, Bytes>>>>,
+            len: usize,
+        },
+        /// The actually readable variant, produced by cloning the Initial variant.
+        ///
+        /// The sdk currently always clones once, even without retry policy.
+        Actual {
+            #[pin]
+            inner: tokio_util::compat::Compat<tokio_util::io::StreamReader<S, Bytes>>,
+            len: usize,
+            read_any: bool,
+        },
+        /// Most likely unneeded, but left to make life easier, in case more clones are added.
+        Cloned {
+            len_was: usize,
+        }
+    }
+}
+
+impl<S> NonSeekableStream<S>
+where
+    S: Stream<Item = std::io::Result<Bytes>> + Send + Sync + 'static,
+{
+    fn new(inner: S, len: usize) -> NonSeekableStream<S> {
+        use tokio_util::compat::TokioAsyncReadCompatExt;
+
+        let inner = tokio_util::io::StreamReader::new(inner).compat();
+        let inner = Some(inner);
+        let inner = std::sync::Mutex::new(inner);
+        NonSeekableStream::Initial { inner, len }
+    }
+}
+
+impl<S> std::fmt::Debug for NonSeekableStream<S> {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        match self {
+            Self::Initial { len, .. } => f.debug_struct("Initial").field("len", len).finish(),
+            Self::Actual { len, .. } => f.debug_struct("Actual").field("len", len).finish(),
+            Self::Cloned { len_was, .. } => f.debug_struct("Cloned").field("len", len_was).finish(),
+        }
+    }
+}
+
+impl<S> futures::io::AsyncRead for NonSeekableStream<S>
+where
+    S: Stream<Item = std::io::Result<Bytes>>,
+{
+    fn poll_read(
+        self: std::pin::Pin<&mut Self>,
+        cx: &mut std::task::Context<'_>,
+        buf: &mut [u8],
+    ) -> std::task::Poll<std::io::Result<usize>> {
+        match self.project() {
+            NonSeekableStreamProj::Actual {
+                inner, read_any, ..
+            } => {
+                *read_any = true;
+                inner.poll_read(cx, buf)
+            }
+            // NonSeekableStream::Initial does not support reading because it is just much easier
+            // to have the mutex in place where one does not poll the contents, or that's how it
+            // seemed originally. If there is a version upgrade which changes the cloning, then
+            // that support needs to be hacked in.
+            //
+            // including {self:?} into the message would be useful, but unsure how to unproject.
+            _ => std::task::Poll::Ready(Err(std::io::Error::new(
+                std::io::ErrorKind::Other,
+                "cloned or initial values cannot be read",
+            ))),
+        }
+    }
+}
+
+impl<S> Clone for NonSeekableStream<S> {
+    /// Weird clone implementation exists to support the sdk doing cloning before issuing the first
+    /// request, see type documentation.
+    fn clone(&self) -> Self {
+        use NonSeekableStream::*;
+
+        match self {
+            Initial { inner, len } => {
+                if let Some(inner) = inner.lock().unwrap().take() {
+                    Actual {
+                        inner,
+                        len: *len,
+                        read_any: false,
+                    }
+                } else {
+                    Self::Cloned { len_was: *len }
+                }
+            }
+            Actual { len, .. } => Cloned { len_was: *len },
+            Cloned { len_was } => Cloned { len_was: *len_was },
+        }
+    }
+}
+
+#[async_trait::async_trait]
+impl<S> azure_core::SeekableStream for NonSeekableStream<S>
+where
+    S: Stream<Item = std::io::Result<Bytes>> + Unpin + Send + Sync + 'static,
+{
+    async fn reset(&mut self) -> azure_core::error::Result<()> {
+        use NonSeekableStream::*;
+
+        let msg = match self {
+            Initial { inner, .. } => {
+                if inner.get_mut().unwrap().is_some() {
+                    return Ok(());
+                } else {
+                    "reset after first clone is not supported"
+                }
+            }
+            Actual { read_any, .. } if !*read_any => return Ok(()),
+            Actual { .. } => "reset after reading is not supported",
+            Cloned { .. } => "reset after second clone is not supported",
+        };
+        Err(azure_core::error::Error::new(
+            azure_core::error::ErrorKind::Io,
+            std::io::Error::new(std::io::ErrorKind::Other, msg),
+        ))
+    }
+
+    // Note: it is not documented if this should be the total or remaining length, total passes the
+    // tests.
+    fn len(&self) -> usize {
+        use NonSeekableStream::*;
+        match self {
+            Initial { len, .. } => *len,
+            Actual { len, .. } => *len,
+            Cloned { len_was, .. } => *len_was,
+        }
+    }
+}

libs/remote_storage/src/lib.rs

@@ -19,8 +19,10 @@ use std::{collections::HashMap, fmt::Debug, num::NonZeroUsize, pin::Pin, sync::A
 use anyhow::{bail, Context};
 use camino::{Utf8Path, Utf8PathBuf};
 
+use bytes::Bytes;
+use futures::stream::Stream;
 use serde::{Deserialize, Serialize};
-use tokio::{io, sync::Semaphore};
+use tokio::sync::Semaphore;
 use toml_edit::Item;
 use tracing::info;
 
@@ -179,7 +181,7 @@ pub trait RemoteStorage: Send + Sync + 'static {
     /// Streams the local file contents into remote into the remote storage entry.
     async fn upload(
         &self,
-        from: impl io::AsyncRead + Unpin + Send + Sync + 'static,
+        from: impl Stream<Item = std::io::Result<Bytes>> + Send + Sync + 'static,
         // S3 PUT request requires the content length to be specified,
         // otherwise it starts to fail with the concurrent connection count increasing.
         data_size_bytes: usize,
@@ -206,7 +208,7 @@ pub trait RemoteStorage: Send + Sync + 'static {
 }
 
 pub struct Download {
-    pub download_stream: Pin<Box<dyn io::AsyncRead + Unpin + Send + Sync>>,
+    pub download_stream: Pin<Box<dyn Stream<Item = std::io::Result<Bytes>> + Unpin + Send + Sync>>,
     /// Extra key-value data, associated with the current remote file.
     pub metadata: Option<StorageMetadata>,
 }
@@ -300,7 +302,7 @@ impl GenericRemoteStorage {
 
     pub async fn upload(
         &self,
-        from: impl io::AsyncRead + Unpin + Send + Sync + 'static,
+        from: impl Stream<Item = std::io::Result<Bytes>> + Send + Sync + 'static,
         data_size_bytes: usize,
         to: &RemotePath,
         metadata: Option<StorageMetadata>,
@@ -398,7 +400,7 @@ impl GenericRemoteStorage {
     /// this path is used for the remote object id conversion only.
     pub async fn upload_storage_object(
         &self,
-        from: impl tokio::io::AsyncRead + Unpin + Send + Sync + 'static,
+        from: impl Stream<Item = std::io::Result<Bytes>> + Send + Sync + 'static,
         from_size_bytes: usize,
         to: &RemotePath,
     ) -> anyhow::Result<()> {

libs/remote_storage/src/local_fs.rs

@@ -7,11 +7,14 @@
 use std::{borrow::Cow, future::Future, io::ErrorKind, pin::Pin};
 
 use anyhow::{bail, ensure, Context};
+use bytes::Bytes;
 use camino::{Utf8Path, Utf8PathBuf};
+use futures::stream::Stream;
 use tokio::{
     fs,
     io::{self, AsyncReadExt, AsyncSeekExt, AsyncWriteExt},
 };
+use tokio_util::io::ReaderStream;
 use tracing::*;
 use utils::{crashsafe::path_with_suffix_extension, fs_ext::is_directory_empty};
 
@@ -99,27 +102,35 @@ impl LocalFs {
         };
 
         // If we were given a directory, we may use it as our starting point.
-        // Otherwise, we must go up to the parent directory.  This is because
+        // Otherwise, we must go up to the first ancestor dir that exists.  This is because
         // S3 object list prefixes can be arbitrary strings, but when reading
         // the local filesystem we need a directory to start calling read_dir on.
         let mut initial_dir = full_path.clone();
-        match fs::metadata(full_path.clone()).await {
-            Ok(meta) => {
-                if !meta.is_dir() {
+        loop {
+            // Did we make it to the root?
+            if initial_dir.parent().is_none() {
+                anyhow::bail!("list_files: failed to find valid ancestor dir for {full_path}");
+            }
+
+            match fs::metadata(initial_dir.clone()).await {
+                Ok(meta) if meta.is_dir() => {
+                    // We found a directory, break
+                    break;
+                }
+                Ok(_meta) => {
                     // It's not a directory: strip back to the parent
                     initial_dir.pop();
                 }
-            }
-            Err(e) if e.kind() == ErrorKind::NotFound => {
-                // It's not a file that exists: strip the prefix back to the parent directory
-                initial_dir.pop();
-            }
-            Err(e) => {
-                // Unexpected I/O error
-                anyhow::bail!(e)
+                Err(e) if e.kind() == ErrorKind::NotFound => {
+                    // It's not a file that exists: strip the prefix back to the parent directory
+                    initial_dir.pop();
+                }
+                Err(e) => {
+                    // Unexpected I/O error
+                    anyhow::bail!(e)
+                }
             }
         }
-
         // Note that Utf8PathBuf starts_with only considers full path segments, but
         // object prefixes are arbitrary strings, so we need the strings for doing
         // starts_with later.
@@ -211,7 +222,7 @@ impl RemoteStorage for LocalFs {
 
     async fn upload(
         &self,
-        data: impl io::AsyncRead + Unpin + Send + Sync + 'static,
+        data: impl Stream<Item = std::io::Result<Bytes>> + Send + Sync,
         data_size_bytes: usize,
         to: &RemotePath,
         metadata: Option<StorageMetadata>,
@@ -244,9 +255,12 @@ impl RemoteStorage for LocalFs {
         );
 
         let from_size_bytes = data_size_bytes as u64;
+        let data = tokio_util::io::StreamReader::new(data);
+        let data = std::pin::pin!(data);
         let mut buffer_to_read = data.take(from_size_bytes);
 
-        let bytes_read = io::copy(&mut buffer_to_read, &mut destination)
+        // alternatively we could just write the bytes to a file, but local_fs is a testing utility
+        let bytes_read = io::copy_buf(&mut buffer_to_read, &mut destination)
             .await
             .with_context(|| {
                 format!(
@@ -300,7 +314,7 @@ impl RemoteStorage for LocalFs {
     async fn download(&self, from: &RemotePath) -> Result<Download, DownloadError> {
         let target_path = from.with_base(&self.storage_root);
         if file_exists(&target_path).map_err(DownloadError::BadInput)? {
-            let source = io::BufReader::new(
+            let source = ReaderStream::new(
                 fs::OpenOptions::new()
                     .read(true)
                     .open(&target_path)
@@ -340,16 +354,14 @@ impl RemoteStorage for LocalFs {
         }
         let target_path = from.with_base(&self.storage_root);
         if file_exists(&target_path).map_err(DownloadError::BadInput)? {
-            let mut source = io::BufReader::new(
-                fs::OpenOptions::new()
-                    .read(true)
-                    .open(&target_path)
-                    .await
-                    .with_context(|| {
-                        format!("Failed to open source file {target_path:?} to use in the download")
-                    })
-                    .map_err(DownloadError::Other)?,
-            );
+            let mut source = tokio::fs::OpenOptions::new()
+                .read(true)
+                .open(&target_path)
+                .await
+                .with_context(|| {
+                    format!("Failed to open source file {target_path:?} to use in the download")
+                })
+                .map_err(DownloadError::Other)?;
             source
                 .seek(io::SeekFrom::Start(start_inclusive))
                 .await
@@ -363,11 +375,13 @@ impl RemoteStorage for LocalFs {
             Ok(match end_exclusive {
                 Some(end_exclusive) => Download {
                     metadata,
-                    download_stream: Box::pin(source.take(end_exclusive - start_inclusive)),
+                    download_stream: Box::pin(ReaderStream::new(
+                        source.take(end_exclusive - start_inclusive),
+                    )),
                 },
                 None => Download {
                     metadata,
-                    download_stream: Box::pin(source),
+                    download_stream: Box::pin(ReaderStream::new(source)),
                 },
             })
         } else {
@@ -467,7 +481,9 @@ fn file_exists(file_path: &Utf8Path) -> anyhow::Result<bool> {
 mod fs_tests {
     use super::*;
 
+    use bytes::Bytes;
     use camino_tempfile::tempdir;
+    use futures_util::Stream;
     use std::{collections::HashMap, io::Write};
 
     async fn read_and_assert_remote_file_contents(
@@ -477,7 +493,7 @@ mod fs_tests {
         remote_storage_path: &RemotePath,
         expected_metadata: Option<&StorageMetadata>,
     ) -> anyhow::Result<String> {
-        let mut download = storage
+        let download = storage
             .download(remote_storage_path)
             .await
             .map_err(|e| anyhow::anyhow!("Download failed: {e}"))?;
@@ -486,13 +502,9 @@ mod fs_tests {
             "Unexpected metadata returned for the downloaded file"
         );
 
-        let mut contents = String::new();
-        download
-            .download_stream
-            .read_to_string(&mut contents)
-            .await
-            .context("Failed to read remote file contents into string")?;
-        Ok(contents)
+        let contents = aggregate(download.download_stream).await?;
+
+        String::from_utf8(contents).map_err(anyhow::Error::new)
     }
 
     #[tokio::test]
@@ -521,25 +533,26 @@ mod fs_tests {
         let storage = create_storage()?;
 
         let id = RemotePath::new(Utf8Path::new("dummy"))?;
-        let content = std::io::Cursor::new(b"12345");
+        let content = Bytes::from_static(b"12345");
+        let content = move || futures::stream::once(futures::future::ready(Ok(content.clone())));
 
         // Check that you get an error if the size parameter doesn't match the actual
         // size of the stream.
         storage
-            .upload(Box::new(content.clone()), 0, &id, None)
+            .upload(content(), 0, &id, None)
             .await
             .expect_err("upload with zero size succeeded");
         storage
-            .upload(Box::new(content.clone()), 4, &id, None)
+            .upload(content(), 4, &id, None)
             .await
             .expect_err("upload with too short size succeeded");
         storage
-            .upload(Box::new(content.clone()), 6, &id, None)
+            .upload(content(), 6, &id, None)
             .await
             .expect_err("upload with too large size succeeded");
 
         // Correct size is 5, this should succeed.
-        storage.upload(Box::new(content), 5, &id, None).await?;
+        storage.upload(content(), 5, &id, None).await?;
 
         Ok(())
     }
@@ -587,7 +600,7 @@ mod fs_tests {
         let uploaded_bytes = dummy_contents(upload_name).into_bytes();
         let (first_part_local, second_part_local) = uploaded_bytes.split_at(3);
 
-        let mut first_part_download = storage
+        let first_part_download = storage
             .download_byte_range(&upload_target, 0, Some(first_part_local.len() as u64))
             .await?;
         assert!(
@@ -595,21 +608,13 @@ mod fs_tests {
             "No metadata should be returned for no metadata upload"
         );
 
-        let mut first_part_remote = io::BufWriter::new(std::io::Cursor::new(Vec::new()));
-        io::copy(
-            &mut first_part_download.download_stream,
-            &mut first_part_remote,
-        )
-        .await?;
-        first_part_remote.flush().await?;
-        let first_part_remote = first_part_remote.into_inner().into_inner();
+        let first_part_remote = aggregate(first_part_download.download_stream).await?;
         assert_eq!(
-            first_part_local,
-            first_part_remote.as_slice(),
+            first_part_local, first_part_remote,
             "First part bytes should be returned when requested"
         );
 
-        let mut second_part_download = storage
+        let second_part_download = storage
             .download_byte_range(
                 &upload_target,
                 first_part_local.len() as u64,
@@ -621,17 +626,9 @@ mod fs_tests {
             "No metadata should be returned for no metadata upload"
         );
 
-        let mut second_part_remote = io::BufWriter::new(std::io::Cursor::new(Vec::new()));
-        io::copy(
-            &mut second_part_download.download_stream,
-            &mut second_part_remote,
-        )
-        .await?;
-        second_part_remote.flush().await?;
-        let second_part_remote = second_part_remote.into_inner().into_inner();
+        let second_part_remote = aggregate(second_part_download.download_stream).await?;
         assert_eq!(
-            second_part_local,
-            second_part_remote.as_slice(),
+            second_part_local, second_part_remote,
             "Second part bytes should be returned when requested"
         );
 
@@ -721,17 +718,10 @@ mod fs_tests {
         let uploaded_bytes = dummy_contents(upload_name).into_bytes();
         let (first_part_local, _) = uploaded_bytes.split_at(3);
 
-        let mut partial_download_with_metadata = storage
+        let partial_download_with_metadata = storage
             .download_byte_range(&upload_target, 0, Some(first_part_local.len() as u64))
             .await?;
-        let mut first_part_remote = io::BufWriter::new(std::io::Cursor::new(Vec::new()));
-        io::copy(
-            &mut partial_download_with_metadata.download_stream,
-            &mut first_part_remote,
-        )
-        .await?;
-        first_part_remote.flush().await?;
-        let first_part_remote = first_part_remote.into_inner().into_inner();
+        let first_part_remote = aggregate(partial_download_with_metadata.download_stream).await?;
         assert_eq!(
             first_part_local,
             first_part_remote.as_slice(),
@@ -807,16 +797,16 @@ mod fs_tests {
                 )
             })?;
 
-        storage
-            .upload(Box::new(file), size, &relative_path, metadata)
-            .await?;
+        let file = tokio_util::io::ReaderStream::new(file);
+
+        storage.upload(file, size, &relative_path, metadata).await?;
         Ok(relative_path)
     }
 
     async fn create_file_for_upload(
         path: &Utf8Path,
         contents: &str,
-    ) -> anyhow::Result<(io::BufReader<fs::File>, usize)> {
+    ) -> anyhow::Result<(fs::File, usize)> {
         std::fs::create_dir_all(path.parent().unwrap())?;
         let mut file_for_writing = std::fs::OpenOptions::new()
             .write(true)
@@ -826,7 +816,7 @@ mod fs_tests {
         drop(file_for_writing);
         let file_size = path.metadata()?.len() as usize;
         Ok((
-            io::BufReader::new(fs::OpenOptions::new().read(true).open(&path).await?),
+            fs::OpenOptions::new().read(true).open(&path).await?,
             file_size,
         ))
     }
@@ -840,4 +830,16 @@ mod fs_tests {
         files.sort_by(|a, b| a.0.cmp(&b.0));
         Ok(files)
     }
+
+    async fn aggregate(
+        stream: impl Stream<Item = std::io::Result<Bytes>>,
+    ) -> anyhow::Result<Vec<u8>> {
+        use futures::stream::StreamExt;
+        let mut out = Vec::new();
+        let mut stream = std::pin::pin!(stream);
+        while let Some(res) = stream.next().await {
+            out.extend_from_slice(&res?[..]);
+        }
+        Ok(out)
+    }
 }

libs/remote_storage/src/s3_bucket.rs

@@ -4,9 +4,14 @@
 //! allowing multiple api users to independently work with the same S3 bucket, if
 //! their bucket prefixes are both specified and different.
 
-use std::{borrow::Cow, sync::Arc};
+use std::{
+    borrow::Cow,
+    pin::Pin,
+    sync::Arc,
+    task::{Context, Poll},
+};
 
-use anyhow::Context;
+use anyhow::Context as _;
 use aws_config::{
     environment::credentials::EnvironmentVariableCredentialsProvider,
     imds::credentials::ImdsCredentialsProvider,
@@ -14,23 +19,24 @@ use aws_config::{
     provider_config::ProviderConfig,
     retry::{RetryConfigBuilder, RetryMode},
     web_identity_token::WebIdentityTokenCredentialsProvider,
+    BehaviorVersion,
 };
-use aws_credential_types::cache::CredentialsCache;
+use aws_credential_types::provider::SharedCredentialsProvider;
 use aws_sdk_s3::{
-    config::{AsyncSleep, Config, Region, SharedAsyncSleep},
+    config::{AsyncSleep, Builder, IdentityCache, Region, SharedAsyncSleep},
     error::SdkError,
     operation::get_object::GetObjectError,
-    primitives::ByteStream,
     types::{Delete, ObjectIdentifier},
     Client,
 };
 use aws_smithy_async::rt::sleep::TokioSleep;
-use aws_smithy_http::body::SdkBody;
+
+use aws_smithy_types::body::SdkBody;
+use aws_smithy_types::byte_stream::ByteStream;
+use bytes::Bytes;
+use futures::stream::Stream;
 use hyper::Body;
 use scopeguard::ScopeGuard;
-use tokio::io::{self, AsyncRead};
-use tokio_util::io::ReaderStream;
-use tracing::debug;
 
 use super::StorageMetadata;
 use crate::{
@@ -61,7 +67,7 @@ struct GetObjectRequest {
 impl S3Bucket {
     /// Creates the S3 storage, errors if incorrect AWS S3 configuration provided.
     pub fn new(aws_config: &S3Config) -> anyhow::Result<Self> {
-        debug!(
+        tracing::debug!(
             "Creating s3 remote storage for S3 bucket {}",
             aws_config.bucket_name
         );
@@ -78,7 +84,6 @@ impl S3Bucket {
             // needed to access remote extensions bucket
             .or_else("token", {
                 let provider_conf = ProviderConfig::without_region().with_region(region.clone());
-
                 WebIdentityTokenCredentialsProvider::builder()
                     .configure(&provider_conf)
                     .build()
@@ -98,18 +103,20 @@ impl S3Bucket {
             .set_max_attempts(Some(1))
             .set_mode(Some(RetryMode::Adaptive));
 
-        let mut config_builder = Config::builder()
+        let mut config_builder = Builder::default()
+            .behavior_version(BehaviorVersion::v2023_11_09())
             .region(region)
-            .credentials_cache(CredentialsCache::lazy())
-            .credentials_provider(credentials_provider)
-            .sleep_impl(SharedAsyncSleep::from(sleep_impl))
-            .retry_config(retry_config.build());
+            .identity_cache(IdentityCache::lazy().build())
+            .credentials_provider(SharedCredentialsProvider::new(credentials_provider))
+            .retry_config(retry_config.build())
+            .sleep_impl(SharedAsyncSleep::from(sleep_impl));
 
         if let Some(custom_endpoint) = aws_config.endpoint.clone() {
             config_builder = config_builder
                 .endpoint_url(custom_endpoint)
                 .force_path_style(true);
         }
+
         let client = Client::from_conf(config_builder.build());
 
         let prefix_in_bucket = aws_config.prefix_in_bucket.as_deref().map(|prefix| {
@@ -222,12 +229,15 @@ impl S3Bucket {
         match get_object {
             Ok(object_output) => {
                 let metadata = object_output.metadata().cloned().map(StorageMetadata);
+
+                let body = object_output.body;
+                let body = ByteStreamAsStream::from(body);
+                let body = PermitCarrying::new(permit, body);
+                let body = TimedDownload::new(started_at, body);
+
                 Ok(Download {
                     metadata,
-                    download_stream: Box::pin(io::BufReader::new(TimedDownload::new(
-                        started_at,
-                        RatelimitedAsyncRead::new(permit, object_output.body.into_async_read()),
-                    ))),
+                    download_stream: Box::pin(body),
                 })
             }
             Err(SdkError::ServiceError(e)) if matches!(e.err(), GetObjectError::NoSuchKey(_)) => {
@@ -240,29 +250,55 @@ impl S3Bucket {
     }
 }
 
+pin_project_lite::pin_project! {
+    struct ByteStreamAsStream {
+        #[pin]
+        inner: aws_smithy_types::byte_stream::ByteStream
+    }
+}
+
+impl From<aws_smithy_types::byte_stream::ByteStream> for ByteStreamAsStream {
+    fn from(inner: aws_smithy_types::byte_stream::ByteStream) -> Self {
+        ByteStreamAsStream { inner }
+    }
+}
+
+impl Stream for ByteStreamAsStream {
+    type Item = std::io::Result<Bytes>;
+
+    fn poll_next(self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<Option<Self::Item>> {
+        // this does the std::io::ErrorKind::Other conversion
+        self.project().inner.poll_next(cx).map_err(|x| x.into())
+    }
+
+    // cannot implement size_hint because inner.size_hint is remaining size in bytes, which makes
+    // sense and Stream::size_hint does not really
+}
+
 pin_project_lite::pin_project! {
     /// An `AsyncRead` adapter which carries a permit for the lifetime of the value.
-    struct RatelimitedAsyncRead<S> {
+    struct PermitCarrying<S> {
         permit: tokio::sync::OwnedSemaphorePermit,
         #[pin]
         inner: S,
     }
 }
 
-impl<S: AsyncRead> RatelimitedAsyncRead<S> {
+impl<S> PermitCarrying<S> {
     fn new(permit: tokio::sync::OwnedSemaphorePermit, inner: S) -> Self {
-        RatelimitedAsyncRead { permit, inner }
+        Self { permit, inner }
     }
 }
 
-impl<S: AsyncRead> AsyncRead for RatelimitedAsyncRead<S> {
-    fn poll_read(
-        self: std::pin::Pin<&mut Self>,
-        cx: &mut std::task::Context<'_>,
-        buf: &mut io::ReadBuf<'_>,
-    ) -> std::task::Poll<std::io::Result<()>> {
-        let this = self.project();
-        this.inner.poll_read(cx, buf)
+impl<S: Stream<Item = std::io::Result<Bytes>>> Stream for PermitCarrying<S> {
+    type Item = <S as Stream>::Item;
+
+    fn poll_next(self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<Option<Self::Item>> {
+        self.project().inner.poll_next(cx)
+    }
+
+    fn size_hint(&self) -> (usize, Option<usize>) {
+        self.inner.size_hint()
     }
 }
 
@@ -282,7 +318,7 @@ pin_project_lite::pin_project! {
     }
 }
 
-impl<S: AsyncRead> TimedDownload<S> {
+impl<S> TimedDownload<S> {
     fn new(started_at: std::time::Instant, inner: S) -> Self {
         TimedDownload {
             started_at,
@@ -292,25 +328,26 @@ impl<S: AsyncRead> TimedDownload<S> {
     }
 }
 
-impl<S: AsyncRead> AsyncRead for TimedDownload<S> {
-    fn poll_read(
-        self: std::pin::Pin<&mut Self>,
-        cx: &mut std::task::Context<'_>,
-        buf: &mut io::ReadBuf<'_>,
-    ) -> std::task::Poll<std::io::Result<()>> {
+impl<S: Stream<Item = std::io::Result<Bytes>>> Stream for TimedDownload<S> {
+    type Item = <S as Stream>::Item;
+
+    fn poll_next(self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<Option<Self::Item>> {
+        use std::task::ready;
+
         let this = self.project();
-        let before = buf.filled().len();
-        let read = std::task::ready!(this.inner.poll_read(cx, buf));
 
-        let read_eof = buf.filled().len() == before;
-
-        match read {
-            Ok(()) if read_eof => *this.outcome = AttemptOutcome::Ok,
-            Ok(()) => { /* still in progress */ }
-            Err(_) => *this.outcome = AttemptOutcome::Err,
+        let res = ready!(this.inner.poll_next(cx));
+        match &res {
+            Some(Ok(_)) => {}
+            Some(Err(_)) => *this.outcome = metrics::AttemptOutcome::Err,
+            None => *this.outcome = metrics::AttemptOutcome::Ok,
         }
 
-        std::task::Poll::Ready(read)
+        Poll::Ready(res)
+    }
+
+    fn size_hint(&self) -> (usize, Option<usize>) {
+        self.inner.size_hint()
     }
 }
 
@@ -371,11 +408,11 @@ impl RemoteStorage for S3Bucket {
 
             let response = response?;
 
-            let keys = response.contents().unwrap_or_default();
+            let keys = response.contents();
             let empty = Vec::new();
             let prefixes = response.common_prefixes.as_ref().unwrap_or(&empty);
 
-            tracing::info!("list: {} prefixes, {} keys", prefixes.len(), keys.len());
+            tracing::debug!("list: {} prefixes, {} keys", prefixes.len(), keys.len());
 
             for object in keys {
                 let object_path = object.key().expect("response does not contain a key");
@@ -400,7 +437,7 @@ impl RemoteStorage for S3Bucket {
 
     async fn upload(
         &self,
-        from: impl io::AsyncRead + Unpin + Send + Sync + 'static,
+        from: impl Stream<Item = std::io::Result<Bytes>> + Send + Sync + 'static,
         from_size_bytes: usize,
         to: &RemotePath,
         metadata: Option<StorageMetadata>,
@@ -410,8 +447,8 @@ impl RemoteStorage for S3Bucket {
 
         let started_at = start_measuring_requests(kind);
 
-        let body = Body::wrap_stream(ReaderStream::new(from));
-        let bytes_stream = ByteStream::new(SdkBody::from(body));
+        let body = Body::wrap_stream(from);
+        let bytes_stream = ByteStream::new(SdkBody::from_body_0_4(body));
 
         let res = self
             .client
@@ -474,7 +511,7 @@ impl RemoteStorage for S3Bucket {
         for path in paths {
             let obj_id = ObjectIdentifier::builder()
                 .set_key(Some(self.relative_path_to_s3_object(path)))
-                .build();
+                .build()?;
             delete_objects.push(obj_id);
         }
 
@@ -485,7 +522,11 @@ impl RemoteStorage for S3Bucket {
                 .client
                 .delete_objects()
                 .bucket(self.bucket_name.clone())
-                .delete(Delete::builder().set_objects(Some(chunk.to_vec())).build())
+                .delete(
+                    Delete::builder()
+                        .set_objects(Some(chunk.to_vec()))
+                        .build()?,
+                )
                 .send()
                 .await;
 

libs/remote_storage/src/simulate_failures.rs

@@ -1,6 +1,8 @@
 //! This module provides a wrapper around a real RemoteStorage implementation that
 //! causes the first N attempts at each upload or download operatio to fail. For
 //! testing purposes.
+use bytes::Bytes;
+use futures::stream::Stream;
 use std::collections::hash_map::Entry;
 use std::collections::HashMap;
 use std::sync::Mutex;
@@ -108,7 +110,7 @@ impl RemoteStorage for UnreliableWrapper {
 
     async fn upload(
         &self,
-        data: impl tokio::io::AsyncRead + Unpin + Send + Sync + 'static,
+        data: impl Stream<Item = std::io::Result<Bytes>> + Send + Sync + 'static,
         // S3 PUT request requires the content length to be specified,
         // otherwise it starts to fail with the concurrent connection count increasing.
         data_size_bytes: usize,

libs/remote_storage/tests/test_real_azure.rs

@@ -7,7 +7,9 @@ use std::sync::Arc;
 use std::time::UNIX_EPOCH;
 
 use anyhow::Context;
+use bytes::Bytes;
 use camino::Utf8Path;
+use futures::stream::Stream;
 use once_cell::sync::OnceCell;
 use remote_storage::{
     AzureConfig, Download, GenericRemoteStorage, RemotePath, RemoteStorageConfig, RemoteStorageKind,
@@ -180,23 +182,14 @@ async fn azure_delete_objects_works(ctx: &mut MaybeEnabledAzure) -> anyhow::Resu
     let path3 = RemotePath::new(Utf8Path::new(format!("{}/path3", ctx.base_prefix).as_str()))
         .with_context(|| "RemotePath conversion")?;
 
-    let data1 = "remote blob data1".as_bytes();
-    let data1_len = data1.len();
-    let data2 = "remote blob data2".as_bytes();
-    let data2_len = data2.len();
-    let data3 = "remote blob data3".as_bytes();
-    let data3_len = data3.len();
-    ctx.client
-        .upload(std::io::Cursor::new(data1), data1_len, &path1, None)
-        .await?;
+    let (data, len) = upload_stream("remote blob data1".as_bytes().into());
+    ctx.client.upload(data, len, &path1, None).await?;
 
-    ctx.client
-        .upload(std::io::Cursor::new(data2), data2_len, &path2, None)
-        .await?;
+    let (data, len) = upload_stream("remote blob data2".as_bytes().into());
+    ctx.client.upload(data, len, &path2, None).await?;
 
-    ctx.client
-        .upload(std::io::Cursor::new(data3), data3_len, &path3, None)
-        .await?;
+    let (data, len) = upload_stream("remote blob data3".as_bytes().into());
+    ctx.client.upload(data, len, &path3, None).await?;
 
     ctx.client.delete_objects(&[path1, path2]).await?;
 
@@ -219,53 +212,56 @@ async fn azure_upload_download_works(ctx: &mut MaybeEnabledAzure) -> anyhow::Res
     let path = RemotePath::new(Utf8Path::new(format!("{}/file", ctx.base_prefix).as_str()))
         .with_context(|| "RemotePath conversion")?;
 
-    let data = "remote blob data here".as_bytes();
-    let data_len = data.len() as u64;
+    let orig = bytes::Bytes::from_static("remote blob data here".as_bytes());
 
-    ctx.client
-        .upload(std::io::Cursor::new(data), data.len(), &path, None)
-        .await?;
+    let (data, len) = wrap_stream(orig.clone());
 
-    async fn download_and_compare(mut dl: Download) -> anyhow::Result<Vec<u8>> {
+    ctx.client.upload(data, len, &path, None).await?;
+
+    async fn download_and_compare(dl: Download) -> anyhow::Result<Vec<u8>> {
         let mut buf = Vec::new();
-        tokio::io::copy(&mut dl.download_stream, &mut buf).await?;
+        tokio::io::copy_buf(
+            &mut tokio_util::io::StreamReader::new(dl.download_stream),
+            &mut buf,
+        )
+        .await?;
         Ok(buf)
     }
     // Normal download request
     let dl = ctx.client.download(&path).await?;
     let buf = download_and_compare(dl).await?;
-    assert_eq!(buf, data);
+    assert_eq!(&buf, &orig);
 
     // Full range (end specified)
     let dl = ctx
         .client
-        .download_byte_range(&path, 0, Some(data_len))
+        .download_byte_range(&path, 0, Some(len as u64))
         .await?;
     let buf = download_and_compare(dl).await?;
-    assert_eq!(buf, data);
+    assert_eq!(&buf, &orig);
 
     // partial range (end specified)
     let dl = ctx.client.download_byte_range(&path, 4, Some(10)).await?;
     let buf = download_and_compare(dl).await?;
-    assert_eq!(buf, data[4..10]);
+    assert_eq!(&buf, &orig[4..10]);
 
     // partial range (end beyond real end)
     let dl = ctx
         .client
-        .download_byte_range(&path, 8, Some(data_len * 100))
+        .download_byte_range(&path, 8, Some(len as u64 * 100))
         .await?;
     let buf = download_and_compare(dl).await?;
-    assert_eq!(buf, data[8..]);
+    assert_eq!(&buf, &orig[8..]);
 
     // Partial range (end unspecified)
     let dl = ctx.client.download_byte_range(&path, 4, None).await?;
     let buf = download_and_compare(dl).await?;
-    assert_eq!(buf, data[4..]);
+    assert_eq!(&buf, &orig[4..]);
 
     // Full range (end unspecified)
     let dl = ctx.client.download_byte_range(&path, 0, None).await?;
     let buf = download_and_compare(dl).await?;
-    assert_eq!(buf, data);
+    assert_eq!(&buf, &orig);
 
     debug!("Cleanup: deleting file at path {path:?}");
     ctx.client
@@ -504,11 +500,8 @@ async fn upload_azure_data(
             let blob_path = blob_prefix.join(Utf8Path::new(&format!("blob_{i}")));
             debug!("Creating remote item {i} at path {blob_path:?}");
 
-            let data = format!("remote blob data {i}").into_bytes();
-            let data_len = data.len();
-            task_client
-                .upload(std::io::Cursor::new(data), data_len, &blob_path, None)
-                .await?;
+            let (data, len) = upload_stream(format!("remote blob data {i}").into_bytes().into());
+            task_client.upload(data, len, &blob_path, None).await?;
 
             Ok::<_, anyhow::Error>((blob_prefix, blob_path))
         });
@@ -589,11 +582,8 @@ async fn upload_simple_azure_data(
             .with_context(|| format!("{blob_path:?} to RemotePath conversion"))?;
             debug!("Creating remote item {i} at path {blob_path:?}");
 
-            let data = format!("remote blob data {i}").into_bytes();
-            let data_len = data.len();
-            task_client
-                .upload(std::io::Cursor::new(data), data_len, &blob_path, None)
-                .await?;
+            let (data, len) = upload_stream(format!("remote blob data {i}").into_bytes().into());
+            task_client.upload(data, len, &blob_path, None).await?;
 
             Ok::<_, anyhow::Error>(blob_path)
         });
@@ -622,3 +612,32 @@ async fn upload_simple_azure_data(
         ControlFlow::Continue(uploaded_blobs)
     }
 }
+
+// FIXME: copypasted from test_real_s3, can't remember how to share a module which is not compiled
+// to binary
+fn upload_stream(
+    content: std::borrow::Cow<'static, [u8]>,
+) -> (
+    impl Stream<Item = std::io::Result<Bytes>> + Send + Sync + 'static,
+    usize,
+) {
+    use std::borrow::Cow;
+
+    let content = match content {
+        Cow::Borrowed(x) => Bytes::from_static(x),
+        Cow::Owned(vec) => Bytes::from(vec),
+    };
+    wrap_stream(content)
+}
+
+fn wrap_stream(
+    content: bytes::Bytes,
+) -> (
+    impl Stream<Item = std::io::Result<Bytes>> + Send + Sync + 'static,
+    usize,
+) {
+    let len = content.len();
+    let content = futures::future::ready(Ok(content));
+
+    (futures::stream::once(content), len)
+}

libs/remote_storage/tests/test_real_s3.rs

@@ -7,7 +7,9 @@ use std::sync::Arc;
 use std::time::UNIX_EPOCH;
 
 use anyhow::Context;
+use bytes::Bytes;
 use camino::Utf8Path;
+use futures::stream::Stream;
 use once_cell::sync::OnceCell;
 use remote_storage::{
     GenericRemoteStorage, RemotePath, RemoteStorageConfig, RemoteStorageKind, S3Config,
@@ -176,23 +178,14 @@ async fn s3_delete_objects_works(ctx: &mut MaybeEnabledS3) -> anyhow::Result<()>
     let path3 = RemotePath::new(Utf8Path::new(format!("{}/path3", ctx.base_prefix).as_str()))
         .with_context(|| "RemotePath conversion")?;
 
-    let data1 = "remote blob data1".as_bytes();
-    let data1_len = data1.len();
-    let data2 = "remote blob data2".as_bytes();
-    let data2_len = data2.len();
-    let data3 = "remote blob data3".as_bytes();
-    let data3_len = data3.len();
-    ctx.client
-        .upload(std::io::Cursor::new(data1), data1_len, &path1, None)
-        .await?;
+    let (data, len) = upload_stream("remote blob data1".as_bytes().into());
+    ctx.client.upload(data, len, &path1, None).await?;
 
-    ctx.client
-        .upload(std::io::Cursor::new(data2), data2_len, &path2, None)
-        .await?;
+    let (data, len) = upload_stream("remote blob data2".as_bytes().into());
+    ctx.client.upload(data, len, &path2, None).await?;
 
-    ctx.client
-        .upload(std::io::Cursor::new(data3), data3_len, &path3, None)
-        .await?;
+    let (data, len) = upload_stream("remote blob data3".as_bytes().into());
+    ctx.client.upload(data, len, &path3, None).await?;
 
     ctx.client.delete_objects(&[path1, path2]).await?;
 
@@ -432,11 +425,9 @@ async fn upload_s3_data(
             let blob_path = blob_prefix.join(Utf8Path::new(&format!("blob_{i}")));
             debug!("Creating remote item {i} at path {blob_path:?}");
 
-            let data = format!("remote blob data {i}").into_bytes();
-            let data_len = data.len();
-            task_client
-                .upload(std::io::Cursor::new(data), data_len, &blob_path, None)
-                .await?;
+            let (data, data_len) =
+                upload_stream(format!("remote blob data {i}").into_bytes().into());
+            task_client.upload(data, data_len, &blob_path, None).await?;
 
             Ok::<_, anyhow::Error>((blob_prefix, blob_path))
         });
@@ -517,11 +508,9 @@ async fn upload_simple_s3_data(
             .with_context(|| format!("{blob_path:?} to RemotePath conversion"))?;
             debug!("Creating remote item {i} at path {blob_path:?}");
 
-            let data = format!("remote blob data {i}").into_bytes();
-            let data_len = data.len();
-            task_client
-                .upload(std::io::Cursor::new(data), data_len, &blob_path, None)
-                .await?;
+            let (data, data_len) =
+                upload_stream(format!("remote blob data {i}").into_bytes().into());
+            task_client.upload(data, data_len, &blob_path, None).await?;
 
             Ok::<_, anyhow::Error>(blob_path)
         });
@@ -550,3 +539,30 @@ async fn upload_simple_s3_data(
         ControlFlow::Continue(uploaded_blobs)
     }
 }
+
+fn upload_stream(
+    content: std::borrow::Cow<'static, [u8]>,
+) -> (
+    impl Stream<Item = std::io::Result<Bytes>> + Send + Sync + 'static,
+    usize,
+) {
+    use std::borrow::Cow;
+
+    let content = match content {
+        Cow::Borrowed(x) => Bytes::from_static(x),
+        Cow::Owned(vec) => Bytes::from(vec),
+    };
+    wrap_stream(content)
+}
+
+fn wrap_stream(
+    content: bytes::Bytes,
+) -> (
+    impl Stream<Item = std::io::Result<Bytes>> + Send + Sync + 'static,
+    usize,
+) {
+    let len = content.len();
+    let content = futures::future::ready(Ok(content));
+
+    (futures::stream::once(content), len)
+}

libs/utils/src/completion.rs

@@ -1,16 +1,14 @@
-use std::sync::Arc;
-
-use tokio::sync::{mpsc, Mutex};
+use tokio_util::task::{task_tracker::TaskTrackerToken, TaskTracker};
 
 /// While a reference is kept around, the associated [`Barrier::wait`] will wait.
 ///
 /// Can be cloned, moved and kept around in futures as "guard objects".
 #[derive(Clone)]
-pub struct Completion(mpsc::Sender<()>);
+pub struct Completion(TaskTrackerToken);
 
 /// Barrier will wait until all clones of [`Completion`] have been dropped.
 #[derive(Clone)]
-pub struct Barrier(Arc<Mutex<mpsc::Receiver<()>>>);
+pub struct Barrier(TaskTracker);
 
 impl Default for Barrier {
     fn default() -> Self {
@@ -21,7 +19,7 @@ impl Default for Barrier {
 
 impl Barrier {
     pub async fn wait(self) {
-        self.0.lock().await.recv().await;
+        self.0.wait().await;
     }
 
     pub async fn maybe_wait(barrier: Option<Barrier>) {
@@ -33,8 +31,7 @@ impl Barrier {
 
 impl PartialEq for Barrier {
     fn eq(&self, other: &Self) -> bool {
-        // we don't use dyn so this is good
-        Arc::ptr_eq(&self.0, &other.0)
+        TaskTracker::ptr_eq(&self.0, &other.0)
     }
 }
 
@@ -42,8 +39,10 @@ impl Eq for Barrier {}
 
 /// Create new Guard and Barrier pair.
 pub fn channel() -> (Completion, Barrier) {
-    let (tx, rx) = mpsc::channel::<()>(1);
-    let rx = Mutex::new(rx);
-    let rx = Arc::new(rx);
-    (Completion(tx), Barrier(rx))
+    let tracker = TaskTracker::new();
+    // otherwise wait never exits
+    tracker.close();
+
+    let token = tracker.token();
+    (Completion(token), Barrier(tracker))
 }

libs/utils/src/generation.rs

@@ -152,3 +152,16 @@ impl Debug for Generation {
         }
     }
 }
+
+#[cfg(test)]
+mod test {
+    use super::*;
+
+    #[test]
+    fn generation_gt() {
+        // Important that a None generation compares less than a valid one, during upgrades from
+        // pre-generation systems.
+        assert!(Generation::none() < Generation::new(0));
+        assert!(Generation::none() < Generation::new(1));
+    }
+}

libs/utils/src/simple_rcu.rs

@@ -1,10 +1,10 @@
 //!
 //! RCU stands for Read-Copy-Update. It's a synchronization mechanism somewhat
 //! similar to a lock, but it allows readers to "hold on" to an old value of RCU
-//! without blocking writers, and allows writing a new values without blocking
-//! readers. When you update the new value, the new value is immediately visible
+//! without blocking writers, and allows writing a new value without blocking
+//! readers. When you update the value, the new value is immediately visible
 //! to new readers, but the update waits until all existing readers have
-//! finishe, so that no one sees the old value anymore.
+//! finished, so that on return, no one sees the old value anymore.
 //!
 //! This implementation isn't wait-free; it uses an RwLock that is held for a
 //! short duration when the value is read or updated.
@@ -26,6 +26,7 @@
 //! Increment the value by one, and wait for old readers to finish:
 //!
 //! ```
+//! # async fn dox() {
 //! # let rcu = utils::simple_rcu::Rcu::new(1);
 //! let write_guard = rcu.lock_for_write();
 //!
@@ -36,15 +37,17 @@
 //!
 //! // Concurrent reads and writes are now possible again. Wait for all the readers
 //! // that still observe the old value to finish.
-//! waitlist.wait();
+//! waitlist.wait().await;
+//! # }
 //! ```
 //!
 #![warn(missing_docs)]
 
 use std::ops::Deref;
-use std::sync::mpsc::{sync_channel, Receiver, SyncSender};
 use std::sync::{Arc, Weak};
-use std::sync::{Mutex, RwLock, RwLockWriteGuard};
+use std::sync::{RwLock, RwLockWriteGuard};
+
+use tokio::sync::watch;
 
 ///
 /// Rcu allows multiple readers to read and hold onto a value without blocking
@@ -68,22 +71,21 @@ struct RcuCell<V> {
     value: V,
 
     /// A dummy channel. We never send anything to this channel. The point is
-    /// that when the RcuCell is dropped, any cloned Senders will be notified
+    /// that when the RcuCell is dropped, any subscribed Receivers will be notified
     /// that the channel is closed. Updaters can use this to wait out until the
     /// RcuCell has been dropped, i.e. until the old value is no longer in use.
     ///
-    /// We never do anything with the receiver, we just need to hold onto it so
-    /// that the Senders will be notified when it's dropped. But because it's
-    /// not Sync, we need a Mutex on it.
-    watch: (SyncSender<()>, Mutex<Receiver<()>>),
+    /// We never send anything to this, we just need to hold onto it so that the
+    /// Receivers will be notified when it's dropped.
+    watch: watch::Sender<()>,
 }
 
 impl<V> RcuCell<V> {
     fn new(value: V) -> Self {
-        let (watch_sender, watch_receiver) = sync_channel(0);
+        let (watch_sender, _) = watch::channel(());
         RcuCell {
             value,
-            watch: (watch_sender, Mutex::new(watch_receiver)),
+            watch: watch_sender,
         }
     }
 }
@@ -141,10 +143,10 @@ impl<V> Deref for RcuReadGuard<V> {
 ///
 /// Write guard returned by `write`
 ///
-/// NB: Holding this guard blocks all concurrent `read` and `write` calls, so
-/// it should only be held for a short duration!
+/// NB: Holding this guard blocks all concurrent `read` and `write` calls, so it should only be
+/// held for a short duration!
 ///
-/// Calling `store` consumes the guard, making new reads and new writes possible
+/// Calling [`Self::store_and_unlock`] consumes the guard, making new reads and new writes possible
 /// again.
 ///
 pub struct RcuWriteGuard<'a, V> {
@@ -179,7 +181,7 @@ impl<'a, V> RcuWriteGuard<'a, V> {
             // the watches for any that do.
             self.inner.old_cells.retain(|weak| {
                 if let Some(cell) = weak.upgrade() {
-                    watches.push(cell.watch.0.clone());
+                    watches.push(cell.watch.subscribe());
                     true
                 } else {
                     false
@@ -193,20 +195,20 @@ impl<'a, V> RcuWriteGuard<'a, V> {
 ///
 /// List of readers who can still see old values.
 ///
-pub struct RcuWaitList(Vec<SyncSender<()>>);
+pub struct RcuWaitList(Vec<watch::Receiver<()>>);
 
 impl RcuWaitList {
     ///
     /// Wait for old readers to finish.
     ///
-    pub fn wait(mut self) {
+    pub async fn wait(mut self) {
         // after all the old_cells are no longer in use, we're done
         for w in self.0.iter_mut() {
             // This will block until the Receiver is closed. That happens when
             // the RcuCell is dropped.
             #[allow(clippy::single_match)]
-            match w.send(()) {
-                Ok(_) => panic!("send() unexpectedly succeeded on dummy channel"),
+            match w.changed().await {
+                Ok(_) => panic!("changed() unexpectedly succeeded on dummy channel"),
                 Err(_) => {
                     // closed, which means that the cell has been dropped, and
                     // its value is no longer in use
@@ -220,11 +222,10 @@ impl RcuWaitList {
 mod tests {
     use super::*;
     use std::sync::{Arc, Mutex};
-    use std::thread::{sleep, spawn};
     use std::time::Duration;
 
-    #[test]
-    fn two_writers() {
+    #[tokio::test]
+    async fn two_writers() {
         let rcu = Rcu::new(1);
 
         let read1 = rcu.read();
@@ -248,33 +249,35 @@ mod tests {
         assert_eq!(*read1, 1);
 
         let log = Arc::new(Mutex::new(Vec::new()));
-        // Wait for the old readers to finish in separate threads.
+        // Wait for the old readers to finish in separate tasks.
         let log_clone = Arc::clone(&log);
-        let thread2 = spawn(move || {
-            wait2.wait();
+        let task2 = tokio::spawn(async move {
+            wait2.wait().await;
             log_clone.lock().unwrap().push("wait2 done");
         });
         let log_clone = Arc::clone(&log);
-        let thread3 = spawn(move || {
-            wait3.wait();
+        let task3 = tokio::spawn(async move {
+            wait3.wait().await;
             log_clone.lock().unwrap().push("wait3 done");
         });
 
         // without this sleep the test can pass on accident if the writer is slow
-        sleep(Duration::from_millis(500));
+        tokio::time::sleep(Duration::from_millis(100)).await;
 
         // Release first reader. This allows first write to finish, but calling
-        // wait() on the second one would still block.
+        // wait() on the 'task3' would still block.
         log.lock().unwrap().push("dropping read1");
         drop(read1);
-        thread2.join().unwrap();
+        task2.await.unwrap();
 
-        sleep(Duration::from_millis(500));
+        assert!(!task3.is_finished());
+
+        tokio::time::sleep(Duration::from_millis(100)).await;
 
         // Release second reader, and finish second writer.
         log.lock().unwrap().push("dropping read2");
         drop(read2);
-        thread3.join().unwrap();
+        task3.await.unwrap();
 
         assert_eq!(
             log.lock().unwrap().as_slice(),

pageserver/Cargo.toml

@@ -51,6 +51,7 @@ regex.workspace = true
 scopeguard.workspace = true
 serde.workspace = true
 serde_json = { workspace = true, features = ["raw_value"] }
+serde_path_to_error.workspace = true
 serde_with.workspace = true
 signal-hook.workspace = true
 smallvec = { workspace = true, features = ["write"] }

pageserver/benches/bench_layer_map.rs

@@ -3,6 +3,7 @@ use pageserver::repository::Key;
 use pageserver::tenant::layer_map::LayerMap;
 use pageserver::tenant::storage_layer::LayerFileName;
 use pageserver::tenant::storage_layer::PersistentLayerDesc;
+use pageserver_api::shard::TenantShardId;
 use rand::prelude::{SeedableRng, SliceRandom, StdRng};
 use std::cmp::{max, min};
 use std::fs::File;
@@ -211,7 +212,7 @@ fn bench_sequential(c: &mut Criterion) {
         let i32 = (i as u32) % 100;
         let zero = Key::from_hex("000000000000000000000000000000000000").unwrap();
         let layer = PersistentLayerDesc::new_img(
-            TenantId::generate(),
+            TenantShardId::unsharded(TenantId::generate()),
             TimelineId::generate(),
             zero.add(10 * i32)..zero.add(10 * i32 + 1),
             Lsn(i),

pageserver/ctl/src/layers.rs

@@ -1,13 +1,15 @@
 use std::path::{Path, PathBuf};
 
 use anyhow::Result;
-use camino::Utf8Path;
+use camino::{Utf8Path, Utf8PathBuf};
 use clap::Subcommand;
 use pageserver::context::{DownloadBehavior, RequestContext};
 use pageserver::task_mgr::TaskKind;
 use pageserver::tenant::block_io::BlockCursor;
 use pageserver::tenant::disk_btree::DiskBtreeReader;
 use pageserver::tenant::storage_layer::delta_layer::{BlobRef, Summary};
+use pageserver::tenant::storage_layer::{delta_layer, image_layer};
+use pageserver::tenant::storage_layer::{DeltaLayer, ImageLayer};
 use pageserver::tenant::{TENANTS_SEGMENT_NAME, TIMELINES_SEGMENT_NAME};
 use pageserver::{page_cache, virtual_file};
 use pageserver::{
@@ -20,6 +22,7 @@ use pageserver::{
 };
 use std::fs;
 use utils::bin_ser::BeSer;
+use utils::id::{TenantId, TimelineId};
 
 use crate::layer_map_analyzer::parse_filename;
 
@@ -45,6 +48,13 @@ pub(crate) enum LayerCmd {
         /// The id from list-layer command
         id: usize,
     },
+    RewriteSummary {
+        layer_file_path: Utf8PathBuf,
+        #[clap(long)]
+        new_tenant_id: Option<TenantId>,
+        #[clap(long)]
+        new_timeline_id: Option<TimelineId>,
+    },
 }
 
 async fn read_delta_file(path: impl AsRef<Path>, ctx: &RequestContext) -> Result<()> {
@@ -100,6 +110,7 @@ pub(crate) async fn main(cmd: &LayerCmd) -> Result<()> {
                     println!("- timeline {}", timeline.file_name().to_string_lossy());
                 }
             }
+            Ok(())
         }
         LayerCmd::ListLayer {
             path,
@@ -128,6 +139,7 @@ pub(crate) async fn main(cmd: &LayerCmd) -> Result<()> {
                     idx += 1;
                 }
             }
+            Ok(())
         }
         LayerCmd::DumpLayer {
             path,
@@ -168,7 +180,63 @@ pub(crate) async fn main(cmd: &LayerCmd) -> Result<()> {
                     idx += 1;
                 }
             }
+            Ok(())
+        }
+        LayerCmd::RewriteSummary {
+            layer_file_path,
+            new_tenant_id,
+            new_timeline_id,
+        } => {
+            pageserver::virtual_file::init(10);
+            pageserver::page_cache::init(100);
+
+            let ctx = RequestContext::new(TaskKind::DebugTool, DownloadBehavior::Error);
+
+            macro_rules! rewrite_closure {
+                ($($summary_ty:tt)*) => {{
+                    |summary| $($summary_ty)* {
+                        tenant_id: new_tenant_id.unwrap_or(summary.tenant_id),
+                        timeline_id: new_timeline_id.unwrap_or(summary.timeline_id),
+                        ..summary
+                    }
+                }};
+            }
+
+            let res = ImageLayer::rewrite_summary(
+                layer_file_path,
+                rewrite_closure!(image_layer::Summary),
+                &ctx,
+            )
+            .await;
+            match res {
+                Ok(()) => {
+                    println!("Successfully rewrote summary of image layer {layer_file_path}");
+                    return Ok(());
+                }
+                Err(image_layer::RewriteSummaryError::MagicMismatch) => (), // fallthrough
+                Err(image_layer::RewriteSummaryError::Other(e)) => {
+                    return Err(e);
+                }
+            }
+
+            let res = DeltaLayer::rewrite_summary(
+                layer_file_path,
+                rewrite_closure!(delta_layer::Summary),
+                &ctx,
+            )
+            .await;
+            match res {
+                Ok(()) => {
+                    println!("Successfully rewrote summary of delta layer {layer_file_path}");
+                    return Ok(());
+                }
+                Err(delta_layer::RewriteSummaryError::MagicMismatch) => (), // fallthrough
+                Err(delta_layer::RewriteSummaryError::Other(e)) => {
+                    return Err(e);
+                }
+            }
+
+            anyhow::bail!("not an image or delta layer: {layer_file_path}");
         }
     }
-    Ok(())
 }

pageserver/src/bin/pageserver.rs

@@ -402,15 +402,11 @@ fn start_pageserver(
     let (init_remote_done_tx, init_remote_done_rx) = utils::completion::channel();
     let (init_done_tx, init_done_rx) = utils::completion::channel();
 
-    let (init_logical_size_done_tx, init_logical_size_done_rx) = utils::completion::channel();
-
     let (background_jobs_can_start, background_jobs_barrier) = utils::completion::channel();
 
     let order = pageserver::InitializationOrder {
         initial_tenant_load_remote: Some(init_done_tx),
         initial_tenant_load: Some(init_remote_done_tx),
-        initial_logical_size_can_start: init_done_rx.clone(),
-        initial_logical_size_attempt: Some(init_logical_size_done_tx),
         background_jobs_can_start: background_jobs_barrier.clone(),
     };
 
@@ -429,7 +425,6 @@ fn start_pageserver(
     let tenant_manager = Arc::new(tenant_manager);
 
     BACKGROUND_RUNTIME.spawn({
-        let init_done_rx = init_done_rx;
         let shutdown_pageserver = shutdown_pageserver.clone();
         let drive_init = async move {
             // NOTE: unlike many futures in pageserver, this one is cancellation-safe
@@ -464,7 +459,7 @@ fn start_pageserver(
             });
 
             let WaitForPhaseResult {
-                timeout_remaining: timeout,
+                timeout_remaining: _timeout,
                 skipped: init_load_skipped,
             } = wait_for_phase("initial_tenant_load", init_load_done, timeout).await;
 
@@ -472,26 +467,6 @@ fn start_pageserver(
 
             scopeguard::ScopeGuard::into_inner(guard);
 
-            let guard = scopeguard::guard_on_success((), |_| {
-                tracing::info!("Cancelled before initial logical sizes completed")
-            });
-
-            let logical_sizes_done = std::pin::pin!(async {
-                init_logical_size_done_rx.wait().await;
-                startup_checkpoint(
-                    started_startup_at,
-                    "initial_logical_sizes",
-                    "Initial logical sizes completed",
-                );
-            });
-
-            let WaitForPhaseResult {
-                timeout_remaining: _,
-                skipped: logical_sizes_skipped,
-            } = wait_for_phase("initial_logical_sizes", logical_sizes_done, timeout).await;
-
-            scopeguard::ScopeGuard::into_inner(guard);
-
             // allow background jobs to start: we either completed prior stages, or they reached timeout
             // and were skipped.  It is important that we do not let them block background jobs indefinitely,
             // because things like consumption metrics for billing are blocked by this barrier.
@@ -514,9 +489,6 @@ fn start_pageserver(
             if let Some(f) = init_load_skipped {
                 f.await;
             }
-            if let Some(f) = logical_sizes_skipped {
-                f.await;
-            }
             scopeguard::ScopeGuard::into_inner(guard);
 
             startup_checkpoint(started_startup_at, "complete", "Startup complete");
@@ -587,7 +559,6 @@ fn start_pageserver(
     }
 
     if let Some(metric_collection_endpoint) = &conf.metric_collection_endpoint {
-        let background_jobs_barrier = background_jobs_barrier;
         let metrics_ctx = RequestContext::todo_child(
             TaskKind::MetricsCollection,
             // This task itself shouldn't download anything.

pageserver/src/config.rs

@@ -5,6 +5,7 @@
 //! See also `settings.md` for better description on every parameter.
 
 use anyhow::{anyhow, bail, ensure, Context, Result};
+use pageserver_api::shard::TenantShardId;
 use remote_storage::{RemotePath, RemoteStorageConfig};
 use serde::de::IntoDeserializer;
 use std::env;
@@ -25,7 +26,7 @@ use toml_edit::{Document, Item};
 use camino::{Utf8Path, Utf8PathBuf};
 use postgres_backend::AuthType;
 use utils::{
-    id::{NodeId, TenantId, TimelineId},
+    id::{NodeId, TimelineId},
     logging::LogFormat,
 };
 
@@ -628,12 +629,13 @@ impl PageServerConf {
         self.deletion_prefix().join(format!("header-{VERSION:02x}"))
     }
 
-    pub fn tenant_path(&self, tenant_id: &TenantId) -> Utf8PathBuf {
-        self.tenants_path().join(tenant_id.to_string())
+    pub fn tenant_path(&self, tenant_shard_id: &TenantShardId) -> Utf8PathBuf {
+        self.tenants_path().join(tenant_shard_id.to_string())
     }
 
-    pub fn tenant_ignore_mark_file_path(&self, tenant_id: &TenantId) -> Utf8PathBuf {
-        self.tenant_path(tenant_id).join(IGNORED_TENANT_FILE_NAME)
+    pub fn tenant_ignore_mark_file_path(&self, tenant_shard_id: &TenantShardId) -> Utf8PathBuf {
+        self.tenant_path(tenant_shard_id)
+            .join(IGNORED_TENANT_FILE_NAME)
     }
 
     /// Points to a place in pageserver's local directory,
@@ -641,47 +643,53 @@ impl PageServerConf {
     ///
     /// Legacy: superseded by tenant_location_config_path.  Eventually
     /// remove this function.
-    pub fn tenant_config_path(&self, tenant_id: &TenantId) -> Utf8PathBuf {
-        self.tenant_path(tenant_id).join(TENANT_CONFIG_NAME)
+    pub fn tenant_config_path(&self, tenant_shard_id: &TenantShardId) -> Utf8PathBuf {
+        self.tenant_path(tenant_shard_id).join(TENANT_CONFIG_NAME)
     }
 
-    pub fn tenant_location_config_path(&self, tenant_id: &TenantId) -> Utf8PathBuf {
-        self.tenant_path(tenant_id)
+    pub fn tenant_location_config_path(&self, tenant_shard_id: &TenantShardId) -> Utf8PathBuf {
+        self.tenant_path(tenant_shard_id)
             .join(TENANT_LOCATION_CONFIG_NAME)
     }
 
-    pub fn timelines_path(&self, tenant_id: &TenantId) -> Utf8PathBuf {
-        self.tenant_path(tenant_id).join(TIMELINES_SEGMENT_NAME)
+    pub fn timelines_path(&self, tenant_shard_id: &TenantShardId) -> Utf8PathBuf {
+        self.tenant_path(tenant_shard_id)
+            .join(TIMELINES_SEGMENT_NAME)
     }
 
-    pub fn timeline_path(&self, tenant_id: &TenantId, timeline_id: &TimelineId) -> Utf8PathBuf {
-        self.timelines_path(tenant_id).join(timeline_id.to_string())
+    pub fn timeline_path(
+        &self,
+        tenant_shard_id: &TenantShardId,
+        timeline_id: &TimelineId,
+    ) -> Utf8PathBuf {
+        self.timelines_path(tenant_shard_id)
+            .join(timeline_id.to_string())
     }
 
     pub fn timeline_uninit_mark_file_path(
         &self,
-        tenant_id: TenantId,
+        tenant_shard_id: TenantShardId,
         timeline_id: TimelineId,
     ) -> Utf8PathBuf {
         path_with_suffix_extension(
-            self.timeline_path(&tenant_id, &timeline_id),
+            self.timeline_path(&tenant_shard_id, &timeline_id),
             TIMELINE_UNINIT_MARK_SUFFIX,
         )
     }
 
     pub fn timeline_delete_mark_file_path(
         &self,
-        tenant_id: TenantId,
+        tenant_shard_id: TenantShardId,
         timeline_id: TimelineId,
     ) -> Utf8PathBuf {
         path_with_suffix_extension(
-            self.timeline_path(&tenant_id, &timeline_id),
+            self.timeline_path(&tenant_shard_id, &timeline_id),
             TIMELINE_DELETE_MARK_SUFFIX,
         )
     }
 
-    pub fn tenant_deleted_mark_file_path(&self, tenant_id: &TenantId) -> Utf8PathBuf {
-        self.tenant_path(tenant_id)
+    pub fn tenant_deleted_mark_file_path(&self, tenant_shard_id: &TenantShardId) -> Utf8PathBuf {
+        self.tenant_path(tenant_shard_id)
             .join(TENANT_DELETED_MARKER_FILE_NAME)
     }
 
@@ -691,20 +699,24 @@ impl PageServerConf {
 
     pub fn trace_path(
         &self,
-        tenant_id: &TenantId,
+        tenant_shard_id: &TenantShardId,
         timeline_id: &TimelineId,
         connection_id: &ConnectionId,
     ) -> Utf8PathBuf {
         self.traces_path()
-            .join(tenant_id.to_string())
+            .join(tenant_shard_id.to_string())
             .join(timeline_id.to_string())
             .join(connection_id.to_string())
     }
 
     /// Points to a place in pageserver's local directory,
     /// where certain timeline's metadata file should be located.
-    pub fn metadata_path(&self, tenant_id: &TenantId, timeline_id: &TimelineId) -> Utf8PathBuf {
-        self.timeline_path(tenant_id, timeline_id)
+    pub fn metadata_path(
+        &self,
+        tenant_shard_id: &TenantShardId,
+        timeline_id: &TimelineId,
+    ) -> Utf8PathBuf {
+        self.timeline_path(tenant_shard_id, timeline_id)
             .join(METADATA_FILE_NAME)
     }
 
@@ -767,7 +779,7 @@ impl PageServerConf {
                     builder.remote_storage_config(RemoteStorageConfig::from_toml(item)?)
                 }
                 "tenant_config" => {
-                    t_conf = Self::parse_toml_tenant_conf(item)?;
+                    t_conf = TenantConfOpt::try_from(item.to_owned()).context(format!("failed to parse: '{key}'"))?;
                 }
                 "id" => builder.id(NodeId(parse_toml_u64(key, item)?)),
                 "broker_endpoint" => builder.broker_endpoint(parse_toml_string(key, item)?.parse().context("failed to parse broker endpoint")?),
@@ -841,114 +853,10 @@ impl PageServerConf {
         Ok(conf)
     }
 
-    // subroutine of parse_and_validate to parse `[tenant_conf]` section
-
-    pub fn parse_toml_tenant_conf(item: &toml_edit::Item) -> Result<TenantConfOpt> {
-        let mut t_conf: TenantConfOpt = Default::default();
-        if let Some(checkpoint_distance) = item.get("checkpoint_distance") {
-            t_conf.checkpoint_distance =
-                Some(parse_toml_u64("checkpoint_distance", checkpoint_distance)?);
-        }
-
-        if let Some(checkpoint_timeout) = item.get("checkpoint_timeout") {
-            t_conf.checkpoint_timeout = Some(parse_toml_duration(
-                "checkpoint_timeout",
-                checkpoint_timeout,
-            )?);
-        }
-
-        if let Some(compaction_target_size) = item.get("compaction_target_size") {
-            t_conf.compaction_target_size = Some(parse_toml_u64(
-                "compaction_target_size",
-                compaction_target_size,
-            )?);
-        }
-
-        if let Some(compaction_period) = item.get("compaction_period") {
-            t_conf.compaction_period =
-                Some(parse_toml_duration("compaction_period", compaction_period)?);
-        }
-
-        if let Some(compaction_threshold) = item.get("compaction_threshold") {
-            t_conf.compaction_threshold =
-                Some(parse_toml_u64("compaction_threshold", compaction_threshold)?.try_into()?);
-        }
-
-        if let Some(image_creation_threshold) = item.get("image_creation_threshold") {
-            t_conf.image_creation_threshold = Some(
-                parse_toml_u64("image_creation_threshold", image_creation_threshold)?.try_into()?,
-            );
-        }
-
-        if let Some(gc_horizon) = item.get("gc_horizon") {
-            t_conf.gc_horizon = Some(parse_toml_u64("gc_horizon", gc_horizon)?);
-        }
-
-        if let Some(gc_period) = item.get("gc_period") {
-            t_conf.gc_period = Some(parse_toml_duration("gc_period", gc_period)?);
-        }
-
-        if let Some(pitr_interval) = item.get("pitr_interval") {
-            t_conf.pitr_interval = Some(parse_toml_duration("pitr_interval", pitr_interval)?);
-        }
-        if let Some(walreceiver_connect_timeout) = item.get("walreceiver_connect_timeout") {
-            t_conf.walreceiver_connect_timeout = Some(parse_toml_duration(
-                "walreceiver_connect_timeout",
-                walreceiver_connect_timeout,
-            )?);
-        }
-        if let Some(lagging_wal_timeout) = item.get("lagging_wal_timeout") {
-            t_conf.lagging_wal_timeout = Some(parse_toml_duration(
-                "lagging_wal_timeout",
-                lagging_wal_timeout,
-            )?);
-        }
-        if let Some(max_lsn_wal_lag) = item.get("max_lsn_wal_lag") {
-            t_conf.max_lsn_wal_lag =
-                Some(deserialize_from_item("max_lsn_wal_lag", max_lsn_wal_lag)?);
-        }
-        if let Some(trace_read_requests) = item.get("trace_read_requests") {
-            t_conf.trace_read_requests =
-                Some(trace_read_requests.as_bool().with_context(|| {
-                    "configure option trace_read_requests is not a bool".to_string()
-                })?);
-        }
-
-        if let Some(eviction_policy) = item.get("eviction_policy") {
-            t_conf.eviction_policy = Some(
-                deserialize_from_item("eviction_policy", eviction_policy)
-                    .context("parse eviction_policy")?,
-            );
-        }
-
-        if let Some(item) = item.get("min_resident_size_override") {
-            t_conf.min_resident_size_override = Some(
-                deserialize_from_item("min_resident_size_override", item)
-                    .context("parse min_resident_size_override")?,
-            );
-        }
-
-        if let Some(item) = item.get("evictions_low_residence_duration_metric_threshold") {
-            t_conf.evictions_low_residence_duration_metric_threshold = Some(parse_toml_duration(
-                "evictions_low_residence_duration_metric_threshold",
-                item,
-            )?);
-        }
-
-        if let Some(gc_feedback) = item.get("gc_feedback") {
-            t_conf.gc_feedback = Some(
-                gc_feedback
-                    .as_bool()
-                    .with_context(|| "configure option gc_feedback is not a bool".to_string())?,
-            );
-        }
-
-        Ok(t_conf)
-    }
-
     #[cfg(test)]
     pub fn test_repo_dir(test_name: &str) -> Utf8PathBuf {
-        Utf8PathBuf::from(format!("../tmp_check/test_{test_name}"))
+        let test_output_dir = std::env::var("TEST_OUTPUT").unwrap_or("../tmp_check".into());
+        Utf8PathBuf::from(format!("{test_output_dir}/test_{test_name}"))
     }
 
     pub fn dummy_conf(repo_dir: Utf8PathBuf) -> Self {
@@ -1417,6 +1325,37 @@ trace_read_requests = {trace_read_requests}"#,
         Ok(())
     }
 
+    #[test]
+    fn parse_incorrect_tenant_config() -> anyhow::Result<()> {
+        let config_string = r#"
+            [tenant_config]
+            checkpoint_distance = -1 # supposed to be an u64
+        "#
+        .to_string();
+
+        let toml: Document = config_string.parse()?;
+        let item = toml.get("tenant_config").unwrap();
+        let error = TenantConfOpt::try_from(item.to_owned()).unwrap_err();
+
+        let expected_error_str = "checkpoint_distance: invalid value: integer `-1`, expected u64";
+        assert_eq!(error.to_string(), expected_error_str);
+
+        Ok(())
+    }
+
+    #[test]
+    fn parse_override_tenant_config() -> anyhow::Result<()> {
+        let config_string = r#"tenant_config={ min_resident_size_override =  400 }"#.to_string();
+
+        let toml: Document = config_string.parse()?;
+        let item = toml.get("tenant_config").unwrap();
+        let conf = TenantConfOpt::try_from(item.to_owned()).unwrap();
+
+        assert_eq!(conf.min_resident_size_override, Some(400));
+
+        Ok(())
+    }
+
     #[test]
     fn eviction_pageserver_config_parse() -> anyhow::Result<()> {
         let tempdir = tempdir()?;

pageserver/src/consumption_metrics.rs

@@ -281,12 +281,18 @@ async fn calculate_synthetic_size_worker(
                 // By using the same limiter, we centralize metrics collection for "start" and "finished" counters,
                 // which turns out is really handy to understand the system.
                 if let Err(e) = tenant.calculate_synthetic_size(cause, cancel, ctx).await {
-                    if let Some(PageReconstructError::Cancelled) =
-                        e.downcast_ref::<PageReconstructError>()
-                    {
-                        return Ok(());
+                    // this error can be returned if timeline is shutting down, but it does not
+                    // mean the synthetic size worker should terminate. we do not need any checks
+                    // in this function because `mgr::get_tenant` will error out after shutdown has
+                    // progressed to shutting down tenants.
+                    let is_cancelled = matches!(
+                        e.downcast_ref::<PageReconstructError>(),
+                        Some(PageReconstructError::Cancelled)
+                    );
+
+                    if !is_cancelled {
+                        error!("failed to calculate synthetic size for tenant {tenant_id}: {e:#}");
                     }
-                    error!("failed to calculate synthetic size for tenant {tenant_id}: {e:#}");
                 }
             }
         }
@@ -299,7 +305,7 @@ async fn calculate_synthetic_size_worker(
 
         let res = tokio::time::timeout_at(
             started_at + synthetic_size_calculation_interval,
-            task_mgr::shutdown_token().cancelled(),
+            cancel.cancelled(),
         )
         .await;
         if res.is_ok() {

pageserver/src/consumption_metrics/metrics.rs

@@ -1,8 +1,8 @@
-use crate::context::RequestContext;
-use anyhow::Context;
+use crate::{context::RequestContext, tenant::timeline::logical_size::CurrentLogicalSize};
 use chrono::{DateTime, Utc};
 use consumption_metrics::EventType;
 use futures::stream::StreamExt;
+use pageserver_api::shard::ShardNumber;
 use std::{sync::Arc, time::SystemTime};
 use utils::{
     id::{TenantId, TimelineId},
@@ -229,6 +229,11 @@ where
     while let Some((tenant_id, tenant)) = tenants.next().await {
         let mut tenant_resident_size = 0;
 
+        // Sharded tenants report all consumption metrics from shard zero
+        if tenant.tenant_shard_id().shard_number != ShardNumber(0) {
+            continue;
+        }
+
         for timeline in tenant.list_timelines() {
             let timeline_id = timeline.timeline_id;
 
@@ -351,14 +356,17 @@ impl TimelineSnapshot {
             let last_record_lsn = t.get_last_record_lsn();
 
             let current_exact_logical_size = {
-                let span = tracing::info_span!("collect_metrics_iteration", tenant_id = %t.tenant_id, timeline_id = %t.timeline_id);
-                let res = span
-                    .in_scope(|| t.get_current_logical_size(ctx))
-                    .context("get_current_logical_size");
-                match res? {
+                let span = tracing::info_span!("collect_metrics_iteration", tenant_id = %t.tenant_shard_id.tenant_id, timeline_id = %t.timeline_id);
+                let size = span.in_scope(|| {
+                    t.get_current_logical_size(
+                        crate::tenant::timeline::GetLogicalSizePriority::Background,
+                        ctx,
+                    )
+                });
+                match size {
                     // Only send timeline logical size when it is fully calculated.
-                    (size, is_exact) if is_exact => Some(size),
-                    (_, _) => None,
+                    CurrentLogicalSize::Exact(ref size) => Some(size.into()),
+                    CurrentLogicalSize::Approximate(_) => None,
                 }
             };
 

pageserver/src/control_plane_client.rs

@@ -1,16 +1,15 @@
 use std::collections::HashMap;
 
-use pageserver_api::control_api::{
-    ReAttachRequest, ReAttachResponse, ValidateRequest, ValidateRequestTenant, ValidateResponse,
+use pageserver_api::{
+    control_api::{
+        ReAttachRequest, ReAttachResponse, ValidateRequest, ValidateRequestTenant, ValidateResponse,
+    },
+    shard::TenantShardId,
 };
 use serde::{de::DeserializeOwned, Serialize};
 use tokio_util::sync::CancellationToken;
 use url::Url;
-use utils::{
-    backoff,
-    generation::Generation,
-    id::{NodeId, TenantId},
-};
+use utils::{backoff, generation::Generation, id::NodeId};
 
 use crate::config::PageServerConf;
 
@@ -31,11 +30,11 @@ pub enum RetryForeverError {
 
 #[async_trait::async_trait]
 pub trait ControlPlaneGenerationsApi {
-    async fn re_attach(&self) -> Result<HashMap<TenantId, Generation>, RetryForeverError>;
+    async fn re_attach(&self) -> Result<HashMap<TenantShardId, Generation>, RetryForeverError>;
     async fn validate(
         &self,
-        tenants: Vec<(TenantId, Generation)>,
-    ) -> Result<HashMap<TenantId, bool>, RetryForeverError>;
+        tenants: Vec<(TenantShardId, Generation)>,
+    ) -> Result<HashMap<TenantShardId, bool>, RetryForeverError>;
 }
 
 impl ControlPlaneClient {
@@ -127,7 +126,7 @@ impl ControlPlaneClient {
 #[async_trait::async_trait]
 impl ControlPlaneGenerationsApi for ControlPlaneClient {
     /// Block until we get a successful response, or error out if we are shut down
-    async fn re_attach(&self) -> Result<HashMap<TenantId, Generation>, RetryForeverError> {
+    async fn re_attach(&self) -> Result<HashMap<TenantShardId, Generation>, RetryForeverError> {
         let re_attach_path = self
             .base_url
             .join("re-attach")
@@ -154,8 +153,8 @@ impl ControlPlaneGenerationsApi for ControlPlaneClient {
     /// Block until we get a successful response, or error out if we are shut down
     async fn validate(
         &self,
-        tenants: Vec<(TenantId, Generation)>,
-    ) -> Result<HashMap<TenantId, bool>, RetryForeverError> {
+        tenants: Vec<(TenantShardId, Generation)>,
+    ) -> Result<HashMap<TenantShardId, bool>, RetryForeverError> {
         let re_attach_path = self
             .base_url
             .join("validate")

pageserver/src/deletion_queue.rs

@@ -10,11 +10,12 @@ use crate::control_plane_client::ControlPlaneGenerationsApi;
 use crate::metrics;
 use crate::tenant::remote_timeline_client::remote_layer_path;
 use crate::tenant::remote_timeline_client::remote_timeline_path;
+use crate::tenant::remote_timeline_client::LayerFileMetadata;
 use crate::virtual_file::MaybeFatalIo;
 use crate::virtual_file::VirtualFile;
 use anyhow::Context;
 use camino::Utf8PathBuf;
-use hex::FromHex;
+use pageserver_api::shard::TenantShardId;
 use remote_storage::{GenericRemoteStorage, RemotePath};
 use serde::Deserialize;
 use serde::Serialize;
@@ -25,7 +26,7 @@ use tracing::Instrument;
 use tracing::{self, debug, error};
 use utils::crashsafe::path_with_suffix_extension;
 use utils::generation::Generation;
-use utils::id::{TenantId, TimelineId};
+use utils::id::TimelineId;
 use utils::lsn::AtomicLsn;
 use utils::lsn::Lsn;
 
@@ -159,11 +160,10 @@ pub struct DeletionQueueClient {
     lsn_table: Arc<std::sync::RwLock<VisibleLsnUpdates>>,
 }
 
-#[derive(Debug, Serialize, Deserialize)]
+#[derive(Debug, Serialize, Deserialize, PartialEq, Eq)]
 struct TenantDeletionList {
     /// For each Timeline, a list of key fragments to append to the timeline remote path
     /// when reconstructing a full key
-    #[serde(serialize_with = "to_hex_map", deserialize_with = "from_hex_map")]
     timelines: HashMap<TimelineId, Vec<String>>,
 
     /// The generation in which this deletion was emitted: note that this may not be the
@@ -178,43 +178,11 @@ impl TenantDeletionList {
     }
 }
 
-/// For HashMaps using a `hex` compatible key, where we would like to encode the key as a string
-fn to_hex_map<S, V, I>(input: &HashMap<I, V>, serializer: S) -> Result<S::Ok, S::Error>
-where
-    S: serde::Serializer,
-    V: Serialize,
-    I: AsRef<[u8]>,
-{
-    let transformed = input.iter().map(|(k, v)| (hex::encode(k), v));
-
-    transformed
-        .collect::<HashMap<String, &V>>()
-        .serialize(serializer)
-}
-
-/// For HashMaps using a FromHex key, where we would like to decode the key
-fn from_hex_map<'de, D, V, I>(deserializer: D) -> Result<HashMap<I, V>, D::Error>
-where
-    D: serde::de::Deserializer<'de>,
-    V: Deserialize<'de>,
-    I: FromHex + std::hash::Hash + Eq,
-{
-    let hex_map = HashMap::<String, V>::deserialize(deserializer)?;
-    hex_map
-        .into_iter()
-        .map(|(k, v)| {
-            I::from_hex(k)
-                .map(|k| (k, v))
-                .map_err(|_| serde::de::Error::custom("Invalid hex ID"))
-        })
-        .collect()
-}
-
 /// Files ending with this suffix will be ignored and erased
 /// during recovery as startup.
 const TEMP_SUFFIX: &str = "tmp";
 
-#[derive(Debug, Serialize, Deserialize)]
+#[derive(Debug, Serialize, Deserialize, PartialEq, Eq)]
 struct DeletionList {
     /// Serialization version, for future use
     version: u8,
@@ -226,8 +194,7 @@ struct DeletionList {
     /// nested HashMaps by TenantTimelineID.  Each Tenant only appears once
     /// with one unique generation ID: if someone tries to push a second generation
     /// ID for the same tenant, we will start a new DeletionList.
-    #[serde(serialize_with = "to_hex_map", deserialize_with = "from_hex_map")]
-    tenants: HashMap<TenantId, TenantDeletionList>,
+    tenants: HashMap<TenantShardId, TenantDeletionList>,
 
     /// Avoid having to walk `tenants` to calculate the number of keys in
     /// the nested deletion lists
@@ -299,7 +266,7 @@ impl DeletionList {
     /// deletion list.
     fn push(
         &mut self,
-        tenant: &TenantId,
+        tenant: &TenantShardId,
         timeline: &TimelineId,
         generation: Generation,
         objects: &mut Vec<RemotePath>,
@@ -391,7 +358,7 @@ struct TenantLsnState {
 
 #[derive(Default)]
 struct VisibleLsnUpdates {
-    tenants: HashMap<TenantId, TenantLsnState>,
+    tenants: HashMap<TenantShardId, TenantLsnState>,
 }
 
 impl VisibleLsnUpdates {
@@ -448,7 +415,7 @@ impl DeletionQueueClient {
 
     pub(crate) fn recover(
         &self,
-        attached_tenants: HashMap<TenantId, Generation>,
+        attached_tenants: HashMap<TenantShardId, Generation>,
     ) -> Result<(), DeletionQueueError> {
         self.do_push(
             &self.tx,
@@ -465,7 +432,7 @@ impl DeletionQueueClient {
     /// backend will later wake up and notice that the tenant's generation requires validation.
     pub(crate) async fn update_remote_consistent_lsn(
         &self,
-        tenant_id: TenantId,
+        tenant_shard_id: TenantShardId,
         timeline_id: TimelineId,
         current_generation: Generation,
         lsn: Lsn,
@@ -476,10 +443,13 @@ impl DeletionQueueClient {
             .write()
             .expect("Lock should never be poisoned");
 
-        let tenant_entry = locked.tenants.entry(tenant_id).or_insert(TenantLsnState {
-            timelines: HashMap::new(),
-            generation: current_generation,
-        });
+        let tenant_entry = locked
+            .tenants
+            .entry(tenant_shard_id)
+            .or_insert(TenantLsnState {
+                timelines: HashMap::new(),
+                generation: current_generation,
+            });
 
         if tenant_entry.generation != current_generation {
             // Generation might have changed if we were detached and then re-attached: in this case,
@@ -506,28 +476,29 @@ impl DeletionQueueClient {
     /// generations in `layers` are the generations in which those layers were written.
     pub(crate) async fn push_layers(
         &self,
-        tenant_id: TenantId,
+        tenant_shard_id: TenantShardId,
         timeline_id: TimelineId,
         current_generation: Generation,
-        layers: Vec<(LayerFileName, Generation)>,
+        layers: Vec<(LayerFileName, LayerFileMetadata)>,
     ) -> Result<(), DeletionQueueError> {
         if current_generation.is_none() {
             debug!("Enqueuing deletions in legacy mode, skipping queue");
 
             let mut layer_paths = Vec::new();
-            for (layer, generation) in layers {
+            for (layer, meta) in layers {
                 layer_paths.push(remote_layer_path(
-                    &tenant_id,
+                    &tenant_shard_id.tenant_id,
                     &timeline_id,
+                    meta.shard,
                     &layer,
-                    generation,
+                    meta.generation,
                 ));
             }
             self.push_immediate(layer_paths).await?;
             return self.flush_immediate().await;
         }
 
-        self.push_layers_sync(tenant_id, timeline_id, current_generation, layers)
+        self.push_layers_sync(tenant_shard_id, timeline_id, current_generation, layers)
     }
 
     /// When a Tenant has a generation, push_layers is always synchronous because
@@ -537,10 +508,10 @@ impl DeletionQueueClient {
     /// support (`<https://github.com/neondatabase/neon/issues/5395>`)
     pub(crate) fn push_layers_sync(
         &self,
-        tenant_id: TenantId,
+        tenant_shard_id: TenantShardId,
         timeline_id: TimelineId,
         current_generation: Generation,
-        layers: Vec<(LayerFileName, Generation)>,
+        layers: Vec<(LayerFileName, LayerFileMetadata)>,
     ) -> Result<(), DeletionQueueError> {
         metrics::DELETION_QUEUE
             .keys_submitted
@@ -548,7 +519,7 @@ impl DeletionQueueClient {
         self.do_push(
             &self.tx,
             ListWriterQueueMessage::Delete(DeletionOp {
-                tenant_id,
+                tenant_shard_id,
                 timeline_id,
                 layers,
                 generation: current_generation,
@@ -751,6 +722,7 @@ impl DeletionQueue {
 mod test {
     use camino::Utf8Path;
     use hex_literal::hex;
+    use pageserver_api::shard::ShardIndex;
     use std::{io::ErrorKind, time::Duration};
     use tracing::info;
 
@@ -815,12 +787,12 @@ mod test {
         }
 
         fn set_latest_generation(&self, gen: Generation) {
-            let tenant_id = self.harness.tenant_id;
+            let tenant_shard_id = self.harness.tenant_shard_id;
             self.mock_control_plane
                 .latest_generation
                 .lock()
                 .unwrap()
-                .insert(tenant_id, gen);
+                .insert(tenant_shard_id, gen);
         }
 
         /// Returns remote layer file name, suitable for use in assert_remote_files
@@ -829,8 +801,8 @@ mod test {
             file_name: LayerFileName,
             gen: Generation,
         ) -> anyhow::Result<String> {
-            let tenant_id = self.harness.tenant_id;
-            let relative_remote_path = remote_timeline_path(&tenant_id, &TIMELINE_ID);
+            let tenant_shard_id = self.harness.tenant_shard_id;
+            let relative_remote_path = remote_timeline_path(&tenant_shard_id, &TIMELINE_ID);
             let remote_timeline_path = self.remote_fs_dir.join(relative_remote_path.get_path());
             std::fs::create_dir_all(&remote_timeline_path)?;
             let remote_layer_file_name = format!("{}{}", file_name, gen.get_suffix());
@@ -848,7 +820,7 @@ mod test {
 
     #[derive(Debug, Clone)]
     struct MockControlPlane {
-        pub latest_generation: std::sync::Arc<std::sync::Mutex<HashMap<TenantId, Generation>>>,
+        pub latest_generation: std::sync::Arc<std::sync::Mutex<HashMap<TenantShardId, Generation>>>,
     }
 
     impl MockControlPlane {
@@ -862,20 +834,20 @@ mod test {
     #[async_trait::async_trait]
     impl ControlPlaneGenerationsApi for MockControlPlane {
         #[allow(clippy::diverging_sub_expression)] // False positive via async_trait
-        async fn re_attach(&self) -> Result<HashMap<TenantId, Generation>, RetryForeverError> {
+        async fn re_attach(&self) -> Result<HashMap<TenantShardId, Generation>, RetryForeverError> {
             unimplemented!()
         }
         async fn validate(
             &self,
-            tenants: Vec<(TenantId, Generation)>,
-        ) -> Result<HashMap<TenantId, bool>, RetryForeverError> {
+            tenants: Vec<(TenantShardId, Generation)>,
+        ) -> Result<HashMap<TenantShardId, bool>, RetryForeverError> {
             let mut result = HashMap::new();
 
             let latest_generation = self.latest_generation.lock().unwrap();
 
-            for (tenant_id, generation) in tenants {
-                if let Some(latest) = latest_generation.get(&tenant_id) {
-                    result.insert(tenant_id, *latest == generation);
+            for (tenant_shard_id, generation) in tenants {
+                if let Some(latest) = latest_generation.get(&tenant_shard_id) {
+                    result.insert(tenant_shard_id, *latest == generation);
                 }
             }
 
@@ -979,10 +951,10 @@ mod test {
         client.recover(HashMap::new())?;
 
         let layer_file_name_1: LayerFileName = "000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__00000000016B59D8-00000000016B5A51".parse().unwrap();
-        let tenant_id = ctx.harness.tenant_id;
+        let tenant_shard_id = ctx.harness.tenant_shard_id;
 
         let content: Vec<u8> = "victim1 contents".into();
-        let relative_remote_path = remote_timeline_path(&tenant_id, &TIMELINE_ID);
+        let relative_remote_path = remote_timeline_path(&tenant_shard_id, &TIMELINE_ID);
         let remote_timeline_path = ctx.remote_fs_dir.join(relative_remote_path.get_path());
         let deletion_prefix = ctx.harness.conf.deletion_prefix();
 
@@ -990,6 +962,8 @@ mod test {
         // we delete, and the generation of the running Tenant.
         let layer_generation = Generation::new(0xdeadbeef);
         let now_generation = Generation::new(0xfeedbeef);
+        let layer_metadata =
+            LayerFileMetadata::new(0xf00, layer_generation, ShardIndex::unsharded());
 
         let remote_layer_file_name_1 =
             format!("{}{}", layer_file_name_1, layer_generation.get_suffix());
@@ -1010,10 +984,10 @@ mod test {
         info!("Pushing");
         client
             .push_layers(
-                tenant_id,
+                tenant_shard_id,
                 TIMELINE_ID,
                 now_generation,
-                [(layer_file_name_1.clone(), layer_generation)].to_vec(),
+                [(layer_file_name_1.clone(), layer_metadata)].to_vec(),
             )
             .await?;
         assert_remote_files(&[&remote_layer_file_name_1], &remote_timeline_path);
@@ -1052,11 +1026,13 @@ mod test {
         let stale_generation = latest_generation.previous();
         // Generation that our example layer file was written with
         let layer_generation = stale_generation.previous();
+        let layer_metadata =
+            LayerFileMetadata::new(0xf00, layer_generation, ShardIndex::unsharded());
 
         ctx.set_latest_generation(latest_generation);
 
-        let tenant_id = ctx.harness.tenant_id;
-        let relative_remote_path = remote_timeline_path(&tenant_id, &TIMELINE_ID);
+        let tenant_shard_id = ctx.harness.tenant_shard_id;
+        let relative_remote_path = remote_timeline_path(&tenant_shard_id, &TIMELINE_ID);
         let remote_timeline_path = ctx.remote_fs_dir.join(relative_remote_path.get_path());
 
         // Initial state: a remote layer exists
@@ -1066,10 +1042,10 @@ mod test {
         tracing::debug!("Pushing...");
         client
             .push_layers(
-                tenant_id,
+                tenant_shard_id,
                 TIMELINE_ID,
                 stale_generation,
-                [(EXAMPLE_LAYER_NAME.clone(), layer_generation)].to_vec(),
+                [(EXAMPLE_LAYER_NAME.clone(), layer_metadata.clone())].to_vec(),
             )
             .await?;
 
@@ -1081,10 +1057,10 @@ mod test {
         tracing::debug!("Pushing...");
         client
             .push_layers(
-                tenant_id,
+                tenant_shard_id,
                 TIMELINE_ID,
                 latest_generation,
-                [(EXAMPLE_LAYER_NAME.clone(), layer_generation)].to_vec(),
+                [(EXAMPLE_LAYER_NAME.clone(), layer_metadata.clone())].to_vec(),
             )
             .await?;
 
@@ -1103,14 +1079,16 @@ mod test {
         let client = ctx.deletion_queue.new_client();
         client.recover(HashMap::new())?;
 
-        let tenant_id = ctx.harness.tenant_id;
+        let tenant_shard_id = ctx.harness.tenant_shard_id;
 
-        let relative_remote_path = remote_timeline_path(&tenant_id, &TIMELINE_ID);
+        let relative_remote_path = remote_timeline_path(&tenant_shard_id, &TIMELINE_ID);
         let remote_timeline_path = ctx.remote_fs_dir.join(relative_remote_path.get_path());
         let deletion_prefix = ctx.harness.conf.deletion_prefix();
 
         let layer_generation = Generation::new(0xdeadbeef);
         let now_generation = Generation::new(0xfeedbeef);
+        let layer_metadata =
+            LayerFileMetadata::new(0xf00, layer_generation, ShardIndex::unsharded());
 
         // Inject a deletion in the generation before generation_now: after restart,
         // this deletion should _not_ get executed (only the immediately previous
@@ -1119,10 +1097,10 @@ mod test {
             ctx.write_remote_layer(EXAMPLE_LAYER_NAME, layer_generation)?;
         client
             .push_layers(
-                tenant_id,
+                tenant_shard_id,
                 TIMELINE_ID,
                 now_generation.previous(),
-                [(EXAMPLE_LAYER_NAME.clone(), layer_generation)].to_vec(),
+                [(EXAMPLE_LAYER_NAME.clone(), layer_metadata.clone())].to_vec(),
             )
             .await?;
 
@@ -1133,10 +1111,10 @@ mod test {
             ctx.write_remote_layer(EXAMPLE_LAYER_NAME_ALT, layer_generation)?;
         client
             .push_layers(
-                tenant_id,
+                tenant_shard_id,
                 TIMELINE_ID,
                 now_generation,
-                [(EXAMPLE_LAYER_NAME_ALT.clone(), layer_generation)].to_vec(),
+                [(EXAMPLE_LAYER_NAME_ALT.clone(), layer_metadata.clone())].to_vec(),
             )
             .await?;
 
@@ -1164,7 +1142,7 @@ mod test {
         drop(client);
         ctx.restart().await;
         let client = ctx.deletion_queue.new_client();
-        client.recover(HashMap::from([(tenant_id, now_generation)]))?;
+        client.recover(HashMap::from([(tenant_shard_id, now_generation)]))?;
 
         info!("Flush-executing");
         client.flush_execute().await?;
@@ -1226,12 +1204,13 @@ pub(crate) mod mock {
                 match msg {
                     ListWriterQueueMessage::Delete(op) => {
                         let mut objects = op.objects;
-                        for (layer, generation) in op.layers {
+                        for (layer, meta) in op.layers {
                             objects.push(remote_layer_path(
-                                &op.tenant_id,
+                                &op.tenant_shard_id.tenant_id,
                                 &op.timeline_id,
+                                meta.shard,
                                 &layer,
-                                generation,
+                                meta.generation,
                             ));
                         }
 
@@ -1311,4 +1290,34 @@ pub(crate) mod mock {
             }
         }
     }
+
+    /// Test round-trip serialization/deserialization, and test stability of the format
+    /// vs. a static expected string for the serialized version.
+    #[test]
+    fn deletion_list_serialization() -> anyhow::Result<()> {
+        let tenant_id = "ad6c1a56f5680419d3a16ff55d97ec3c"
+            .to_string()
+            .parse::<TenantShardId>()?;
+        let timeline_id = "be322c834ed9e709e63b5c9698691910"
+            .to_string()
+            .parse::<TimelineId>()?;
+        let generation = Generation::new(123);
+
+        let object =
+            RemotePath::from_string(&format!("tenants/{tenant_id}/timelines/{timeline_id}/foo"))?;
+        let mut objects = [object].to_vec();
+
+        let mut example = DeletionList::new(1);
+        example.push(&tenant_id, &timeline_id, generation, &mut objects);
+
+        let encoded = serde_json::to_string(&example)?;
+
+        let expected = "{\"version\":1,\"sequence\":1,\"tenants\":{\"ad6c1a56f5680419d3a16ff55d97ec3c\":{\"timelines\":{\"be322c834ed9e709e63b5c9698691910\":[\"foo\"]},\"generation\":123}},\"size\":1}".to_string();
+        assert_eq!(encoded, expected);
+
+        let decoded = serde_json::from_str::<DeletionList>(&encoded)?;
+        assert_eq!(example, decoded);
+
+        Ok(())
+    }
 }

pageserver/src/deletion_queue/list_writer.rs

@@ -19,6 +19,7 @@ use std::collections::HashMap;
 use std::fs::create_dir_all;
 use std::time::Duration;
 
+use pageserver_api::shard::TenantShardId;
 use regex::Regex;
 use remote_storage::RemotePath;
 use tokio_util::sync::CancellationToken;
@@ -26,13 +27,13 @@ use tracing::debug;
 use tracing::info;
 use tracing::warn;
 use utils::generation::Generation;
-use utils::id::TenantId;
 use utils::id::TimelineId;
 
 use crate::config::PageServerConf;
 use crate::deletion_queue::TEMP_SUFFIX;
 use crate::metrics;
 use crate::tenant::remote_timeline_client::remote_layer_path;
+use crate::tenant::remote_timeline_client::LayerFileMetadata;
 use crate::tenant::storage_layer::LayerFileName;
 use crate::virtual_file::on_fatal_io_error;
 use crate::virtual_file::MaybeFatalIo;
@@ -53,22 +54,22 @@ const FRONTEND_FLUSHING_TIMEOUT: Duration = Duration::from_millis(100);
 
 #[derive(Debug)]
 pub(super) struct DeletionOp {
-    pub(super) tenant_id: TenantId,
+    pub(super) tenant_shard_id: TenantShardId,
     pub(super) timeline_id: TimelineId,
     // `layers` and `objects` are both just lists of objects.  `layers` is used if you do not
     // have a config object handy to project it to a remote key, and need the consuming worker
     // to do it for you.
-    pub(super) layers: Vec<(LayerFileName, Generation)>,
+    pub(super) layers: Vec<(LayerFileName, LayerFileMetadata)>,
     pub(super) objects: Vec<RemotePath>,
 
-    /// The _current_ generation of the Tenant attachment in which we are enqueuing
+    /// The _current_ generation of the Tenant shard attachment in which we are enqueuing
     /// this deletion.
     pub(super) generation: Generation,
 }
 
 #[derive(Debug)]
 pub(super) struct RecoverOp {
-    pub(super) attached_tenants: HashMap<TenantId, Generation>,
+    pub(super) attached_tenants: HashMap<TenantShardId, Generation>,
 }
 
 #[derive(Debug)]
@@ -205,7 +206,7 @@ impl ListWriter {
 
     async fn recover(
         &mut self,
-        attached_tenants: HashMap<TenantId, Generation>,
+        attached_tenants: HashMap<TenantShardId, Generation>,
     ) -> Result<(), anyhow::Error> {
         debug!(
             "recovering with {} attached tenants",
@@ -308,10 +309,21 @@ impl ListWriter {
                 // generation was issued to another node in the interval while we restarted,
                 // then we may treat deletion lists from the previous generation as if they
                 // belong to our currently attached generation, and proceed to validate & execute.
-                for (tenant_id, tenant_list) in &mut deletion_list.tenants {
-                    if let Some(attached_gen) = attached_tenants.get(tenant_id) {
+                for (tenant_shard_id, tenant_list) in &mut deletion_list.tenants {
+                    if let Some(attached_gen) = attached_tenants.get(tenant_shard_id) {
                         if attached_gen.previous() == tenant_list.generation {
+                            info!(
+                                seq=%s, tenant_id=%tenant_shard_id.tenant_id,
+                                shard_id=%tenant_shard_id.shard_slug(),
+                                old_gen=?tenant_list.generation, new_gen=?attached_gen,
+                                "Updating gen on recovered list");
                             tenant_list.generation = *attached_gen;
+                        } else {
+                            info!(
+                                seq=%s, tenant_id=%tenant_shard_id.tenant_id,
+                                shard_id=%tenant_shard_id.shard_slug(),
+                                old_gen=?tenant_list.generation, new_gen=?attached_gen,
+                                "Encountered stale generation on recovered list");
                         }
                     }
                 }
@@ -387,25 +399,26 @@ impl ListWriter {
                     );
 
                     let mut layer_paths = Vec::new();
-                    for (layer, generation) in op.layers {
+                    for (layer, meta) in op.layers {
                         layer_paths.push(remote_layer_path(
-                            &op.tenant_id,
+                            &op.tenant_shard_id.tenant_id,
                             &op.timeline_id,
+                            meta.shard,
                             &layer,
-                            generation,
+                            meta.generation,
                         ));
                     }
                     layer_paths.extend(op.objects);
 
                     if !self.pending.push(
-                        &op.tenant_id,
+                        &op.tenant_shard_id,
                         &op.timeline_id,
                         op.generation,
                         &mut layer_paths,
                     ) {
                         self.flush().await;
                         let retry_succeeded = self.pending.push(
-                            &op.tenant_id,
+                            &op.tenant_shard_id,
                             &op.timeline_id,
                             op.generation,
                             &mut layer_paths,

pageserver/src/deletion_queue/validator.rs

@@ -178,7 +178,14 @@ where
                 .unwrap_or(false);
 
             if valid && *validated_generation == tenant_lsn_state.generation {
-                for (_timeline_id, pending_lsn) in tenant_lsn_state.timelines {
+                for (timeline_id, pending_lsn) in tenant_lsn_state.timelines {
+                    tracing::debug!(
+                        %tenant_id,
+                        %timeline_id,
+                        current = %pending_lsn.result_slot.load(),
+                        projected = %pending_lsn.projected,
+                        "advancing validated remote_consistent_lsn",
+                    );
                     pending_lsn.result_slot.store(pending_lsn.projected);
                 }
             } else {

pageserver/src/disk_usage_eviction_task.rs

@@ -310,7 +310,7 @@ pub async fn disk_usage_eviction_task_iteration_impl<U: Usage>(
                 .unwrap()
                 .as_micros(),
             partition,
-            desc.tenant_id,
+            desc.tenant_shard_id,
             desc.timeline_id,
             candidate.layer,
         );
@@ -380,7 +380,7 @@ pub async fn disk_usage_eviction_task_iteration_impl<U: Usage>(
     let limit = Arc::new(tokio::sync::Semaphore::new(1000.max(max_batch_size)));
 
     for (timeline, batch) in batched {
-        let tenant_id = timeline.tenant_id;
+        let tenant_shard_id = timeline.tenant_shard_id;
         let timeline_id = timeline.timeline_id;
         let batch_size =
             u32::try_from(batch.len()).expect("batch size limited to u32::MAX during partitioning");
@@ -431,7 +431,7 @@ pub async fn disk_usage_eviction_task_iteration_impl<U: Usage>(
                 (evicted_bytes, evictions_failed)
             }
         }
-        .instrument(tracing::info_span!("evict_batch", %tenant_id, %timeline_id, batch_size));
+        .instrument(tracing::info_span!("evict_batch", tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), %timeline_id, batch_size));
 
         js.spawn(evict);
 
@@ -572,7 +572,7 @@ async fn collect_eviction_candidates(
                 continue;
             }
             let info = tl.get_local_layers_for_disk_usage_eviction().await;
-            debug!(tenant_id=%tl.tenant_id, timeline_id=%tl.timeline_id, "timeline resident layers count: {}", info.resident_layers.len());
+            debug!(tenant_id=%tl.tenant_shard_id.tenant_id, shard_id=%tl.tenant_shard_id.shard_slug(), timeline_id=%tl.timeline_id, "timeline resident layers count: {}", info.resident_layers.len());
             tenant_candidates.extend(
                 info.resident_layers
                     .into_iter()

pageserver/src/http/openapi_spec.yml

@@ -624,6 +624,99 @@ paths:
                 $ref: "#/components/schemas/ServiceUnavailableError"
 
 
+  /v1/tenant/{tenant_id}/location_config:
+    parameters:
+      - name: tenant_id
+        in: path
+        required: true
+        schema:
+          type: string
+          format: hex
+      - name: flush_ms
+        in: query
+        required: false
+        schema:
+          type: integer
+    put:
+      description: |
+        Configures a _tenant location_, that is how a particular pageserver handles
+        a particular tenant.  This includes _attached_ tenants, i.e. those ingesting WAL
+        and page service requests, and _secondary_ tenants, i.e. those which are just keeping
+        a warm cache in anticipation of transitioning to attached state in the future.
+
+        This is a declarative, idempotent API: there are not separate endpoints
+        for different tenant location configurations.  Rather, this single endpoint accepts
+        a description of the desired location configuration, and makes whatever changes
+        are required to reach that state.
+
+        In imperative terms, this API is used to attach and detach tenants, and
+        to transition tenants to and from secondary mode.
+
+        This is a synchronous API: there is no 202 response.  State transitions should always
+        be fast (milliseconds), with the exception of requests setting `flush_ms`, in which case
+        the caller controls the runtime of the request.
+
+        In some state transitions, it makes sense to flush dirty data to remote storage: this includes transitions
+        to AttachedStale and Detached.  Flushing is never necessary for correctness, but is an
+        important optimization when doing migrations.  The `flush_ms` parameter controls whether
+        flushing should be attempted, and how much time is allowed for flushing.  If the time limit expires,
+        the requested transition will continue without waiting for any outstanding data to flush.  Callers
+        should use a duration which is substantially less than their HTTP client's request
+        timeout.  It is safe to supply flush_ms irrespective of the request body: in state transitions
+        where flushing doesn't make sense, the server will ignore it.
+
+        It is safe to retry requests, but if one receives a 409 or 503 response, it is not
+        useful to retry aggressively: there is probably an existing request still ongoing.
+      requestBody:
+        required: false
+        content:
+          application/json:
+            schema:
+              $ref: "#/components/schemas/TenantLocationConfigRequest"
+      responses:
+        "200":
+          description: Tenant is now in requested state
+        "503":
+          description: Tenant's state cannot be changed right now.  Wait a few seconds and retry.
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/Error"
+        "401":
+          description: Unauthorized Error
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/UnauthorizedError"
+        "403":
+          description: Forbidden Error
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/ForbiddenError"
+        "409":
+          description: |
+            The tenant is already known to Pageserver in some way,
+            and hence this `/attach` call has been rejected.
+
+            Some examples of how this can happen:
+            - tenant was created on this pageserver
+            - tenant attachment was started by an earlier call to `/attach`.
+
+            Callers should poll the tenant status's `attachment_status` field,
+            like for status 202. See the longer description for `POST /attach`
+            for details.
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/ConflictError"
+        "500":
+          description: Generic operation error
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/Error"
+
   /v1/tenant/{tenant_id}/detach:
     parameters:
       - name: tenant_id
@@ -935,6 +1028,9 @@ paths:
                   format: hex
                 pg_version:
                   type: integer
+                existing_initdb_timeline_id:
+                  type: string
+                  format: hex
       responses:
         "201":
           description: TimelineInfo
@@ -1274,6 +1370,31 @@ components:
             tenant_id:
               type: string
               format: hex
+    TenantLocationConfigRequest:
+      type: object
+      required:
+        - tenant_id
+      properties:
+        tenant_id:
+          type: string
+          format: hex
+        mode:
+          type: string
+          enum: ["AttachedSingle", "AttachedMulti", "AttachedStale", "Secondary", "Detached"]
+          description: Mode of functionality that this pageserver will run in for this tenant.
+        generation:
+          type: integer
+          description: Attachment generation number, mandatory when `mode` is an attached state
+        secondary_conf:
+          $ref: '#/components/schemas/SecondaryConfig'
+        tenant_conf:
+          $ref: '#/components/schemas/TenantConfig'
+    SecondaryConfig:
+      type: object
+      properties:
+        warm:
+          type: boolean
+          description: Whether to poll remote storage for layers to download.  If false, secondary locations don't download anything.
     TenantConfig:
       type: object
       properties:

pageserver/src/http/routes.rs

@@ -4,6 +4,7 @@
 use std::collections::HashMap;
 use std::str::FromStr;
 use std::sync::Arc;
+use std::time::Duration;
 
 use anyhow::{anyhow, Context, Result};
 use enumset::EnumSet;
@@ -337,13 +338,8 @@ async fn build_timeline_info_common(
         Lsn(0) => None,
         lsn @ Lsn(_) => Some(lsn),
     };
-    let current_logical_size = match timeline.get_current_logical_size(ctx) {
-        Ok((size, _)) => Some(size),
-        Err(err) => {
-            error!("Timeline info creation failed to get current logical size: {err:?}");
-            None
-        }
-    };
+    let current_logical_size =
+        timeline.get_current_logical_size(tenant::timeline::GetLogicalSizePriority::User, ctx);
     let current_physical_size = Some(timeline.layer_size_sum().await);
     let state = timeline.current_state();
     let remote_consistent_lsn_projected = timeline
@@ -356,7 +352,8 @@ async fn build_timeline_info_common(
     let walreceiver_status = timeline.walreceiver_status();
 
     let info = TimelineInfo {
-        tenant_id: timeline.tenant_id,
+        // TODO(sharding): add a shard_id field, or make tenant_id into a tenant_shard_id
+        tenant_id: timeline.tenant_shard_id.tenant_id,
         timeline_id: timeline.timeline_id,
         ancestor_timeline_id,
         ancestor_lsn,
@@ -366,7 +363,11 @@ async fn build_timeline_info_common(
         last_record_lsn,
         prev_record_lsn: Some(timeline.get_prev_record_lsn()),
         latest_gc_cutoff_lsn: *timeline.get_latest_gc_cutoff_lsn(),
-        current_logical_size,
+        current_logical_size: current_logical_size.size_dont_care_about_accuracy(),
+        current_logical_size_is_accurate: match current_logical_size.accuracy() {
+            tenant::timeline::logical_size::Accuracy::Approximate => false,
+            tenant::timeline::logical_size::Accuracy::Exact => true,
+        },
         current_physical_size,
         current_logical_size_non_incremental: None,
         timeline_dir_layer_file_size_sum: None,
@@ -439,6 +440,7 @@ async fn timeline_create_handler(
             request_data.ancestor_timeline_id.map(TimelineId::from),
             request_data.ancestor_start_lsn,
             request_data.pg_version.unwrap_or(crate::DEFAULT_PG_VERSION),
+            request_data.existing_initdb_timeline_id,
             state.broker_client.clone(),
             &ctx,
         )
@@ -707,6 +709,26 @@ async fn tenant_detach_handler(
     json_response(StatusCode::OK, ())
 }
 
+async fn tenant_reset_handler(
+    request: Request<Body>,
+    _cancel: CancellationToken,
+) -> Result<Response<Body>, ApiError> {
+    let tenant_shard_id: TenantShardId = parse_request_param(&request, "tenant_shard_id")?;
+    check_permission(&request, Some(tenant_shard_id.tenant_id))?;
+
+    let drop_cache: Option<bool> = parse_query_param(&request, "drop_cache")?;
+
+    let ctx = RequestContext::new(TaskKind::MgmtRequest, DownloadBehavior::Warn);
+    let state = get_state(&request);
+    state
+        .tenant_manager
+        .reset_tenant(tenant_shard_id, drop_cache.unwrap_or(false), ctx)
+        .await
+        .map_err(ApiError::InternalServerError)?;
+
+    json_response(StatusCode::OK, ())
+}
+
 async fn tenant_load_handler(
     mut request: Request<Body>,
     _cancel: CancellationToken,
@@ -822,7 +844,7 @@ async fn tenant_delete_handler(
     mgr::delete_tenant(state.conf, state.remote_storage.clone(), tenant_shard_id)
         .instrument(info_span!("tenant_delete_handler",
             tenant_id = %tenant_shard_id.tenant_id,
-            shard = tenant_shard_id.shard_slug()
+            shard = %tenant_shard_id.shard_slug()
         ))
         .await?;
 
@@ -1157,6 +1179,7 @@ async fn put_tenant_location_config_handler(
     let tenant_shard_id: TenantShardId = parse_request_param(&request, "tenant_shard_id")?;
 
     let request_data: TenantLocationConfigRequest = json_request(&mut request).await?;
+    let flush = parse_query_param(&request, "flush_ms")?.map(Duration::from_millis);
     check_permission(&request, Some(tenant_shard_id.tenant_id))?;
 
     let ctx = RequestContext::new(TaskKind::MgmtRequest, DownloadBehavior::Warn);
@@ -1170,7 +1193,7 @@ async fn put_tenant_location_config_handler(
             mgr::detach_tenant(conf, tenant_shard_id, true, &state.deletion_queue_client)
                 .instrument(info_span!("tenant_detach",
                     tenant_id = %tenant_shard_id.tenant_id,
-                    shard = tenant_shard_id.shard_slug()
+                    shard = %tenant_shard_id.shard_slug()
                 ))
                 .await
         {
@@ -1189,7 +1212,7 @@ async fn put_tenant_location_config_handler(
 
     state
         .tenant_manager
-        .upsert_location(tenant_shard_id, location_conf, &ctx)
+        .upsert_location(tenant_shard_id, location_conf, flush, &ctx)
         .await
         // TODO: badrequest assumes the caller was asking for something unreasonable, but in
         // principle we might have hit something like concurrent API calls to the same tenant,
@@ -1825,6 +1848,9 @@ pub fn make_router(
         .post("/v1/tenant/:tenant_id/detach", |r| {
             api_handler(r, tenant_detach_handler)
         })
+        .post("/v1/tenant/:tenant_shard_id/reset", |r| {
+            api_handler(r, tenant_reset_handler)
+        })
         .post("/v1/tenant/:tenant_id/load", |r| {
             api_handler(r, tenant_load_handler)
         })

pageserver/src/import_datadir.rs

@@ -2,17 +2,18 @@
 //! Import data and WAL from a PostgreSQL data directory and WAL segments into
 //! a neon Timeline.
 //!
+use std::io::SeekFrom;
 use std::path::{Path, PathBuf};
-use std::pin::Pin;
-use std::task::{self, Poll};
 
 use anyhow::{bail, ensure, Context, Result};
+use async_compression::tokio::bufread::ZstdDecoder;
 use async_compression::{tokio::write::ZstdEncoder, zstd::CParameter, Level};
 use bytes::Bytes;
 use camino::Utf8Path;
 use futures::StreamExt;
 use nix::NixPath;
-use tokio::io::{AsyncRead, AsyncReadExt, AsyncWrite, AsyncWriteExt};
+use tokio::fs::{File, OpenOptions};
+use tokio::io::{AsyncBufRead, AsyncRead, AsyncReadExt, AsyncSeekExt, AsyncWriteExt};
 use tokio_tar::Archive;
 use tokio_tar::Builder;
 use tokio_tar::HeaderMode;
@@ -628,70 +629,16 @@ async fn read_all_bytes(reader: &mut (impl AsyncRead + Unpin)) -> Result<Bytes>
     Ok(Bytes::from(buf))
 }
 
-/// An in-memory buffer implementing `AsyncWrite`, inserting yields every now and then
-///
-/// The number of yields is bounded by above by the number of times poll_write is called,
-/// so calling it with 8 KB chunks and 8 MB chunks gives the same number of yields in total.
-/// This is an explicit choice as the `YieldingVec` is meant to give the async executor
-/// breathing room between units of CPU intensive preparation of buffers to be written.
-/// Once a write call is issued, the whole buffer has been prepared already, so there is no
-/// gain in splitting up the memcopy further.
-struct YieldingVec {
-    yield_budget: usize,
-    // the buffer written into
-    buf: Vec<u8>,
-}
+pub async fn create_tar_zst(pgdata_path: &Utf8Path, tmp_path: &Utf8Path) -> Result<(File, u64)> {
+    let file = OpenOptions::new()
+        .create(true)
+        .truncate(true)
+        .read(true)
+        .write(true)
+        .open(&tmp_path)
+        .await
+        .with_context(|| format!("tempfile creation {tmp_path}"))?;
 
-impl YieldingVec {
-    fn new() -> Self {
-        Self {
-            yield_budget: 0,
-            buf: Vec::new(),
-        }
-    }
-    // Whether we should yield for a read operation of given size
-    fn should_yield(&mut self, add_buf_len: usize) -> bool {
-        // Set this limit to a small value so that we are a
-        // good async citizen and yield repeatedly (but not
-        // too often for many small writes to cause many yields)
-        const YIELD_DIST: usize = 1024;
-
-        let target_buf_len = self.buf.len() + add_buf_len;
-        let ret = self.yield_budget / YIELD_DIST < target_buf_len / YIELD_DIST;
-        if self.yield_budget < target_buf_len {
-            self.yield_budget += add_buf_len;
-        }
-        ret
-    }
-}
-
-impl AsyncWrite for YieldingVec {
-    fn poll_write(
-        mut self: Pin<&mut Self>,
-        cx: &mut task::Context<'_>,
-        buf: &[u8],
-    ) -> Poll<std::io::Result<usize>> {
-        if self.should_yield(buf.len()) {
-            cx.waker().wake_by_ref();
-            return Poll::Pending;
-        }
-        self.get_mut().buf.extend_from_slice(buf);
-        Poll::Ready(Ok(buf.len()))
-    }
-
-    fn poll_flush(self: Pin<&mut Self>, _cx: &mut task::Context<'_>) -> Poll<std::io::Result<()>> {
-        Poll::Ready(Ok(()))
-    }
-
-    fn poll_shutdown(
-        self: Pin<&mut Self>,
-        _cx: &mut task::Context<'_>,
-    ) -> Poll<std::io::Result<()>> {
-        Poll::Ready(Ok(()))
-    }
-}
-
-pub async fn create_tar_zst(pgdata_path: &Utf8Path) -> Result<Vec<u8>> {
     let mut paths = Vec::new();
     for entry in WalkDir::new(pgdata_path) {
         let entry = entry?;
@@ -706,7 +653,7 @@ pub async fn create_tar_zst(pgdata_path: &Utf8Path) -> Result<Vec<u8>> {
     // Do a sort to get a more consistent listing
     paths.sort_unstable();
     let zstd = ZstdEncoder::with_quality_and_params(
-        YieldingVec::new(),
+        file,
         Level::Default,
         &[CParameter::enable_long_distance_matching(true)],
     );
@@ -724,11 +671,22 @@ pub async fn create_tar_zst(pgdata_path: &Utf8Path) -> Result<Vec<u8>> {
     }
     let mut zstd = builder.into_inner().await?;
     zstd.shutdown().await?;
-    let compressed = zstd.into_inner();
-    let compressed_len = compressed.buf.len();
-    const INITDB_TAR_ZST_WARN_LIMIT: usize = 2_000_000;
+    let mut compressed = zstd.into_inner();
+    let compressed_len = compressed.metadata().await?.len();
+    const INITDB_TAR_ZST_WARN_LIMIT: u64 = 2 * 1024 * 1024;
     if compressed_len > INITDB_TAR_ZST_WARN_LIMIT {
         warn!("compressed {INITDB_PATH} size of {compressed_len} is above limit {INITDB_TAR_ZST_WARN_LIMIT}.");
     }
-    Ok(compressed.buf)
+    compressed.seek(SeekFrom::Start(0)).await?;
+    Ok((compressed, compressed_len))
+}
+
+pub async fn extract_tar_zst(
+    pgdata_path: &Utf8Path,
+    tar_zst: impl AsyncBufRead + Unpin,
+) -> Result<()> {
+    let tar = Box::pin(ZstdDecoder::new(tar_zst));
+    let mut archive = Archive::new(tar);
+    archive.unpack(pgdata_path).await?;
+    Ok(())
 }

pageserver/src/lib.rs

@@ -186,13 +186,6 @@ pub struct InitializationOrder {
     /// Each initial tenant load task carries this until completion.
     pub initial_tenant_load: Option<utils::completion::Completion>,
 
-    /// Barrier for when we can start initial logical size calculations.
-    pub initial_logical_size_can_start: utils::completion::Barrier,
-
-    /// Each timeline owns a clone of this to be consumed on the initial logical size calculation
-    /// attempt. It is important to drop this once the attempt has completed.
-    pub initial_logical_size_attempt: Option<utils::completion::Completion>,
-
     /// Barrier for when we can start any background jobs.
     ///
     /// This can be broken up later on, but right now there is just one class of a background job.
@@ -212,7 +205,7 @@ async fn timed<Fut: std::future::Future>(
     match tokio::time::timeout(warn_at, &mut fut).await {
         Ok(ret) => {
             tracing::info!(
-                task = name,
+                stage = name,
                 elapsed_ms = started.elapsed().as_millis(),
                 "completed"
             );
@@ -220,7 +213,7 @@ async fn timed<Fut: std::future::Future>(
         }
         Err(_) => {
             tracing::info!(
-                task = name,
+                stage = name,
                 elapsed_ms = started.elapsed().as_millis(),
                 "still waiting, taking longer than expected..."
             );
@@ -229,7 +222,7 @@ async fn timed<Fut: std::future::Future>(
 
             // this has a global allowed_errors
             tracing::warn!(
-                task = name,
+                stage = name,
                 elapsed_ms = started.elapsed().as_millis(),
                 "completed, took longer than expected"
             );

pageserver/src/metrics.rs

@@ -7,6 +7,7 @@ use metrics::{
     HistogramVec, IntCounter, IntCounterVec, IntGauge, IntGaugeVec, UIntGauge, UIntGaugeVec,
 };
 use once_cell::sync::Lazy;
+use pageserver_api::shard::TenantShardId;
 use strum::{EnumCount, IntoEnumIterator, VariantNames};
 use strum_macros::{EnumVariantNames, IntoStaticStr};
 use utils::id::{TenantId, TimelineId};
@@ -284,6 +285,63 @@ pub static PAGE_CACHE_SIZE: Lazy<PageCacheSizeMetrics> = Lazy::new(|| PageCacheS
     },
 });
 
+pub(crate) mod page_cache_eviction_metrics {
+    use std::num::NonZeroUsize;
+
+    use metrics::{register_int_counter_vec, IntCounter, IntCounterVec};
+    use once_cell::sync::Lazy;
+
+    #[derive(Clone, Copy)]
+    pub(crate) enum Outcome {
+        FoundSlotUnused { iters: NonZeroUsize },
+        FoundSlotEvicted { iters: NonZeroUsize },
+        ItersExceeded { iters: NonZeroUsize },
+    }
+
+    static ITERS_TOTAL_VEC: Lazy<IntCounterVec> = Lazy::new(|| {
+        register_int_counter_vec!(
+            "pageserver_page_cache_find_victim_iters_total",
+            "Counter for the number of iterations in the find_victim loop",
+            &["outcome"],
+        )
+        .expect("failed to define a metric")
+    });
+
+    static CALLS_VEC: Lazy<IntCounterVec> = Lazy::new(|| {
+        register_int_counter_vec!(
+            "pageserver_page_cache_find_victim_calls",
+            "Incremented at the end of each find_victim() call.\
+             Filter by outcome to get e.g., eviction rate.",
+            &["outcome"]
+        )
+        .unwrap()
+    });
+
+    pub(crate) fn observe(outcome: Outcome) {
+        macro_rules! dry {
+            ($label:literal, $iters:expr) => {{
+                static LABEL: &'static str = $label;
+                static ITERS_TOTAL: Lazy<IntCounter> =
+                    Lazy::new(|| ITERS_TOTAL_VEC.with_label_values(&[LABEL]));
+                static CALLS: Lazy<IntCounter> =
+                    Lazy::new(|| CALLS_VEC.with_label_values(&[LABEL]));
+                ITERS_TOTAL.inc_by(($iters.get()) as u64);
+                CALLS.inc();
+            }};
+        }
+        match outcome {
+            Outcome::FoundSlotUnused { iters } => dry!("found_empty", iters),
+            Outcome::FoundSlotEvicted { iters } => {
+                dry!("found_evicted", iters)
+            }
+            Outcome::ItersExceeded { iters } => {
+                dry!("err_iters_exceeded", iters);
+                super::page_cache_errors_inc(super::PageCacheErrorKind::EvictIterLimit);
+            }
+        }
+    }
+}
+
 pub(crate) static PAGE_CACHE_ACQUIRE_PINNED_SLOT_TIME: Lazy<Histogram> = Lazy::new(|| {
     register_histogram!(
         "pageserver_page_cache_acquire_pinned_slot_seconds",
@@ -293,14 +351,6 @@ pub(crate) static PAGE_CACHE_ACQUIRE_PINNED_SLOT_TIME: Lazy<Histogram> = Lazy::n
     .expect("failed to define a metric")
 });
 
-pub(crate) static PAGE_CACHE_FIND_VICTIMS_ITERS_TOTAL: Lazy<IntCounter> = Lazy::new(|| {
-    register_int_counter!(
-        "pageserver_page_cache_find_victim_iters_total",
-        "Counter for the number of iterations in the find_victim loop",
-    )
-    .expect("failed to define a metric")
-});
-
 static PAGE_CACHE_ERRORS: Lazy<IntCounterVec> = Lazy::new(|| {
     register_int_counter_vec!(
         "page_cache_errors_total",
@@ -402,6 +452,129 @@ static CURRENT_LOGICAL_SIZE: Lazy<UIntGaugeVec> = Lazy::new(|| {
     .expect("failed to define current logical size metric")
 });
 
+pub(crate) mod initial_logical_size {
+    use metrics::{register_int_counter, register_int_counter_vec, IntCounter, IntCounterVec};
+    use once_cell::sync::Lazy;
+
+    pub(crate) struct StartCalculation(IntCounterVec);
+    pub(crate) static START_CALCULATION: Lazy<StartCalculation> = Lazy::new(|| {
+        StartCalculation(
+            register_int_counter_vec!(
+                "pageserver_initial_logical_size_start_calculation",
+                "Incremented each time we start an initial logical size calculation attempt. \
+                 The `circumstances` label provides some additional details.",
+                &["attempt", "circumstances"]
+            )
+            .unwrap(),
+        )
+    });
+
+    struct DropCalculation {
+        first: IntCounter,
+        retry: IntCounter,
+    }
+
+    static DROP_CALCULATION: Lazy<DropCalculation> = Lazy::new(|| {
+        let vec = register_int_counter_vec!(
+            "pageserver_initial_logical_size_drop_calculation",
+            "Incremented each time we abort a started size calculation attmpt.",
+            &["attempt"]
+        )
+        .unwrap();
+        DropCalculation {
+            first: vec.with_label_values(&["first"]),
+            retry: vec.with_label_values(&["retry"]),
+        }
+    });
+
+    pub(crate) struct Calculated {
+        pub(crate) births: IntCounter,
+        pub(crate) deaths: IntCounter,
+    }
+
+    pub(crate) static CALCULATED: Lazy<Calculated> = Lazy::new(|| Calculated {
+        births: register_int_counter!(
+            "pageserver_initial_logical_size_finish_calculation",
+            "Incremented every time we finish calculation of initial logical size.\
+             If everything is working well, this should happen at most once per Timeline object."
+        )
+        .unwrap(),
+        deaths: register_int_counter!(
+            "pageserver_initial_logical_size_drop_finished_calculation",
+            "Incremented when we drop a finished initial logical size calculation result.\
+             Mainly useful to turn pageserver_initial_logical_size_finish_calculation into a gauge."
+        )
+        .unwrap(),
+    });
+
+    pub(crate) struct OngoingCalculationGuard {
+        inc_drop_calculation: Option<IntCounter>,
+    }
+
+    #[derive(strum_macros::IntoStaticStr)]
+    pub(crate) enum StartCircumstances {
+        EmptyInitial,
+        SkippedConcurrencyLimiter,
+        AfterBackgroundTasksRateLimit,
+    }
+
+    impl StartCalculation {
+        pub(crate) fn first(&self, circumstances: StartCircumstances) -> OngoingCalculationGuard {
+            let circumstances_label: &'static str = circumstances.into();
+            self.0.with_label_values(&["first", circumstances_label]);
+            OngoingCalculationGuard {
+                inc_drop_calculation: Some(DROP_CALCULATION.first.clone()),
+            }
+        }
+        pub(crate) fn retry(&self, circumstances: StartCircumstances) -> OngoingCalculationGuard {
+            let circumstances_label: &'static str = circumstances.into();
+            self.0.with_label_values(&["retry", circumstances_label]);
+            OngoingCalculationGuard {
+                inc_drop_calculation: Some(DROP_CALCULATION.retry.clone()),
+            }
+        }
+    }
+
+    impl Drop for OngoingCalculationGuard {
+        fn drop(&mut self) {
+            if let Some(counter) = self.inc_drop_calculation.take() {
+                counter.inc();
+            }
+        }
+    }
+
+    impl OngoingCalculationGuard {
+        pub(crate) fn calculation_result_saved(mut self) -> FinishedCalculationGuard {
+            drop(self.inc_drop_calculation.take());
+            CALCULATED.births.inc();
+            FinishedCalculationGuard {
+                inc_on_drop: CALCULATED.deaths.clone(),
+            }
+        }
+    }
+
+    pub(crate) struct FinishedCalculationGuard {
+        inc_on_drop: IntCounter,
+    }
+
+    impl Drop for FinishedCalculationGuard {
+        fn drop(&mut self) {
+            self.inc_on_drop.inc();
+        }
+    }
+
+    // context: https://github.com/neondatabase/neon/issues/5963
+    pub(crate) static TIMELINES_WHERE_WALRECEIVER_GOT_APPROXIMATE_SIZE: Lazy<IntCounter> =
+        Lazy::new(|| {
+            register_int_counter!(
+                "pageserver_initial_logical_size_timelines_where_walreceiver_got_approximate_size",
+                "Counter for the following event: walreceiver calls\
+                 Timeline::get_current_logical_size() and it returns `Approximate` for the first time."
+            )
+            .unwrap()
+        });
+}
+
 pub(crate) static TENANT_STATE_METRIC: Lazy<UIntGaugeVec> = Lazy::new(|| {
     register_uint_gauge_vec!(
         "pageserver_tenant_states_count",
@@ -650,6 +823,7 @@ const STORAGE_IO_TIME_BUCKETS: &[f64] = &[
 )]
 pub(crate) enum StorageIoOperation {
     Open,
+    OpenAfterReplace,
     Close,
     CloseByReplace,
     Read,
@@ -663,6 +837,7 @@ impl StorageIoOperation {
     pub fn as_str(&self) -> &'static str {
         match self {
             StorageIoOperation::Open => "open",
+            StorageIoOperation::OpenAfterReplace => "open-after-replace",
             StorageIoOperation::Close => "close",
             StorageIoOperation::CloseByReplace => "close-by-replace",
             StorageIoOperation::Read => "read",
@@ -717,6 +892,25 @@ pub(crate) static STORAGE_IO_SIZE: Lazy<IntGaugeVec> = Lazy::new(|| {
     .expect("failed to define a metric")
 });
 
+pub(crate) mod virtual_file_descriptor_cache {
+    use super::*;
+
+    pub(crate) static SIZE_MAX: Lazy<UIntGauge> = Lazy::new(|| {
+        register_uint_gauge!(
+            "pageserver_virtual_file_descriptor_cache_size_max",
+            "Maximum number of open file descriptors in the cache."
+        )
+        .unwrap()
+    });
+
+    // SIZE_CURRENT: derive it like so:
+    // ```
+    // sum (pageserver_io_operations_seconds_count{operation=~"^(open|open-after-replace)$")
+    // -ignoring(operation)
+    // sum(pageserver_io_operations_seconds_count{operation=~"^(close|close-by-replace)$"}
+    // ```
+}
+
 #[derive(Debug)]
 struct GlobalAndPerTimelineHistogram {
     global: Histogram,
@@ -1043,6 +1237,30 @@ pub(crate) static DELETION_QUEUE: Lazy<DeletionQueueMetrics> = Lazy::new(|| {
 }
 });
 
+pub(crate) struct WalIngestMetrics {
+    pub(crate) records_received: IntCounter,
+    pub(crate) records_committed: IntCounter,
+    pub(crate) records_filtered: IntCounter,
+}
+
+pub(crate) static WAL_INGEST: Lazy<WalIngestMetrics> = Lazy::new(|| WalIngestMetrics {
+    records_received: register_int_counter!(
+        "pageserver_wal_ingest_records_received",
+        "Number of WAL records received from safekeepers"
+    )
+    .expect("failed to define a metric"),
+    records_committed: register_int_counter!(
+        "pageserver_wal_ingest_records_committed",
+        "Number of WAL records which resulted in writes to pageserver storage"
+    )
+    .expect("failed to define a metric"),
+    records_filtered: register_int_counter!(
+        "pageserver_wal_ingest_records_filtered",
+        "Number of WAL records filtered out due to sharding"
+    )
+    .expect("failed to define a metric"),
+});
+
 #[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
 pub enum RemoteOpKind {
     Upload,
@@ -1252,9 +1470,20 @@ pub(crate) static WAL_REDO_RECORD_COUNTER: Lazy<IntCounter> = Lazy::new(|| {
     .unwrap()
 });
 
+pub(crate) static WAL_REDO_PROCESS_LAUNCH_DURATION_HISTOGRAM: Lazy<Histogram> = Lazy::new(|| {
+    register_histogram!(
+        "pageserver_wal_redo_process_launch_duration",
+        "Histogram of the duration of successful WalRedoProcess::launch calls",
+        redo_histogram_time_buckets!(),
+    )
+    .expect("failed to define a metric")
+});
+
 pub(crate) struct WalRedoProcessCounters {
     pub(crate) started: IntCounter,
     pub(crate) killed_by_cause: enum_map::EnumMap<WalRedoKillCause, IntCounter>,
+    pub(crate) active_stderr_logger_tasks_started: IntCounter,
+    pub(crate) active_stderr_logger_tasks_finished: IntCounter,
 }
 
 #[derive(Debug, enum_map::Enum, strum_macros::IntoStaticStr)]
@@ -1278,6 +1507,19 @@ impl Default for WalRedoProcessCounters {
             &["cause"],
         )
         .unwrap();
+
+        let active_stderr_logger_tasks_started = register_int_counter!(
+            "pageserver_walredo_stderr_logger_tasks_started_total",
+            "Number of active walredo stderr logger tasks that have started",
+        )
+        .unwrap();
+
+        let active_stderr_logger_tasks_finished = register_int_counter!(
+            "pageserver_walredo_stderr_logger_tasks_finished_total",
+            "Number of active walredo stderr logger tasks that have finished",
+        )
+        .unwrap();
+
         Self {
             started,
             killed_by_cause: EnumMap::from_array(std::array::from_fn(|i| {
@@ -1285,6 +1527,8 @@ impl Default for WalRedoProcessCounters {
                 let cause_str: &'static str = cause.into();
                 killed.with_label_values(&[cause_str])
             })),
+            active_stderr_logger_tasks_started,
+            active_stderr_logger_tasks_finished,
         }
     }
 }
@@ -1571,9 +1815,9 @@ pub struct RemoteTimelineClientMetrics {
 }
 
 impl RemoteTimelineClientMetrics {
-    pub fn new(tenant_id: &TenantId, timeline_id: &TimelineId) -> Self {
+    pub fn new(tenant_shard_id: &TenantShardId, timeline_id: &TimelineId) -> Self {
         RemoteTimelineClientMetrics {
-            tenant_id: tenant_id.to_string(),
+            tenant_id: tenant_shard_id.tenant_id.to_string(),
             timeline_id: timeline_id.to_string(),
             calls_unfinished_gauge: Mutex::new(HashMap::default()),
             bytes_started_counter: Mutex::new(HashMap::default()),
@@ -1944,6 +2188,8 @@ pub fn preinitialize_metrics() {
     // Tenant manager stats
     Lazy::force(&TENANT_MANAGER);
 
+    Lazy::force(&crate::tenant::storage_layer::layer::LAYER_IMPL_METRICS);
+
     // countervecs
     [&BACKGROUND_LOOP_PERIOD_OVERRUN_COUNT]
         .into_iter()
@@ -1961,6 +2207,7 @@ pub fn preinitialize_metrics() {
         &WAL_REDO_TIME,
         &WAL_REDO_RECORDS_HISTOGRAM,
         &WAL_REDO_BYTES_HISTOGRAM,
+        &WAL_REDO_PROCESS_LAUNCH_DURATION_HISTOGRAM,
     ]
     .into_iter()
     .for_each(|h| {

pageserver/src/page_cache.rs

@@ -88,7 +88,11 @@ use utils::{
     lsn::Lsn,
 };
 
-use crate::{context::RequestContext, metrics::PageCacheSizeMetrics, repository::Key};
+use crate::{
+    context::RequestContext,
+    metrics::{page_cache_eviction_metrics, PageCacheSizeMetrics},
+    repository::Key,
+};
 
 static PAGE_CACHE: OnceCell<PageCache> = OnceCell::new();
 const TEST_PAGE_CACHE_SIZE: usize = 50;
@@ -897,8 +901,10 @@ impl PageCache {
                             // Note that just yielding to tokio during iteration without such
                             // priority boosting is likely counter-productive. We'd just give more opportunities
                             // for B to bump usage count, further starving A.
-                            crate::metrics::page_cache_errors_inc(
-                                crate::metrics::PageCacheErrorKind::EvictIterLimit,
+                            page_cache_eviction_metrics::observe(
+                                page_cache_eviction_metrics::Outcome::ItersExceeded {
+                                    iters: iters.try_into().unwrap(),
+                                },
                             );
                             anyhow::bail!("exceeded evict iter limit");
                         }
@@ -909,8 +915,18 @@ impl PageCache {
                     // remove mapping for old buffer
                     self.remove_mapping(old_key);
                     inner.key = None;
+                    page_cache_eviction_metrics::observe(
+                        page_cache_eviction_metrics::Outcome::FoundSlotEvicted {
+                            iters: iters.try_into().unwrap(),
+                        },
+                    );
+                } else {
+                    page_cache_eviction_metrics::observe(
+                        page_cache_eviction_metrics::Outcome::FoundSlotUnused {
+                            iters: iters.try_into().unwrap(),
+                        },
+                    );
                 }
-                crate::metrics::PAGE_CACHE_FIND_VICTIMS_ITERS_TOTAL.inc_by(iters as u64);
                 return Ok((slot_idx, inner));
             }
         }

pageserver/src/page_service.rs

@@ -53,21 +53,23 @@ use crate::context::{DownloadBehavior, RequestContext};
 use crate::import_datadir::import_wal_from_tar;
 use crate::metrics;
 use crate::metrics::LIVE_CONNECTIONS_COUNT;
+use crate::pgdatadir_mapping::rel_block_to_key;
 use crate::task_mgr;
 use crate::task_mgr::TaskKind;
 use crate::tenant::debug_assert_current_span_has_tenant_and_timeline_id;
 use crate::tenant::mgr;
 use crate::tenant::mgr::get_active_tenant_with_timeout;
 use crate::tenant::mgr::GetActiveTenantError;
+use crate::tenant::mgr::ShardSelector;
 use crate::tenant::Timeline;
 use crate::trace::Tracer;
 
 use postgres_ffi::pg_constants::DEFAULTTABLESPACE_OID;
 use postgres_ffi::BLCKSZ;
 
-// How long we may block waiting for a [`TenantSlot::InProgress`]` and/or a [`Tenant`] which
+// How long we may wait for a [`TenantSlot::InProgress`]` and/or a [`Tenant`] which
 // is not yet in state [`TenantState::Active`].
-const ACTIVE_TENANT_TIMEOUT: Duration = Duration::from_millis(5000);
+const ACTIVE_TENANT_TIMEOUT: Duration = Duration::from_millis(30000);
 
 /// Read the end of a tar archive.
 ///
@@ -399,18 +401,25 @@ impl PageServerHandler {
     {
         debug_assert_current_span_has_tenant_and_timeline_id();
 
-        // Make request tracer if needed
+        // Note that since one connection may contain getpage requests that target different
+        // shards (e.g. during splitting when the compute is not yet aware of the split), the tenant
+        // that we look up here may not be the one that serves all the actual requests: we will double
+        // check the mapping of key->shard later before calling into Timeline for getpage requests.
         let tenant = mgr::get_active_tenant_with_timeout(
             tenant_id,
+            ShardSelector::First,
             ACTIVE_TENANT_TIMEOUT,
             &task_mgr::shutdown_token(),
         )
         .await?;
+
+        // Make request tracer if needed
         let mut tracer = if tenant.get_trace_read_requests() {
             let connection_id = ConnectionId::generate();
-            let path = tenant
-                .conf
-                .trace_path(&tenant_id, &timeline_id, &connection_id);
+            let path =
+                tenant
+                    .conf
+                    .trace_path(&tenant.tenant_shard_id(), &timeline_id, &connection_id);
             Some(Tracer::new(path))
         } else {
             None
@@ -562,6 +571,7 @@ impl PageServerHandler {
         info!("creating new timeline");
         let tenant = get_active_tenant_with_timeout(
             tenant_id,
+            ShardSelector::Zero,
             ACTIVE_TENANT_TIMEOUT,
             &task_mgr::shutdown_token(),
         )
@@ -624,7 +634,7 @@ impl PageServerHandler {
         debug_assert_current_span_has_tenant_and_timeline_id();
 
         let timeline = self
-            .get_active_tenant_timeline(tenant_id, timeline_id)
+            .get_active_tenant_timeline(tenant_id, timeline_id, ShardSelector::Zero)
             .await?;
         let last_record_lsn = timeline.get_last_record_lsn();
         if last_record_lsn != start_lsn {
@@ -803,9 +813,49 @@ impl PageServerHandler {
         }
         */
 
-        let page = timeline
-            .get_rel_page_at_lsn(req.rel, req.blkno, lsn, req.latest, ctx)
-            .await?;
+        let key = rel_block_to_key(req.rel, req.blkno);
+        let page = if timeline.get_shard_identity().is_key_local(&key) {
+            timeline
+                .get_rel_page_at_lsn(req.rel, req.blkno, lsn, req.latest, ctx)
+                .await?
+        } else {
+            // The Tenant shard we looked up at connection start does not hold this particular
+            // key: look for other shards in this tenant.  This scenario occurs if a pageserver
+            // has multiple shards for the same tenant.
+            //
+            // TODO: optimize this (https://github.com/neondatabase/neon/pull/6037)
+            let timeline = match self
+                .get_active_tenant_timeline(
+                    timeline.tenant_shard_id.tenant_id,
+                    timeline.timeline_id,
+                    ShardSelector::Page(key),
+                )
+                .await
+            {
+                Ok(t) => t,
+                Err(GetActiveTimelineError::Tenant(GetActiveTenantError::NotFound(_))) => {
+                    // We already know this tenant exists in general, because we resolved it at
+                    // start of connection.  Getting a NotFound here indicates that the shard containing
+                    // the requested page is not present on this node.
+
+                    // TODO: this should be some kind of structured error that the client will understand,
+                    // so that it can block until its config is updated: this error is expected in the case
+                    // that the Tenant's shards' placements are being updated and the client hasn't been
+                    // informed yet.
+                    //
+                    // https://github.com/neondatabase/neon/issues/6038
+                    return Err(anyhow::anyhow!("Request routed to wrong shard"));
+                }
+                Err(e) => return Err(e.into()),
+            };
+
+            // Take a GateGuard for the duration of this request.  If we were using our main Timeline object,
+            // the GateGuard was already held over the whole connection.
+            let _timeline_guard = timeline.gate.enter().map_err(|_| QueryError::Shutdown)?;
+            timeline
+                .get_rel_page_at_lsn(req.rel, req.blkno, lsn, req.latest, ctx)
+                .await?
+        };
 
         Ok(PagestreamBeMessage::GetPage(PagestreamGetPageResponse {
             page,
@@ -834,7 +884,7 @@ impl PageServerHandler {
 
         // check that the timeline exists
         let timeline = self
-            .get_active_tenant_timeline(tenant_id, timeline_id)
+            .get_active_tenant_timeline(tenant_id, timeline_id, ShardSelector::Zero)
             .await?;
         let latest_gc_cutoff_lsn = timeline.get_latest_gc_cutoff_lsn();
         if let Some(lsn) = lsn {
@@ -940,9 +990,11 @@ impl PageServerHandler {
         &self,
         tenant_id: TenantId,
         timeline_id: TimelineId,
+        selector: ShardSelector,
     ) -> Result<Arc<Timeline>, GetActiveTimelineError> {
         let tenant = get_active_tenant_with_timeout(
             tenant_id,
+            selector,
             ACTIVE_TENANT_TIMEOUT,
             &task_mgr::shutdown_token(),
         )
@@ -1116,7 +1168,7 @@ where
 
             self.check_permission(Some(tenant_id))?;
             let timeline = self
-                .get_active_tenant_timeline(tenant_id, timeline_id)
+                .get_active_tenant_timeline(tenant_id, timeline_id, ShardSelector::Zero)
                 .await?;
 
             let end_of_timeline = timeline.get_last_record_rlsn();
@@ -1303,6 +1355,7 @@ where
 
             let tenant = get_active_tenant_with_timeout(
                 tenant_id,
+                ShardSelector::Zero,
                 ACTIVE_TENANT_TIMEOUT,
                 &task_mgr::shutdown_token(),
             )

pageserver/src/pgdatadir_mapping.rs

@@ -13,6 +13,7 @@ use crate::repository::*;
 use crate::walrecord::NeonWalRecord;
 use anyhow::Context;
 use bytes::{Buf, Bytes};
+use pageserver_api::key::is_rel_block_key;
 use pageserver_api::reltag::{RelTag, SlruKind};
 use postgres_ffi::relfile_utils::{FSM_FORKNUM, VISIBILITYMAP_FORKNUM};
 use postgres_ffi::BLCKSZ;
@@ -282,6 +283,10 @@ impl Timeline {
     }
 
     /// Get a list of all existing relations in given tablespace and database.
+    ///
+    /// # Cancel-Safety
+    ///
+    /// This method is cancellation-safe.
     pub async fn list_rels(
         &self,
         spcnode: Oid,
@@ -630,6 +635,10 @@ impl Timeline {
     ///
     /// Only relation blocks are counted currently. That excludes metadata,
     /// SLRUs, twophase files etc.
+    ///
+    /// # Cancel-Safety
+    ///
+    /// This method is cancellation-safe.
     pub async fn get_current_logical_size_non_incremental(
         &self,
         lsn: Lsn,
@@ -1314,7 +1323,7 @@ impl<'a> DatadirModification<'a> {
         // Flush relation and  SLRU data blocks, keep metadata.
         let mut retained_pending_updates = HashMap::new();
         for (key, value) in self.pending_updates.drain() {
-            if is_rel_block_key(key) || is_slru_block_key(key) {
+            if is_rel_block_key(&key) || is_slru_block_key(key) {
                 // This bails out on first error without modifying pending_updates.
                 // That's Ok, cf this function's doc comment.
                 writer.put(key, self.lsn, &value, ctx).await?;
@@ -1359,6 +1368,10 @@ impl<'a> DatadirModification<'a> {
         Ok(())
     }
 
+    pub(crate) fn is_empty(&self) -> bool {
+        self.pending_updates.is_empty() && self.pending_deletions.is_empty()
+    }
+
     // Internal helper functions to batch the modifications
 
     async fn get(&self, key: Key, ctx: &RequestContext) -> Result<Bytes, PageReconstructError> {
@@ -1570,7 +1583,7 @@ fn rel_dir_to_key(spcnode: Oid, dbnode: Oid) -> Key {
     }
 }
 
-fn rel_block_to_key(rel: RelTag, blknum: BlockNumber) -> Key {
+pub(crate) fn rel_block_to_key(rel: RelTag, blknum: BlockNumber) -> Key {
     Key {
         field1: 0x00,
         field2: rel.spcnode,
@@ -1769,10 +1782,6 @@ pub fn key_to_rel_block(key: Key) -> anyhow::Result<(RelTag, BlockNumber)> {
     })
 }
 
-fn is_rel_block_key(key: Key) -> bool {
-    key.field1 == 0x00 && key.field4 != 0
-}
-
 pub fn is_rel_fsm_block_key(key: Key) -> bool {
     key.field1 == 0x00 && key.field4 != 0 && key.field5 == FSM_FORKNUM && key.field6 != 0xffffffff
 }

pageserver/src/repository.rs

@@ -138,6 +138,14 @@ pub struct GcResult {
 
     #[serde(serialize_with = "serialize_duration_as_millis")]
     pub elapsed: Duration,
+
+    /// The layers which were garbage collected.
+    ///
+    /// Used in `/v1/tenant/:tenant_id/timeline/:timeline_id/do_gc` to wait for the layers to be
+    /// dropped in tests.
+    #[cfg(feature = "testing")]
+    #[serde(skip)]
+    pub(crate) doomed_layers: Vec<crate::tenant::storage_layer::Layer>,
 }
 
 // helper function for `GcResult`, serializing a `Duration` as an integer number of milliseconds
@@ -158,5 +166,11 @@ impl AddAssign for GcResult {
         self.layers_removed += other.layers_removed;
 
         self.elapsed += other.elapsed;
+
+        #[cfg(feature = "testing")]
+        {
+            let mut other = other;
+            self.doomed_layers.append(&mut other.doomed_layers);
+        }
     }
 }

pageserver/src/tenant/config.rs

@@ -8,9 +8,12 @@
 //! We cannot use global or default config instead, because wrong settings
 //! may lead to a data loss.
 //!
-use anyhow::Context;
+use anyhow::bail;
 use pageserver_api::models;
+use pageserver_api::shard::{ShardCount, ShardIdentity, ShardNumber, ShardStripeSize};
+use serde::de::IntoDeserializer;
 use serde::{Deserialize, Serialize};
+use serde_json::Value;
 use std::num::NonZeroU64;
 use std::time::Duration;
 use utils::generation::Generation;
@@ -88,6 +91,14 @@ pub(crate) struct LocationConf {
     /// The location-specific part of the configuration, describes the operating
     /// mode of this pageserver for this tenant.
     pub(crate) mode: LocationMode,
+
+    /// The detailed shard identity.  This structure is already scoped within
+    /// a TenantShardId, but we need the full ShardIdentity to enable calculating
+    /// key->shard mappings.
+    #[serde(default = "ShardIdentity::unsharded")]
+    #[serde(skip_serializing_if = "ShardIdentity::is_unsharded")]
+    pub(crate) shard: ShardIdentity,
+
     /// The pan-cluster tenant configuration, the same on all locations
     pub(crate) tenant_conf: TenantConfOpt,
 }
@@ -160,6 +171,8 @@ impl LocationConf {
                 generation,
                 attach_mode: AttachmentMode::Single,
             }),
+            // Legacy configuration loads are always from tenants created before sharding existed.
+            shard: ShardIdentity::unsharded(),
             tenant_conf,
         }
     }
@@ -187,6 +200,7 @@ impl LocationConf {
 
         fn get_generation(conf: &'_ models::LocationConfig) -> Result<Generation, anyhow::Error> {
             conf.generation
+                .map(Generation::new)
                 .ok_or_else(|| anyhow::anyhow!("Generation must be set when attaching"))
         }
 
@@ -226,7 +240,21 @@ impl LocationConf {
             }
         };
 
-        Ok(Self { mode, tenant_conf })
+        let shard = if conf.shard_count == 0 {
+            ShardIdentity::unsharded()
+        } else {
+            ShardIdentity::new(
+                ShardNumber(conf.shard_number),
+                ShardCount(conf.shard_count),
+                ShardStripeSize(conf.shard_stripe_size),
+            )?
+        };
+
+        Ok(Self {
+            shard,
+            mode,
+            tenant_conf,
+        })
     }
 }
 
@@ -241,6 +269,7 @@ impl Default for LocationConf {
                 attach_mode: AttachmentMode::Single,
             }),
             tenant_conf: TenantConfOpt::default(),
+            shard: ShardIdentity::unsharded(),
         }
     }
 }
@@ -494,105 +523,49 @@ impl Default for TenantConf {
     }
 }
 
-// Helper function to standardize the error messages we produce on bad durations
-//
-// Intended to be used with anyhow's `with_context`, e.g.:
-//
-//   let value = result.with_context(bad_duration("name", &value))?;
-//
-fn bad_duration<'a>(field_name: &'static str, value: &'a str) -> impl 'a + Fn() -> String {
-    move || format!("Cannot parse `{field_name}` duration {value:?}")
-}
-
 impl TryFrom<&'_ models::TenantConfig> for TenantConfOpt {
     type Error = anyhow::Error;
 
     fn try_from(request_data: &'_ models::TenantConfig) -> Result<Self, Self::Error> {
-        let mut tenant_conf = TenantConfOpt::default();
+        // Convert the request_data to a JSON Value
+        let json_value: Value = serde_json::to_value(request_data)?;
 
-        if let Some(gc_period) = &request_data.gc_period {
-            tenant_conf.gc_period = Some(
-                humantime::parse_duration(gc_period)
-                    .with_context(bad_duration("gc_period", gc_period))?,
-            );
-        }
-        tenant_conf.gc_horizon = request_data.gc_horizon;
-        tenant_conf.image_creation_threshold = request_data.image_creation_threshold;
+        // Create a Deserializer from the JSON Value
+        let deserializer = json_value.into_deserializer();
 
-        if let Some(pitr_interval) = &request_data.pitr_interval {
-            tenant_conf.pitr_interval = Some(
-                humantime::parse_duration(pitr_interval)
-                    .with_context(bad_duration("pitr_interval", pitr_interval))?,
-            );
-        }
-
-        if let Some(walreceiver_connect_timeout) = &request_data.walreceiver_connect_timeout {
-            tenant_conf.walreceiver_connect_timeout = Some(
-                humantime::parse_duration(walreceiver_connect_timeout).with_context(
-                    bad_duration("walreceiver_connect_timeout", walreceiver_connect_timeout),
-                )?,
-            );
-        }
-        if let Some(lagging_wal_timeout) = &request_data.lagging_wal_timeout {
-            tenant_conf.lagging_wal_timeout = Some(
-                humantime::parse_duration(lagging_wal_timeout)
-                    .with_context(bad_duration("lagging_wal_timeout", lagging_wal_timeout))?,
-            );
-        }
-        if let Some(max_lsn_wal_lag) = request_data.max_lsn_wal_lag {
-            tenant_conf.max_lsn_wal_lag = Some(max_lsn_wal_lag);
-        }
-        if let Some(trace_read_requests) = request_data.trace_read_requests {
-            tenant_conf.trace_read_requests = Some(trace_read_requests);
-        }
-
-        tenant_conf.checkpoint_distance = request_data.checkpoint_distance;
-        if let Some(checkpoint_timeout) = &request_data.checkpoint_timeout {
-            tenant_conf.checkpoint_timeout = Some(
-                humantime::parse_duration(checkpoint_timeout)
-                    .with_context(bad_duration("checkpoint_timeout", checkpoint_timeout))?,
-            );
-        }
-
-        tenant_conf.compaction_target_size = request_data.compaction_target_size;
-        tenant_conf.compaction_threshold = request_data.compaction_threshold;
-
-        if let Some(compaction_period) = &request_data.compaction_period {
-            tenant_conf.compaction_period = Some(
-                humantime::parse_duration(compaction_period)
-                    .with_context(bad_duration("compaction_period", compaction_period))?,
-            );
-        }
-
-        if let Some(eviction_policy) = &request_data.eviction_policy {
-            tenant_conf.eviction_policy = Some(
-                serde::Deserialize::deserialize(eviction_policy)
-                    .context("parse field `eviction_policy`")?,
-            );
-        }
-
-        tenant_conf.min_resident_size_override = request_data.min_resident_size_override;
-
-        if let Some(evictions_low_residence_duration_metric_threshold) =
-            &request_data.evictions_low_residence_duration_metric_threshold
-        {
-            tenant_conf.evictions_low_residence_duration_metric_threshold = Some(
-                humantime::parse_duration(evictions_low_residence_duration_metric_threshold)
-                    .with_context(bad_duration(
-                        "evictions_low_residence_duration_metric_threshold",
-                        evictions_low_residence_duration_metric_threshold,
-                    ))?,
-            );
-        }
-        tenant_conf.gc_feedback = request_data.gc_feedback;
+        // Use serde_path_to_error to deserialize the JSON Value into TenantConfOpt
+        let tenant_conf: TenantConfOpt = serde_path_to_error::deserialize(deserializer)?;
 
         Ok(tenant_conf)
     }
 }
 
+impl TryFrom<toml_edit::Item> for TenantConfOpt {
+    type Error = anyhow::Error;
+
+    fn try_from(item: toml_edit::Item) -> Result<Self, Self::Error> {
+        match item {
+            toml_edit::Item::Value(value) => {
+                let d = value.into_deserializer();
+                return serde_path_to_error::deserialize(d)
+                    .map_err(|e| anyhow::anyhow!("{}: {}", e.path(), e.inner().message()));
+            }
+            toml_edit::Item::Table(table) => {
+                let deserializer = toml_edit::de::Deserializer::new(table.into());
+                return serde_path_to_error::deserialize(deserializer)
+                    .map_err(|e| anyhow::anyhow!("{}: {}", e.path(), e.inner().message()));
+            }
+            _ => {
+                bail!("expected non-inline table but found {item}")
+            }
+        }
+    }
+}
+
 #[cfg(test)]
 mod tests {
     use super::*;
+    use models::TenantConfig;
 
     #[test]
     fn de_serializing_pageserver_config_omits_empty_values() {
@@ -609,4 +582,38 @@ mod tests {
         assert_eq!(json_form, "{\"gc_horizon\":42}");
         assert_eq!(small_conf, serde_json::from_str(&json_form).unwrap());
     }
+
+    #[test]
+    fn test_try_from_models_tenant_config_err() {
+        let tenant_config = models::TenantConfig {
+            lagging_wal_timeout: Some("5a".to_string()),
+            ..TenantConfig::default()
+        };
+
+        let tenant_conf_opt = TenantConfOpt::try_from(&tenant_config);
+
+        assert!(
+            tenant_conf_opt.is_err(),
+            "Suceeded to convert TenantConfig to TenantConfOpt"
+        );
+
+        let expected_error_str =
+            "lagging_wal_timeout: invalid value: string \"5a\", expected a duration";
+        assert_eq!(tenant_conf_opt.unwrap_err().to_string(), expected_error_str);
+    }
+
+    #[test]
+    fn test_try_from_models_tenant_config_success() {
+        let tenant_config = models::TenantConfig {
+            lagging_wal_timeout: Some("5s".to_string()),
+            ..TenantConfig::default()
+        };
+
+        let tenant_conf_opt = TenantConfOpt::try_from(&tenant_config).unwrap();
+
+        assert_eq!(
+            tenant_conf_opt.lagging_wal_timeout,
+            Some(Duration::from_secs(5))
+        );
+    }
 }

pageserver/src/tenant/delete.rs

@@ -2,22 +2,19 @@ use std::sync::Arc;
 
 use anyhow::Context;
 use camino::{Utf8Path, Utf8PathBuf};
-use pageserver_api::models::TenantState;
+use pageserver_api::{models::TenantState, shard::TenantShardId};
 use remote_storage::{GenericRemoteStorage, RemotePath};
 use tokio::sync::OwnedMutexGuard;
 use tokio_util::sync::CancellationToken;
-use tracing::{error, instrument, warn, Instrument, Span};
+use tracing::{error, instrument, Instrument, Span};
 
-use utils::{
-    backoff, completion, crashsafe, fs_ext,
-    id::{TenantId, TimelineId},
-};
+use utils::{backoff, completion, crashsafe, fs_ext, id::TimelineId};
 
 use crate::{
     config::PageServerConf,
     context::RequestContext,
     task_mgr::{self, TaskKind},
-    InitializationOrder,
+    tenant::mgr::{TenantSlot, TenantsMapRemoveResult},
 };
 
 use super::{
@@ -59,10 +56,10 @@ type DeletionGuard = tokio::sync::OwnedMutexGuard<DeleteTenantFlow>;
 
 fn remote_tenant_delete_mark_path(
     conf: &PageServerConf,
-    tenant_id: &TenantId,
+    tenant_shard_id: &TenantShardId,
 ) -> anyhow::Result<RemotePath> {
     let tenant_remote_path = conf
-        .tenant_path(tenant_id)
+        .tenant_path(tenant_shard_id)
         .strip_prefix(&conf.workdir)
         .context("Failed to strip workdir prefix")
         .and_then(RemotePath::new)
@@ -73,15 +70,17 @@ fn remote_tenant_delete_mark_path(
 async fn create_remote_delete_mark(
     conf: &PageServerConf,
     remote_storage: &GenericRemoteStorage,
-    tenant_id: &TenantId,
+    tenant_shard_id: &TenantShardId,
 ) -> Result<(), DeleteTenantError> {
-    let remote_mark_path = remote_tenant_delete_mark_path(conf, tenant_id)?;
+    let remote_mark_path = remote_tenant_delete_mark_path(conf, tenant_shard_id)?;
 
     let data: &[u8] = &[];
     backoff::retry(
         || async {
+            let data = bytes::Bytes::from_static(data);
+            let stream = futures::stream::once(futures::future::ready(Ok(data)));
             remote_storage
-                .upload(data, 0, &remote_mark_path, None)
+                .upload(stream, 0, &remote_mark_path, None)
                 .await
         },
         |_e| false,
@@ -99,9 +98,9 @@ async fn create_remote_delete_mark(
 
 async fn create_local_delete_mark(
     conf: &PageServerConf,
-    tenant_id: &TenantId,
+    tenant_shard_id: &TenantShardId,
 ) -> Result<(), DeleteTenantError> {
-    let marker_path = conf.tenant_deleted_mark_file_path(tenant_id);
+    let marker_path = conf.tenant_deleted_mark_file_path(tenant_shard_id);
 
     // Note: we're ok to replace existing file.
     let _ = std::fs::OpenOptions::new()
@@ -170,10 +169,10 @@ async fn ensure_timelines_dir_empty(timelines_path: &Utf8Path) -> Result<(), Del
 async fn remove_tenant_remote_delete_mark(
     conf: &PageServerConf,
     remote_storage: Option<&GenericRemoteStorage>,
-    tenant_id: &TenantId,
+    tenant_shard_id: &TenantShardId,
 ) -> Result<(), DeleteTenantError> {
     if let Some(remote_storage) = remote_storage {
-        let path = remote_tenant_delete_mark_path(conf, tenant_id)?;
+        let path = remote_tenant_delete_mark_path(conf, tenant_shard_id)?;
         backoff::retry(
             || async { remote_storage.delete(&path).await },
             |_e| false,
@@ -192,7 +191,7 @@ async fn remove_tenant_remote_delete_mark(
 // Cleanup fs traces: tenant config, timelines dir local delete mark, tenant dir
 async fn cleanup_remaining_fs_traces(
     conf: &PageServerConf,
-    tenant_id: &TenantId,
+    tenant_shard_id: &TenantShardId,
 ) -> Result<(), DeleteTenantError> {
     let rm = |p: Utf8PathBuf, is_dir: bool| async move {
         if is_dir {
@@ -204,8 +203,8 @@ async fn cleanup_remaining_fs_traces(
         .with_context(|| format!("failed to delete {p}"))
     };
 
-    rm(conf.tenant_config_path(tenant_id), false).await?;
-    rm(conf.tenant_location_config_path(tenant_id), false).await?;
+    rm(conf.tenant_config_path(tenant_shard_id), false).await?;
+    rm(conf.tenant_location_config_path(tenant_shard_id), false).await?;
 
     fail::fail_point!("tenant-delete-before-remove-timelines-dir", |_| {
         Err(anyhow::anyhow!(
@@ -213,7 +212,7 @@ async fn cleanup_remaining_fs_traces(
         ))?
     });
 
-    rm(conf.timelines_path(tenant_id), true).await?;
+    rm(conf.timelines_path(tenant_shard_id), true).await?;
 
     fail::fail_point!("tenant-delete-before-remove-deleted-mark", |_| {
         Err(anyhow::anyhow!(
@@ -227,14 +226,14 @@ async fn cleanup_remaining_fs_traces(
     // to be reordered later and thus missed if a crash occurs.
     // Note that we dont need to sync after mark file is removed
     // because we can tolerate the case when mark file reappears on startup.
-    let tenant_path = &conf.tenant_path(tenant_id);
+    let tenant_path = &conf.tenant_path(tenant_shard_id);
     if tenant_path.exists() {
-        crashsafe::fsync_async(&conf.tenant_path(tenant_id))
+        crashsafe::fsync_async(&conf.tenant_path(tenant_shard_id))
             .await
             .context("fsync_pre_mark_remove")?;
     }
 
-    rm(conf.tenant_deleted_mark_file_path(tenant_id), false).await?;
+    rm(conf.tenant_deleted_mark_file_path(tenant_shard_id), false).await?;
 
     fail::fail_point!("tenant-delete-before-remove-tenant-dir", |_| {
         Err(anyhow::anyhow!(
@@ -242,7 +241,7 @@ async fn cleanup_remaining_fs_traces(
         ))?
     });
 
-    rm(conf.tenant_path(tenant_id), true).await?;
+    rm(conf.tenant_path(tenant_shard_id), true).await?;
 
     Ok(())
 }
@@ -287,6 +286,8 @@ impl DeleteTenantFlow {
     ) -> Result<(), DeleteTenantError> {
         span::debug_assert_current_span_has_tenant_id();
 
+        pausable_failpoint!("tenant-delete-before-run");
+
         let mut guard = Self::prepare(&tenant).await?;
 
         if let Err(e) = Self::run_inner(&mut guard, conf, remote_storage.as_ref(), &tenant).await {
@@ -321,7 +322,7 @@ impl DeleteTenantFlow {
         // Though sounds scary, different mark name?
         // Detach currently uses remove_dir_all so in case of a crash we can end up in a weird state.
         if let Some(remote_storage) = &remote_storage {
-            create_remote_delete_mark(conf, remote_storage, &tenant.tenant_id)
+            create_remote_delete_mark(conf, remote_storage, &tenant.tenant_shard_id)
                 .await
                 .context("remote_mark")?
         }
@@ -332,7 +333,7 @@ impl DeleteTenantFlow {
             ))?
         });
 
-        create_local_delete_mark(conf, &tenant.tenant_id)
+        create_local_delete_mark(conf, &tenant.tenant_shard_id)
             .await
             .context("local delete mark")?;
 
@@ -361,17 +362,27 @@ impl DeleteTenantFlow {
         conf: &'static PageServerConf,
         remote_mark_exists: bool,
         tenant: &Tenant,
-    ) -> Option<DeletionGuard> {
-        let tenant_id = tenant.tenant_id;
-
-        if remote_mark_exists || conf.tenant_deleted_mark_file_path(&tenant_id).exists() {
+    ) -> Result<Option<DeletionGuard>, DeleteTenantError> {
+        let acquire = |t: &Tenant| {
             Some(
-                Arc::clone(&tenant.delete_progress)
+                Arc::clone(&t.delete_progress)
                     .try_lock_owned()
                     .expect("we're the only owner during init"),
             )
+        };
+
+        if remote_mark_exists {
+            return Ok(acquire(tenant));
+        }
+
+        // Check local mark first, if its there there is no need to go to s3 to check whether remote one exists.
+        if conf
+            .tenant_deleted_mark_file_path(&tenant.tenant_shard_id)
+            .exists()
+        {
+            Ok(acquire(tenant))
         } else {
-            None
+            Ok(None)
         }
     }
 
@@ -380,7 +391,6 @@ impl DeleteTenantFlow {
         tenant: &Arc<Tenant>,
         preload: Option<TenantPreload>,
         tenants: &'static std::sync::RwLock<TenantsMap>,
-        init_order: Option<InitializationOrder>,
         ctx: &RequestContext,
     ) -> Result<(), DeleteTenantError> {
         let (_, progress) = completion::channel();
@@ -390,10 +400,7 @@ impl DeleteTenantFlow {
             .await
             .expect("cant be stopping or broken");
 
-        tenant
-            .attach(init_order, preload, ctx)
-            .await
-            .context("attach")?;
+        tenant.attach(preload, ctx).await.context("attach")?;
 
         Self::background(
             guard,
@@ -451,12 +458,12 @@ impl DeleteTenantFlow {
         tenants: &'static std::sync::RwLock<TenantsMap>,
         tenant: Arc<Tenant>,
     ) {
-        let tenant_id = tenant.tenant_id;
+        let tenant_shard_id = tenant.tenant_shard_id;
 
         task_mgr::spawn(
             task_mgr::BACKGROUND_RUNTIME.handle(),
             TaskKind::TimelineDeletionWorker,
-            Some(tenant_id),
+            Some(tenant_shard_id.tenant_id),
             None,
             "tenant_delete",
             false,
@@ -470,7 +477,7 @@ impl DeleteTenantFlow {
                 Ok(())
             }
             .instrument({
-                let span = tracing::info_span!(parent: None, "delete_tenant", tenant_id=%tenant_id);
+                let span = tracing::info_span!(parent: None, "delete_tenant", tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug());
                 span.follows_from(Span::current());
                 span
             }),
@@ -508,7 +515,7 @@ impl DeleteTenantFlow {
             }
         }
 
-        let timelines_path = conf.timelines_path(&tenant.tenant_id);
+        let timelines_path = conf.timelines_path(&tenant.tenant_shard_id);
         // May not exist if we fail in cleanup_remaining_fs_traces after removing it
         if timelines_path.exists() {
             // sanity check to guard against layout changes
@@ -517,7 +524,8 @@ impl DeleteTenantFlow {
                 .context("timelines dir not empty")?;
         }
 
-        remove_tenant_remote_delete_mark(conf, remote_storage.as_ref(), &tenant.tenant_id).await?;
+        remove_tenant_remote_delete_mark(conf, remote_storage.as_ref(), &tenant.tenant_shard_id)
+            .await?;
 
         fail::fail_point!("tenant-delete-before-cleanup-remaining-fs-traces", |_| {
             Err(anyhow::anyhow!(
@@ -525,21 +533,73 @@ impl DeleteTenantFlow {
             ))?
         });
 
-        cleanup_remaining_fs_traces(conf, &tenant.tenant_id)
+        cleanup_remaining_fs_traces(conf, &tenant.tenant_shard_id)
             .await
             .context("cleanup_remaining_fs_traces")?;
 
         {
-            let mut locked = tenants.write().unwrap();
-            if locked.remove(&tenant.tenant_id).is_none() {
-                warn!("Tenant got removed from tenants map during deletion");
-            };
+            pausable_failpoint!("tenant-delete-before-map-remove");
 
-            // FIXME: we should not be modifying this from outside of mgr.rs.
-            // This will go away when we simplify deletion (https://github.com/neondatabase/neon/issues/5080)
-            crate::metrics::TENANT_MANAGER
-                .tenant_slots
-                .set(locked.len() as u64);
+            // This block is simply removing the TenantSlot for this tenant.  It requires a loop because
+            // we might conflict with a TenantSlot::InProgress marker and need to wait for it.
+            //
+            // This complexity will go away when we simplify how deletion works:
+            // https://github.com/neondatabase/neon/issues/5080
+            loop {
+                // Under the TenantMap lock, try to remove the tenant.  We usually succeed, but if
+                // we encounter an InProgress marker, yield the barrier it contains and wait on it.
+                let barrier = {
+                    let mut locked = tenants.write().unwrap();
+                    let removed = locked.remove(&tenant.tenant_shard_id.tenant_id);
+
+                    // FIXME: we should not be modifying this from outside of mgr.rs.
+                    // This will go away when we simplify deletion (https://github.com/neondatabase/neon/issues/5080)
+                    crate::metrics::TENANT_MANAGER
+                        .tenant_slots
+                        .set(locked.len() as u64);
+
+                    match removed {
+                        TenantsMapRemoveResult::Occupied(TenantSlot::Attached(tenant)) => {
+                            match tenant.current_state() {
+                                TenantState::Stopping { .. } | TenantState::Broken { .. } => {
+                                    // Expected: we put the tenant into stopping state before we start deleting it
+                                }
+                                state => {
+                                    // Unexpected state
+                                    tracing::warn!(
+                                        "Tenant in unexpected state {state} after deletion"
+                                    );
+                                }
+                            }
+                            break;
+                        }
+                        TenantsMapRemoveResult::Occupied(TenantSlot::Secondary) => {
+                            // This is unexpected: this secondary tenants should not have been created, and we
+                            // are not in a position to shut it down from here.
+                            tracing::warn!("Tenant transitioned to secondary mode while deleting!");
+                            break;
+                        }
+                        TenantsMapRemoveResult::Occupied(TenantSlot::InProgress(_)) => {
+                            unreachable!("TenantsMap::remove handles InProgress separately, should never return it here");
+                        }
+                        TenantsMapRemoveResult::Vacant => {
+                            tracing::warn!(
+                                "Tenant removed from TenantsMap before deletion completed"
+                            );
+                            break;
+                        }
+                        TenantsMapRemoveResult::InProgress(barrier) => {
+                            // An InProgress entry was found, we must wait on its barrier
+                            barrier
+                        }
+                    }
+                };
+
+                tracing::info!(
+                    "Waiting for competing operation to complete before deleting state for tenant"
+                );
+                barrier.wait().await;
+            }
         }
 
         *guard = Self::Finished;

pageserver/src/tenant/ephemeral_file.rs

@@ -7,18 +7,19 @@ use crate::page_cache::{self, PAGE_SZ};
 use crate::tenant::block_io::{BlockCursor, BlockLease, BlockReader};
 use crate::virtual_file::VirtualFile;
 use camino::Utf8PathBuf;
+use pageserver_api::shard::TenantShardId;
 use std::cmp::min;
 use std::fs::OpenOptions;
 use std::io::{self, ErrorKind};
 use std::ops::DerefMut;
 use std::sync::atomic::AtomicU64;
 use tracing::*;
-use utils::id::{TenantId, TimelineId};
+use utils::id::TimelineId;
 
 pub struct EphemeralFile {
     page_cache_file_id: page_cache::FileId,
 
-    _tenant_id: TenantId,
+    _tenant_shard_id: TenantShardId,
     _timeline_id: TimelineId,
     file: VirtualFile,
     len: u64,
@@ -31,7 +32,7 @@ pub struct EphemeralFile {
 impl EphemeralFile {
     pub async fn create(
         conf: &PageServerConf,
-        tenant_id: TenantId,
+        tenant_shard_id: TenantShardId,
         timeline_id: TimelineId,
     ) -> Result<EphemeralFile, io::Error> {
         static NEXT_FILENAME: AtomicU64 = AtomicU64::new(1);
@@ -39,7 +40,7 @@ impl EphemeralFile {
             NEXT_FILENAME.fetch_add(1, std::sync::atomic::Ordering::Relaxed);
 
         let filename = conf
-            .timeline_path(&tenant_id, &timeline_id)
+            .timeline_path(&tenant_shard_id, &timeline_id)
             .join(Utf8PathBuf::from(format!(
                 "ephemeral-{filename_disambiguator}"
             )));
@@ -52,7 +53,7 @@ impl EphemeralFile {
 
         Ok(EphemeralFile {
             page_cache_file_id: page_cache::next_file_id(),
-            _tenant_id: tenant_id,
+            _tenant_shard_id: tenant_shard_id,
             _timeline_id: timeline_id,
             file,
             len: 0,
@@ -282,7 +283,7 @@ mod tests {
     ) -> Result<
         (
             &'static PageServerConf,
-            TenantId,
+            TenantShardId,
             TimelineId,
             RequestContext,
         ),
@@ -295,13 +296,13 @@ mod tests {
         // OK in a test.
         let conf: &'static PageServerConf = Box::leak(Box::new(conf));
 
-        let tenant_id = TenantId::from_str("11000000000000000000000000000000").unwrap();
+        let tenant_shard_id = TenantShardId::from_str("11000000000000000000000000000000").unwrap();
         let timeline_id = TimelineId::from_str("22000000000000000000000000000000").unwrap();
-        fs::create_dir_all(conf.timeline_path(&tenant_id, &timeline_id))?;
+        fs::create_dir_all(conf.timeline_path(&tenant_shard_id, &timeline_id))?;
 
         let ctx = RequestContext::new(TaskKind::UnitTest, DownloadBehavior::Error);
 
-        Ok((conf, tenant_id, timeline_id, ctx))
+        Ok((conf, tenant_shard_id, timeline_id, ctx))
     }
 
     #[tokio::test]

pageserver/src/tenant/metadata.rs

@@ -11,15 +11,12 @@
 use std::io::{self};
 
 use anyhow::{ensure, Context};
+use pageserver_api::shard::TenantShardId;
 use serde::{de::Error, Deserialize, Serialize, Serializer};
 use thiserror::Error;
 use utils::bin_ser::SerializeError;
 use utils::crashsafe::path_with_suffix_extension;
-use utils::{
-    bin_ser::BeSer,
-    id::{TenantId, TimelineId},
-    lsn::Lsn,
-};
+use utils::{bin_ser::BeSer, id::TimelineId, lsn::Lsn};
 
 use crate::config::PageServerConf;
 use crate::virtual_file::VirtualFile;
@@ -272,14 +269,14 @@ impl Serialize for TimelineMetadata {
 }
 
 /// Save timeline metadata to file
-#[tracing::instrument(skip_all, fields(%tenant_id, %timeline_id))]
+#[tracing::instrument(skip_all, fields(%tenant_id=tenant_shard_id.tenant_id, %shard_id=tenant_shard_id.shard_slug(), %timeline_id))]
 pub async fn save_metadata(
     conf: &'static PageServerConf,
-    tenant_id: &TenantId,
+    tenant_shard_id: &TenantShardId,
     timeline_id: &TimelineId,
     data: &TimelineMetadata,
 ) -> anyhow::Result<()> {
-    let path = conf.metadata_path(tenant_id, timeline_id);
+    let path = conf.metadata_path(tenant_shard_id, timeline_id);
     let temp_path = path_with_suffix_extension(&path, TEMP_FILE_SUFFIX);
     let metadata_bytes = data.to_bytes().context("serialize metadata")?;
     VirtualFile::crashsafe_overwrite(&path, &temp_path, &metadata_bytes)
@@ -299,10 +296,10 @@ pub enum LoadMetadataError {
 
 pub fn load_metadata(
     conf: &'static PageServerConf,
-    tenant_id: &TenantId,
+    tenant_shard_id: &TenantShardId,
     timeline_id: &TimelineId,
 ) -> Result<TimelineMetadata, LoadMetadataError> {
-    let metadata_path = conf.metadata_path(tenant_id, timeline_id);
+    let metadata_path = conf.metadata_path(tenant_shard_id, timeline_id);
     let metadata_bytes = std::fs::read(metadata_path)?;
 
     Ok(TimelineMetadata::from_bytes(&metadata_bytes)?)

pageserver/src/tenant/mgr.rs

@@ -2,7 +2,8 @@
 //! page server.
 
 use camino::{Utf8DirEntry, Utf8Path, Utf8PathBuf};
-use pageserver_api::shard::TenantShardId;
+use pageserver_api::key::Key;
+use pageserver_api::shard::{ShardIdentity, ShardNumber, TenantShardId};
 use rand::{distributions::Alphanumeric, Rng};
 use std::borrow::Cow;
 use std::collections::{BTreeMap, HashMap};
@@ -29,7 +30,9 @@ use crate::control_plane_client::{
 use crate::deletion_queue::DeletionQueueClient;
 use crate::metrics::TENANT_MANAGER as METRICS;
 use crate::task_mgr::{self, TaskKind};
-use crate::tenant::config::{AttachmentMode, LocationConf, LocationMode, TenantConfOpt};
+use crate::tenant::config::{
+    AttachedLocationConfig, AttachmentMode, LocationConf, LocationMode, TenantConfOpt,
+};
 use crate::tenant::delete::DeleteTenantFlow;
 use crate::tenant::span::debug_assert_current_span_has_tenant_id;
 use crate::tenant::{create_tenant_files, AttachedTenantConf, SpawnMode, Tenant, TenantState};
@@ -122,6 +125,24 @@ fn exactly_one_or_none<'a>(
     }
 }
 
+pub(crate) enum TenantsMapRemoveResult {
+    Occupied(TenantSlot),
+    Vacant,
+    InProgress(utils::completion::Barrier),
+}
+
+/// When resolving a TenantId to a shard, we may be looking for the 0th
+/// shard, or we might be looking for whichever shard holds a particular page.
+pub(crate) enum ShardSelector {
+    /// Only return the 0th shard, if it is present.  If a non-0th shard is present,
+    /// ignore it.
+    Zero,
+    /// Pick the first shard we find for the TenantId
+    First,
+    /// Pick the shard that holds this key
+    Page(Key),
+}
+
 impl TenantsMap {
     /// Convenience function for typical usage, where we want to get a `Tenant` object, for
     /// working with attached tenants.  If the TenantId is in the map but in Secondary state,
@@ -136,12 +157,71 @@ impl TenantsMap {
         }
     }
 
-    pub(crate) fn remove(&mut self, tenant_id: &TenantId) -> Option<TenantSlot> {
+    /// A page service client sends a TenantId, and to look up the correct Tenant we must
+    /// resolve this to a fully qualified TenantShardId.
+    fn resolve_shard(
+        &self,
+        tenant_id: &TenantId,
+        selector: ShardSelector,
+    ) -> Option<TenantShardId> {
+        let mut want_shard = None;
         match self {
             TenantsMap::Initializing => None,
+            TenantsMap::Open(m) | TenantsMap::ShuttingDown(m) => {
+                for slot in m.range(TenantShardId::tenant_range(*tenant_id)) {
+                    match selector {
+                        ShardSelector::First => return Some(*slot.0),
+                        ShardSelector::Zero if slot.0.shard_number == ShardNumber(0) => {
+                            return Some(*slot.0)
+                        }
+                        ShardSelector::Page(key) => {
+                            if let Some(tenant) = slot.1.get_attached() {
+                                // First slot we see for this tenant, calculate the expected shard number
+                                // for the key: we will use this for checking if this and subsequent
+                                // slots contain the key, rather than recalculating the hash each time.
+                                if want_shard.is_none() {
+                                    want_shard = Some(tenant.shard_identity.get_shard_number(&key));
+                                }
+
+                                if Some(tenant.shard_identity.number) == want_shard {
+                                    return Some(*slot.0);
+                                }
+                            } else {
+                                continue;
+                            }
+                        }
+                        _ => continue,
+                    }
+                }
+
+                // Fall through: we didn't find an acceptable shard
+                None
+            }
+        }
+    }
+
+    /// Only for use from DeleteTenantFlow.  This method directly removes a TenantSlot from the map.
+    ///
+    /// The normal way to remove a tenant is using a SlotGuard, which will gracefully remove the guarded
+    /// slot if the enclosed tenant is shutdown.
+    pub(crate) fn remove(&mut self, tenant_id: &TenantId) -> TenantsMapRemoveResult {
+        use std::collections::btree_map::Entry;
+        match self {
+            TenantsMap::Initializing => TenantsMapRemoveResult::Vacant,
             TenantsMap::Open(m) | TenantsMap::ShuttingDown(m) => {
                 let key = exactly_one_or_none(m, tenant_id).map(|(k, _)| *k);
-                key.and_then(|key| m.remove(&key))
+                match key {
+                    Some(key) => match m.entry(key) {
+                        Entry::Occupied(entry) => match entry.get() {
+                            TenantSlot::InProgress(barrier) => {
+                                TenantsMapRemoveResult::InProgress(barrier.clone())
+                            }
+                            _ => TenantsMapRemoveResult::Occupied(entry.remove()),
+                        },
+                        Entry::Vacant(_entry) => TenantsMapRemoveResult::Vacant,
+                    },
+                    None => TenantsMapRemoveResult::Vacant,
+                }
             }
         }
     }
@@ -190,49 +270,6 @@ async fn safe_rename_tenant_dir(path: impl AsRef<Utf8Path>) -> std::io::Result<U
 static TENANTS: Lazy<std::sync::RwLock<TenantsMap>> =
     Lazy::new(|| std::sync::RwLock::new(TenantsMap::Initializing));
 
-/// Create a directory, including parents.  This does no fsyncs and makes
-/// no guarantees about the persistence of the resulting metadata: for
-/// use when creating dirs for use as cache.
-async fn unsafe_create_dir_all(path: &Utf8PathBuf) -> std::io::Result<()> {
-    let mut dirs_to_create = Vec::new();
-    let mut path: &Utf8Path = path.as_ref();
-
-    // Figure out which directories we need to create.
-    loop {
-        let meta = tokio::fs::metadata(path).await;
-        match meta {
-            Ok(metadata) if metadata.is_dir() => break,
-            Ok(_) => {
-                return Err(std::io::Error::new(
-                    std::io::ErrorKind::AlreadyExists,
-                    format!("non-directory found in path: {path}"),
-                ));
-            }
-            Err(ref e) if e.kind() == std::io::ErrorKind::NotFound => {}
-            Err(e) => return Err(e),
-        }
-
-        dirs_to_create.push(path);
-
-        match path.parent() {
-            Some(parent) => path = parent,
-            None => {
-                return Err(std::io::Error::new(
-                    std::io::ErrorKind::InvalidInput,
-                    format!("can't find parent of path '{path}'"),
-                ));
-            }
-        }
-    }
-
-    // Create directories from parent to child.
-    for &path in dirs_to_create.iter().rev() {
-        tokio::fs::create_dir(path).await?;
-    }
-
-    Ok(())
-}
-
 /// The TenantManager is responsible for storing and mutating the collection of all tenants
 /// that this pageserver process has state for.  Every Tenant and SecondaryTenant instance
 /// lives inside the TenantManager.
@@ -250,8 +287,8 @@ pub struct TenantManager {
 }
 
 fn emergency_generations(
-    tenant_confs: &HashMap<TenantId, anyhow::Result<LocationConf>>,
-) -> HashMap<TenantId, Generation> {
+    tenant_confs: &HashMap<TenantShardId, anyhow::Result<LocationConf>>,
+) -> HashMap<TenantShardId, Generation> {
     tenant_confs
         .iter()
         .filter_map(|(tid, lc)| {
@@ -271,10 +308,10 @@ fn emergency_generations(
 
 async fn init_load_generations(
     conf: &'static PageServerConf,
-    tenant_confs: &HashMap<TenantId, anyhow::Result<LocationConf>>,
+    tenant_confs: &HashMap<TenantShardId, anyhow::Result<LocationConf>>,
     resources: &TenantSharedResources,
     cancel: &CancellationToken,
-) -> anyhow::Result<Option<HashMap<TenantId, Generation>>> {
+) -> anyhow::Result<Option<HashMap<TenantShardId, Generation>>> {
     let generations = if conf.control_plane_emergency_mode {
         error!(
             "Emergency mode!  Tenants will be attached unsafely using their last known generation"
@@ -317,7 +354,7 @@ async fn init_load_generations(
 fn load_tenant_config(
     conf: &'static PageServerConf,
     dentry: Utf8DirEntry,
-) -> anyhow::Result<Option<(TenantId, anyhow::Result<LocationConf>)>> {
+) -> anyhow::Result<Option<(TenantShardId, anyhow::Result<LocationConf>)>> {
     let tenant_dir_path = dentry.path().to_path_buf();
     if crate::is_temporary(&tenant_dir_path) {
         info!("Found temporary tenant directory, removing: {tenant_dir_path}");
@@ -353,10 +390,10 @@ fn load_tenant_config(
         return Ok(None);
     }
 
-    let tenant_id = match tenant_dir_path
+    let tenant_shard_id = match tenant_dir_path
         .file_name()
         .unwrap_or_default()
-        .parse::<TenantId>()
+        .parse::<TenantShardId>()
     {
         Ok(id) => id,
         Err(_) => {
@@ -366,8 +403,8 @@ fn load_tenant_config(
     };
 
     Ok(Some((
-        tenant_id,
-        Tenant::load_tenant_config(conf, &tenant_id),
+        tenant_shard_id,
+        Tenant::load_tenant_config(conf, &tenant_shard_id),
     )))
 }
 
@@ -378,7 +415,7 @@ fn load_tenant_config(
 /// seconds even on reasonably fast drives.
 async fn init_load_tenant_configs(
     conf: &'static PageServerConf,
-) -> anyhow::Result<HashMap<TenantId, anyhow::Result<LocationConf>>> {
+) -> anyhow::Result<HashMap<TenantShardId, anyhow::Result<LocationConf>>> {
     let tenants_dir = conf.tenants_path();
 
     let dentries = tokio::task::spawn_blocking(move || -> anyhow::Result<Vec<Utf8DirEntry>> {
@@ -428,19 +465,19 @@ pub async fn init_tenant_mgr(
         init_load_generations(conf, &tenant_configs, &resources, &cancel).await?;
 
     // Construct `Tenant` objects and start them running
-    for (tenant_id, location_conf) in tenant_configs {
-        let tenant_dir_path = conf.tenant_path(&tenant_id);
+    for (tenant_shard_id, location_conf) in tenant_configs {
+        let tenant_dir_path = conf.tenant_path(&tenant_shard_id);
 
         let mut location_conf = match location_conf {
             Ok(l) => l,
             Err(e) => {
-                warn!(%tenant_id, "Marking tenant broken, failed to {e:#}");
+                warn!(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), "Marking tenant broken, failed to {e:#}");
 
                 tenants.insert(
-                    TenantShardId::unsharded(tenant_id),
+                    tenant_shard_id,
                     TenantSlot::Attached(Tenant::create_broken_tenant(
                         conf,
-                        tenant_id,
+                        tenant_shard_id,
                         format!("{}", e),
                     )),
                 );
@@ -451,7 +488,7 @@ pub async fn init_tenant_mgr(
         let generation = if let Some(generations) = &tenant_generations {
             // We have a generation map: treat it as the authority for whether
             // this tenant is really attached.
-            if let Some(gen) = generations.get(&tenant_id) {
+            if let Some(gen) = generations.get(&tenant_shard_id) {
                 *gen
             } else {
                 match &location_conf.mode {
@@ -459,8 +496,8 @@ pub async fn init_tenant_mgr(
                         // We do not require the control plane's permission for secondary mode
                         // tenants, because they do no remote writes and hence require no
                         // generation number
-                        info!(%tenant_id, "Loaded tenant in secondary mode");
-                        tenants.insert(TenantShardId::unsharded(tenant_id), TenantSlot::Secondary);
+                        info!(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), "Loaded tenant in secondary mode");
+                        tenants.insert(tenant_shard_id, TenantSlot::Secondary);
                     }
                     LocationMode::Attached(_) => {
                         // TODO: augment re-attach API to enable the control plane to
@@ -468,9 +505,9 @@ pub async fn init_tenant_mgr(
                         // away local state, we can gracefully fall back to secondary here, if the control
                         // plane tells us so.
                         // (https://github.com/neondatabase/neon/issues/5377)
-                        info!(%tenant_id, "Detaching tenant, control plane omitted it in re-attach response");
+                        info!(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), "Detaching tenant, control plane omitted it in re-attach response");
                         if let Err(e) = safe_remove_tenant_dir_all(&tenant_dir_path).await {
-                            error!(%tenant_id,
+                            error!(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(),
                                 "Failed to remove detached tenant directory '{tenant_dir_path}': {e:?}",
                             );
                         }
@@ -482,21 +519,23 @@ pub async fn init_tenant_mgr(
         } else {
             // Legacy mode: no generation information, any tenant present
             // on local disk may activate
-            info!(%tenant_id, "Starting tenant in legacy mode, no generation",);
+            info!(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), "Starting tenant in legacy mode, no generation",);
             Generation::none()
         };
 
         // Presence of a generation number implies attachment: attach the tenant
         // if it wasn't already, and apply the generation number.
         location_conf.attach_in_generation(generation);
-        Tenant::persist_tenant_config(conf, &tenant_id, &location_conf).await?;
+        Tenant::persist_tenant_config(conf, &tenant_shard_id, &location_conf).await?;
 
+        let shard_identity = location_conf.shard;
         match tenant_spawn(
             conf,
-            tenant_id,
+            tenant_shard_id,
             &tenant_dir_path,
             resources.clone(),
             AttachedTenantConf::try_from(location_conf)?,
+            shard_identity,
             Some(init_order.clone()),
             &TENANTS,
             SpawnMode::Normal,
@@ -509,7 +548,7 @@ pub async fn init_tenant_mgr(
                 );
             }
             Err(e) => {
-                error!(%tenant_id, "Failed to start tenant: {e:#}");
+                error!(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), "Failed to start tenant: {e:#}");
             }
         }
     }
@@ -533,10 +572,11 @@ pub async fn init_tenant_mgr(
 #[allow(clippy::too_many_arguments)]
 pub(crate) fn tenant_spawn(
     conf: &'static PageServerConf,
-    tenant_id: TenantId,
+    tenant_shard_id: TenantShardId,
     tenant_path: &Utf8Path,
     resources: TenantSharedResources,
     location_conf: AttachedTenantConf,
+    shard_identity: ShardIdentity,
     init_order: Option<InitializationOrder>,
     tenants: &'static std::sync::RwLock<TenantsMap>,
     mode: SpawnMode,
@@ -557,18 +597,25 @@ pub(crate) fn tenant_spawn(
         "Cannot load tenant from empty directory {tenant_path:?}"
     );
 
-    let tenant_ignore_mark = conf.tenant_ignore_mark_file_path(&tenant_id);
+    let tenant_ignore_mark = conf.tenant_ignore_mark_file_path(&tenant_shard_id);
     anyhow::ensure!(
-        !conf.tenant_ignore_mark_file_path(&tenant_id).exists(),
+        !conf.tenant_ignore_mark_file_path(&tenant_shard_id).exists(),
         "Cannot load tenant, ignore mark found at {tenant_ignore_mark:?}"
     );
 
-    info!("Attaching tenant {tenant_id}");
+    info!(
+        tenant_id = %tenant_shard_id.tenant_id,
+        shard_id = %tenant_shard_id.shard_slug(),
+        generation = ?location_conf.location.generation,
+        attach_mode = ?location_conf.location.attach_mode,
+        "Attaching tenant"
+    );
     let tenant = match Tenant::spawn(
         conf,
-        tenant_id,
+        tenant_shard_id,
         resources,
         location_conf,
+        shard_identity,
         init_order,
         tenants,
         mode,
@@ -576,8 +623,8 @@ pub(crate) fn tenant_spawn(
     ) {
         Ok(tenant) => tenant,
         Err(e) => {
-            error!("Failed to spawn tenant {tenant_id}, reason: {e:#}");
-            Tenant::create_broken_tenant(conf, tenant_id, format!("{e:#}"))
+            error!("Failed to spawn tenant {tenant_shard_id}, reason: {e:#}");
+            Tenant::create_broken_tenant(conf, tenant_shard_id, format!("{e:#}"))
         }
     };
 
@@ -732,19 +779,20 @@ pub(crate) async fn create_tenant(
     ctx: &RequestContext,
 ) -> Result<Arc<Tenant>, TenantMapInsertError> {
     let location_conf = LocationConf::attached_single(tenant_conf, generation);
+    info!("Creating tenant at location {location_conf:?}");
 
     let slot_guard =
         tenant_map_acquire_slot(&tenant_shard_id, TenantSlotAcquireMode::MustNotExist)?;
-    // TODO(sharding): make local paths shard-aware
-    let tenant_path =
-        super::create_tenant_files(conf, &location_conf, &tenant_shard_id.tenant_id).await?;
+    let tenant_path = super::create_tenant_files(conf, &location_conf, &tenant_shard_id).await?;
 
+    let shard_identity = location_conf.shard;
     let created_tenant = tenant_spawn(
         conf,
-        tenant_shard_id.tenant_id,
+        tenant_shard_id,
         &tenant_path,
         resources,
         AttachedTenantConf::try_from(location_conf)?,
+        shard_identity,
         None,
         &TENANTS,
         SpawnMode::Create,
@@ -781,8 +829,9 @@ pub(crate) async fn set_new_tenant_config(
     // API to use is the location_config/ endpoint, which lets the caller provide
     // the full LocationConf.
     let location_conf = LocationConf::attached_single(new_tenant_conf, tenant.generation);
+    let tenant_shard_id = TenantShardId::unsharded(tenant_id);
 
-    Tenant::persist_tenant_config(conf, &tenant_id, &location_conf)
+    Tenant::persist_tenant_config(conf, &tenant_shard_id, &location_conf)
         .await
         .map_err(SetNewTenantConfigError::Persist)?;
     tenant.set_new_tenant_config(new_tenant_conf);
@@ -792,8 +841,6 @@ pub(crate) async fn set_new_tenant_config(
 impl TenantManager {
     /// Gets the attached tenant from the in-memory data, erroring if it's absent, in secondary mode, or is not fitting to the query.
     /// `active_only = true` allows to query only tenants that are ready for operations, erroring on other kinds of tenants.
-    ///
-    /// This method is cancel-safe.
     pub(crate) fn get_attached_tenant_shard(
         &self,
         tenant_shard_id: TenantShardId,
@@ -838,10 +885,12 @@ impl TenantManager {
         Ok(())
     }
 
+    #[instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug()))]
     pub(crate) async fn upsert_location(
         &self,
         tenant_shard_id: TenantShardId,
         new_location_config: LocationConf,
+        flush: Option<Duration>,
         ctx: &RequestContext,
     ) -> Result<(), anyhow::Error> {
         debug_assert_current_span_has_tenant_id();
@@ -850,7 +899,7 @@ impl TenantManager {
         // Special case fast-path for updates to Tenant: if our upsert is only updating configuration,
         // then we do not need to set the slot to InProgress, we can just call into the
         // existng tenant.
-        {
+        let modify_tenant = {
             let locked = self.tenants.read().unwrap();
             let peek_slot =
                 tenant_map_peek_slot(&locked, &tenant_shard_id, TenantSlotPeekMode::Write)?;
@@ -861,22 +910,50 @@ impl TenantManager {
                         // take our fast path and just provide the updated configuration
                         // to the tenant.
                         tenant.set_new_location_config(AttachedTenantConf::try_from(
-                            new_location_config,
+                            new_location_config.clone(),
                         )?);
 
-                        // Persist the new config in the background, to avoid holding up any
-                        // locks while we do so.
-                        // TODO
-
-                        return Ok(());
+                        Some(tenant.clone())
                     } else {
                         // Different generations, fall through to general case
+                        None
                     }
                 }
                 _ => {
                     // Not an Attached->Attached transition, fall through to general case
+                    None
                 }
             }
+        };
+
+        // Fast-path continued: having dropped out of the self.tenants lock, do the async
+        // phase of waiting for flush, before returning.
+        if let Some(tenant) = modify_tenant {
+            // Transition to AttachedStale means we may well hold a valid generation
+            // still, and have been requested to go stale as part of a migration.  If
+            // the caller set `flush`, then flush to remote storage.
+            if let LocationMode::Attached(AttachedLocationConfig {
+                generation: _,
+                attach_mode: AttachmentMode::Stale,
+            }) = &new_location_config.mode
+            {
+                if let Some(flush_timeout) = flush {
+                    match tokio::time::timeout(flush_timeout, tenant.flush_remote()).await {
+                        Ok(Err(e)) => {
+                            return Err(e);
+                        }
+                        Ok(Ok(_)) => return Ok(()),
+                        Err(_) => {
+                            tracing::warn!(
+                                timeout_ms = flush_timeout.as_millis(),
+                                "Timed out waiting for flush to remote storage, proceeding anyway."
+                            )
+                        }
+                    }
+                }
+            }
+
+            return Ok(());
         }
 
         // General case for upserts to TenantsMap, excluding the case above: we will substitute an
@@ -915,55 +992,44 @@ impl TenantManager {
             slot_guard.drop_old_value().expect("We just shut it down");
         }
 
-        // TODO(sharding): make local paths sharding-aware
-        let tenant_path = self.conf.tenant_path(&tenant_shard_id.tenant_id);
+        let tenant_path = self.conf.tenant_path(&tenant_shard_id);
 
         let new_slot = match &new_location_config.mode {
             LocationMode::Secondary(_) => {
                 // Directory doesn't need to be fsync'd because if we crash it can
                 // safely be recreated next time this tenant location is configured.
-                unsafe_create_dir_all(&tenant_path)
+                tokio::fs::create_dir_all(&tenant_path)
                     .await
                     .with_context(|| format!("Creating {tenant_path}"))?;
 
-                // TODO(sharding): make local paths sharding-aware
-                Tenant::persist_tenant_config(
-                    self.conf,
-                    &tenant_shard_id.tenant_id,
-                    &new_location_config,
-                )
-                .await
-                .map_err(SetNewTenantConfigError::Persist)?;
+                Tenant::persist_tenant_config(self.conf, &tenant_shard_id, &new_location_config)
+                    .await
+                    .map_err(SetNewTenantConfigError::Persist)?;
 
                 TenantSlot::Secondary
             }
             LocationMode::Attached(_attach_config) => {
-                // TODO(sharding): make local paths sharding-aware
-                let timelines_path = self.conf.timelines_path(&tenant_shard_id.tenant_id);
+                let timelines_path = self.conf.timelines_path(&tenant_shard_id);
 
                 // Directory doesn't need to be fsync'd because we do not depend on
                 // it to exist after crashes: it may be recreated when tenant is
                 // re-attached, see https://github.com/neondatabase/neon/issues/5550
-                unsafe_create_dir_all(&timelines_path)
+                tokio::fs::create_dir_all(&tenant_path)
                     .await
                     .with_context(|| format!("Creating {timelines_path}"))?;
 
-                // TODO(sharding): make local paths sharding-aware
-                Tenant::persist_tenant_config(
-                    self.conf,
-                    &tenant_shard_id.tenant_id,
-                    &new_location_config,
-                )
-                .await
-                .map_err(SetNewTenantConfigError::Persist)?;
+                Tenant::persist_tenant_config(self.conf, &tenant_shard_id, &new_location_config)
+                    .await
+                    .map_err(SetNewTenantConfigError::Persist)?;
 
-                // TODO(sharding): make spawn sharding-aware
+                let shard_identity = new_location_config.shard;
                 let tenant = tenant_spawn(
                     self.conf,
-                    tenant_shard_id.tenant_id,
+                    tenant_shard_id,
                     &tenant_path,
                     self.resources.clone(),
                     AttachedTenantConf::try_from(new_location_config)?,
+                    shard_identity,
                     None,
                     self.tenants,
                     SpawnMode::Normal,
@@ -978,6 +1044,81 @@ impl TenantManager {
 
         Ok(())
     }
+
+    /// Resetting a tenant is equivalent to detaching it, then attaching it again with the same
+    /// LocationConf that was last used to attach it.  Optionally, the local file cache may be
+    /// dropped before re-attaching.
+    ///
+    /// This is not part of a tenant's normal lifecycle: it is used for debug/support, in situations
+    /// where an issue is identified that would go away with a restart of the tenant.
+    ///
+    /// This does not have any special "force" shutdown of a tenant: it relies on the tenant's tasks
+    /// to respect the cancellation tokens used in normal shutdown().
+    #[instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), %drop_cache))]
+    pub(crate) async fn reset_tenant(
+        &self,
+        tenant_shard_id: TenantShardId,
+        drop_cache: bool,
+        ctx: RequestContext,
+    ) -> anyhow::Result<()> {
+        let mut slot_guard = tenant_map_acquire_slot(&tenant_shard_id, TenantSlotAcquireMode::Any)?;
+        let Some(old_slot) = slot_guard.get_old_value() else {
+            anyhow::bail!("Tenant not found when trying to reset");
+        };
+
+        let Some(tenant) = old_slot.get_attached() else {
+            slot_guard.revert();
+            anyhow::bail!("Tenant is not in attached state");
+        };
+
+        let (_guard, progress) = utils::completion::channel();
+        match tenant.shutdown(progress, false).await {
+            Ok(()) => {
+                slot_guard.drop_old_value()?;
+            }
+            Err(_barrier) => {
+                slot_guard.revert();
+                anyhow::bail!("Cannot reset Tenant, already shutting down");
+            }
+        }
+
+        let tenant_path = self.conf.tenant_path(&tenant_shard_id);
+        let timelines_path = self.conf.timelines_path(&tenant_shard_id);
+        let config = Tenant::load_tenant_config(self.conf, &tenant_shard_id)?;
+
+        if drop_cache {
+            tracing::info!("Dropping local file cache");
+
+            match tokio::fs::read_dir(&timelines_path).await {
+                Err(e) => {
+                    tracing::warn!("Failed to list timelines while dropping cache: {}", e);
+                }
+                Ok(mut entries) => {
+                    while let Some(entry) = entries.next_entry().await? {
+                        tokio::fs::remove_dir_all(entry.path()).await?;
+                    }
+                }
+            }
+        }
+
+        let shard_identity = config.shard;
+        let tenant = tenant_spawn(
+            self.conf,
+            tenant_shard_id,
+            &tenant_path,
+            self.resources.clone(),
+            AttachedTenantConf::try_from(config)?,
+            shard_identity,
+            None,
+            self.tenants,
+            SpawnMode::Normal,
+            &ctx,
+        )?;
+
+        slot_guard.upsert(TenantSlot::Attached(tenant))?;
+
+        Ok(())
+    }
 }
 
 #[derive(Debug, thiserror::Error)]
@@ -1062,6 +1203,7 @@ pub(crate) enum GetActiveTenantError {
 /// then wait for up to `timeout` (minus however long we waited for the slot).
 pub(crate) async fn get_active_tenant_with_timeout(
     tenant_id: TenantId,
+    shard_selector: ShardSelector,
     timeout: Duration,
     cancel: &CancellationToken,
 ) -> Result<Arc<Tenant>, GetActiveTenantError> {
@@ -1070,15 +1212,17 @@ pub(crate) async fn get_active_tenant_with_timeout(
         Tenant(Arc<Tenant>),
     }
 
-    // TODO(sharding): make page service interface sharding-aware (page service should apply ShardIdentity to the key
-    // to decide which shard services the request)
-    let tenant_shard_id = TenantShardId::unsharded(tenant_id);
-
     let wait_start = Instant::now();
     let deadline = wait_start + timeout;
 
-    let wait_for = {
+    let (wait_for, tenant_shard_id) = {
         let locked = TENANTS.read().unwrap();
+
+        // Resolve TenantId to TenantShardId
+        let tenant_shard_id = locked.resolve_shard(&tenant_id, shard_selector).ok_or(
+            GetActiveTenantError::NotFound(GetTenantError::NotFound(tenant_id)),
+        )?;
+
         let peek_slot = tenant_map_peek_slot(&locked, &tenant_shard_id, TenantSlotPeekMode::Read)
             .map_err(GetTenantError::MapState)?;
         match peek_slot {
@@ -1088,7 +1232,7 @@ pub(crate) async fn get_active_tenant_with_timeout(
                         // Fast path: we don't need to do any async waiting.
                         return Ok(tenant.clone());
                     }
-                    _ => WaitFor::Tenant(tenant.clone()),
+                    _ => (WaitFor::Tenant(tenant.clone()), tenant_shard_id),
                 }
             }
             Some(TenantSlot::Secondary) => {
@@ -1096,7 +1240,9 @@ pub(crate) async fn get_active_tenant_with_timeout(
                     tenant_id,
                 )))
             }
-            Some(TenantSlot::InProgress(barrier)) => WaitFor::Barrier(barrier.clone()),
+            Some(TenantSlot::InProgress(barrier)) => {
+                (WaitFor::Barrier(barrier.clone()), tenant_shard_id)
+            }
             None => {
                 return Err(GetActiveTenantError::NotFound(GetTenantError::NotFound(
                     tenant_id,
@@ -1181,8 +1327,7 @@ pub(crate) async fn delete_tenant(
     // See https://github.com/neondatabase/neon/issues/5080
 
     // TODO(sharding): make delete API sharding-aware
-    let mut slot_guard =
-        tenant_map_acquire_slot(&tenant_shard_id, TenantSlotAcquireMode::MustExist)?;
+    let slot_guard = tenant_map_acquire_slot(&tenant_shard_id, TenantSlotAcquireMode::MustExist)?;
 
     // unwrap is safe because we used MustExist mode when acquiring
     let tenant = match slot_guard.get_old_value().as_ref().unwrap() {
@@ -1262,8 +1407,7 @@ async fn detach_tenant0(
     deletion_queue_client: &DeletionQueueClient,
 ) -> Result<Utf8PathBuf, TenantStateError> {
     let tenant_dir_rename_operation = |tenant_id_to_clean: TenantShardId| async move {
-        // TODO(sharding): make local path helpers shard-aware
-        let local_tenant_directory = conf.tenant_path(&tenant_id_to_clean.tenant_id);
+        let local_tenant_directory = conf.tenant_path(&tenant_id_to_clean);
         safe_rename_tenant_dir(&local_tenant_directory)
             .await
             .with_context(|| format!("local tenant directory {local_tenant_directory:?} rename"))
@@ -1288,8 +1432,7 @@ async fn detach_tenant0(
             Err(TenantStateError::SlotError(TenantSlotError::NotFound(_)))
         )
     {
-        // TODO(sharding): make local paths sharding-aware
-        let tenant_ignore_mark = conf.tenant_ignore_mark_file_path(&tenant_shard_id.tenant_id);
+        let tenant_ignore_mark = conf.tenant_ignore_mark_file_path(&tenant_shard_id);
         if tenant_ignore_mark.exists() {
             info!("Detaching an ignored tenant");
             let tmp_path = tenant_dir_rename_operation(tenant_shard_id)
@@ -1318,9 +1461,9 @@ pub(crate) async fn load_tenant(
 
     let slot_guard =
         tenant_map_acquire_slot(&tenant_shard_id, TenantSlotAcquireMode::MustNotExist)?;
-    let tenant_path = conf.tenant_path(&tenant_id);
+    let tenant_path = conf.tenant_path(&tenant_shard_id);
 
-    let tenant_ignore_mark = conf.tenant_ignore_mark_file_path(&tenant_id);
+    let tenant_ignore_mark = conf.tenant_ignore_mark_file_path(&tenant_shard_id);
     if tenant_ignore_mark.exists() {
         std::fs::remove_file(&tenant_ignore_mark).with_context(|| {
             format!(
@@ -1336,17 +1479,19 @@ pub(crate) async fn load_tenant(
     };
 
     let mut location_conf =
-        Tenant::load_tenant_config(conf, &tenant_id).map_err(TenantMapInsertError::Other)?;
+        Tenant::load_tenant_config(conf, &tenant_shard_id).map_err(TenantMapInsertError::Other)?;
     location_conf.attach_in_generation(generation);
 
-    Tenant::persist_tenant_config(conf, &tenant_id, &location_conf).await?;
+    Tenant::persist_tenant_config(conf, &tenant_shard_id, &location_conf).await?;
 
+    let shard_identity = location_conf.shard;
     let new_tenant = tenant_spawn(
         conf,
-        tenant_id,
+        tenant_shard_id,
         &tenant_path,
         resources,
         AttachedTenantConf::try_from(location_conf)?,
+        shard_identity,
         None,
         &TENANTS,
         SpawnMode::Normal,
@@ -1374,7 +1519,7 @@ async fn ignore_tenant0(
     let tenant_shard_id = TenantShardId::unsharded(tenant_id);
 
     remove_tenant_from_memory(tenants, tenant_shard_id, async {
-        let ignore_mark_file = conf.tenant_ignore_mark_file_path(&tenant_id);
+        let ignore_mark_file = conf.tenant_ignore_mark_file_path(&tenant_shard_id);
         fs::File::create(&ignore_mark_file)
             .await
             .context("Failed to create ignore mark file")
@@ -1432,16 +1577,18 @@ pub(crate) async fn attach_tenant(
     let slot_guard =
         tenant_map_acquire_slot(&tenant_shard_id, TenantSlotAcquireMode::MustNotExist)?;
     let location_conf = LocationConf::attached_single(tenant_conf, generation);
-    let tenant_dir = create_tenant_files(conf, &location_conf, &tenant_id).await?;
+    let tenant_dir = create_tenant_files(conf, &location_conf, &tenant_shard_id).await?;
     // TODO: tenant directory remains on disk if we bail out from here on.
     //       See https://github.com/neondatabase/neon/issues/4233
 
+    let shard_identity = location_conf.shard;
     let attached_tenant = tenant_spawn(
         conf,
-        tenant_id,
+        tenant_shard_id,
         &tenant_dir,
         resources,
         AttachedTenantConf::try_from(location_conf)?,
+        shard_identity,
         None,
         &TENANTS,
         SpawnMode::Normal,
@@ -1507,9 +1654,10 @@ pub enum TenantSlotUpsertError {
     MapState(#[from] TenantMapError),
 }
 
-#[derive(Debug)]
+#[derive(Debug, thiserror::Error)]
 enum TenantSlotDropError {
     /// It is only legal to drop a TenantSlot if its contents are fully shut down
+    #[error("Tenant was not shut down")]
     NotShutdown,
 }
 
@@ -1569,9 +1717,9 @@ impl SlotGuard {
         }
     }
 
-    /// Take any value that was present in the slot before we acquired ownership
+    /// Get any value that was present in the slot before we acquired ownership
     /// of it: in state transitions, this will be the old state.
-    fn get_old_value(&mut self) -> &Option<TenantSlot> {
+    fn get_old_value(&self) -> &Option<TenantSlot> {
         &self.old_value
     }
 
@@ -1789,7 +1937,7 @@ fn tenant_map_acquire_slot_impl(
     METRICS.tenant_slot_writes.inc();
 
     let mut locked = tenants.write().unwrap();
-    let span = tracing::info_span!("acquire_slot", tenant_id=%tenant_shard_id.tenant_id, shard=tenant_shard_id.shard_slug());
+    let span = tracing::info_span!("acquire_slot", tenant_id=%tenant_shard_id.tenant_id, shard = %tenant_shard_id.shard_slug());
     let _guard = span.enter();
 
     let m = match &mut *locked {
@@ -1954,6 +2102,9 @@ pub(crate) async fn immediate_gc(
         .with_context(|| format!("tenant {tenant_id}"))
         .map_err(|e| ApiError::NotFound(e.into()))?;
 
+    // TODO(sharding): make callers of this function shard-aware
+    let tenant_shard_id = TenantShardId::unsharded(tenant_id);
+
     let gc_horizon = gc_req.gc_horizon.unwrap_or_else(|| tenant.get_gc_horizon());
     // Use tenant's pitr setting
     let pitr = tenant.get_pitr_interval();
@@ -1961,6 +2112,7 @@ pub(crate) async fn immediate_gc(
     // Run in task_mgr to avoid race with tenant_detach operation
     let ctx = ctx.detached_child(TaskKind::GarbageCollector, DownloadBehavior::Download);
     let (task_done, wait_task_done) = tokio::sync::oneshot::channel();
+    // TODO: spawning is redundant now, need to hold the gate
     task_mgr::spawn(
         &tokio::runtime::Handle::current(),
         TaskKind::GarbageCollector,
@@ -1970,12 +2122,40 @@ pub(crate) async fn immediate_gc(
         false,
         async move {
             fail::fail_point!("immediate_gc_task_pre");
-            let result = tenant
+
+            #[allow(unused_mut)]
+            let mut result = tenant
                 .gc_iteration(Some(timeline_id), gc_horizon, pitr, &cancel, &ctx)
-                .instrument(info_span!("manual_gc", %tenant_id, %timeline_id))
+                .instrument(info_span!("manual_gc", tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), %timeline_id))
                 .await;
                 // FIXME: `gc_iteration` can return an error for multiple reasons; we should handle it
                 // better once the types support it.
+
+            #[cfg(feature = "testing")]
+            {
+                if let Ok(result) = result.as_mut() {
+                    // why not futures unordered? it seems it needs very much the same task structure
+                    // but would only run on single task.
+                    let mut js = tokio::task::JoinSet::new();
+                    for layer in std::mem::take(&mut result.doomed_layers) {
+                        js.spawn(layer.wait_drop());
+                    }
+                    tracing::info!(total = js.len(), "starting to wait for the gc'd layers to be dropped");
+                    while let Some(res) = js.join_next().await {
+                        res.expect("wait_drop should not panic");
+                    }
+                }
+
+                let timeline = tenant.get_timeline(timeline_id, false).ok();
+                let rtc = timeline.as_ref().and_then(|x| x.remote_client.as_ref());
+
+                if let Some(rtc) = rtc {
+                    // layer drops schedule actions on remote timeline client to actually do the
+                    // deletions; don't care just exit fast about the shutdown error
+                    drop(rtc.wait_completion().await);
+                }
+            }
+
             match task_done.send(result) {
                 Ok(_) => (),
                 Err(result) => error!("failed to send gc result: {result:?}"),

pageserver/src/tenant/remote_timeline_client.rs

@@ -188,6 +188,8 @@ use anyhow::Context;
 use camino::Utf8Path;
 use chrono::{NaiveDateTime, Utc};
 
+pub(crate) use download::download_initdb_tar_zst;
+use pageserver_api::shard::{ShardIndex, TenantShardId};
 use scopeguard::ScopeGuard;
 use tokio_util::sync::CancellationToken;
 pub(crate) use upload::upload_initdb_dir;
@@ -252,6 +254,9 @@ pub(crate) const FAILED_UPLOAD_WARN_THRESHOLD: u32 = 3;
 
 pub(crate) const INITDB_PATH: &str = "initdb.tar.zst";
 
+/// Default buffer size when interfacing with [`tokio::fs::File`].
+pub(crate) const BUFFER_SIZE: usize = 32 * 1024;
+
 pub enum MaybeDeletedIndexPart {
     IndexPart(IndexPart),
     Deleted(IndexPart),
@@ -300,7 +305,7 @@ pub struct RemoteTimelineClient {
 
     runtime: tokio::runtime::Handle,
 
-    tenant_id: TenantId,
+    tenant_shard_id: TenantShardId,
     timeline_id: TimelineId,
     generation: Generation,
 
@@ -324,7 +329,7 @@ impl RemoteTimelineClient {
         remote_storage: GenericRemoteStorage,
         deletion_queue_client: DeletionQueueClient,
         conf: &'static PageServerConf,
-        tenant_id: TenantId,
+        tenant_shard_id: TenantShardId,
         timeline_id: TimelineId,
         generation: Generation,
     ) -> RemoteTimelineClient {
@@ -336,13 +341,16 @@ impl RemoteTimelineClient {
             } else {
                 BACKGROUND_RUNTIME.handle().clone()
             },
-            tenant_id,
+            tenant_shard_id,
             timeline_id,
             generation,
             storage_impl: remote_storage,
             deletion_queue_client,
             upload_queue: Mutex::new(UploadQueue::Uninitialized),
-            metrics: Arc::new(RemoteTimelineClientMetrics::new(&tenant_id, &timeline_id)),
+            metrics: Arc::new(RemoteTimelineClientMetrics::new(
+                &tenant_shard_id,
+                &timeline_id,
+            )),
         }
     }
 
@@ -463,13 +471,13 @@ impl RemoteTimelineClient {
 
         let index_part = download::download_index_part(
             &self.storage_impl,
-            &self.tenant_id,
+            &self.tenant_shard_id,
             &self.timeline_id,
             self.generation,
             cancel,
         )
         .measure_remote_op(
-            self.tenant_id,
+            self.tenant_shard_id.tenant_id,
             self.timeline_id,
             RemoteOpFileKind::Index,
             RemoteOpKind::Download,
@@ -505,13 +513,13 @@ impl RemoteTimelineClient {
             download::download_layer_file(
                 self.conf,
                 &self.storage_impl,
-                self.tenant_id,
+                self.tenant_shard_id,
                 self.timeline_id,
                 layer_file_name,
                 layer_metadata,
             )
             .measure_remote_op(
-                self.tenant_id,
+                self.tenant_shard_id.tenant_id,
                 self.timeline_id,
                 RemoteOpFileKind::Layer,
                 RemoteOpKind::Download,
@@ -657,10 +665,10 @@ impl RemoteTimelineClient {
         let mut guard = self.upload_queue.lock().unwrap();
         let upload_queue = guard.initialized_mut()?;
 
-        let with_generations =
+        let with_metadata =
             self.schedule_unlinking_of_layers_from_index_part0(upload_queue, names.iter().cloned());
 
-        self.schedule_deletion_of_unlinked0(upload_queue, with_generations);
+        self.schedule_deletion_of_unlinked0(upload_queue, with_metadata);
 
         // Launch the tasks immediately, if possible
         self.launch_queued_tasks(upload_queue);
@@ -695,7 +703,7 @@ impl RemoteTimelineClient {
         self: &Arc<Self>,
         upload_queue: &mut UploadQueueInitialized,
         names: I,
-    ) -> Vec<(LayerFileName, Generation)>
+    ) -> Vec<(LayerFileName, LayerFileMetadata)>
     where
         I: IntoIterator<Item = LayerFileName>,
     {
@@ -703,16 +711,17 @@ impl RemoteTimelineClient {
         // so we don't need update it. Just serialize it.
         let metadata = upload_queue.latest_metadata.clone();
 
-        // Decorate our list of names with each name's generation, dropping
-        // names that are unexpectedly missing from our metadata.
-        let with_generations: Vec<_> = names
+        // Decorate our list of names with each name's metadata, dropping
+        // names that are unexpectedly missing from our metadata.  This metadata
+        // is later used when physically deleting layers, to construct key paths.
+        let with_metadata: Vec<_> = names
             .into_iter()
             .filter_map(|name| {
                 let meta = upload_queue.latest_files.remove(&name);
 
                 if let Some(meta) = meta {
                     upload_queue.latest_files_changes_since_metadata_upload_scheduled += 1;
-                    Some((name, meta.generation))
+                    Some((name, meta))
                 } else {
                     // This can only happen if we forgot to to schedule the file upload
                     // before scheduling the delete. Log it because it is a rare/strange
@@ -725,9 +734,10 @@ impl RemoteTimelineClient {
             .collect();
 
         #[cfg(feature = "testing")]
-        for (name, gen) in &with_generations {
-            if let Some(unexpected) = upload_queue.dangling_files.insert(name.to_owned(), *gen) {
-                if &unexpected == gen {
+        for (name, metadata) in &with_metadata {
+            let gen = metadata.generation;
+            if let Some(unexpected) = upload_queue.dangling_files.insert(name.to_owned(), gen) {
+                if unexpected == gen {
                     tracing::error!("{name} was unlinked twice with same generation");
                 } else {
                     tracing::error!("{name} was unlinked twice with different generations {gen:?} and {unexpected:?}");
@@ -742,14 +752,14 @@ impl RemoteTimelineClient {
             self.schedule_index_upload(upload_queue, metadata);
         }
 
-        with_generations
+        with_metadata
     }
 
     /// Schedules deletion for layer files which have previously been unlinked from the
     /// `index_part.json` with [`Self::schedule_gc_update`] or [`Self::schedule_compaction_update`].
     pub(crate) fn schedule_deletion_of_unlinked(
         self: &Arc<Self>,
-        layers: Vec<(LayerFileName, Generation)>,
+        layers: Vec<(LayerFileName, LayerFileMetadata)>,
     ) -> anyhow::Result<()> {
         let mut guard = self.upload_queue.lock().unwrap();
         let upload_queue = guard.initialized_mut()?;
@@ -762,16 +772,22 @@ impl RemoteTimelineClient {
     fn schedule_deletion_of_unlinked0(
         self: &Arc<Self>,
         upload_queue: &mut UploadQueueInitialized,
-        with_generations: Vec<(LayerFileName, Generation)>,
+        with_metadata: Vec<(LayerFileName, LayerFileMetadata)>,
     ) {
-        for (name, gen) in &with_generations {
-            info!("scheduling deletion of layer {}{}", name, gen.get_suffix());
+        for (name, meta) in &with_metadata {
+            info!(
+                "scheduling deletion of layer {}{} (shard {})",
+                name,
+                meta.generation.get_suffix(),
+                meta.shard
+            );
         }
 
         #[cfg(feature = "testing")]
-        for (name, gen) in &with_generations {
+        for (name, meta) in &with_metadata {
+            let gen = meta.generation;
             match upload_queue.dangling_files.remove(name) {
-                Some(same) if &same == gen => { /* expected */ }
+                Some(same) if same == gen => { /* expected */ }
                 Some(other) => {
                     tracing::error!("{name} was unlinked with {other:?} but deleted with {gen:?}");
                 }
@@ -783,7 +799,7 @@ impl RemoteTimelineClient {
 
         // schedule the actual deletions
         let op = UploadOp::Delete(Delete {
-            layers: with_generations,
+            layers: with_metadata,
         });
         self.calls_unfinished_metric_begin(&op);
         upload_queue.queued_operations.push_back(op);
@@ -812,10 +828,8 @@ impl RemoteTimelineClient {
         Ok(())
     }
 
-    ///
     /// Wait for all previously scheduled uploads/deletions to complete
-    ///
-    pub async fn wait_completion(self: &Arc<Self>) -> anyhow::Result<()> {
+    pub(crate) async fn wait_completion(self: &Arc<Self>) -> anyhow::Result<()> {
         let mut receiver = {
             let mut guard = self.upload_queue.lock().unwrap();
             let upload_queue = guard.initialized_mut()?;
@@ -825,6 +839,7 @@ impl RemoteTimelineClient {
         if receiver.changed().await.is_err() {
             anyhow::bail!("wait_completion aborted because upload queue was stopped");
         }
+
         Ok(())
     }
 
@@ -851,6 +866,56 @@ impl RemoteTimelineClient {
         receiver
     }
 
+    /// Wait for all previously scheduled operations to complete, and then stop.
+    ///
+    /// Not cancellation safe
+    pub(crate) async fn shutdown(self: &Arc<Self>) -> Result<(), StopError> {
+        // On cancellation the queue is left in ackward state of refusing new operations but
+        // proper stop is yet to be called. On cancel the original or some later task must call
+        // `stop` or `shutdown`.
+        let sg = scopeguard::guard((), |_| {
+            tracing::error!("RemoteTimelineClient::shutdown was cancelled; this should not happen, do not make this into an allowed_error")
+        });
+
+        let fut = {
+            let mut guard = self.upload_queue.lock().unwrap();
+            let upload_queue = match &mut *guard {
+                UploadQueue::Stopped(_) => return Ok(()),
+                UploadQueue::Uninitialized => return Err(StopError::QueueUninitialized),
+                UploadQueue::Initialized(ref mut init) => init,
+            };
+
+            // if the queue is already stuck due to a shutdown operation which was cancelled, then
+            // just don't add more of these as they would never complete.
+            //
+            // TODO: if launch_queued_tasks were to be refactored to accept a &mut UploadQueue
+            // in every place we would not have to jump through this hoop, and this method could be
+            // made cancellable.
+            if !upload_queue.shutting_down {
+                upload_queue.shutting_down = true;
+                upload_queue.queued_operations.push_back(UploadOp::Shutdown);
+                // this operation is not counted similar to Barrier
+
+                self.launch_queued_tasks(upload_queue);
+            }
+
+            upload_queue.shutdown_ready.clone().acquire_owned()
+        };
+
+        let res = fut.await;
+
+        scopeguard::ScopeGuard::into_inner(sg);
+
+        match res {
+            Ok(_permit) => unreachable!("shutdown_ready should not have been added permits"),
+            Err(_closed) => {
+                // expected
+            }
+        }
+
+        self.stop()
+    }
+
     /// Set the deleted_at field in the remote index file.
     ///
     /// This fails if the upload queue has not been `stop()`ed.
@@ -902,7 +967,7 @@ impl RemoteTimelineClient {
             || {
                 upload::upload_index_part(
                     &self.storage_impl,
-                    &self.tenant_id,
+                    &self.tenant_shard_id,
                     &self.timeline_id,
                     self.generation,
                     &index_part_with_deleted_at,
@@ -960,8 +1025,9 @@ impl RemoteTimelineClient {
                 .drain()
                 .map(|(file_name, meta)| {
                     remote_layer_path(
-                        &self.tenant_id,
+                        &self.tenant_shard_id.tenant_id,
                         &self.timeline_id,
+                        meta.shard,
                         &file_name,
                         meta.generation,
                     )
@@ -974,7 +1040,7 @@ impl RemoteTimelineClient {
 
         // Do not delete index part yet, it is needed for possible retry. If we remove it first
         // and retry will arrive to different pageserver there wont be any traces of it on remote storage
-        let timeline_storage_path = remote_timeline_path(&self.tenant_id, &self.timeline_id);
+        let timeline_storage_path = remote_timeline_path(&self.tenant_shard_id, &self.timeline_id);
 
         // Execute all pending deletions, so that when we proceed to do a list_prefixes below, we aren't
         // taking the burden of listing all the layers that we already know we should delete.
@@ -1010,12 +1076,22 @@ impl RemoteTimelineClient {
             .unwrap_or(
                 // No generation-suffixed indices, assume we are dealing with
                 // a legacy index.
-                remote_index_path(&self.tenant_id, &self.timeline_id, Generation::none()),
+                remote_index_path(&self.tenant_shard_id, &self.timeline_id, Generation::none()),
             );
 
         let remaining_layers: Vec<RemotePath> = remaining
             .into_iter()
-            .filter(|p| p!= &latest_index)
+            .filter(|p| {
+                if p == &latest_index {
+                    return false;
+                }
+                if let Some(name) = p.object_name() {
+                    if name == INITDB_PATH {
+                        return false;
+                    }
+                }
+                true
+            })
             .inspect(|path| {
                 if let Some(name) = path.object_name() {
                     info!(%name, "deleting a file not referenced from index_part.json");
@@ -1081,7 +1157,9 @@ impl RemoteTimelineClient {
                     upload_queue.num_inprogress_deletions == upload_queue.inprogress_tasks.len()
                 }
 
-                UploadOp::Barrier(_) => upload_queue.inprogress_tasks.is_empty(),
+                UploadOp::Barrier(_) | UploadOp::Shutdown => {
+                    upload_queue.inprogress_tasks.is_empty()
+                }
             };
 
             // If we cannot launch this task, don't look any further.
@@ -1094,6 +1172,13 @@ impl RemoteTimelineClient {
                 break;
             }
 
+            if let UploadOp::Shutdown = next_op {
+                // leave the op in the queue but do not start more tasks; it will be dropped when
+                // the stop is called.
+                upload_queue.shutdown_ready.close();
+                break;
+            }
+
             // We can launch this task. Remove it from the queue first.
             let next_op = upload_queue.queued_operations.pop_front().unwrap();
 
@@ -1114,6 +1199,7 @@ impl RemoteTimelineClient {
                     sender.send_replace(());
                     continue;
                 }
+                UploadOp::Shutdown => unreachable!("shutdown is intentionally never popped off"),
             };
 
             // Assign unique ID to this task
@@ -1132,12 +1218,12 @@ impl RemoteTimelineClient {
 
             // Spawn task to perform the task
             let self_rc = Arc::clone(self);
-            let tenant_id = self.tenant_id;
+            let tenant_shard_id = self.tenant_shard_id;
             let timeline_id = self.timeline_id;
             task_mgr::spawn(
                 &self.runtime,
                 TaskKind::RemoteUploadTask,
-                Some(self.tenant_id),
+                Some(self.tenant_shard_id.tenant_id),
                 Some(self.timeline_id),
                 "remote upload",
                 false,
@@ -1145,7 +1231,7 @@ impl RemoteTimelineClient {
                     self_rc.perform_upload_task(task).await;
                     Ok(())
                 }
-                .instrument(info_span!(parent: None, "remote_upload", %tenant_id, %timeline_id, %upload_task_id)),
+                .instrument(info_span!(parent: None, "remote_upload", tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), %timeline_id, %upload_task_id)),
             );
 
             // Loop back to process next task
@@ -1197,7 +1283,7 @@ impl RemoteTimelineClient {
                         self.generation,
                     )
                     .measure_remote_op(
-                        self.tenant_id,
+                        self.tenant_shard_id.tenant_id,
                         self.timeline_id,
                         RemoteOpFileKind::Layer,
                         RemoteOpKind::Upload,
@@ -1217,13 +1303,13 @@ impl RemoteTimelineClient {
 
                     let res = upload::upload_index_part(
                         &self.storage_impl,
-                        &self.tenant_id,
+                        &self.tenant_shard_id,
                         &self.timeline_id,
                         self.generation,
                         index_part,
                     )
                     .measure_remote_op(
-                        self.tenant_id,
+                        self.tenant_shard_id.tenant_id,
                         self.timeline_id,
                         RemoteOpFileKind::Index,
                         RemoteOpKind::Upload,
@@ -1243,7 +1329,7 @@ impl RemoteTimelineClient {
                     pausable_failpoint!("before-delete-layer-pausable");
                     self.deletion_queue_client
                         .push_layers(
-                            self.tenant_id,
+                            self.tenant_shard_id,
                             self.timeline_id,
                             self.generation,
                             delete.layers.clone(),
@@ -1251,10 +1337,10 @@ impl RemoteTimelineClient {
                         .await
                         .map_err(|e| anyhow::anyhow!(e))
                 }
-                UploadOp::Barrier(_) => {
+                unexpected @ UploadOp::Barrier(_) | unexpected @ UploadOp::Shutdown => {
                     // unreachable. Barrier operations are handled synchronously in
                     // launch_queued_tasks
-                    warn!("unexpected Barrier operation in perform_upload_task");
+                    warn!("unexpected {unexpected:?} operation in perform_upload_task");
                     break;
                 }
             };
@@ -1348,7 +1434,7 @@ impl RemoteTimelineClient {
                     upload_queue.num_inprogress_deletions -= 1;
                     None
                 }
-                UploadOp::Barrier(_) => unreachable!(),
+                UploadOp::Barrier(..) | UploadOp::Shutdown => unreachable!(),
             };
 
             // Launch any queued tasks that were unblocked by this one.
@@ -1362,7 +1448,7 @@ impl RemoteTimelineClient {
             // data safety guarantees (see docs/rfcs/025-generation-numbers.md)
             self.deletion_queue_client
                 .update_remote_consistent_lsn(
-                    self.tenant_id,
+                    self.tenant_shard_id,
                     self.timeline_id,
                     self.generation,
                     lsn,
@@ -1403,7 +1489,7 @@ impl RemoteTimelineClient {
                     reason: "should we track deletes? positive or negative sign?",
                 },
             ),
-            UploadOp::Barrier(_) => {
+            UploadOp::Barrier(..) | UploadOp::Shutdown => {
                 // we do not account these
                 return None;
             }
@@ -1429,10 +1515,13 @@ impl RemoteTimelineClient {
     }
 
     /// Close the upload queue for new operations and cancel queued operations.
+    ///
+    /// Use [`RemoteTimelineClient::shutdown`] for graceful stop.
+    ///
     /// In-progress operations will still be running after this function returns.
     /// Use `task_mgr::shutdown_tasks(None, Some(self.tenant_id), Some(timeline_id))`
     /// to wait for them to complete, after calling this function.
-    pub fn stop(&self) -> Result<(), StopError> {
+    pub(crate) fn stop(&self) -> Result<(), StopError> {
         // Whichever *task* for this RemoteTimelineClient grabs the mutex first will transition the queue
         // into stopped state, thereby dropping all off the queued *ops* which haven't become *tasks* yet.
         // The other *tasks* will come here and observe an already shut down queue and hence simply wrap up their business.
@@ -1470,6 +1559,8 @@ impl RemoteTimelineClient {
                         queued_operations: VecDeque::default(),
                         #[cfg(feature = "testing")]
                         dangling_files: HashMap::default(),
+                        shutting_down: false,
+                        shutdown_ready: Arc::new(tokio::sync::Semaphore::new(0)),
                     };
 
                     let upload_queue = std::mem::replace(
@@ -1515,24 +1606,32 @@ impl RemoteTimelineClient {
     }
 }
 
-pub fn remote_timelines_path(tenant_id: &TenantId) -> RemotePath {
-    let path = format!("tenants/{tenant_id}/{TIMELINES_SEGMENT_NAME}");
+pub fn remote_timelines_path(tenant_shard_id: &TenantShardId) -> RemotePath {
+    let path = format!("tenants/{tenant_shard_id}/{TIMELINES_SEGMENT_NAME}");
     RemotePath::from_string(&path).expect("Failed to construct path")
 }
 
-pub fn remote_timeline_path(tenant_id: &TenantId, timeline_id: &TimelineId) -> RemotePath {
-    remote_timelines_path(tenant_id).join(Utf8Path::new(&timeline_id.to_string()))
+pub fn remote_timeline_path(
+    tenant_shard_id: &TenantShardId,
+    timeline_id: &TimelineId,
+) -> RemotePath {
+    remote_timelines_path(tenant_shard_id).join(Utf8Path::new(&timeline_id.to_string()))
 }
 
+/// Note that the shard component of a remote layer path is _not_ always the same
+/// as in the TenantShardId of the caller: tenants may reference layers from a different
+/// ShardIndex.  Use the ShardIndex from the layer's metadata.
 pub fn remote_layer_path(
     tenant_id: &TenantId,
     timeline_id: &TimelineId,
+    shard: ShardIndex,
     layer_file_name: &LayerFileName,
     generation: Generation,
 ) -> RemotePath {
     // Generation-aware key format
     let path = format!(
-        "tenants/{tenant_id}/{TIMELINES_SEGMENT_NAME}/{timeline_id}/{0}{1}",
+        "tenants/{tenant_id}{0}/{TIMELINES_SEGMENT_NAME}/{timeline_id}/{1}{2}",
+        shard.get_suffix(),
         layer_file_name.file_name(),
         generation.get_suffix()
     );
@@ -1548,12 +1647,12 @@ pub fn remote_initdb_archive_path(tenant_id: &TenantId, timeline_id: &TimelineId
 }
 
 pub fn remote_index_path(
-    tenant_id: &TenantId,
+    tenant_shard_id: &TenantShardId,
     timeline_id: &TimelineId,
     generation: Generation,
 ) -> RemotePath {
     RemotePath::from_string(&format!(
-        "tenants/{tenant_id}/{TIMELINES_SEGMENT_NAME}/{timeline_id}/{0}{1}",
+        "tenants/{tenant_shard_id}/{TIMELINES_SEGMENT_NAME}/{timeline_id}/{0}{1}",
         IndexPart::FILE_NAME,
         generation.get_suffix()
     ))
@@ -1695,14 +1794,14 @@ mod tests {
             Arc::new(RemoteTimelineClient {
                 conf: self.harness.conf,
                 runtime: tokio::runtime::Handle::current(),
-                tenant_id: self.harness.tenant_id,
+                tenant_shard_id: self.harness.tenant_shard_id,
                 timeline_id: TIMELINE_ID,
                 generation,
                 storage_impl: self.harness.remote_storage.clone(),
                 deletion_queue_client: self.harness.deletion_queue.new_client(),
                 upload_queue: Mutex::new(UploadQueue::Uninitialized),
                 metrics: Arc::new(RemoteTimelineClientMetrics::new(
-                    &self.harness.tenant_id,
+                    &self.harness.tenant_shard_id,
                     &TIMELINE_ID,
                 )),
             })
@@ -1778,6 +1877,7 @@ mod tests {
         println!("remote_timeline_dir: {remote_timeline_dir}");
 
         let generation = harness.generation;
+        let shard = harness.shard;
 
         // Create a couple of dummy files,  schedule upload for them
 
@@ -1794,7 +1894,7 @@ mod tests {
                 harness.conf,
                 &timeline,
                 name,
-                LayerFileMetadata::new(contents.len() as u64, generation),
+                LayerFileMetadata::new(contents.len() as u64, generation, shard),
             )
         }).collect::<Vec<_>>();
 
@@ -1943,7 +2043,7 @@ mod tests {
             harness.conf,
             &timeline,
             layer_file_name_1.clone(),
-            LayerFileMetadata::new(content_1.len() as u64, harness.generation),
+            LayerFileMetadata::new(content_1.len() as u64, harness.generation, harness.shard),
         );
 
         #[derive(Debug, PartialEq, Clone, Copy)]
@@ -2029,7 +2129,12 @@ mod tests {
         std::fs::create_dir_all(remote_timeline_dir).expect("creating test dir should work");
 
         let index_path = test_state.harness.remote_fs_dir.join(
-            remote_index_path(&test_state.harness.tenant_id, &TIMELINE_ID, generation).get_path(),
+            remote_index_path(
+                &test_state.harness.tenant_shard_id,
+                &TIMELINE_ID,
+                generation,
+            )
+            .get_path(),
         );
         eprintln!("Writing {index_path}");
         std::fs::write(&index_path, index_part_bytes).unwrap();

pageserver/src/tenant/remote_timeline_client/download.rs

@@ -8,10 +8,12 @@ use std::future::Future;
 use std::time::Duration;
 
 use anyhow::{anyhow, Context};
-use camino::Utf8Path;
-use tokio::fs;
-use tokio::io::AsyncWriteExt;
+use camino::{Utf8Path, Utf8PathBuf};
+use pageserver_api::shard::TenantShardId;
+use tokio::fs::{self, File, OpenOptions};
+use tokio::io::{AsyncSeekExt, AsyncWriteExt};
 use tokio_util::sync::CancellationToken;
+use tracing::warn;
 use utils::{backoff, crashsafe};
 
 use crate::config::PageServerConf;
@@ -19,14 +21,15 @@ use crate::tenant::remote_timeline_client::{remote_layer_path, remote_timelines_
 use crate::tenant::storage_layer::LayerFileName;
 use crate::tenant::timeline::span::debug_assert_current_span_has_tenant_and_timeline_id;
 use crate::tenant::Generation;
+use crate::TEMP_FILE_SUFFIX;
 use remote_storage::{DownloadError, GenericRemoteStorage, ListingMode};
 use utils::crashsafe::path_with_suffix_extension;
-use utils::id::{TenantId, TimelineId};
+use utils::id::TimelineId;
 
 use super::index::{IndexPart, LayerFileMetadata};
 use super::{
-    parse_remote_index_path, remote_index_path, FAILED_DOWNLOAD_WARN_THRESHOLD,
-    FAILED_REMOTE_OP_RETRIES,
+    parse_remote_index_path, remote_index_path, remote_initdb_archive_path,
+    FAILED_DOWNLOAD_WARN_THRESHOLD, FAILED_REMOTE_OP_RETRIES, INITDB_PATH,
 };
 
 static MAX_DOWNLOAD_DURATION: Duration = Duration::from_secs(120);
@@ -39,7 +42,7 @@ static MAX_DOWNLOAD_DURATION: Duration = Duration::from_secs(120);
 pub async fn download_layer_file<'a>(
     conf: &'static PageServerConf,
     storage: &'a GenericRemoteStorage,
-    tenant_id: TenantId,
+    tenant_shard_id: TenantShardId,
     timeline_id: TimelineId,
     layer_file_name: &'a LayerFileName,
     layer_metadata: &'a LayerFileMetadata,
@@ -47,12 +50,13 @@ pub async fn download_layer_file<'a>(
     debug_assert_current_span_has_tenant_and_timeline_id();
 
     let local_path = conf
-        .timeline_path(&tenant_id, &timeline_id)
+        .timeline_path(&tenant_shard_id, &timeline_id)
         .join(layer_file_name.file_name());
 
     let remote_path = remote_layer_path(
-        &tenant_id,
+        &tenant_shard_id.tenant_id,
         &timeline_id,
+        layer_metadata.shard,
         layer_file_name,
         layer_metadata.generation,
     );
@@ -71,12 +75,11 @@ pub async fn download_layer_file<'a>(
 
     let (mut destination_file, bytes_amount) = download_retry(
         || async {
-            // TODO: this doesn't use the cached fd for some reason?
-            let mut destination_file = fs::File::create(&temp_file_path)
+            let destination_file = tokio::fs::File::create(&temp_file_path)
                 .await
                 .with_context(|| format!("create a destination file for layer '{temp_file_path}'"))
                 .map_err(DownloadError::Other)?;
-            let mut download = storage
+            let download = storage
                 .download(&remote_path)
                 .await
                 .with_context(|| {
@@ -86,9 +89,14 @@ pub async fn download_layer_file<'a>(
                 })
                 .map_err(DownloadError::Other)?;
 
+            let mut destination_file =
+                tokio::io::BufWriter::with_capacity(super::BUFFER_SIZE, destination_file);
+
+            let mut reader = tokio_util::io::StreamReader::new(download.download_stream);
+
             let bytes_amount = tokio::time::timeout(
                 MAX_DOWNLOAD_DURATION,
-                tokio::io::copy(&mut download.download_stream, &mut destination_file),
+                tokio::io::copy_buf(&mut reader, &mut destination_file),
             )
             .await
             .map_err(|e| DownloadError::Other(anyhow::anyhow!("Timed out  {:?}", e)))?
@@ -99,6 +107,8 @@ pub async fn download_layer_file<'a>(
             })
             .map_err(DownloadError::Other)?;
 
+            let destination_file = destination_file.into_inner();
+
             Ok((destination_file, bytes_amount))
         },
         &format!("download {remote_path:?}"),
@@ -169,10 +179,10 @@ pub fn is_temp_download_file(path: &Utf8Path) -> bool {
 /// List timelines of given tenant in remote storage
 pub async fn list_remote_timelines(
     storage: &GenericRemoteStorage,
-    tenant_id: TenantId,
+    tenant_shard_id: TenantShardId,
     cancel: CancellationToken,
 ) -> anyhow::Result<(HashSet<TimelineId>, HashSet<String>)> {
-    let remote_path = remote_timelines_path(&tenant_id);
+    let remote_path = remote_timelines_path(&tenant_shard_id);
 
     fail::fail_point!("storage-sync-list-remote-timelines", |_| {
         anyhow::bail!("storage-sync-list-remote-timelines");
@@ -180,7 +190,7 @@ pub async fn list_remote_timelines(
 
     let listing = download_retry_forever(
         || storage.list(Some(&remote_path), ListingMode::WithDelimiter),
-        &format!("list timelines for {tenant_id}"),
+        &format!("list timelines for {tenant_shard_id}"),
         cancel,
     )
     .await?;
@@ -190,7 +200,7 @@ pub async fn list_remote_timelines(
 
     for timeline_remote_storage_key in listing.prefixes {
         let object_name = timeline_remote_storage_key.object_name().ok_or_else(|| {
-            anyhow::anyhow!("failed to get timeline id for remote tenant {tenant_id}")
+            anyhow::anyhow!("failed to get timeline id for remote tenant {tenant_shard_id}")
         })?;
 
         match object_name.parse::<TimelineId>() {
@@ -211,25 +221,27 @@ pub async fn list_remote_timelines(
 
 async fn do_download_index_part(
     storage: &GenericRemoteStorage,
-    tenant_id: &TenantId,
+    tenant_shard_id: &TenantShardId,
     timeline_id: &TimelineId,
     index_generation: Generation,
     cancel: CancellationToken,
 ) -> Result<IndexPart, DownloadError> {
-    let remote_path = remote_index_path(tenant_id, timeline_id, index_generation);
+    use futures::stream::StreamExt;
+
+    let remote_path = remote_index_path(tenant_shard_id, timeline_id, index_generation);
 
     let index_part_bytes = download_retry_forever(
         || async {
-            let mut index_part_download = storage.download(&remote_path).await?;
+            let index_part_download = storage.download(&remote_path).await?;
 
             let mut index_part_bytes = Vec::new();
-            tokio::io::copy(
-                &mut index_part_download.download_stream,
-                &mut index_part_bytes,
-            )
-            .await
-            .with_context(|| format!("download index part at {remote_path:?}"))
-            .map_err(DownloadError::Other)?;
+            let mut stream = std::pin::pin!(index_part_download.download_stream);
+            while let Some(chunk) = stream.next().await {
+                let chunk = chunk
+                    .with_context(|| format!("download index part at {remote_path:?}"))
+                    .map_err(DownloadError::Other)?;
+                index_part_bytes.extend_from_slice(&chunk[..]);
+            }
             Ok(index_part_bytes)
         },
         &format!("download {remote_path:?}"),
@@ -252,7 +264,7 @@ async fn do_download_index_part(
 #[tracing::instrument(skip_all, fields(generation=?my_generation))]
 pub(super) async fn download_index_part(
     storage: &GenericRemoteStorage,
-    tenant_id: &TenantId,
+    tenant_shard_id: &TenantShardId,
     timeline_id: &TimelineId,
     my_generation: Generation,
     cancel: CancellationToken,
@@ -261,8 +273,14 @@ pub(super) async fn download_index_part(
 
     if my_generation.is_none() {
         // Operating without generations: just fetch the generation-less path
-        return do_download_index_part(storage, tenant_id, timeline_id, my_generation, cancel)
-            .await;
+        return do_download_index_part(
+            storage,
+            tenant_shard_id,
+            timeline_id,
+            my_generation,
+            cancel,
+        )
+        .await;
     }
 
     // Stale case: If we were intentionally attached in a stale generation, there may already be a remote
@@ -271,7 +289,7 @@ pub(super) async fn download_index_part(
     // This is an optimization to avoid doing the listing for the general case below.
     let res = do_download_index_part(
         storage,
-        tenant_id,
+        tenant_shard_id,
         timeline_id,
         my_generation,
         cancel.clone(),
@@ -298,7 +316,7 @@ pub(super) async fn download_index_part(
     // This is an optimization to avoid doing the listing for the general case below.
     let res = do_download_index_part(
         storage,
-        tenant_id,
+        tenant_shard_id,
         timeline_id,
         my_generation.previous(),
         cancel.clone(),
@@ -320,8 +338,9 @@ pub(super) async fn download_index_part(
     }
 
     // General case/fallback: if there is no index at my_generation or prev_generation, then list all index_part.json
-    // objects, and select the highest one with a generation <= my_generation.
-    let index_prefix = remote_index_path(tenant_id, timeline_id, Generation::none());
+    // objects, and select the highest one with a generation <= my_generation.  Constructing the prefix is equivalent
+    // to constructing a full index path with no generation, because the generation is a suffix.
+    let index_prefix = remote_index_path(tenant_shard_id, timeline_id, Generation::none());
     let indices = backoff::retry(
         || async { storage.list_files(Some(&index_prefix)).await },
         |_| false,
@@ -347,18 +366,93 @@ pub(super) async fn download_index_part(
     match max_previous_generation {
         Some(g) => {
             tracing::debug!("Found index_part in generation {g:?}");
-            do_download_index_part(storage, tenant_id, timeline_id, g, cancel).await
+            do_download_index_part(storage, tenant_shard_id, timeline_id, g, cancel).await
         }
         None => {
             // Migration from legacy pre-generation state: we have a generation but no prior
             // attached pageservers did.  Try to load from a no-generation path.
-            tracing::info!("No index_part.json* found");
-            do_download_index_part(storage, tenant_id, timeline_id, Generation::none(), cancel)
-                .await
+            tracing::debug!("No index_part.json* found");
+            do_download_index_part(
+                storage,
+                tenant_shard_id,
+                timeline_id,
+                Generation::none(),
+                cancel,
+            )
+            .await
         }
     }
 }
 
+pub(crate) async fn download_initdb_tar_zst(
+    conf: &'static PageServerConf,
+    storage: &GenericRemoteStorage,
+    tenant_shard_id: &TenantShardId,
+    timeline_id: &TimelineId,
+) -> Result<(Utf8PathBuf, File), DownloadError> {
+    debug_assert_current_span_has_tenant_and_timeline_id();
+
+    let remote_path = remote_initdb_archive_path(&tenant_shard_id.tenant_id, timeline_id);
+
+    let timeline_path = conf.timelines_path(tenant_shard_id);
+
+    if !timeline_path.exists() {
+        tokio::fs::create_dir_all(&timeline_path)
+            .await
+            .with_context(|| format!("timeline dir creation {timeline_path}"))
+            .map_err(DownloadError::Other)?;
+    }
+    let temp_path = timeline_path.join(format!(
+        "{INITDB_PATH}.download-{timeline_id}.{TEMP_FILE_SUFFIX}"
+    ));
+
+    let file = download_retry(
+        || async {
+            let file = OpenOptions::new()
+                .create(true)
+                .truncate(true)
+                .read(true)
+                .write(true)
+                .open(&temp_path)
+                .await
+                .with_context(|| format!("tempfile creation {temp_path}"))
+                .map_err(DownloadError::Other)?;
+
+            let download = storage.download(&remote_path).await?;
+            let mut download = tokio_util::io::StreamReader::new(download.download_stream);
+            let mut writer = tokio::io::BufWriter::with_capacity(8 * 1024, file);
+
+            tokio::io::copy_buf(&mut download, &mut writer)
+                .await
+                .with_context(|| format!("download initdb.tar.zst at {remote_path:?}"))
+                .map_err(DownloadError::Other)?;
+
+            let mut file = writer.into_inner();
+
+            file.seek(std::io::SeekFrom::Start(0))
+                .await
+                .with_context(|| format!("rewinding initdb.tar.zst at: {remote_path:?}"))
+                .map_err(DownloadError::Other)?;
+
+            Ok(file)
+        },
+        &format!("download {remote_path}"),
+    )
+    .await
+    .map_err(|e| {
+        // Do a best-effort attempt at deleting the temporary file upon encountering an error.
+        // We don't have async here nor do we want to pile on any extra errors.
+        if let Err(e) = std::fs::remove_file(&temp_path) {
+            if e.kind() != std::io::ErrorKind::NotFound {
+                warn!("error deleting temporary file {temp_path}: {e}");
+            }
+        }
+        e
+    })?;
+
+    Ok((temp_path, file))
+}
+
 /// Helper function to handle retries for a download operation.
 ///
 /// Remote operations can fail due to rate limits (IAM, S3), spurious network

pageserver/src/tenant/remote_timeline_client/index.rs

@@ -12,6 +12,7 @@ use crate::tenant::metadata::TimelineMetadata;
 use crate::tenant::storage_layer::LayerFileName;
 use crate::tenant::upload_queue::UploadQueueInitialized;
 use crate::tenant::Generation;
+use pageserver_api::shard::ShardIndex;
 
 use utils::lsn::Lsn;
 
@@ -25,6 +26,8 @@ pub struct LayerFileMetadata {
     file_size: u64,
 
     pub(crate) generation: Generation,
+
+    pub(crate) shard: ShardIndex,
 }
 
 impl From<&'_ IndexLayerMetadata> for LayerFileMetadata {
@@ -32,15 +35,17 @@ impl From<&'_ IndexLayerMetadata> for LayerFileMetadata {
         LayerFileMetadata {
             file_size: other.file_size,
             generation: other.generation,
+            shard: other.shard,
         }
     }
 }
 
 impl LayerFileMetadata {
-    pub fn new(file_size: u64, generation: Generation) -> Self {
+    pub fn new(file_size: u64, generation: Generation, shard: ShardIndex) -> Self {
         LayerFileMetadata {
             file_size,
             generation,
+            shard,
         }
     }
 
@@ -161,6 +166,10 @@ pub struct IndexLayerMetadata {
     #[serde(default = "Generation::none")]
     #[serde(skip_serializing_if = "Generation::is_none")]
     pub generation: Generation,
+
+    #[serde(default = "ShardIndex::unsharded")]
+    #[serde(skip_serializing_if = "ShardIndex::is_unsharded")]
+    pub shard: ShardIndex,
 }
 
 impl From<LayerFileMetadata> for IndexLayerMetadata {
@@ -168,6 +177,7 @@ impl From<LayerFileMetadata> for IndexLayerMetadata {
         IndexLayerMetadata {
             file_size: other.file_size,
             generation: other.generation,
+            shard: other.shard,
         }
     }
 }
@@ -195,13 +205,15 @@ mod tests {
             layer_metadata: HashMap::from([
                 ("000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__0000000001696070-00000000016960E9".parse().unwrap(), IndexLayerMetadata {
                     file_size: 25600000,
-                    generation: Generation::none()
+                    generation: Generation::none(),
+                    shard: ShardIndex::unsharded()
                 }),
                 ("000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__00000000016B59D8-00000000016B5A51".parse().unwrap(), IndexLayerMetadata {
                     // serde_json should always parse this but this might be a double with jq for
                     // example.
                     file_size: 9007199254741001,
-                    generation: Generation::none()
+                    generation: Generation::none(),
+                    shard: ShardIndex::unsharded()
                 })
             ]),
             disk_consistent_lsn: "0/16960E8".parse::<Lsn>().unwrap(),
@@ -233,13 +245,15 @@ mod tests {
             layer_metadata: HashMap::from([
                 ("000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__0000000001696070-00000000016960E9".parse().unwrap(), IndexLayerMetadata {
                     file_size: 25600000,
-                    generation: Generation::none()
+                    generation: Generation::none(),
+                    shard: ShardIndex::unsharded()
                 }),
                 ("000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__00000000016B59D8-00000000016B5A51".parse().unwrap(), IndexLayerMetadata {
                     // serde_json should always parse this but this might be a double with jq for
                     // example.
                     file_size: 9007199254741001,
-                    generation: Generation::none()
+                    generation: Generation::none(),
+                    shard: ShardIndex::unsharded()
                 })
             ]),
             disk_consistent_lsn: "0/16960E8".parse::<Lsn>().unwrap(),
@@ -272,13 +286,15 @@ mod tests {
             layer_metadata: HashMap::from([
                 ("000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__0000000001696070-00000000016960E9".parse().unwrap(), IndexLayerMetadata {
                     file_size: 25600000,
-                    generation: Generation::none()
+                    generation: Generation::none(),
+                    shard: ShardIndex::unsharded()
                 }),
                 ("000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__00000000016B59D8-00000000016B5A51".parse().unwrap(), IndexLayerMetadata {
                     // serde_json should always parse this but this might be a double with jq for
                     // example.
                     file_size: 9007199254741001,
-                    generation: Generation::none()
+                    generation: Generation::none(),
+                    shard: ShardIndex::unsharded()
                 })
             ]),
             disk_consistent_lsn: "0/16960E8".parse::<Lsn>().unwrap(),
@@ -354,19 +370,21 @@ mod tests {
             layer_metadata: HashMap::from([
                 ("000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__0000000001696070-00000000016960E9".parse().unwrap(), IndexLayerMetadata {
                     file_size: 25600000,
-                    generation: Generation::none()
+                    generation: Generation::none(),
+                    shard: ShardIndex::unsharded()
                 }),
                 ("000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__00000000016B59D8-00000000016B5A51".parse().unwrap(), IndexLayerMetadata {
                     // serde_json should always parse this but this might be a double with jq for
                     // example.
                     file_size: 9007199254741001,
-                    generation: Generation::none()
+                    generation: Generation::none(),
+                    shard: ShardIndex::unsharded()
                 })
             ]),
             disk_consistent_lsn: "0/16960E8".parse::<Lsn>().unwrap(),
             metadata: TimelineMetadata::from_bytes(&[113,11,159,210,0,54,0,4,0,0,0,0,1,105,96,232,1,0,0,0,0,1,105,96,112,0,0,0,0,0,0,0,0,0,0,0,0,0,1,105,96,112,0,0,0,0,1,105,96,112,0,0,0,14,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0]).unwrap(),
             deleted_at: Some(chrono::NaiveDateTime::parse_from_str(
-                "2023-07-31T09:00:00.123000000", "%Y-%m-%dT%H:%M:%S.%f").unwrap())
+                "2023-07-31T09:00:00.123000000", "%Y-%m-%dT%H:%M:%S.%f").unwrap()),
         };
 
         let part = IndexPart::from_s3_bytes(example.as_bytes()).unwrap();

pageserver/src/tenant/remote_timeline_client/upload.rs

@@ -1,11 +1,11 @@
 //! Helper functions to upload files to remote storage with a RemoteStorage
 
 use anyhow::{bail, Context};
-use bytes::Bytes;
 use camino::Utf8Path;
 use fail::fail_point;
+use pageserver_api::shard::TenantShardId;
 use std::io::ErrorKind;
-use tokio::fs;
+use tokio::fs::{self, File};
 
 use super::Generation;
 use crate::{
@@ -24,7 +24,7 @@ use tracing::info;
 /// Serializes and uploads the given index part data to the remote storage.
 pub(super) async fn upload_index_part<'a>(
     storage: &'a GenericRemoteStorage,
-    tenant_id: &TenantId,
+    tenant_shard_id: &TenantShardId,
     timeline_id: &TimelineId,
     generation: Generation,
     index_part: &'a IndexPart,
@@ -40,13 +40,17 @@ pub(super) async fn upload_index_part<'a>(
         .to_s3_bytes()
         .context("serialize index part file into bytes")?;
     let index_part_size = index_part_bytes.len();
-    let index_part_bytes = tokio::io::BufReader::new(std::io::Cursor::new(index_part_bytes));
+    let index_part_bytes = bytes::Bytes::from(index_part_bytes);
 
-    let remote_path = remote_index_path(tenant_id, timeline_id, generation);
+    let remote_path = remote_index_path(tenant_shard_id, timeline_id, generation);
     storage
-        .upload_storage_object(Box::new(index_part_bytes), index_part_size, &remote_path)
+        .upload_storage_object(
+            futures::stream::once(futures::future::ready(Ok(index_part_bytes))),
+            index_part_size,
+            &remote_path,
+        )
         .await
-        .with_context(|| format!("upload index part for '{tenant_id} / {timeline_id}'"))
+        .with_context(|| format!("upload index part for '{tenant_shard_id} / {timeline_id}'"))
 }
 
 /// Attempts to upload given layer files.
@@ -100,8 +104,10 @@ pub(super) async fn upload_timeline_layer<'a>(
     let fs_size = usize::try_from(fs_size)
         .with_context(|| format!("convert {source_path:?} size {fs_size} usize"))?;
 
+    let reader = tokio_util::io::ReaderStream::with_capacity(source_file, super::BUFFER_SIZE);
+
     storage
-        .upload(source_file, fs_size, &storage_path, None)
+        .upload(reader, fs_size, &storage_path, None)
         .await
         .with_context(|| format!("upload layer from local path '{source_path}'"))?;
 
@@ -113,16 +119,16 @@ pub(crate) async fn upload_initdb_dir(
     storage: &GenericRemoteStorage,
     tenant_id: &TenantId,
     timeline_id: &TimelineId,
-    initdb_dir: Bytes,
+    initdb_tar_zst: File,
+    size: u64,
 ) -> anyhow::Result<()> {
     tracing::trace!("uploading initdb dir");
 
-    let size = initdb_dir.len();
-    let bytes = tokio::io::BufReader::new(std::io::Cursor::new(initdb_dir));
+    let file = tokio_util::io::ReaderStream::with_capacity(initdb_tar_zst, super::BUFFER_SIZE);
 
     let remote_path = remote_initdb_archive_path(tenant_id, timeline_id);
     storage
-        .upload_storage_object(bytes, size, &remote_path)
+        .upload_storage_object(file, size as usize, &remote_path)
         .await
         .with_context(|| format!("upload initdb dir for '{tenant_id} / {timeline_id}'"))
 }

pageserver/src/tenant/storage_layer.rs

@@ -2,9 +2,9 @@
 
 pub mod delta_layer;
 mod filename;
-mod image_layer;
+pub mod image_layer;
 mod inmemory_layer;
-mod layer;
+pub(crate) mod layer;
 mod layer_desc;
 
 use crate::context::{AccessStatsBehavior, RequestContext};
@@ -24,10 +24,7 @@ use tracing::warn;
 use utils::history_buffer::HistoryBufferWithDropCounter;
 use utils::rate_limit::RateLimit;
 
-use utils::{
-    id::{TenantId, TimelineId},
-    lsn::Lsn,
-};
+use utils::{id::TimelineId, lsn::Lsn};
 
 pub use delta_layer::{DeltaLayer, DeltaLayerWriter, ValueRef};
 pub use filename::{DeltaFileName, ImageFileName, LayerFileName};
@@ -304,12 +301,14 @@ pub trait AsLayerDesc {
 }
 
 pub mod tests {
+    use pageserver_api::shard::TenantShardId;
+
     use super::*;
 
     impl From<DeltaFileName> for PersistentLayerDesc {
         fn from(value: DeltaFileName) -> Self {
             PersistentLayerDesc::new_delta(
-                TenantId::from_array([0; 16]),
+                TenantShardId::from([0; 18]),
                 TimelineId::from_array([0; 16]),
                 value.key_range,
                 value.lsn_range,
@@ -321,7 +320,7 @@ pub mod tests {
     impl From<ImageFileName> for PersistentLayerDesc {
         fn from(value: ImageFileName) -> Self {
             PersistentLayerDesc::new_img(
-                TenantId::from_array([0; 16]),
+                TenantShardId::from([0; 18]),
                 TimelineId::from_array([0; 16]),
                 value.key_range,
                 value.lsn,

pageserver/src/tenant/storage_layer/delta_layer.rs

@@ -42,6 +42,7 @@ use crate::{DELTA_FILE_MAGIC, STORAGE_FORMAT_VERSION};
 use anyhow::{bail, ensure, Context, Result};
 use camino::{Utf8Path, Utf8PathBuf};
 use pageserver_api::models::LayerAccessKind;
+use pageserver_api::shard::TenantShardId;
 use rand::{distributions::Alphanumeric, Rng};
 use serde::{Deserialize, Serialize};
 use std::fs::File;
@@ -69,13 +70,13 @@ use super::{AsLayerDesc, LayerAccessStats, PersistentLayerDesc, ResidentLayer};
 #[derive(Debug, Serialize, Deserialize, PartialEq, Eq)]
 pub struct Summary {
     /// Magic value to identify this as a neon delta file. Always DELTA_FILE_MAGIC.
-    magic: u16,
-    format_version: u16,
+    pub magic: u16,
+    pub format_version: u16,
 
-    tenant_id: TenantId,
-    timeline_id: TimelineId,
-    key_range: Range<Key>,
-    lsn_range: Range<Lsn>,
+    pub tenant_id: TenantId,
+    pub timeline_id: TimelineId,
+    pub key_range: Range<Key>,
+    pub lsn_range: Range<Lsn>,
 
     /// Block number where the 'index' part of the file begins.
     pub index_start_blk: u32,
@@ -86,7 +87,7 @@ pub struct Summary {
 impl From<&DeltaLayer> for Summary {
     fn from(layer: &DeltaLayer) -> Self {
         Self::expected(
-            layer.desc.tenant_id,
+            layer.desc.tenant_shard_id.tenant_id,
             layer.desc.timeline_id,
             layer.desc.key_range.clone(),
             layer.desc.lsn_range.clone(),
@@ -248,7 +249,7 @@ impl DeltaLayer {
 
     fn temp_path_for(
         conf: &PageServerConf,
-        tenant_id: &TenantId,
+        tenant_shard_id: &TenantShardId,
         timeline_id: &TimelineId,
         key_start: Key,
         lsn_range: &Range<Lsn>,
@@ -259,14 +260,15 @@ impl DeltaLayer {
             .map(char::from)
             .collect();
 
-        conf.timeline_path(tenant_id, timeline_id).join(format!(
-            "{}-XXX__{:016X}-{:016X}.{}.{}",
-            key_start,
-            u64::from(lsn_range.start),
-            u64::from(lsn_range.end),
-            rand_string,
-            TEMP_FILE_SUFFIX,
-        ))
+        conf.timeline_path(tenant_shard_id, timeline_id)
+            .join(format!(
+                "{}-XXX__{:016X}-{:016X}.{}.{}",
+                key_start,
+                u64::from(lsn_range.start),
+                u64::from(lsn_range.end),
+                rand_string,
+                TEMP_FILE_SUFFIX,
+            ))
     }
 
     ///
@@ -318,10 +320,14 @@ impl DeltaLayer {
             .metadata()
             .context("get file metadata to determine size")?;
 
+        // TODO(sharding): we must get the TenantShardId from the path instead of reading the Summary.
+        // we should also validate the path against the Summary, as both should contain the same tenant, timeline, key, lsn.
+        let tenant_shard_id = TenantShardId::unsharded(summary.tenant_id);
+
         Ok(DeltaLayer {
             path: path.to_path_buf(),
             desc: PersistentLayerDesc::new_delta(
-                summary.tenant_id,
+                tenant_shard_id,
                 summary.timeline_id,
                 summary.key_range,
                 summary.lsn_range,
@@ -353,7 +359,7 @@ struct DeltaLayerWriterInner {
     conf: &'static PageServerConf,
     pub path: Utf8PathBuf,
     timeline_id: TimelineId,
-    tenant_id: TenantId,
+    tenant_shard_id: TenantShardId,
 
     key_start: Key,
     lsn_range: Range<Lsn>,
@@ -370,7 +376,7 @@ impl DeltaLayerWriterInner {
     async fn new(
         conf: &'static PageServerConf,
         timeline_id: TimelineId,
-        tenant_id: TenantId,
+        tenant_shard_id: TenantShardId,
         key_start: Key,
         lsn_range: Range<Lsn>,
     ) -> anyhow::Result<Self> {
@@ -380,7 +386,8 @@ impl DeltaLayerWriterInner {
         //
         // Note: This overwrites any existing file. There shouldn't be any.
         // FIXME: throw an error instead?
-        let path = DeltaLayer::temp_path_for(conf, &tenant_id, &timeline_id, key_start, &lsn_range);
+        let path =
+            DeltaLayer::temp_path_for(conf, &tenant_shard_id, &timeline_id, key_start, &lsn_range);
 
         let mut file = VirtualFile::create(&path).await?;
         // make room for the header block
@@ -395,7 +402,7 @@ impl DeltaLayerWriterInner {
             conf,
             path,
             timeline_id,
-            tenant_id,
+            tenant_shard_id,
             key_start,
             lsn_range,
             tree: tree_builder,
@@ -457,7 +464,7 @@ impl DeltaLayerWriterInner {
         let summary = Summary {
             magic: DELTA_FILE_MAGIC,
             format_version: STORAGE_FORMAT_VERSION,
-            tenant_id: self.tenant_id,
+            tenant_id: self.tenant_shard_id.tenant_id,
             timeline_id: self.timeline_id,
             key_range: self.key_start..key_end,
             lsn_range: self.lsn_range.clone(),
@@ -498,7 +505,7 @@ impl DeltaLayerWriterInner {
         // set inner.file here. The first read will have to re-open it.
 
         let desc = PersistentLayerDesc::new_delta(
-            self.tenant_id,
+            self.tenant_shard_id,
             self.timeline_id,
             self.key_start..key_end,
             self.lsn_range.clone(),
@@ -549,14 +556,20 @@ impl DeltaLayerWriter {
     pub async fn new(
         conf: &'static PageServerConf,
         timeline_id: TimelineId,
-        tenant_id: TenantId,
+        tenant_shard_id: TenantShardId,
         key_start: Key,
         lsn_range: Range<Lsn>,
     ) -> anyhow::Result<Self> {
         Ok(Self {
             inner: Some(
-                DeltaLayerWriterInner::new(conf, timeline_id, tenant_id, key_start, lsn_range)
-                    .await?,
+                DeltaLayerWriterInner::new(
+                    conf,
+                    timeline_id,
+                    tenant_shard_id,
+                    key_start,
+                    lsn_range,
+                )
+                .await?,
             ),
         })
     }
@@ -611,6 +624,61 @@ impl Drop for DeltaLayerWriter {
     }
 }
 
+#[derive(thiserror::Error, Debug)]
+pub enum RewriteSummaryError {
+    #[error("magic mismatch")]
+    MagicMismatch,
+    #[error(transparent)]
+    Other(#[from] anyhow::Error),
+}
+
+impl From<std::io::Error> for RewriteSummaryError {
+    fn from(e: std::io::Error) -> Self {
+        Self::Other(anyhow::anyhow!(e))
+    }
+}
+
+impl DeltaLayer {
+    pub async fn rewrite_summary<F>(
+        path: &Utf8Path,
+        rewrite: F,
+        ctx: &RequestContext,
+    ) -> Result<(), RewriteSummaryError>
+    where
+        F: Fn(Summary) -> Summary,
+    {
+        let file = VirtualFile::open_with_options(
+            path,
+            &*std::fs::OpenOptions::new().read(true).write(true),
+        )
+        .await
+        .with_context(|| format!("Failed to open file '{}'", path))?;
+        let file = FileBlockReader::new(file);
+        let summary_blk = file.read_blk(0, ctx).await?;
+        let actual_summary = Summary::des_prefix(summary_blk.as_ref()).context("deserialize")?;
+        let mut file = file.file;
+        if actual_summary.magic != DELTA_FILE_MAGIC {
+            return Err(RewriteSummaryError::MagicMismatch);
+        }
+
+        let new_summary = rewrite(actual_summary);
+
+        let mut buf = smallvec::SmallVec::<[u8; PAGE_SZ]>::new();
+        Summary::ser_into(&new_summary, &mut buf).context("serialize")?;
+        if buf.spilled() {
+            // The code in DeltaLayerWriterInner just warn!()s for this.
+            // It should probably error out as well.
+            return Err(RewriteSummaryError::Other(anyhow::anyhow!(
+                "Used more than one page size for summary buffer: {}",
+                buf.len()
+            )));
+        }
+        file.seek(SeekFrom::Start(0)).await?;
+        file.write_all(&buf).await?;
+        Ok(())
+    }
+}
+
 impl DeltaLayerInner {
     /// Returns nested result following Result<Result<_, OpErr>, Critical>:
     /// - inner has the success or transient failure

pageserver/src/tenant/storage_layer/image_layer.rs

@@ -41,6 +41,7 @@ use bytes::Bytes;
 use camino::{Utf8Path, Utf8PathBuf};
 use hex;
 use pageserver_api::models::LayerAccessKind;
+use pageserver_api::shard::TenantShardId;
 use rand::{distributions::Alphanumeric, Rng};
 use serde::{Deserialize, Serialize};
 use std::fs::File;
@@ -67,27 +68,27 @@ use super::{AsLayerDesc, Layer, PersistentLayerDesc, ResidentLayer};
 /// the 'index' starts at the block indicated by 'index_start_blk'
 ///
 #[derive(Debug, Serialize, Deserialize, PartialEq, Eq)]
-pub(super) struct Summary {
+pub struct Summary {
     /// Magic value to identify this as a neon image file. Always IMAGE_FILE_MAGIC.
-    magic: u16,
-    format_version: u16,
+    pub magic: u16,
+    pub format_version: u16,
 
-    tenant_id: TenantId,
-    timeline_id: TimelineId,
-    key_range: Range<Key>,
-    lsn: Lsn,
+    pub tenant_id: TenantId,
+    pub timeline_id: TimelineId,
+    pub key_range: Range<Key>,
+    pub lsn: Lsn,
 
     /// Block number where the 'index' part of the file begins.
-    index_start_blk: u32,
+    pub index_start_blk: u32,
     /// Block within the 'index', where the B-tree root page is stored
-    index_root_blk: u32,
+    pub index_root_blk: u32,
     // the 'values' part starts after the summary header, on block 1.
 }
 
 impl From<&ImageLayer> for Summary {
     fn from(layer: &ImageLayer) -> Self {
         Self::expected(
-            layer.desc.tenant_id,
+            layer.desc.tenant_shard_id.tenant_id,
             layer.desc.timeline_id,
             layer.desc.key_range.clone(),
             layer.lsn,
@@ -217,7 +218,7 @@ impl ImageLayer {
     fn temp_path_for(
         conf: &PageServerConf,
         timeline_id: TimelineId,
-        tenant_id: TenantId,
+        tenant_shard_id: TenantShardId,
         fname: &ImageFileName,
     ) -> Utf8PathBuf {
         let rand_string: String = rand::thread_rng()
@@ -226,7 +227,7 @@ impl ImageLayer {
             .map(char::from)
             .collect();
 
-        conf.timeline_path(&tenant_id, &timeline_id)
+        conf.timeline_path(&tenant_shard_id, &timeline_id)
             .join(format!("{fname}.{rand_string}.{TEMP_FILE_SUFFIX}"))
     }
 
@@ -276,10 +277,15 @@ impl ImageLayer {
         let metadata = file
             .metadata()
             .context("get file metadata to determine size")?;
+
+        // TODO(sharding): we should get TenantShardId from path.
+        // OR, not at all: any layer we load from disk should also get reconciled with remote IndexPart.
+        let tenant_shard_id = TenantShardId::unsharded(summary.tenant_id);
+
         Ok(ImageLayer {
             path: path.to_path_buf(),
             desc: PersistentLayerDesc::new_img(
-                summary.tenant_id,
+                tenant_shard_id,
                 summary.timeline_id,
                 summary.key_range,
                 summary.lsn,
@@ -296,6 +302,61 @@ impl ImageLayer {
     }
 }
 
+#[derive(thiserror::Error, Debug)]
+pub enum RewriteSummaryError {
+    #[error("magic mismatch")]
+    MagicMismatch,
+    #[error(transparent)]
+    Other(#[from] anyhow::Error),
+}
+
+impl From<std::io::Error> for RewriteSummaryError {
+    fn from(e: std::io::Error) -> Self {
+        Self::Other(anyhow::anyhow!(e))
+    }
+}
+
+impl ImageLayer {
+    pub async fn rewrite_summary<F>(
+        path: &Utf8Path,
+        rewrite: F,
+        ctx: &RequestContext,
+    ) -> Result<(), RewriteSummaryError>
+    where
+        F: Fn(Summary) -> Summary,
+    {
+        let file = VirtualFile::open_with_options(
+            path,
+            &*std::fs::OpenOptions::new().read(true).write(true),
+        )
+        .await
+        .with_context(|| format!("Failed to open file '{}'", path))?;
+        let file = FileBlockReader::new(file);
+        let summary_blk = file.read_blk(0, ctx).await?;
+        let actual_summary = Summary::des_prefix(summary_blk.as_ref()).context("deserialize")?;
+        let mut file = file.file;
+        if actual_summary.magic != IMAGE_FILE_MAGIC {
+            return Err(RewriteSummaryError::MagicMismatch);
+        }
+
+        let new_summary = rewrite(actual_summary);
+
+        let mut buf = smallvec::SmallVec::<[u8; PAGE_SZ]>::new();
+        Summary::ser_into(&new_summary, &mut buf).context("serialize")?;
+        if buf.spilled() {
+            // The code in ImageLayerWriterInner just warn!()s for this.
+            // It should probably error out as well.
+            return Err(RewriteSummaryError::Other(anyhow::anyhow!(
+                "Used more than one page size for summary buffer: {}",
+                buf.len()
+            )));
+        }
+        file.seek(SeekFrom::Start(0)).await?;
+        file.write_all(&buf).await?;
+        Ok(())
+    }
+}
+
 impl ImageLayerInner {
     /// Returns nested result following Result<Result<_, OpErr>, Critical>:
     /// - inner has the success or transient failure
@@ -400,7 +461,7 @@ struct ImageLayerWriterInner {
     conf: &'static PageServerConf,
     path: Utf8PathBuf,
     timeline_id: TimelineId,
-    tenant_id: TenantId,
+    tenant_shard_id: TenantShardId,
     key_range: Range<Key>,
     lsn: Lsn,
 
@@ -415,7 +476,7 @@ impl ImageLayerWriterInner {
     async fn new(
         conf: &'static PageServerConf,
         timeline_id: TimelineId,
-        tenant_id: TenantId,
+        tenant_shard_id: TenantShardId,
         key_range: &Range<Key>,
         lsn: Lsn,
     ) -> anyhow::Result<Self> {
@@ -424,7 +485,7 @@ impl ImageLayerWriterInner {
         let path = ImageLayer::temp_path_for(
             conf,
             timeline_id,
-            tenant_id,
+            tenant_shard_id,
             &ImageFileName {
                 key_range: key_range.clone(),
                 lsn,
@@ -448,7 +509,7 @@ impl ImageLayerWriterInner {
             conf,
             path,
             timeline_id,
-            tenant_id,
+            tenant_shard_id,
             key_range: key_range.clone(),
             lsn,
             tree: tree_builder,
@@ -495,7 +556,7 @@ impl ImageLayerWriterInner {
         let summary = Summary {
             magic: IMAGE_FILE_MAGIC,
             format_version: STORAGE_FORMAT_VERSION,
-            tenant_id: self.tenant_id,
+            tenant_id: self.tenant_shard_id.tenant_id,
             timeline_id: self.timeline_id,
             key_range: self.key_range.clone(),
             lsn: self.lsn,
@@ -521,7 +582,7 @@ impl ImageLayerWriterInner {
             .context("get metadata to determine file size")?;
 
         let desc = PersistentLayerDesc::new_img(
-            self.tenant_id,
+            self.tenant_shard_id,
             self.timeline_id,
             self.key_range.clone(),
             self.lsn,
@@ -577,13 +638,14 @@ impl ImageLayerWriter {
     pub async fn new(
         conf: &'static PageServerConf,
         timeline_id: TimelineId,
-        tenant_id: TenantId,
+        tenant_shard_id: TenantShardId,
         key_range: &Range<Key>,
         lsn: Lsn,
     ) -> anyhow::Result<ImageLayerWriter> {
         Ok(Self {
             inner: Some(
-                ImageLayerWriterInner::new(conf, timeline_id, tenant_id, key_range, lsn).await?,
+                ImageLayerWriterInner::new(conf, timeline_id, tenant_shard_id, key_range, lsn)
+                    .await?,
             ),
         })
     }

pageserver/src/tenant/storage_layer/inmemory_layer.rs

@@ -14,15 +14,11 @@ use crate::tenant::Timeline;
 use crate::walrecord;
 use anyhow::{ensure, Result};
 use pageserver_api::models::InMemoryLayerInfo;
+use pageserver_api::shard::TenantShardId;
 use std::collections::HashMap;
 use std::sync::{Arc, OnceLock};
 use tracing::*;
-use utils::{
-    bin_ser::BeSer,
-    id::{TenantId, TimelineId},
-    lsn::Lsn,
-    vec_map::VecMap,
-};
+use utils::{bin_ser::BeSer, id::TimelineId, lsn::Lsn, vec_map::VecMap};
 // avoid binding to Write (conflicts with std::io::Write)
 // while being able to use std::fmt::Write's methods
 use std::fmt::Write as _;
@@ -33,7 +29,7 @@ use super::{DeltaLayerWriter, ResidentLayer};
 
 pub struct InMemoryLayer {
     conf: &'static PageServerConf,
-    tenant_id: TenantId,
+    tenant_shard_id: TenantShardId,
     timeline_id: TimelineId,
 
     /// This layer contains all the changes from 'start_lsn'. The
@@ -226,17 +222,17 @@ impl InMemoryLayer {
     pub async fn create(
         conf: &'static PageServerConf,
         timeline_id: TimelineId,
-        tenant_id: TenantId,
+        tenant_shard_id: TenantShardId,
         start_lsn: Lsn,
     ) -> Result<InMemoryLayer> {
         trace!("initializing new empty InMemoryLayer for writing on timeline {timeline_id} at {start_lsn}");
 
-        let file = EphemeralFile::create(conf, tenant_id, timeline_id).await?;
+        let file = EphemeralFile::create(conf, tenant_shard_id, timeline_id).await?;
 
         Ok(InMemoryLayer {
             conf,
             timeline_id,
-            tenant_id,
+            tenant_shard_id,
             start_lsn,
             end_lsn: OnceLock::new(),
             inner: RwLock::new(InMemoryLayerInner {
@@ -335,7 +331,7 @@ impl InMemoryLayer {
         let mut delta_layer_writer = DeltaLayerWriter::new(
             self.conf,
             self.timeline_id,
-            self.tenant_id,
+            self.tenant_shard_id,
             Key::MIN,
             self.start_lsn..end_lsn,
         )

pageserver/src/tenant/storage_layer/layer.rs

@@ -3,6 +3,7 @@ use camino::{Utf8Path, Utf8PathBuf};
 use pageserver_api::models::{
     HistoricLayerInfo, LayerAccessKind, LayerResidenceEventReason, LayerResidenceStatus,
 };
+use pageserver_api::shard::ShardIndex;
 use std::ops::Range;
 use std::sync::atomic::{AtomicBool, AtomicUsize, Ordering};
 use std::sync::{Arc, Weak};
@@ -81,7 +82,7 @@ impl Layer {
         metadata: LayerFileMetadata,
     ) -> Self {
         let desc = PersistentLayerDesc::from_filename(
-            timeline.tenant_id,
+            timeline.tenant_shard_id,
             timeline.timeline_id,
             file_name,
             metadata.file_size(),
@@ -96,6 +97,7 @@ impl Layer {
             desc,
             None,
             metadata.generation,
+            metadata.shard,
         )));
 
         debug_assert!(owner.0.needs_download_blocking().unwrap().is_some());
@@ -111,7 +113,7 @@ impl Layer {
         metadata: LayerFileMetadata,
     ) -> ResidentLayer {
         let desc = PersistentLayerDesc::from_filename(
-            timeline.tenant_id,
+            timeline.tenant_shard_id,
             timeline.timeline_id,
             file_name,
             metadata.file_size(),
@@ -136,6 +138,7 @@ impl Layer {
                 desc,
                 Some(inner),
                 metadata.generation,
+                metadata.shard,
             )
         }));
 
@@ -179,6 +182,7 @@ impl Layer {
                 desc,
                 Some(inner),
                 timeline.generation,
+                timeline.get_shard_index(),
             )
         }));
 
@@ -218,14 +222,18 @@ impl Layer {
     ///
     /// [gc]: [`RemoteTimelineClient::schedule_gc_update`]
     /// [compaction]: [`RemoteTimelineClient::schedule_compaction_update`]
-    pub(crate) fn garbage_collect_on_drop(&self) {
-        self.0.garbage_collect_on_drop();
+    pub(crate) fn delete_on_drop(&self) {
+        self.0.delete_on_drop();
     }
 
     /// Return data needed to reconstruct given page at LSN.
     ///
     /// It is up to the caller to collect more data from the previous layer and
     /// perform WAL redo, if necessary.
+    ///
+    /// # Cancellation-Safety
+    ///
+    /// This method is cancellation-safe.
     pub(crate) async fn get_value_reconstruct_data(
         &self,
         key: Key,
@@ -322,6 +330,24 @@ impl Layer {
 
         Ok(())
     }
+
+    /// Waits until this layer has been dropped (and if needed, local file deletion and remote
+    /// deletion scheduling has completed).
+    ///
+    /// Does not start local deletion, use [`Self::delete_on_drop`] for that
+    /// separatedly.
+    #[cfg(feature = "testing")]
+    pub(crate) fn wait_drop(&self) -> impl std::future::Future<Output = ()> + 'static {
+        let mut rx = self.0.status.subscribe();
+
+        async move {
+            loop {
+                if let Err(tokio::sync::broadcast::error::RecvError::Closed) = rx.recv().await {
+                    break;
+                }
+            }
+        }
+    }
 }
 
 /// The download-ness ([`DownloadedLayer`]) can be either resident or wanted evicted.
@@ -397,8 +423,8 @@ struct LayerInner {
     /// Initialization and deinitialization are done while holding a permit.
     inner: heavier_once_cell::OnceCell<ResidentOrWantedEvicted>,
 
-    /// Do we want to garbage collect this when `LayerInner` is dropped
-    wanted_garbage_collected: AtomicBool,
+    /// Do we want to delete locally and remotely this when `LayerInner` is dropped
+    wanted_deleted: AtomicBool,
 
     /// Do we want to evict this layer as soon as possible? After being set to `true`, all accesses
     /// will try to downgrade [`ResidentOrWantedEvicted`], which will eventually trigger
@@ -412,10 +438,6 @@ struct LayerInner {
     version: AtomicUsize,
 
     /// Allow subscribing to when the layer actually gets evicted.
-    ///
-    /// If in future we need to implement "wait until layer instances are gone and done", carrying
-    /// this over to the gc spawn_blocking from LayerInner::drop will do the trick, and adding a
-    /// method for "wait_gc" which will wait to this being closed.
     status: tokio::sync::broadcast::Sender<Status>,
 
     /// Counter for exponential backoff with the download
@@ -426,6 +448,15 @@ struct LayerInner {
     /// For loaded layers (resident or evicted) this comes from [`LayerFileMetadata::generation`],
     /// for created layers from [`Timeline::generation`].
     generation: Generation,
+
+    /// The shard of this Layer.
+    ///
+    /// For layers created in this process, this will always be the [`ShardIndex`] of the
+    /// current `ShardIdentity`` (TODO: add link once it's introduced).
+    ///
+    /// For loaded layers, this may be some other value if the tenant has undergone
+    /// a shard split since the layer was originally written.
+    shard: ShardIndex,
 }
 
 impl std::fmt::Display for LayerInner {
@@ -448,24 +479,28 @@ enum Status {
 
 impl Drop for LayerInner {
     fn drop(&mut self) {
-        if !*self.wanted_garbage_collected.get_mut() {
+        if !*self.wanted_deleted.get_mut() {
             // should we try to evict if the last wish was for eviction?
             // feels like there's some hazard of overcrowding near shutdown near by, but we don't
             // run drops during shutdown (yet)
             return;
         }
 
-        let span = tracing::info_span!(parent: None, "layer_gc", tenant_id = %self.layer_desc().tenant_id, timeline_id = %self.layer_desc().timeline_id);
+        let span = tracing::info_span!(parent: None, "layer_delete", tenant_id = %self.layer_desc().tenant_shard_id.tenant_id, shard_id=%self.layer_desc().tenant_shard_id.shard_slug(), timeline_id = %self.layer_desc().timeline_id);
 
         let path = std::mem::take(&mut self.path);
         let file_name = self.layer_desc().filename();
-        let gen = self.generation;
         let file_size = self.layer_desc().file_size;
         let timeline = self.timeline.clone();
+        let meta = self.metadata();
+        let status = self.status.clone();
 
         crate::task_mgr::BACKGROUND_RUNTIME.spawn_blocking(move || {
             let _g = span.entered();
 
+            // carry this until we are finished for [`Layer::wait_drop`] support
+            let _status = status;
+
             let removed = match std::fs::remove_file(path) {
                 Ok(()) => true,
                 Err(e) if e.kind() == std::io::ErrorKind::NotFound => {
@@ -478,8 +513,8 @@ impl Drop for LayerInner {
                     false
                 }
                 Err(e) => {
-                    tracing::error!("failed to remove garbage collected layer: {e}");
-                    LAYER_IMPL_METRICS.inc_gc_removes_failed();
+                    tracing::error!("failed to remove wanted deleted layer: {e}");
+                    LAYER_IMPL_METRICS.inc_delete_removes_failed();
                     false
                 }
             };
@@ -489,7 +524,7 @@ impl Drop for LayerInner {
                     timeline.metrics.resident_physical_size_sub(file_size);
                 }
                 if let Some(remote_client) = timeline.remote_client.as_ref() {
-                    let res = remote_client.schedule_deletion_of_unlinked(vec![(file_name, gen)]);
+                    let res = remote_client.schedule_deletion_of_unlinked(vec![(file_name, meta)]);
 
                     if let Err(e) = res {
                         // test_timeline_deletion_with_files_stuck_in_upload_queue is good at
@@ -501,15 +536,15 @@ impl Drop for LayerInner {
                         } else {
                             tracing::warn!("scheduling deletion on drop failed: {e:#}");
                         }
-                        LAYER_IMPL_METRICS.inc_gcs_failed(GcFailed::DeleteSchedulingFailed);
+                        LAYER_IMPL_METRICS.inc_deletes_failed(DeleteFailed::DeleteSchedulingFailed);
                     } else {
-                        LAYER_IMPL_METRICS.inc_completed_gcs();
+                        LAYER_IMPL_METRICS.inc_completed_deletes();
                     }
                 }
             } else {
                 // no need to nag that timeline is gone: under normal situation on
                 // task_mgr::remove_tenant_from_memory the timeline is gone before we get dropped.
-                LAYER_IMPL_METRICS.inc_gcs_failed(GcFailed::TimelineGone);
+                LAYER_IMPL_METRICS.inc_deletes_failed(DeleteFailed::TimelineGone);
             }
         });
     }
@@ -523,9 +558,10 @@ impl LayerInner {
         desc: PersistentLayerDesc,
         downloaded: Option<Arc<DownloadedLayer>>,
         generation: Generation,
+        shard: ShardIndex,
     ) -> Self {
         let path = conf
-            .timeline_path(&timeline.tenant_id, &timeline.timeline_id)
+            .timeline_path(&timeline.tenant_shard_id, &timeline.timeline_id)
             .join(desc.filename().to_string());
 
         let (inner, version) = if let Some(inner) = downloaded {
@@ -543,26 +579,24 @@ impl LayerInner {
             timeline: Arc::downgrade(timeline),
             have_remote_client: timeline.remote_client.is_some(),
             access_stats,
-            wanted_garbage_collected: AtomicBool::new(false),
+            wanted_deleted: AtomicBool::new(false),
             wanted_evicted: AtomicBool::new(false),
             inner,
             version: AtomicUsize::new(version),
             status: tokio::sync::broadcast::channel(1).0,
             consecutive_failures: AtomicUsize::new(0),
             generation,
+            shard,
         }
     }
 
-    fn garbage_collect_on_drop(&self) {
-        let res = self.wanted_garbage_collected.compare_exchange(
-            false,
-            true,
-            Ordering::Release,
-            Ordering::Relaxed,
-        );
+    fn delete_on_drop(&self) {
+        let res =
+            self.wanted_deleted
+                .compare_exchange(false, true, Ordering::Release, Ordering::Relaxed);
 
         if res.is_ok() {
-            LAYER_IMPL_METRICS.inc_started_gcs();
+            LAYER_IMPL_METRICS.inc_started_deletes();
         }
     }
 
@@ -630,6 +664,10 @@ impl LayerInner {
                 // disable any scheduled but not yet running eviction deletions for this
                 let next_version = 1 + self.version.fetch_add(1, Ordering::Relaxed);
 
+                // count cancellations, which currently remain largely unexpected
+                let init_cancelled =
+                    scopeguard::guard((), |_| LAYER_IMPL_METRICS.inc_init_cancelled());
+
                 // no need to make the evict_and_wait wait for the actual download to complete
                 drop(self.status.send(Status::Downloaded));
 
@@ -638,6 +676,8 @@ impl LayerInner {
                     .upgrade()
                     .ok_or_else(|| DownloadError::TimelineShutdown)?;
 
+                // FIXME: grab a gate
+
                 let can_ever_evict = timeline.remote_client.as_ref().is_some();
 
                 // check if we really need to be downloaded; could have been already downloaded by a
@@ -698,6 +738,8 @@ impl LayerInner {
                     tracing::info!(waiters, "completing the on-demand download for other tasks");
                 }
 
+                scopeguard::ScopeGuard::into_inner(init_cancelled);
+
                 Ok((ResidentOrWantedEvicted::Resident(res), permit))
             };
 
@@ -795,7 +837,7 @@ impl LayerInner {
         crate::task_mgr::spawn(
             &tokio::runtime::Handle::current(),
             crate::task_mgr::TaskKind::RemoteDownloadTask,
-            Some(self.desc.tenant_id),
+            Some(self.desc.tenant_shard_id.tenant_id),
             Some(self.desc.timeline_id),
             &task_name,
             false,
@@ -826,14 +868,13 @@ impl LayerInner {
                     match res {
                         (Ok(()), _) => {
                             // our caller is cancellation safe so this is fine; if someone
-                            // else requests the layer, they'll find it already downloaded
-                            // or redownload.
+                            // else requests the layer, they'll find it already downloaded.
                             //
-                            // however, could be that we should consider marking the layer
-                            // for eviction? alas, cannot: because only DownloadedLayer
-                            // will handle that.
-                            tracing::info!("layer file download completed after requester had cancelled");
-                            LAYER_IMPL_METRICS.inc_download_completed_without_requester();
+                            // See counter [`LayerImplMetrics::inc_init_needed_no_download`]
+                            //
+                            // FIXME(#6028): however, could be that we should consider marking the
+                            // layer for eviction? alas, cannot: because only DownloadedLayer will
+                            // handle that.
                         },
                         (Err(e), _) => {
                             // our caller is cancellation safe, but we might be racing with
@@ -953,14 +994,17 @@ impl LayerInner {
 
     /// `DownloadedLayer` is being dropped, so it calls this method.
     fn on_downloaded_layer_drop(self: Arc<LayerInner>, version: usize) {
-        let gc = self.wanted_garbage_collected.load(Ordering::Acquire);
+        let delete = self.wanted_deleted.load(Ordering::Acquire);
         let evict = self.wanted_evicted.load(Ordering::Acquire);
         let can_evict = self.have_remote_client;
 
-        if gc {
-            // do nothing now, only in LayerInner::drop
+        if delete {
+            // do nothing now, only in LayerInner::drop -- this was originally implemented because
+            // we could had already scheduled the deletion at the time.
+            //
+            // FIXME: this is not true anymore, we can safely evict wanted deleted files.
         } else if can_evict && evict {
-            let span = tracing::info_span!(parent: None, "layer_evict", tenant_id = %self.desc.tenant_id, timeline_id = %self.desc.timeline_id, layer=%self, %version);
+            let span = tracing::info_span!(parent: None, "layer_evict", tenant_id = %self.desc.tenant_shard_id.tenant_id, shard_id = %self.desc.tenant_shard_id.shard_slug(), timeline_id = %self.desc.timeline_id, layer=%self, %version);
 
             // downgrade for queueing, in case there's a tear down already ongoing we should not
             // hold it alive.
@@ -973,7 +1017,7 @@ impl LayerInner {
             crate::task_mgr::BACKGROUND_RUNTIME.spawn_blocking(move || {
                 let _g = span.entered();
 
-                // if LayerInner is already dropped here, do nothing because the garbage collection
+                // if LayerInner is already dropped here, do nothing because the delete on drop
                 // has already ran while we were in queue
                 let Some(this) = this.upgrade() else {
                     LAYER_IMPL_METRICS.inc_eviction_cancelled(EvictionCancelled::LayerGone);
@@ -1077,7 +1121,7 @@ impl LayerInner {
     }
 
     fn metadata(&self) -> LayerFileMetadata {
-        LayerFileMetadata::new(self.desc.file_size, self.generation)
+        LayerFileMetadata::new(self.desc.file_size, self.generation, self.shard)
     }
 }
 
@@ -1192,7 +1236,7 @@ impl DownloadedLayer {
 
             let res = if owner.desc.is_delta {
                 let summary = Some(delta_layer::Summary::expected(
-                    owner.desc.tenant_id,
+                    owner.desc.tenant_shard_id.tenant_id,
                     owner.desc.timeline_id,
                     owner.desc.key_range.clone(),
                     owner.desc.lsn_range.clone(),
@@ -1203,7 +1247,7 @@ impl DownloadedLayer {
             } else {
                 let lsn = owner.desc.image_layer_lsn();
                 let summary = Some(image_layer::Summary::expected(
-                    owner.desc.tenant_id,
+                    owner.desc.tenant_shard_id.tenant_id,
                     owner.desc.timeline_id,
                     owner.desc.key_range.clone(),
                     lsn,
@@ -1364,35 +1408,37 @@ impl From<ResidentLayer> for Layer {
     }
 }
 
-use metrics::{IntCounter, IntCounterVec};
+use metrics::IntCounter;
 
-struct LayerImplMetrics {
+pub(crate) struct LayerImplMetrics {
     started_evictions: IntCounter,
     completed_evictions: IntCounter,
-    cancelled_evictions: IntCounterVec,
+    cancelled_evictions: enum_map::EnumMap<EvictionCancelled, IntCounter>,
 
-    started_gcs: IntCounter,
-    completed_gcs: IntCounter,
-    failed_gcs: IntCounterVec,
+    started_deletes: IntCounter,
+    completed_deletes: IntCounter,
+    failed_deletes: enum_map::EnumMap<DeleteFailed, IntCounter>,
 
-    rare_counters: IntCounterVec,
+    rare_counters: enum_map::EnumMap<RareEvent, IntCounter>,
+    inits_cancelled: metrics::core::GenericCounter<metrics::core::AtomicU64>,
 }
 
 impl Default for LayerImplMetrics {
     fn default() -> Self {
-        let evictions = metrics::register_int_counter_vec!(
-            "pageserver_layer_evictions_count",
-            "Evictions started and completed in the Layer implementation",
-            &["state"]
+        use enum_map::Enum;
+
+        // reminder: these will be pageserver_layer_* with "_total" suffix
+
+        let started_evictions = metrics::register_int_counter!(
+            "pageserver_layer_started_evictions",
+            "Evictions started in the Layer implementation"
+        )
+        .unwrap();
+        let completed_evictions = metrics::register_int_counter!(
+            "pageserver_layer_completed_evictions",
+            "Evictions completed in the Layer implementation"
         )
         .unwrap();
-
-        let started_evictions = evictions
-            .get_metric_with_label_values(&["started"])
-            .unwrap();
-        let completed_evictions = evictions
-            .get_metric_with_label_values(&["completed"])
-            .unwrap();
 
         let cancelled_evictions = metrics::register_int_counter_vec!(
             "pageserver_layer_cancelled_evictions_count",
@@ -1401,23 +1447,36 @@ impl Default for LayerImplMetrics {
         )
         .unwrap();
 
-        let gcs = metrics::register_int_counter_vec!(
-            "pageserver_layer_gcs_count",
-            "Garbage collections started and completed in the Layer implementation",
-            &["state"]
+        let cancelled_evictions = enum_map::EnumMap::from_array(std::array::from_fn(|i| {
+            let reason = EvictionCancelled::from_usize(i);
+            let s = reason.as_str();
+            cancelled_evictions.with_label_values(&[s])
+        }));
+
+        let started_deletes = metrics::register_int_counter!(
+            "pageserver_layer_started_deletes",
+            "Deletions on drop pending in the Layer implementation"
+        )
+        .unwrap();
+        let completed_deletes = metrics::register_int_counter!(
+            "pageserver_layer_completed_deletes",
+            "Deletions on drop completed in the Layer implementation"
         )
         .unwrap();
 
-        let started_gcs = gcs.get_metric_with_label_values(&["pending"]).unwrap();
-        let completed_gcs = gcs.get_metric_with_label_values(&["completed"]).unwrap();
-
-        let failed_gcs = metrics::register_int_counter_vec!(
-            "pageserver_layer_failed_gcs_count",
-            "Different reasons for garbage collections to have failed",
+        let failed_deletes = metrics::register_int_counter_vec!(
+            "pageserver_layer_failed_deletes_count",
+            "Different reasons for deletions on drop to have failed",
             &["reason"]
         )
         .unwrap();
 
+        let failed_deletes = enum_map::EnumMap::from_array(std::array::from_fn(|i| {
+            let reason = DeleteFailed::from_usize(i);
+            let s = reason.as_str();
+            failed_deletes.with_label_values(&[s])
+        }));
+
         let rare_counters = metrics::register_int_counter_vec!(
             "pageserver_layer_assumed_rare_count",
             "Times unexpected or assumed rare event happened",
@@ -1425,16 +1484,29 @@ impl Default for LayerImplMetrics {
         )
         .unwrap();
 
+        let rare_counters = enum_map::EnumMap::from_array(std::array::from_fn(|i| {
+            let event = RareEvent::from_usize(i);
+            let s = event.as_str();
+            rare_counters.with_label_values(&[s])
+        }));
+
+        let inits_cancelled = metrics::register_int_counter!(
+            "pageserver_layer_inits_cancelled_count",
+            "Times Layer initialization was cancelled",
+        )
+        .unwrap();
+
         Self {
             started_evictions,
             completed_evictions,
             cancelled_evictions,
 
-            started_gcs,
-            completed_gcs,
-            failed_gcs,
+            started_deletes,
+            completed_deletes,
+            failed_deletes,
 
             rare_counters,
+            inits_cancelled,
         }
     }
 }
@@ -1447,57 +1519,33 @@ impl LayerImplMetrics {
         self.completed_evictions.inc();
     }
     fn inc_eviction_cancelled(&self, reason: EvictionCancelled) {
-        self.cancelled_evictions
-            .get_metric_with_label_values(&[reason.as_str()])
-            .unwrap()
-            .inc()
+        self.cancelled_evictions[reason].inc()
     }
 
-    fn inc_started_gcs(&self) {
-        self.started_gcs.inc();
+    fn inc_started_deletes(&self) {
+        self.started_deletes.inc();
     }
-    fn inc_completed_gcs(&self) {
-        self.completed_gcs.inc();
+    fn inc_completed_deletes(&self) {
+        self.completed_deletes.inc();
     }
-    fn inc_gcs_failed(&self, reason: GcFailed) {
-        self.failed_gcs
-            .get_metric_with_label_values(&[reason.as_str()])
-            .unwrap()
-            .inc();
+    fn inc_deletes_failed(&self, reason: DeleteFailed) {
+        self.failed_deletes[reason].inc();
     }
 
-    /// Counted separatedly from failed gcs because we will complete the gc attempt regardless of
-    /// failure to delete local file.
-    fn inc_gc_removes_failed(&self) {
-        self.rare_counters
-            .get_metric_with_label_values(&["gc_remove_failed"])
-            .unwrap()
-            .inc();
+    /// Counted separatedly from failed layer deletes because we will complete the layer deletion
+    /// attempt regardless of failure to delete local file.
+    fn inc_delete_removes_failed(&self) {
+        self.rare_counters[RareEvent::RemoveOnDropFailed].inc();
     }
 
-    /// Expected rare because requires a race with `evict_blocking` and
-    /// `get_or_maybe_download`.
+    /// Expected rare because requires a race with `evict_blocking` and `get_or_maybe_download`.
     fn inc_retried_get_or_maybe_download(&self) {
-        self.rare_counters
-            .get_metric_with_label_values(&["retried_gomd"])
-            .unwrap()
-            .inc();
+        self.rare_counters[RareEvent::RetriedGetOrMaybeDownload].inc();
     }
 
-    /// Expected rare because cancellations are unexpected
-    fn inc_download_completed_without_requester(&self) {
-        self.rare_counters
-            .get_metric_with_label_values(&["download_completed_without"])
-            .unwrap()
-            .inc();
-    }
-
-    /// Expected rare because cancellations are unexpected
+    /// Expected rare because cancellations are unexpected, and failures are unexpected
     fn inc_download_failed_without_requester(&self) {
-        self.rare_counters
-            .get_metric_with_label_values(&["download_failed_without"])
-            .unwrap()
-            .inc();
+        self.rare_counters[RareEvent::DownloadFailedWithoutRequester].inc();
     }
 
     /// The Weak in ResidentOrWantedEvicted::WantedEvicted was successfully upgraded.
@@ -1505,37 +1553,30 @@ impl LayerImplMetrics {
     /// If this counter is always zero, we should replace ResidentOrWantedEvicted type with an
     /// Option.
     fn inc_raced_wanted_evicted_accesses(&self) {
-        self.rare_counters
-            .get_metric_with_label_values(&["raced_wanted_evicted"])
-            .unwrap()
-            .inc();
+        self.rare_counters[RareEvent::UpgradedWantedEvicted].inc();
     }
 
-    /// These are only expected for [`Self::inc_download_completed_without_requester`] amount when
+    /// These are only expected for [`Self::inc_init_cancelled`] amount when
     /// running with remote storage.
     fn inc_init_needed_no_download(&self) {
-        self.rare_counters
-            .get_metric_with_label_values(&["init_needed_no_download"])
-            .unwrap()
-            .inc();
+        self.rare_counters[RareEvent::InitWithoutDownload].inc();
     }
 
     /// Expected rare because all layer files should be readable and good
     fn inc_permanent_loading_failures(&self) {
-        self.rare_counters
-            .get_metric_with_label_values(&["permanent_loading_failure"])
-            .unwrap()
-            .inc();
+        self.rare_counters[RareEvent::PermanentLoadingFailure].inc();
     }
 
     fn inc_broadcast_lagged(&self) {
-        self.rare_counters
-            .get_metric_with_label_values(&["broadcast_lagged"])
-            .unwrap()
-            .inc();
+        self.rare_counters[RareEvent::EvictAndWaitLagged].inc();
+    }
+
+    fn inc_init_cancelled(&self) {
+        self.inits_cancelled.inc()
     }
 }
 
+#[derive(enum_map::Enum)]
 enum EvictionCancelled {
     LayerGone,
     TimelineGone,
@@ -1564,19 +1605,47 @@ impl EvictionCancelled {
     }
 }
 
-enum GcFailed {
+#[derive(enum_map::Enum)]
+enum DeleteFailed {
     TimelineGone,
     DeleteSchedulingFailed,
 }
 
-impl GcFailed {
+impl DeleteFailed {
     fn as_str(&self) -> &'static str {
         match self {
-            GcFailed::TimelineGone => "timeline_gone",
-            GcFailed::DeleteSchedulingFailed => "delete_scheduling_failed",
+            DeleteFailed::TimelineGone => "timeline_gone",
+            DeleteFailed::DeleteSchedulingFailed => "delete_scheduling_failed",
         }
     }
 }
 
-static LAYER_IMPL_METRICS: once_cell::sync::Lazy<LayerImplMetrics> =
+#[derive(enum_map::Enum)]
+enum RareEvent {
+    RemoveOnDropFailed,
+    RetriedGetOrMaybeDownload,
+    DownloadFailedWithoutRequester,
+    UpgradedWantedEvicted,
+    InitWithoutDownload,
+    PermanentLoadingFailure,
+    EvictAndWaitLagged,
+}
+
+impl RareEvent {
+    fn as_str(&self) -> &'static str {
+        use RareEvent::*;
+
+        match self {
+            RemoveOnDropFailed => "remove_on_drop_failed",
+            RetriedGetOrMaybeDownload => "retried_gomd",
+            DownloadFailedWithoutRequester => "download_failed_without",
+            UpgradedWantedEvicted => "raced_wanted_evicted",
+            InitWithoutDownload => "init_needed_no_download",
+            PermanentLoadingFailure => "permanent_loading_failure",
+            EvictAndWaitLagged => "broadcast_lagged",
+        }
+    }
+}
+
+pub(crate) static LAYER_IMPL_METRICS: once_cell::sync::Lazy<LayerImplMetrics> =
     once_cell::sync::Lazy::new(LayerImplMetrics::default);

pageserver/src/tenant/storage_layer/layer_desc.rs

@@ -1,9 +1,7 @@
 use core::fmt::Display;
+use pageserver_api::shard::TenantShardId;
 use std::ops::Range;
-use utils::{
-    id::{TenantId, TimelineId},
-    lsn::Lsn,
-};
+use utils::{id::TimelineId, lsn::Lsn};
 
 use crate::repository::Key;
 
@@ -11,12 +9,15 @@ use super::{DeltaFileName, ImageFileName, LayerFileName};
 
 use serde::{Deserialize, Serialize};
 
+#[cfg(test)]
+use utils::id::TenantId;
+
 /// A unique identifier of a persistent layer. This is different from `LayerDescriptor`, which is only used in the
 /// benchmarks. This struct contains all necessary information to find the image / delta layer. It also provides
 /// a unified way to generate layer information like file name.
 #[derive(Debug, PartialEq, Eq, Clone, Serialize, Deserialize)]
 pub struct PersistentLayerDesc {
-    pub tenant_id: TenantId,
+    pub tenant_shard_id: TenantShardId,
     pub timeline_id: TimelineId,
     /// Range of keys that this layer covers
     pub key_range: Range<Key>,
@@ -56,7 +57,7 @@ impl PersistentLayerDesc {
     #[cfg(test)]
     pub fn new_test(key_range: Range<Key>) -> Self {
         Self {
-            tenant_id: TenantId::generate(),
+            tenant_shard_id: TenantShardId::unsharded(TenantId::generate()),
             timeline_id: TimelineId::generate(),
             key_range,
             lsn_range: Lsn(0)..Lsn(1),
@@ -66,14 +67,14 @@ impl PersistentLayerDesc {
     }
 
     pub fn new_img(
-        tenant_id: TenantId,
+        tenant_shard_id: TenantShardId,
         timeline_id: TimelineId,
         key_range: Range<Key>,
         lsn: Lsn,
         file_size: u64,
     ) -> Self {
         Self {
-            tenant_id,
+            tenant_shard_id,
             timeline_id,
             key_range,
             lsn_range: Self::image_layer_lsn_range(lsn),
@@ -83,14 +84,14 @@ impl PersistentLayerDesc {
     }
 
     pub fn new_delta(
-        tenant_id: TenantId,
+        tenant_shard_id: TenantShardId,
         timeline_id: TimelineId,
         key_range: Range<Key>,
         lsn_range: Range<Lsn>,
         file_size: u64,
     ) -> Self {
         Self {
-            tenant_id,
+            tenant_shard_id,
             timeline_id,
             key_range,
             lsn_range,
@@ -100,18 +101,22 @@ impl PersistentLayerDesc {
     }
 
     pub fn from_filename(
-        tenant_id: TenantId,
+        tenant_shard_id: TenantShardId,
         timeline_id: TimelineId,
         filename: LayerFileName,
         file_size: u64,
     ) -> Self {
         match filename {
             LayerFileName::Image(i) => {
-                Self::new_img(tenant_id, timeline_id, i.key_range, i.lsn, file_size)
-            }
-            LayerFileName::Delta(d) => {
-                Self::new_delta(tenant_id, timeline_id, d.key_range, d.lsn_range, file_size)
+                Self::new_img(tenant_shard_id, timeline_id, i.key_range, i.lsn, file_size)
             }
+            LayerFileName::Delta(d) => Self::new_delta(
+                tenant_shard_id,
+                timeline_id,
+                d.key_range,
+                d.lsn_range,
+                file_size,
+            ),
         }
     }
 
@@ -172,10 +177,6 @@ impl PersistentLayerDesc {
         self.timeline_id
     }
 
-    pub fn get_tenant_id(&self) -> TenantId {
-        self.tenant_id
-    }
-
     /// Does this layer only contain some data for the key-range (incremental),
     /// or does it contain a version of every page? This is important to know
     /// for garbage collecting old layers: an incremental layer depends on
@@ -192,7 +193,7 @@ impl PersistentLayerDesc {
         if self.is_delta {
             println!(
                 "----- delta layer for ten {} tli {} keys {}-{} lsn {}-{} is_incremental {} size {} ----",
-                self.tenant_id,
+                self.tenant_shard_id,
                 self.timeline_id,
                 self.key_range.start,
                 self.key_range.end,
@@ -204,7 +205,7 @@ impl PersistentLayerDesc {
         } else {
             println!(
                 "----- image layer for ten {} tli {} key {}-{} at {} is_incremental {} size {} ----",
-                self.tenant_id,
+                self.tenant_shard_id,
                 self.timeline_id,
                 self.key_range.start,
                 self.key_range.end,

pageserver/src/tenant/tasks.rs

@@ -44,6 +44,7 @@ pub(crate) enum BackgroundLoopKind {
     Eviction,
     ConsumptionMetricsCollectMetrics,
     ConsumptionMetricsSyntheticSizeWorker,
+    InitialLogicalSizeCalculation,
 }
 
 impl BackgroundLoopKind {
@@ -86,7 +87,7 @@ pub fn start_background_loops(
     tenant: &Arc<Tenant>,
     background_jobs_can_start: Option<&completion::Barrier>,
 ) {
-    let tenant_id = tenant.tenant_id;
+    let tenant_id = tenant.tenant_shard_id.tenant_id;
     task_mgr::spawn(
         BACKGROUND_RUNTIME.handle(),
         TaskKind::Compaction,

pageserver/src/tenant/timeline/delete.rs

@@ -4,13 +4,10 @@ use std::{
 };
 
 use anyhow::Context;
-use pageserver_api::models::TimelineState;
+use pageserver_api::{models::TimelineState, shard::TenantShardId};
 use tokio::sync::OwnedMutexGuard;
 use tracing::{debug, error, info, instrument, warn, Instrument, Span};
-use utils::{
-    crashsafe, fs_ext,
-    id::{TenantId, TimelineId},
-};
+use utils::{crashsafe, fs_ext, id::TimelineId};
 
 use crate::{
     config::PageServerConf,
@@ -24,7 +21,6 @@ use crate::{
         },
         CreateTimelineCause, DeleteTimelineError, Tenant,
     },
-    InitializationOrder,
 };
 
 use super::{Timeline, TimelineResources};
@@ -47,7 +43,7 @@ async fn stop_tasks(timeline: &Timeline) -> Result<(), DeleteTimelineError> {
     // Shut down the layer flush task before the remote client, as one depends on the other
     task_mgr::shutdown_tasks(
         Some(TaskKind::LayerFlushTask),
-        Some(timeline.tenant_id),
+        Some(timeline.tenant_shard_id.tenant_id),
         Some(timeline.timeline_id),
     )
     .await;
@@ -73,7 +69,12 @@ async fn stop_tasks(timeline: &Timeline) -> Result<(), DeleteTimelineError> {
     // NB: This and other delete_timeline calls do not run as a task_mgr task,
     //     so, they are not affected by this shutdown_tasks() call.
     info!("waiting for timeline tasks to shutdown");
-    task_mgr::shutdown_tasks(None, Some(timeline.tenant_id), Some(timeline.timeline_id)).await;
+    task_mgr::shutdown_tasks(
+        None,
+        Some(timeline.tenant_shard_id.tenant_id),
+        Some(timeline.timeline_id),
+    )
+    .await;
 
     fail::fail_point!("timeline-delete-before-index-deleted-at", |_| {
         Err(anyhow::anyhow!(
@@ -110,11 +111,11 @@ async fn set_deleted_in_remote_index(timeline: &Timeline) -> Result<(), DeleteTi
     Ok(())
 }
 
-/// Grab the layer_removal_cs lock, and actually perform the deletion.
+/// Grab the compaction and gc locks, and actually perform the deletion.
 ///
-/// This lock prevents prevents GC or compaction from running at the same time.
-/// The GC task doesn't register itself with the timeline it's operating on,
-/// so it might still be running even though we called `shutdown_tasks`.
+/// The locks prevent GC or compaction from running at the same time. The background tasks do not
+/// register themselves with the timeline it's operating on, so it might still be running even
+/// though we called `shutdown_tasks`.
 ///
 /// Note that there are still other race conditions between
 /// GC, compaction and timeline deletion. See
@@ -122,19 +123,24 @@ async fn set_deleted_in_remote_index(timeline: &Timeline) -> Result<(), DeleteTi
 ///
 /// No timeout here, GC & Compaction should be responsive to the
 /// `TimelineState::Stopping` change.
-async fn delete_local_layer_files(
+// pub(super): documentation link
+pub(super) async fn delete_local_layer_files(
     conf: &PageServerConf,
-    tenant_id: TenantId,
+    tenant_shard_id: TenantShardId,
     timeline: &Timeline,
 ) -> anyhow::Result<()> {
-    info!("waiting for layer_removal_cs.lock()");
-    let layer_removal_guard = timeline.layer_removal_cs.lock().await;
-    info!("got layer_removal_cs.lock(), deleting layer files");
+    let guards = async { tokio::join!(timeline.gc_lock.lock(), timeline.compaction_lock.lock()) };
+    let guards = crate::timed(
+        guards,
+        "acquire gc and compaction locks",
+        std::time::Duration::from_secs(5),
+    )
+    .await;
 
     // NB: storage_sync upload tasks that reference these layers have been cancelled
     //     by the caller.
 
-    let local_timeline_directory = conf.timeline_path(&tenant_id, &timeline.timeline_id);
+    let local_timeline_directory = conf.timeline_path(&tenant_shard_id, &timeline.timeline_id);
 
     fail::fail_point!("timeline-delete-before-rm", |_| {
         Err(anyhow::anyhow!("failpoint: timeline-delete-before-rm"))?
@@ -150,8 +156,8 @@ async fn delete_local_layer_files(
     // because of a previous failure/cancellation at/after
     // failpoint timeline-delete-after-rm.
     //
-    // It can also happen if we race with tenant detach, because,
-    // it doesn't grab the layer_removal_cs lock.
+    // ErrorKind::NotFound can also happen if we race with tenant detach, because,
+    // no locks are shared.
     //
     // For now, log and continue.
     // warn! level is technically not appropriate for the
@@ -170,7 +176,7 @@ async fn delete_local_layer_files(
         return Ok(());
     }
 
-    let metadata_path = conf.metadata_path(&tenant_id, &timeline.timeline_id);
+    let metadata_path = conf.metadata_path(&tenant_shard_id, &timeline.timeline_id);
 
     for entry in walkdir::WalkDir::new(&local_timeline_directory).contents_first(true) {
         #[cfg(feature = "testing")]
@@ -219,8 +225,8 @@ async fn delete_local_layer_files(
         .with_context(|| format!("Failed to remove: {}", entry.path().display()))?;
     }
 
-    info!("finished deleting layer files, releasing layer_removal_cs.lock()");
-    drop(layer_removal_guard);
+    info!("finished deleting layer files, releasing locks");
+    drop(guards);
 
     fail::fail_point!("timeline-delete-after-rm", |_| {
         Err(anyhow::anyhow!("failpoint: timeline-delete-after-rm"))?
@@ -245,11 +251,11 @@ async fn delete_remote_layers_and_index(timeline: &Timeline) -> anyhow::Result<(
 // (nothing can fail after its deletion)
 async fn cleanup_remaining_timeline_fs_traces(
     conf: &PageServerConf,
-    tenant_id: TenantId,
+    tenant_shard_id: TenantShardId,
     timeline_id: TimelineId,
 ) -> anyhow::Result<()> {
     // Remove local metadata
-    tokio::fs::remove_file(conf.metadata_path(&tenant_id, &timeline_id))
+    tokio::fs::remove_file(conf.metadata_path(&tenant_shard_id, &timeline_id))
         .await
         .or_else(fs_ext::ignore_not_found)
         .context("remove metadata")?;
@@ -261,7 +267,7 @@ async fn cleanup_remaining_timeline_fs_traces(
     });
 
     // Remove timeline dir
-    tokio::fs::remove_dir(conf.timeline_path(&tenant_id, &timeline_id))
+    tokio::fs::remove_dir(conf.timeline_path(&tenant_shard_id, &timeline_id))
         .await
         .or_else(fs_ext::ignore_not_found)
         .context("timeline dir")?;
@@ -276,7 +282,7 @@ async fn cleanup_remaining_timeline_fs_traces(
     // to be reordered later and thus missed if a crash occurs.
     // Note that we dont need to sync after mark file is removed
     // because we can tolerate the case when mark file reappears on startup.
-    let timeline_path = conf.timelines_path(&tenant_id);
+    let timeline_path = conf.timelines_path(&tenant_shard_id);
     crashsafe::fsync_async(timeline_path)
         .await
         .context("fsync_pre_mark_remove")?;
@@ -284,7 +290,7 @@ async fn cleanup_remaining_timeline_fs_traces(
     // Remove delete mark
     // TODO: once we are confident that no more exist in the field, remove this
     // line.  It cleans up a legacy marker file that might in rare cases be present.
-    tokio::fs::remove_file(conf.timeline_delete_mark_file_path(tenant_id, timeline_id))
+    tokio::fs::remove_file(conf.timeline_delete_mark_file_path(tenant_shard_id, timeline_id))
         .await
         .or_else(fs_ext::ignore_not_found)
         .context("remove delete mark")
@@ -350,7 +356,7 @@ impl DeleteTimelineFlow {
     // NB: If this fails half-way through, and is retried, the retry will go through
     // all the same steps again. Make sure the code here is idempotent, and don't
     // error out if some of the shutdown tasks have already been completed!
-    #[instrument(skip(tenant), fields(tenant_id=%tenant.tenant_id))]
+    #[instrument(skip(tenant), fields(tenant_id=%tenant.tenant_shard_id.tenant_id, shard_id=%tenant.tenant_shard_id.shard_slug()))]
     pub async fn run(
         tenant: &Arc<Tenant>,
         timeline_id: TimelineId,
@@ -400,7 +406,6 @@ impl DeleteTimelineFlow {
         local_metadata: &TimelineMetadata,
         remote_client: Option<RemoteTimelineClient>,
         deletion_queue_client: DeletionQueueClient,
-        init_order: Option<&InitializationOrder>,
     ) -> anyhow::Result<()> {
         // Note: here we even skip populating layer map. Timeline is essentially uninitialized.
         // RemoteTimelineClient is the only functioning part.
@@ -413,7 +418,6 @@ impl DeleteTimelineFlow {
                     remote_client,
                     deletion_queue_client,
                 },
-                init_order,
                 // Important. We dont pass ancestor above because it can be missing.
                 // Thus we need to skip the validation here.
                 CreateTimelineCause::Delete,
@@ -446,7 +450,8 @@ impl DeleteTimelineFlow {
         timeline_id: TimelineId,
     ) -> anyhow::Result<()> {
         let r =
-            cleanup_remaining_timeline_fs_traces(tenant.conf, tenant.tenant_id, timeline_id).await;
+            cleanup_remaining_timeline_fs_traces(tenant.conf, tenant.tenant_shard_id, timeline_id)
+                .await;
         info!("Done");
         r
     }
@@ -517,13 +522,13 @@ impl DeleteTimelineFlow {
         tenant: Arc<Tenant>,
         timeline: Arc<Timeline>,
     ) {
-        let tenant_id = timeline.tenant_id;
+        let tenant_shard_id = timeline.tenant_shard_id;
         let timeline_id = timeline.timeline_id;
 
         task_mgr::spawn(
             task_mgr::BACKGROUND_RUNTIME.handle(),
             TaskKind::TimelineDeletionWorker,
-            Some(tenant_id),
+            Some(tenant_shard_id.tenant_id),
             Some(timeline_id),
             "timeline_delete",
             false,
@@ -536,7 +541,7 @@ impl DeleteTimelineFlow {
             }
             .instrument({
                 let span =
-                    tracing::info_span!(parent: None, "delete_timeline", tenant_id=%tenant_id, timeline_id=%timeline_id);
+                    tracing::info_span!(parent: None, "delete_timeline", tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(),timeline_id=%timeline_id);
                 span.follows_from(Span::current());
                 span
             }),
@@ -549,13 +554,14 @@ impl DeleteTimelineFlow {
         tenant: &Tenant,
         timeline: &Timeline,
     ) -> Result<(), DeleteTimelineError> {
-        delete_local_layer_files(conf, tenant.tenant_id, timeline).await?;
+        delete_local_layer_files(conf, tenant.tenant_shard_id, timeline).await?;
 
         delete_remote_layers_and_index(timeline).await?;
 
         pausable_failpoint!("in_progress_delete");
 
-        cleanup_remaining_timeline_fs_traces(conf, tenant.tenant_id, timeline.timeline_id).await?;
+        cleanup_remaining_timeline_fs_traces(conf, tenant.tenant_shard_id, timeline.timeline_id)
+            .await?;
 
         remove_timeline_from_tenant(tenant, timeline.timeline_id, &guard).await?;
 

pageserver/src/tenant/timeline/eviction_task.rs

@@ -60,9 +60,12 @@ impl Timeline {
         task_mgr::spawn(
             BACKGROUND_RUNTIME.handle(),
             TaskKind::Eviction,
-            Some(self.tenant_id),
+            Some(self.tenant_shard_id.tenant_id),
             Some(self.timeline_id),
-            &format!("layer eviction for {}/{}", self.tenant_id, self.timeline_id),
+            &format!(
+                "layer eviction for {}/{}",
+                self.tenant_shard_id, self.timeline_id
+            ),
             false,
             async move {
                 let cancel = task_mgr::shutdown_token();
@@ -77,7 +80,7 @@ impl Timeline {
         );
     }
 
-    #[instrument(skip_all, fields(tenant_id = %self.tenant_id, timeline_id = %self.timeline_id))]
+    #[instrument(skip_all, fields(tenant_id = %self.tenant_shard_id.tenant_id, shard_id = %self.tenant_shard_id.shard_slug(), timeline_id = %self.timeline_id))]
     async fn eviction_task(self: Arc<Self>, cancel: CancellationToken) {
         use crate::tenant::tasks::random_init_delay;
         {
@@ -296,7 +299,6 @@ impl Timeline {
                     stats.evicted += 1;
                 }
                 Some(Err(EvictionError::NotFound | EvictionError::Downloaded)) => {
-                    // compaction/gc removed the file while we were waiting on layer_removal_cs
                     stats.not_evictable += 1;
                 }
             }
@@ -341,7 +343,7 @@ impl Timeline {
         // Make one of the tenant's timelines draw the short straw and run the calculation.
         // The others wait until the calculation is done so that they take into account the
         // imitated accesses that the winner made.
-        let tenant = match crate::tenant::mgr::get_tenant(self.tenant_id, true) {
+        let tenant = match crate::tenant::mgr::get_tenant(self.tenant_shard_id.tenant_id, true) {
             Ok(t) => t,
             Err(_) => {
                 return ControlFlow::Break(());

pageserver/src/tenant/timeline/init.rs

@@ -13,6 +13,7 @@ use crate::{
 };
 use anyhow::Context;
 use camino::Utf8Path;
+use pageserver_api::shard::ShardIndex;
 use std::{collections::HashMap, str::FromStr};
 use utils::lsn::Lsn;
 
@@ -107,6 +108,7 @@ pub(super) fn reconcile(
     index_part: Option<&IndexPart>,
     disk_consistent_lsn: Lsn,
     generation: Generation,
+    shard: ShardIndex,
 ) -> Vec<(LayerFileName, Result<Decision, DismissedLayer>)> {
     use Decision::*;
 
@@ -118,10 +120,13 @@ pub(super) fn reconcile(
         .map(|(name, file_size)| {
             (
                 name,
-                // The generation here will be corrected to match IndexPart in the merge below, unless
+                // The generation and shard here will be corrected to match IndexPart in the merge below, unless
                 // it is not in IndexPart, in which case using our current generation makes sense
                 // because it will be uploaded in this generation.
-                (Some(LayerFileMetadata::new(file_size, generation)), None),
+                (
+                    Some(LayerFileMetadata::new(file_size, generation, shard)),
+                    None,
+                ),
             )
         })
         .collect::<Collected>();

pageserver/src/tenant/timeline/layer_manager.rs

@@ -1,8 +1,9 @@
 use anyhow::{bail, ensure, Context, Result};
+use pageserver_api::shard::TenantShardId;
 use std::{collections::HashMap, sync::Arc};
 use tracing::trace;
 use utils::{
-    id::{TenantId, TimelineId},
+    id::TimelineId,
     lsn::{AtomicLsn, Lsn},
 };
 
@@ -73,7 +74,7 @@ impl LayerManager {
         last_record_lsn: Lsn,
         conf: &'static PageServerConf,
         timeline_id: TimelineId,
-        tenant_id: TenantId,
+        tenant_shard_id: TenantShardId,
     ) -> Result<Arc<InMemoryLayer>> {
         ensure!(lsn.is_aligned());
 
@@ -109,7 +110,8 @@ impl LayerManager {
                 lsn
             );
 
-            let new_layer = InMemoryLayer::create(conf, timeline_id, tenant_id, start_lsn).await?;
+            let new_layer =
+                InMemoryLayer::create(conf, timeline_id, tenant_shard_id, start_lsn).await?;
             let layer = Arc::new(new_layer);
 
             self.layer_map.open_layer = Some(layer.clone());
@@ -190,7 +192,6 @@ impl LayerManager {
     /// Called when compaction is completed.
     pub(crate) fn finish_compact_l0(
         &mut self,
-        layer_removal_cs: &Arc<tokio::sync::OwnedMutexGuard<()>>,
         compact_from: &[Layer],
         compact_to: &[ResidentLayer],
         metrics: &TimelineMetrics,
@@ -201,25 +202,16 @@ impl LayerManager {
             metrics.record_new_file_metrics(l.layer_desc().file_size);
         }
         for l in compact_from {
-            Self::delete_historic_layer(layer_removal_cs, l, &mut updates, &mut self.layer_fmgr);
+            Self::delete_historic_layer(l, &mut updates, &mut self.layer_fmgr);
         }
         updates.flush();
     }
 
-    /// Called when garbage collect the timeline. Returns a guard that will apply the updates to the layer map.
-    pub(crate) fn finish_gc_timeline(
-        &mut self,
-        layer_removal_cs: &Arc<tokio::sync::OwnedMutexGuard<()>>,
-        gc_layers: Vec<Layer>,
-    ) {
+    /// Called when garbage collect has selected the layers to be removed.
+    pub(crate) fn finish_gc_timeline(&mut self, gc_layers: &[Layer]) {
         let mut updates = self.layer_map.batch_update();
         for doomed_layer in gc_layers {
-            Self::delete_historic_layer(
-                layer_removal_cs,
-                &doomed_layer,
-                &mut updates,
-                &mut self.layer_fmgr,
-            );
+            Self::delete_historic_layer(doomed_layer, &mut updates, &mut self.layer_fmgr);
         }
         updates.flush()
     }
@@ -238,7 +230,6 @@ impl LayerManager {
     /// Remote storage is not affected by this operation.
     fn delete_historic_layer(
         // we cannot remove layers otherwise, since gc and compaction will race
-        _layer_removal_cs: &Arc<tokio::sync::OwnedMutexGuard<()>>,
         layer: &Layer,
         updates: &mut BatchedUpdates<'_>,
         mapping: &mut LayerFileManager<Layer>,
@@ -252,7 +243,7 @@ impl LayerManager {
         //      map index without actually rebuilding the index.
         updates.remove_historic(desc);
         mapping.remove(layer);
-        layer.garbage_collect_on_drop();
+        layer.delete_on_drop();
     }
 
     pub(crate) fn contains(&self, layer: &Layer) -> bool {

pageserver/src/tenant/timeline/logical_size.rs

@@ -1,11 +1,10 @@
 use anyhow::Context;
-use once_cell::sync::OnceCell;
 
-use tokio::sync::Semaphore;
+use once_cell::sync::OnceCell;
+use tokio_util::sync::CancellationToken;
 use utils::lsn::Lsn;
 
-use std::sync::atomic::{AtomicI64, Ordering as AtomicOrdering};
-use std::sync::Arc;
+use std::sync::atomic::{AtomicBool, AtomicI64, Ordering as AtomicOrdering};
 
 /// Internal structure to hold all data needed for logical size calculation.
 ///
@@ -23,10 +22,17 @@ pub(super) struct LogicalSize {
     ///
     /// NOTE: size at a given LSN is constant, but after a restart we will calculate
     /// the initial size at a different LSN.
-    pub initial_logical_size: OnceCell<u64>,
+    pub initial_logical_size: OnceCell<(
+        u64,
+        crate::metrics::initial_logical_size::FinishedCalculationGuard,
+    )>,
 
-    /// Semaphore to track ongoing calculation of `initial_logical_size`.
-    pub initial_size_computation: Arc<tokio::sync::Semaphore>,
+    /// Cancellation for the best-effort logical size calculation.
+    ///
+    /// The token is kept in a once-cell so that we can error out if a higher priority
+    /// request comes in *before* we have started the normal logical size calculation.
+    pub(crate) cancel_wait_for_background_loop_concurrency_limit_semaphore:
+        OnceCell<CancellationToken>,
 
     /// Latest Lsn that has its size uncalculated, could be absent for freshly created timelines.
     pub initial_part_end: Option<Lsn>,
@@ -52,25 +58,57 @@ pub(super) struct LogicalSize {
     /// see `current_logical_size_gauge`. Use the `update_current_logical_size`
     /// to modify this, it will also keep the prometheus metric in sync.
     pub size_added_after_initial: AtomicI64,
+
+    /// For [`crate::metrics::initial_logical_size::TIMELINES_WHERE_WALRECEIVER_GOT_APPROXIMATE_SIZE`].
+    pub(super) did_return_approximate_to_walreceiver: AtomicBool,
 }
 
 /// Normalized current size, that the data in pageserver occupies.
 #[derive(Debug, Clone, Copy)]
-pub(super) enum CurrentLogicalSize {
+pub(crate) enum CurrentLogicalSize {
     /// The size is not yet calculated to the end, this is an intermediate result,
     /// constructed from walreceiver increments and normalized: logical data could delete some objects, hence be negative,
     /// yet total logical size cannot be below 0.
-    Approximate(u64),
+    Approximate(Approximate),
     // Fully calculated logical size, only other future walreceiver increments are changing it, and those changes are
     // available for observation without any calculations.
-    Exact(u64),
+    Exact(Exact),
+}
+
+#[derive(Debug, Copy, Clone, PartialEq, Eq)]
+pub(crate) enum Accuracy {
+    Approximate,
+    Exact,
+}
+
+#[derive(Debug, Clone, Copy)]
+pub(crate) struct Approximate(u64);
+#[derive(Debug, Clone, Copy)]
+pub(crate) struct Exact(u64);
+
+impl From<&Approximate> for u64 {
+    fn from(value: &Approximate) -> Self {
+        value.0
+    }
+}
+
+impl From<&Exact> for u64 {
+    fn from(val: &Exact) -> Self {
+        val.0
+    }
 }
 
 impl CurrentLogicalSize {
-    pub(super) fn size(&self) -> u64 {
-        *match self {
-            Self::Approximate(size) => size,
-            Self::Exact(size) => size,
+    pub(crate) fn size_dont_care_about_accuracy(&self) -> u64 {
+        match self {
+            Self::Approximate(size) => size.into(),
+            Self::Exact(size) => size.into(),
+        }
+    }
+    pub(crate) fn accuracy(&self) -> Accuracy {
+        match self {
+            Self::Approximate(_) => Accuracy::Approximate,
+            Self::Exact(_) => Accuracy::Exact,
         }
     }
 }
@@ -78,36 +116,42 @@ impl CurrentLogicalSize {
 impl LogicalSize {
     pub(super) fn empty_initial() -> Self {
         Self {
-            initial_logical_size: OnceCell::with_value(0),
-            //  initial_logical_size already computed, so, don't admit any calculations
-            initial_size_computation: Arc::new(Semaphore::new(0)),
+            initial_logical_size: OnceCell::with_value((0, {
+                crate::metrics::initial_logical_size::START_CALCULATION
+                    .first(crate::metrics::initial_logical_size::StartCircumstances::EmptyInitial)
+                    .calculation_result_saved()
+            })),
+            cancel_wait_for_background_loop_concurrency_limit_semaphore: OnceCell::new(),
             initial_part_end: None,
             size_added_after_initial: AtomicI64::new(0),
+            did_return_approximate_to_walreceiver: AtomicBool::new(false),
         }
     }
 
     pub(super) fn deferred_initial(compute_to: Lsn) -> Self {
         Self {
             initial_logical_size: OnceCell::new(),
-            initial_size_computation: Arc::new(Semaphore::new(1)),
+            cancel_wait_for_background_loop_concurrency_limit_semaphore: OnceCell::new(),
             initial_part_end: Some(compute_to),
             size_added_after_initial: AtomicI64::new(0),
+            did_return_approximate_to_walreceiver: AtomicBool::new(false),
         }
     }
 
-    pub(super) fn current_size(&self) -> anyhow::Result<CurrentLogicalSize> {
+    pub(super) fn current_size(&self) -> CurrentLogicalSize {
         let size_increment: i64 = self.size_added_after_initial.load(AtomicOrdering::Acquire);
         //                  ^^^ keep this type explicit so that the casts in this function break if
         //                  we change the type.
         match self.initial_logical_size.get() {
-            Some(initial_size) => {
-                initial_size.checked_add_signed(size_increment)
+            Some((initial_size, _)) => {
+                CurrentLogicalSize::Exact(Exact(initial_size.checked_add_signed(size_increment)
                     .with_context(|| format!("Overflow during logical size calculation, initial_size: {initial_size}, size_increment: {size_increment}"))
-                    .map(CurrentLogicalSize::Exact)
+                    .unwrap()))
             }
             None => {
+
                 let non_negative_size_increment = u64::try_from(size_increment).unwrap_or(0);
-                Ok(CurrentLogicalSize::Approximate(non_negative_size_increment))
+                CurrentLogicalSize::Approximate(Approximate(non_negative_size_increment))
             }
         }
     }
@@ -121,7 +165,7 @@ impl LogicalSize {
     /// available for re-use. This doesn't contain the incremental part.
     pub(super) fn initialized_size(&self, lsn: Lsn) -> Option<u64> {
         match self.initial_part_end {
-            Some(v) if v == lsn => self.initial_logical_size.get().copied(),
+            Some(v) if v == lsn => self.initial_logical_size.get().map(|(s, _)| *s),
             _ => None,
         }
     }

pageserver/src/tenant/timeline/uninit.rs

@@ -43,11 +43,11 @@ impl<'t> UninitializedTimeline<'t> {
     /// The caller is responsible for activating the timeline (function `.activate()`).
     pub(crate) fn finish_creation(mut self) -> anyhow::Result<Arc<Timeline>> {
         let timeline_id = self.timeline_id;
-        let tenant_id = self.owning_tenant.tenant_id;
+        let tenant_shard_id = self.owning_tenant.tenant_shard_id;
 
         if self.raw_timeline.is_none() {
             return Err(anyhow::anyhow!(
-                "No timeline for initialization found for {tenant_id}/{timeline_id}"
+                "No timeline for initialization found for {tenant_shard_id}/{timeline_id}"
             ));
         }
 
@@ -61,13 +61,13 @@ impl<'t> UninitializedTimeline<'t> {
 
         anyhow::ensure!(
             new_disk_consistent_lsn.is_valid(),
-            "new timeline {tenant_id}/{timeline_id} has invalid disk_consistent_lsn"
+            "new timeline {tenant_shard_id}/{timeline_id} has invalid disk_consistent_lsn"
         );
 
         let mut timelines = self.owning_tenant.timelines.lock().unwrap();
         match timelines.entry(timeline_id) {
             Entry::Occupied(_) => anyhow::bail!(
-                "Found freshly initialized timeline {tenant_id}/{timeline_id} in the tenant map"
+                "Found freshly initialized timeline {tenant_shard_id}/{timeline_id} in the tenant map"
             ),
             Entry::Vacant(v) => {
                 // after taking here should be no fallible operations, because the drop guard will not
@@ -79,7 +79,7 @@ impl<'t> UninitializedTimeline<'t> {
                 // this should be an assertion.
                 uninit_mark.remove_uninit_mark().with_context(|| {
                     format!(
-                        "Failed to remove uninit mark file for timeline {tenant_id}/{timeline_id}"
+                        "Failed to remove uninit mark file for timeline {tenant_shard_id}/{timeline_id}"
                     )
                 })?;
                 v.insert(Arc::clone(&new_timeline));
@@ -134,7 +134,7 @@ impl<'t> UninitializedTimeline<'t> {
             .with_context(|| {
                 format!(
                     "No raw timeline {}/{} found",
-                    self.owning_tenant.tenant_id, self.timeline_id
+                    self.owning_tenant.tenant_shard_id, self.timeline_id
                 )
             })?
             .0)
@@ -144,7 +144,7 @@ impl<'t> UninitializedTimeline<'t> {
 impl Drop for UninitializedTimeline<'_> {
     fn drop(&mut self) {
         if let Some((_, uninit_mark)) = self.raw_timeline.take() {
-            let _entered = info_span!("drop_uninitialized_timeline", tenant_id = %self.owning_tenant.tenant_id, timeline_id = %self.timeline_id).entered();
+            let _entered = info_span!("drop_uninitialized_timeline", tenant_id = %self.owning_tenant.tenant_shard_id.tenant_id, shard_id = %self.owning_tenant.tenant_shard_id.shard_slug(), timeline_id = %self.timeline_id).entered();
             error!("Timeline got dropped without initializing, cleaning its files");
             cleanup_timeline_directory(uninit_mark);
         }

pageserver/src/tenant/timeline/walreceiver.rs

@@ -71,7 +71,7 @@ impl WalReceiver {
         mut broker_client: BrokerClientChannel,
         ctx: &RequestContext,
     ) -> Self {
-        let tenant_id = timeline.tenant_id;
+        let tenant_id = timeline.tenant_shard_id.tenant_id;
         let timeline_id = timeline.timeline_id;
         let walreceiver_ctx =
             ctx.detached_child(TaskKind::WalReceiverManager, DownloadBehavior::Error);

pageserver/src/tenant/timeline/walreceiver/connection_manager.rs

@@ -75,7 +75,7 @@ pub(super) async fn connection_manager_loop_step(
     }
 
     let id = TenantTimelineId {
-        tenant_id: connection_manager_state.timeline.tenant_id,
+        tenant_id: connection_manager_state.timeline.tenant_shard_id.tenant_id,
         timeline_id: connection_manager_state.timeline.timeline_id,
     };
 
@@ -388,7 +388,7 @@ struct BrokerSkTimeline {
 impl ConnectionManagerState {
     pub(super) fn new(timeline: Arc<Timeline>, conf: WalReceiverConf) -> Self {
         let id = TenantTimelineId {
-            tenant_id: timeline.tenant_id,
+            tenant_id: timeline.tenant_shard_id.tenant_id,
             timeline_id: timeline.timeline_id,
         };
         Self {

pageserver/src/tenant/timeline/walreceiver/walreceiver_connection.rs

@@ -163,7 +163,7 @@ pub(super) async fn handle_walreceiver_connection(
     task_mgr::spawn(
         WALRECEIVER_RUNTIME.handle(),
         TaskKind::WalReceiverConnectionPoller,
-        Some(timeline.tenant_id),
+        Some(timeline.tenant_shard_id.tenant_id),
         Some(timeline.timeline_id),
         "walreceiver connection",
         false,
@@ -396,11 +396,15 @@ pub(super) async fn handle_walreceiver_connection(
 
             // Send the replication feedback message.
             // Regular standby_status_update fields are put into this message.
-            let (timeline_logical_size, _) = timeline
-                .get_current_logical_size(&ctx)
-                .context("Status update creation failed to get current logical size")?;
+            let current_timeline_size = timeline
+                .get_current_logical_size(
+                    crate::tenant::timeline::GetLogicalSizePriority::User,
+                    &ctx,
+                )
+                // FIXME: https://github.com/neondatabase/neon/issues/5963
+                .size_dont_care_about_accuracy();
             let status_update = PageserverFeedback {
-                current_timeline_size: timeline_logical_size,
+                current_timeline_size,
                 last_received_lsn,
                 disk_consistent_lsn,
                 remote_consistent_lsn,

pageserver/src/tenant/upload_queue.rs

@@ -1,6 +1,5 @@
 use super::storage_layer::LayerFileName;
 use super::storage_layer::ResidentLayer;
-use super::Generation;
 use crate::tenant::metadata::TimelineMetadata;
 use crate::tenant::remote_timeline_client::index::IndexPart;
 use crate::tenant::remote_timeline_client::index::LayerFileMetadata;
@@ -15,6 +14,9 @@ use utils::lsn::AtomicLsn;
 use std::sync::atomic::AtomicU32;
 use utils::lsn::Lsn;
 
+#[cfg(feature = "testing")]
+use utils::generation::Generation;
+
 // clippy warns that Uninitialized is much smaller than Initialized, which wastes
 // memory for Uninitialized variants. Doesn't matter in practice, there are not
 // that many upload queues in a running pageserver, and most of them are initialized
@@ -88,6 +90,14 @@ pub(crate) struct UploadQueueInitialized {
     /// bug causing leaks, then it's better to not leave this enabled for production builds.
     #[cfg(feature = "testing")]
     pub(crate) dangling_files: HashMap<LayerFileName, Generation>,
+
+    /// Set to true when we have inserted the `UploadOp::Shutdown` into the `inprogress_tasks`.
+    pub(crate) shutting_down: bool,
+
+    /// Permitless semaphore on which any number of `RemoteTimelineClient::shutdown` futures can
+    /// wait on until one of them stops the queue. The semaphore is closed when
+    /// `RemoteTimelineClient::launch_queued_tasks` encounters `UploadOp::Shutdown`.
+    pub(crate) shutdown_ready: Arc<tokio::sync::Semaphore>,
 }
 
 impl UploadQueueInitialized {
@@ -146,6 +156,8 @@ impl UploadQueue {
             queued_operations: VecDeque::new(),
             #[cfg(feature = "testing")]
             dangling_files: HashMap::new(),
+            shutting_down: false,
+            shutdown_ready: Arc::new(tokio::sync::Semaphore::new(0)),
         };
 
         *self = UploadQueue::Initialized(state);
@@ -193,6 +205,8 @@ impl UploadQueue {
             queued_operations: VecDeque::new(),
             #[cfg(feature = "testing")]
             dangling_files: HashMap::new(),
+            shutting_down: false,
+            shutdown_ready: Arc::new(tokio::sync::Semaphore::new(0)),
         };
 
         *self = UploadQueue::Initialized(state);
@@ -204,7 +218,13 @@ impl UploadQueue {
             UploadQueue::Uninitialized | UploadQueue::Stopped(_) => {
                 anyhow::bail!("queue is in state {}", self.as_str())
             }
-            UploadQueue::Initialized(x) => Ok(x),
+            UploadQueue::Initialized(x) => {
+                if !x.shutting_down {
+                    Ok(x)
+                } else {
+                    anyhow::bail!("queue is shutting down")
+                }
+            }
         }
     }
 
@@ -232,7 +252,7 @@ pub(crate) struct UploadTask {
 /// for timeline deletion, which skips this queue and goes directly to DeletionQueue.
 #[derive(Debug)]
 pub(crate) struct Delete {
-    pub(crate) layers: Vec<(LayerFileName, Generation)>,
+    pub(crate) layers: Vec<(LayerFileName, LayerFileMetadata)>,
 }
 
 #[derive(Debug)]
@@ -248,6 +268,10 @@ pub(crate) enum UploadOp {
 
     /// Barrier. When the barrier operation is reached,
     Barrier(tokio::sync::watch::Sender<()>),
+
+    /// Shutdown; upon encountering this operation no new operations will be spawned, otherwise
+    /// this is the same as a Barrier.
+    Shutdown,
 }
 
 impl std::fmt::Display for UploadOp {
@@ -269,6 +293,7 @@ impl std::fmt::Display for UploadOp {
                 write!(f, "Delete({} layers)", delete.layers.len())
             }
             UploadOp::Barrier(_) => write!(f, "Barrier"),
+            UploadOp::Shutdown => write!(f, "Shutdown"),
         }
     }
 }

pageserver/src/virtual_file.rs

@@ -288,6 +288,9 @@ impl VirtualFile {
         }
         let (handle, mut slot_guard) = get_open_files().find_victim_slot();
 
+        // NB: there is also StorageIoOperation::OpenAfterReplace which is for the case
+        // where our caller doesn't get to use the returned VirtualFile before its
+        // slot gets re-used by someone else.
         let file = STORAGE_IO_TIME_METRIC
             .get(StorageIoOperation::Open)
             .observe_closure_duration(|| open_options.open(path))?;
@@ -311,6 +314,9 @@ impl VirtualFile {
             timeline_id,
         };
 
+        // TODO: Under pressure, it's likely the slot will get re-used and
+        // the underlying file closed before they get around to using it.
+        // => https://github.com/neondatabase/neon/issues/6065
         slot_guard.file.replace(file);
 
         Ok(vfile)
@@ -421,9 +427,12 @@ impl VirtualFile {
         // now locked in write-mode. Find a free slot to put it in.
         let (handle, mut slot_guard) = open_files.find_victim_slot();
 
-        // Open the physical file
+        // Re-open the physical file.
+        // NB: we use StorageIoOperation::OpenAferReplace for this to distinguish this
+        // case from StorageIoOperation::Open. This helps with identifying thrashing
+        // of the virtual file descriptor cache.
         let file = STORAGE_IO_TIME_METRIC
-            .get(StorageIoOperation::Open)
+            .get(StorageIoOperation::OpenAfterReplace)
             .observe_closure_duration(|| self.open_options.open(&self.path))?;
 
         // Perform the requested operation on it
@@ -610,9 +619,11 @@ impl Drop for VirtualFile {
             slot.recently_used.store(false, Ordering::Relaxed);
             // there is also operation "close-by-replace" for closes done on eviction for
             // comparison.
-            STORAGE_IO_TIME_METRIC
-                .get(StorageIoOperation::Close)
-                .observe_closure_duration(|| drop(slot_guard.file.take()));
+            if let Some(fd) = slot_guard.file.take() {
+                STORAGE_IO_TIME_METRIC
+                    .get(StorageIoOperation::Close)
+                    .observe_closure_duration(|| drop(fd));
+            }
         }
     }
 }
@@ -643,6 +654,7 @@ pub fn init(num_slots: usize) {
     if OPEN_FILES.set(OpenFiles::new(num_slots)).is_err() {
         panic!("virtual_file::init called twice");
     }
+    crate::metrics::virtual_file_descriptor_cache::SIZE_MAX.set(num_slots as u64);
 }
 
 const TEST_MAX_FILE_DESCRIPTORS: usize = 10;

pageserver/src/walingest.rs

@@ -21,6 +21,7 @@
 //! redo Postgres process, but some records it can handle directly with
 //! bespoken Rust code.
 
+use pageserver_api::shard::ShardIdentity;
 use postgres_ffi::v14::nonrelfile_utils::clogpage_precedes;
 use postgres_ffi::v14::nonrelfile_utils::slru_may_delete_clogsegment;
 use postgres_ffi::{fsm_logical_to_physical, page_is_new, page_set_lsn};
@@ -30,6 +31,7 @@ use bytes::{Buf, Bytes, BytesMut};
 use tracing::*;
 
 use crate::context::RequestContext;
+use crate::metrics::WAL_INGEST;
 use crate::pgdatadir_mapping::*;
 use crate::tenant::PageReconstructError;
 use crate::tenant::Timeline;
@@ -46,6 +48,7 @@ use postgres_ffi::BLCKSZ;
 use utils::lsn::Lsn;
 
 pub struct WalIngest<'a> {
+    shard: ShardIdentity,
     timeline: &'a Timeline,
 
     checkpoint: CheckPoint,
@@ -65,6 +68,7 @@ impl<'a> WalIngest<'a> {
         trace!("CheckPoint.nextXid = {}", checkpoint.nextXid.value);
 
         Ok(WalIngest {
+            shard: *timeline.get_shard_identity(),
             timeline,
             checkpoint,
             checkpoint_modified: false,
@@ -87,6 +91,8 @@ impl<'a> WalIngest<'a> {
         decoded: &mut DecodedWALRecord,
         ctx: &RequestContext,
     ) -> anyhow::Result<()> {
+        WAL_INGEST.records_received.inc();
+
         modification.lsn = lsn;
         decode_wal_record(recdata, decoded, self.timeline.pg_version)?;
 
@@ -98,266 +104,290 @@ impl<'a> WalIngest<'a> {
             self.checkpoint_modified = true;
         }
 
-        // Heap AM records need some special handling, because they modify VM pages
-        // without registering them with the standard mechanism.
-        if decoded.xl_rmid == pg_constants::RM_HEAP_ID
-            || decoded.xl_rmid == pg_constants::RM_HEAP2_ID
-        {
-            self.ingest_heapam_record(&mut buf, modification, decoded, ctx)
-                .await?;
-        }
-        if decoded.xl_rmid == pg_constants::RM_NEON_ID {
-            self.ingest_neonrmgr_record(&mut buf, modification, decoded, ctx)
-                .await?;
-        }
-        // Handle other special record types
-        if decoded.xl_rmid == pg_constants::RM_SMGR_ID
-            && (decoded.xl_info & pg_constants::XLR_RMGR_INFO_MASK)
-                == pg_constants::XLOG_SMGR_CREATE
-        {
-            let create = XlSmgrCreate::decode(&mut buf);
-            self.ingest_xlog_smgr_create(modification, &create, ctx)
-                .await?;
-        } else if decoded.xl_rmid == pg_constants::RM_SMGR_ID
-            && (decoded.xl_info & pg_constants::XLR_RMGR_INFO_MASK)
-                == pg_constants::XLOG_SMGR_TRUNCATE
-        {
-            let truncate = XlSmgrTruncate::decode(&mut buf);
-            self.ingest_xlog_smgr_truncate(modification, &truncate, ctx)
-                .await?;
-        } else if decoded.xl_rmid == pg_constants::RM_DBASE_ID {
-            debug!(
-                "handle RM_DBASE_ID for Postgres version {:?}",
-                self.timeline.pg_version
-            );
-            if self.timeline.pg_version == 14 {
-                if (decoded.xl_info & pg_constants::XLR_RMGR_INFO_MASK)
-                    == postgres_ffi::v14::bindings::XLOG_DBASE_CREATE
-                {
-                    let createdb = XlCreateDatabase::decode(&mut buf);
-                    debug!("XLOG_DBASE_CREATE v14");
+        match decoded.xl_rmid {
+            pg_constants::RM_HEAP_ID | pg_constants::RM_HEAP2_ID => {
+                // Heap AM records need some special handling, because they modify VM pages
+                // without registering them with the standard mechanism.
+                self.ingest_heapam_record(&mut buf, modification, decoded, ctx)
+                    .await?;
+            }
+            pg_constants::RM_NEON_ID => {
+                self.ingest_neonrmgr_record(&mut buf, modification, decoded, ctx)
+                    .await?;
+            }
+            // Handle other special record types
+            pg_constants::RM_SMGR_ID => {
+                let info = decoded.xl_info & pg_constants::XLR_RMGR_INFO_MASK;
 
-                    self.ingest_xlog_dbase_create(modification, &createdb, ctx)
+                if info == pg_constants::XLOG_SMGR_CREATE {
+                    let create = XlSmgrCreate::decode(&mut buf);
+                    self.ingest_xlog_smgr_create(modification, &create, ctx)
                         .await?;
-                } else if (decoded.xl_info & pg_constants::XLR_RMGR_INFO_MASK)
-                    == postgres_ffi::v14::bindings::XLOG_DBASE_DROP
-                {
-                    let dropdb = XlDropDatabase::decode(&mut buf);
-                    for tablespace_id in dropdb.tablespace_ids {
-                        trace!("Drop db {}, {}", tablespace_id, dropdb.db_id);
-                        modification
-                            .drop_dbdir(tablespace_id, dropdb.db_id, ctx)
-                            .await?;
-                    }
-                }
-            } else if self.timeline.pg_version == 15 {
-                if (decoded.xl_info & pg_constants::XLR_RMGR_INFO_MASK)
-                    == postgres_ffi::v15::bindings::XLOG_DBASE_CREATE_WAL_LOG
-                {
-                    debug!("XLOG_DBASE_CREATE_WAL_LOG: noop");
-                } else if (decoded.xl_info & pg_constants::XLR_RMGR_INFO_MASK)
-                    == postgres_ffi::v15::bindings::XLOG_DBASE_CREATE_FILE_COPY
-                {
-                    // The XLOG record was renamed between v14 and v15,
-                    // but the record format is the same.
-                    // So we can reuse XlCreateDatabase here.
-                    debug!("XLOG_DBASE_CREATE_FILE_COPY");
-                    let createdb = XlCreateDatabase::decode(&mut buf);
-                    self.ingest_xlog_dbase_create(modification, &createdb, ctx)
+                } else if info == pg_constants::XLOG_SMGR_TRUNCATE {
+                    let truncate = XlSmgrTruncate::decode(&mut buf);
+                    self.ingest_xlog_smgr_truncate(modification, &truncate, ctx)
                         .await?;
-                } else if (decoded.xl_info & pg_constants::XLR_RMGR_INFO_MASK)
-                    == postgres_ffi::v15::bindings::XLOG_DBASE_DROP
-                {
-                    let dropdb = XlDropDatabase::decode(&mut buf);
-                    for tablespace_id in dropdb.tablespace_ids {
-                        trace!("Drop db {}, {}", tablespace_id, dropdb.db_id);
-                        modification
-                            .drop_dbdir(tablespace_id, dropdb.db_id, ctx)
-                            .await?;
-                    }
-                }
-            } else if self.timeline.pg_version == 16 {
-                if (decoded.xl_info & pg_constants::XLR_RMGR_INFO_MASK)
-                    == postgres_ffi::v16::bindings::XLOG_DBASE_CREATE_WAL_LOG
-                {
-                    debug!("XLOG_DBASE_CREATE_WAL_LOG: noop");
-                } else if (decoded.xl_info & pg_constants::XLR_RMGR_INFO_MASK)
-                    == postgres_ffi::v16::bindings::XLOG_DBASE_CREATE_FILE_COPY
-                {
-                    // The XLOG record was renamed between v14 and v15,
-                    // but the record format is the same.
-                    // So we can reuse XlCreateDatabase here.
-                    debug!("XLOG_DBASE_CREATE_FILE_COPY");
-                    let createdb = XlCreateDatabase::decode(&mut buf);
-                    self.ingest_xlog_dbase_create(modification, &createdb, ctx)
-                        .await?;
-                } else if (decoded.xl_info & pg_constants::XLR_RMGR_INFO_MASK)
-                    == postgres_ffi::v16::bindings::XLOG_DBASE_DROP
-                {
-                    let dropdb = XlDropDatabase::decode(&mut buf);
-                    for tablespace_id in dropdb.tablespace_ids {
-                        trace!("Drop db {}, {}", tablespace_id, dropdb.db_id);
-                        modification
-                            .drop_dbdir(tablespace_id, dropdb.db_id, ctx)
-                            .await?;
-                    }
                 }
             }
-        } else if decoded.xl_rmid == pg_constants::RM_TBLSPC_ID {
-            trace!("XLOG_TBLSPC_CREATE/DROP is not handled yet");
-        } else if decoded.xl_rmid == pg_constants::RM_CLOG_ID {
-            let info = decoded.xl_info & !pg_constants::XLR_INFO_MASK;
-            if info == pg_constants::CLOG_ZEROPAGE {
-                let pageno = buf.get_u32_le();
-                let segno = pageno / pg_constants::SLRU_PAGES_PER_SEGMENT;
-                let rpageno = pageno % pg_constants::SLRU_PAGES_PER_SEGMENT;
-                self.put_slru_page_image(
-                    modification,
-                    SlruKind::Clog,
-                    segno,
-                    rpageno,
-                    ZERO_PAGE.clone(),
-                    ctx,
-                )
-                .await?;
-            } else {
-                assert!(info == pg_constants::CLOG_TRUNCATE);
-                let xlrec = XlClogTruncate::decode(&mut buf);
-                self.ingest_clog_truncate_record(modification, &xlrec, ctx)
-                    .await?;
-            }
-        } else if decoded.xl_rmid == pg_constants::RM_XACT_ID {
-            let info = decoded.xl_info & pg_constants::XLOG_XACT_OPMASK;
-            if info == pg_constants::XLOG_XACT_COMMIT || info == pg_constants::XLOG_XACT_ABORT {
-                let parsed_xact =
-                    XlXactParsedRecord::decode(&mut buf, decoded.xl_xid, decoded.xl_info);
-                self.ingest_xact_record(
-                    modification,
-                    &parsed_xact,
-                    info == pg_constants::XLOG_XACT_COMMIT,
-                    ctx,
-                )
-                .await?;
-            } else if info == pg_constants::XLOG_XACT_COMMIT_PREPARED
-                || info == pg_constants::XLOG_XACT_ABORT_PREPARED
-            {
-                let parsed_xact =
-                    XlXactParsedRecord::decode(&mut buf, decoded.xl_xid, decoded.xl_info);
-                self.ingest_xact_record(
-                    modification,
-                    &parsed_xact,
-                    info == pg_constants::XLOG_XACT_COMMIT_PREPARED,
-                    ctx,
-                )
-                .await?;
-                // Remove twophase file. see RemoveTwoPhaseFile() in postgres code
-                trace!(
-                    "Drop twophaseFile for xid {} parsed_xact.xid {} here at {}",
-                    decoded.xl_xid,
-                    parsed_xact.xid,
-                    lsn,
-                );
-                modification
-                    .drop_twophase_file(parsed_xact.xid, ctx)
-                    .await?;
-            } else if info == pg_constants::XLOG_XACT_PREPARE {
-                modification
-                    .put_twophase_file(decoded.xl_xid, Bytes::copy_from_slice(&buf[..]), ctx)
-                    .await?;
-            }
-        } else if decoded.xl_rmid == pg_constants::RM_MULTIXACT_ID {
-            let info = decoded.xl_info & pg_constants::XLR_RMGR_INFO_MASK;
+            pg_constants::RM_DBASE_ID => {
+                let info = decoded.xl_info & pg_constants::XLR_RMGR_INFO_MASK;
+                debug!(%info, pg_version=%self.timeline.pg_version, "handle RM_DBASE_ID");
 
-            if info == pg_constants::XLOG_MULTIXACT_ZERO_OFF_PAGE {
-                let pageno = buf.get_u32_le();
-                let segno = pageno / pg_constants::SLRU_PAGES_PER_SEGMENT;
-                let rpageno = pageno % pg_constants::SLRU_PAGES_PER_SEGMENT;
-                self.put_slru_page_image(
-                    modification,
-                    SlruKind::MultiXactOffsets,
-                    segno,
-                    rpageno,
-                    ZERO_PAGE.clone(),
-                    ctx,
-                )
-                .await?;
-            } else if info == pg_constants::XLOG_MULTIXACT_ZERO_MEM_PAGE {
-                let pageno = buf.get_u32_le();
-                let segno = pageno / pg_constants::SLRU_PAGES_PER_SEGMENT;
-                let rpageno = pageno % pg_constants::SLRU_PAGES_PER_SEGMENT;
-                self.put_slru_page_image(
-                    modification,
-                    SlruKind::MultiXactMembers,
-                    segno,
-                    rpageno,
-                    ZERO_PAGE.clone(),
-                    ctx,
-                )
-                .await?;
-            } else if info == pg_constants::XLOG_MULTIXACT_CREATE_ID {
-                let xlrec = XlMultiXactCreate::decode(&mut buf);
-                self.ingest_multixact_create_record(modification, &xlrec)?;
-            } else if info == pg_constants::XLOG_MULTIXACT_TRUNCATE_ID {
-                let xlrec = XlMultiXactTruncate::decode(&mut buf);
-                self.ingest_multixact_truncate_record(modification, &xlrec, ctx)
+                if self.timeline.pg_version == 14 {
+                    if info == postgres_ffi::v14::bindings::XLOG_DBASE_CREATE {
+                        let createdb = XlCreateDatabase::decode(&mut buf);
+                        debug!("XLOG_DBASE_CREATE v14");
+
+                        self.ingest_xlog_dbase_create(modification, &createdb, ctx)
+                            .await?;
+                    } else if info == postgres_ffi::v14::bindings::XLOG_DBASE_DROP {
+                        let dropdb = XlDropDatabase::decode(&mut buf);
+                        for tablespace_id in dropdb.tablespace_ids {
+                            trace!("Drop db {}, {}", tablespace_id, dropdb.db_id);
+                            modification
+                                .drop_dbdir(tablespace_id, dropdb.db_id, ctx)
+                                .await?;
+                        }
+                    }
+                } else if self.timeline.pg_version == 15 {
+                    if info == postgres_ffi::v15::bindings::XLOG_DBASE_CREATE_WAL_LOG {
+                        debug!("XLOG_DBASE_CREATE_WAL_LOG: noop");
+                    } else if info == postgres_ffi::v15::bindings::XLOG_DBASE_CREATE_FILE_COPY {
+                        // The XLOG record was renamed between v14 and v15,
+                        // but the record format is the same.
+                        // So we can reuse XlCreateDatabase here.
+                        debug!("XLOG_DBASE_CREATE_FILE_COPY");
+                        let createdb = XlCreateDatabase::decode(&mut buf);
+                        self.ingest_xlog_dbase_create(modification, &createdb, ctx)
+                            .await?;
+                    } else if info == postgres_ffi::v15::bindings::XLOG_DBASE_DROP {
+                        let dropdb = XlDropDatabase::decode(&mut buf);
+                        for tablespace_id in dropdb.tablespace_ids {
+                            trace!("Drop db {}, {}", tablespace_id, dropdb.db_id);
+                            modification
+                                .drop_dbdir(tablespace_id, dropdb.db_id, ctx)
+                                .await?;
+                        }
+                    }
+                } else if self.timeline.pg_version == 16 {
+                    if info == postgres_ffi::v16::bindings::XLOG_DBASE_CREATE_WAL_LOG {
+                        debug!("XLOG_DBASE_CREATE_WAL_LOG: noop");
+                    } else if info == postgres_ffi::v16::bindings::XLOG_DBASE_CREATE_FILE_COPY {
+                        // The XLOG record was renamed between v14 and v15,
+                        // but the record format is the same.
+                        // So we can reuse XlCreateDatabase here.
+                        debug!("XLOG_DBASE_CREATE_FILE_COPY");
+                        let createdb = XlCreateDatabase::decode(&mut buf);
+                        self.ingest_xlog_dbase_create(modification, &createdb, ctx)
+                            .await?;
+                    } else if info == postgres_ffi::v16::bindings::XLOG_DBASE_DROP {
+                        let dropdb = XlDropDatabase::decode(&mut buf);
+                        for tablespace_id in dropdb.tablespace_ids {
+                            trace!("Drop db {}, {}", tablespace_id, dropdb.db_id);
+                            modification
+                                .drop_dbdir(tablespace_id, dropdb.db_id, ctx)
+                                .await?;
+                        }
+                    }
+                }
+            }
+            pg_constants::RM_TBLSPC_ID => {
+                trace!("XLOG_TBLSPC_CREATE/DROP is not handled yet");
+            }
+            pg_constants::RM_CLOG_ID => {
+                let info = decoded.xl_info & !pg_constants::XLR_INFO_MASK;
+
+                if info == pg_constants::CLOG_ZEROPAGE {
+                    let pageno = buf.get_u32_le();
+                    let segno = pageno / pg_constants::SLRU_PAGES_PER_SEGMENT;
+                    let rpageno = pageno % pg_constants::SLRU_PAGES_PER_SEGMENT;
+                    self.put_slru_page_image(
+                        modification,
+                        SlruKind::Clog,
+                        segno,
+                        rpageno,
+                        ZERO_PAGE.clone(),
+                        ctx,
+                    )
                     .await?;
-            }
-        } else if decoded.xl_rmid == pg_constants::RM_RELMAP_ID {
-            let xlrec = XlRelmapUpdate::decode(&mut buf);
-            self.ingest_relmap_page(modification, &xlrec, decoded, ctx)
-                .await?;
-        } else if decoded.xl_rmid == pg_constants::RM_XLOG_ID {
-            let info = decoded.xl_info & pg_constants::XLR_RMGR_INFO_MASK;
-            if info == pg_constants::XLOG_NEXTOID {
-                let next_oid = buf.get_u32_le();
-                if self.checkpoint.nextOid != next_oid {
-                    self.checkpoint.nextOid = next_oid;
-                    self.checkpoint_modified = true;
+                } else {
+                    assert!(info == pg_constants::CLOG_TRUNCATE);
+                    let xlrec = XlClogTruncate::decode(&mut buf);
+                    self.ingest_clog_truncate_record(modification, &xlrec, ctx)
+                        .await?;
                 }
-            } else if info == pg_constants::XLOG_CHECKPOINT_ONLINE
-                || info == pg_constants::XLOG_CHECKPOINT_SHUTDOWN
-            {
-                let mut checkpoint_bytes = [0u8; SIZEOF_CHECKPOINT];
-                buf.copy_to_slice(&mut checkpoint_bytes);
-                let xlog_checkpoint = CheckPoint::decode(&checkpoint_bytes)?;
-                trace!(
-                    "xlog_checkpoint.oldestXid={}, checkpoint.oldestXid={}",
-                    xlog_checkpoint.oldestXid,
-                    self.checkpoint.oldestXid
-                );
-                if (self
-                    .checkpoint
-                    .oldestXid
-                    .wrapping_sub(xlog_checkpoint.oldestXid) as i32)
-                    < 0
+            }
+            pg_constants::RM_XACT_ID => {
+                let info = decoded.xl_info & pg_constants::XLOG_XACT_OPMASK;
+
+                if info == pg_constants::XLOG_XACT_COMMIT || info == pg_constants::XLOG_XACT_ABORT {
+                    let parsed_xact =
+                        XlXactParsedRecord::decode(&mut buf, decoded.xl_xid, decoded.xl_info);
+                    self.ingest_xact_record(
+                        modification,
+                        &parsed_xact,
+                        info == pg_constants::XLOG_XACT_COMMIT,
+                        ctx,
+                    )
+                    .await?;
+                } else if info == pg_constants::XLOG_XACT_COMMIT_PREPARED
+                    || info == pg_constants::XLOG_XACT_ABORT_PREPARED
                 {
-                    self.checkpoint.oldestXid = xlog_checkpoint.oldestXid;
-                    self.checkpoint_modified = true;
-                }
-            }
-        } else if decoded.xl_rmid == pg_constants::RM_LOGICALMSG_ID {
-            let info = decoded.xl_info & pg_constants::XLR_RMGR_INFO_MASK;
-            if info == pg_constants::XLOG_LOGICAL_MESSAGE {
-                let xlrec = XlLogicalMessage::decode(&mut buf);
-                let prefix = std::str::from_utf8(&buf[0..xlrec.prefix_size - 1])?;
-                let message = &buf[xlrec.prefix_size..xlrec.prefix_size + xlrec.message_size];
-                if prefix == "neon-test" {
-                    // This is a convenient way to make the WAL ingestion pause at
-                    // particular point in the WAL. For more fine-grained control,
-                    // we could peek into the message and only pause if it contains
-                    // a particular string, for example, but this is enough for now.
-                    crate::failpoint_support::sleep_millis_async!(
-                        "wal-ingest-logical-message-sleep"
+                    let parsed_xact =
+                        XlXactParsedRecord::decode(&mut buf, decoded.xl_xid, decoded.xl_info);
+                    self.ingest_xact_record(
+                        modification,
+                        &parsed_xact,
+                        info == pg_constants::XLOG_XACT_COMMIT_PREPARED,
+                        ctx,
+                    )
+                    .await?;
+                    // Remove twophase file. see RemoveTwoPhaseFile() in postgres code
+                    trace!(
+                        "Drop twophaseFile for xid {} parsed_xact.xid {} here at {}",
+                        decoded.xl_xid,
+                        parsed_xact.xid,
+                        lsn,
                     );
-                } else if let Some(path) = prefix.strip_prefix("neon-file:") {
-                    modification.put_file(path, message, ctx).await?;
+                    modification
+                        .drop_twophase_file(parsed_xact.xid, ctx)
+                        .await?;
+                } else if info == pg_constants::XLOG_XACT_PREPARE {
+                    modification
+                        .put_twophase_file(decoded.xl_xid, Bytes::copy_from_slice(&buf[..]), ctx)
+                        .await?;
                 }
             }
+            pg_constants::RM_MULTIXACT_ID => {
+                let info = decoded.xl_info & pg_constants::XLR_RMGR_INFO_MASK;
+
+                if info == pg_constants::XLOG_MULTIXACT_ZERO_OFF_PAGE {
+                    let pageno = buf.get_u32_le();
+                    let segno = pageno / pg_constants::SLRU_PAGES_PER_SEGMENT;
+                    let rpageno = pageno % pg_constants::SLRU_PAGES_PER_SEGMENT;
+                    self.put_slru_page_image(
+                        modification,
+                        SlruKind::MultiXactOffsets,
+                        segno,
+                        rpageno,
+                        ZERO_PAGE.clone(),
+                        ctx,
+                    )
+                    .await?;
+                } else if info == pg_constants::XLOG_MULTIXACT_ZERO_MEM_PAGE {
+                    let pageno = buf.get_u32_le();
+                    let segno = pageno / pg_constants::SLRU_PAGES_PER_SEGMENT;
+                    let rpageno = pageno % pg_constants::SLRU_PAGES_PER_SEGMENT;
+                    self.put_slru_page_image(
+                        modification,
+                        SlruKind::MultiXactMembers,
+                        segno,
+                        rpageno,
+                        ZERO_PAGE.clone(),
+                        ctx,
+                    )
+                    .await?;
+                } else if info == pg_constants::XLOG_MULTIXACT_CREATE_ID {
+                    let xlrec = XlMultiXactCreate::decode(&mut buf);
+                    self.ingest_multixact_create_record(modification, &xlrec)?;
+                } else if info == pg_constants::XLOG_MULTIXACT_TRUNCATE_ID {
+                    let xlrec = XlMultiXactTruncate::decode(&mut buf);
+                    self.ingest_multixact_truncate_record(modification, &xlrec, ctx)
+                        .await?;
+                }
+            }
+            pg_constants::RM_RELMAP_ID => {
+                let xlrec = XlRelmapUpdate::decode(&mut buf);
+                self.ingest_relmap_page(modification, &xlrec, decoded, ctx)
+                    .await?;
+            }
+            pg_constants::RM_XLOG_ID => {
+                let info = decoded.xl_info & pg_constants::XLR_RMGR_INFO_MASK;
+
+                if info == pg_constants::XLOG_NEXTOID {
+                    let next_oid = buf.get_u32_le();
+                    if self.checkpoint.nextOid != next_oid {
+                        self.checkpoint.nextOid = next_oid;
+                        self.checkpoint_modified = true;
+                    }
+                } else if info == pg_constants::XLOG_CHECKPOINT_ONLINE
+                    || info == pg_constants::XLOG_CHECKPOINT_SHUTDOWN
+                {
+                    let mut checkpoint_bytes = [0u8; SIZEOF_CHECKPOINT];
+                    buf.copy_to_slice(&mut checkpoint_bytes);
+                    let xlog_checkpoint = CheckPoint::decode(&checkpoint_bytes)?;
+                    trace!(
+                        "xlog_checkpoint.oldestXid={}, checkpoint.oldestXid={}",
+                        xlog_checkpoint.oldestXid,
+                        self.checkpoint.oldestXid
+                    );
+                    if (self
+                        .checkpoint
+                        .oldestXid
+                        .wrapping_sub(xlog_checkpoint.oldestXid) as i32)
+                        < 0
+                    {
+                        self.checkpoint.oldestXid = xlog_checkpoint.oldestXid;
+                        self.checkpoint_modified = true;
+                    }
+                }
+            }
+            pg_constants::RM_LOGICALMSG_ID => {
+                let info = decoded.xl_info & pg_constants::XLR_RMGR_INFO_MASK;
+
+                if info == pg_constants::XLOG_LOGICAL_MESSAGE {
+                    let xlrec = XlLogicalMessage::decode(&mut buf);
+                    let prefix = std::str::from_utf8(&buf[0..xlrec.prefix_size - 1])?;
+                    let message = &buf[xlrec.prefix_size..xlrec.prefix_size + xlrec.message_size];
+                    if prefix == "neon-test" {
+                        // This is a convenient way to make the WAL ingestion pause at
+                        // particular point in the WAL. For more fine-grained control,
+                        // we could peek into the message and only pause if it contains
+                        // a particular string, for example, but this is enough for now.
+                        crate::failpoint_support::sleep_millis_async!(
+                            "wal-ingest-logical-message-sleep"
+                        );
+                    } else if let Some(path) = prefix.strip_prefix("neon-file:") {
+                        modification.put_file(path, message, ctx).await?;
+                    }
+                }
+            }
+            _x => {
+                // TODO: should probably log & fail here instead of blindly
+                // doing something without understanding the protocol
+            }
         }
 
         // Iterate through all the blocks that the record modifies, and
         // "put" a separate copy of the record for each block.
         for blk in decoded.blocks.iter() {
+            let rel = RelTag {
+                spcnode: blk.rnode_spcnode,
+                dbnode: blk.rnode_dbnode,
+                relnode: blk.rnode_relnode,
+                forknum: blk.forknum,
+            };
+
+            let key = rel_block_to_key(rel, blk.blkno);
+            let key_is_local = self.shard.is_key_local(&key);
+
+            tracing::debug!(
+                lsn=%lsn,
+                key=%key,
+                "ingest: shard decision {} (checkpoint={})",
+                if !key_is_local { "drop" } else { "keep" },
+                self.checkpoint_modified
+            );
+
+            if !key_is_local {
+                if self.shard.is_zero() {
+                    // Shard 0 tracks relation sizes.  Although we will not store this block, we will observe
+                    // its blkno in case it implicitly extends a relation.
+                    self.observe_decoded_block(modification, blk, ctx).await?;
+                }
+
+                continue;
+            }
             self.ingest_decoded_block(modification, lsn, decoded, blk, ctx)
                 .await?;
         }
@@ -370,13 +400,38 @@ impl<'a> WalIngest<'a> {
             self.checkpoint_modified = false;
         }
 
+        if modification.is_empty() {
+            tracing::debug!("ingest: filtered out record @ LSN {lsn}");
+            WAL_INGEST.records_filtered.inc();
+            modification.tline.finish_write(lsn);
+        } else {
+            WAL_INGEST.records_committed.inc();
+            modification.commit(ctx).await?;
+        }
+
         // Now that this record has been fully handled, including updating the
-        // checkpoint data, let the repository know that it is up-to-date to this LSN
-        modification.commit(ctx).await?;
+        // checkpoint data, let the repository know that it is up-to-date to this LSN.
 
         Ok(())
     }
 
+    /// Do not store this block, but observe it for the purposes of updating our relation size state.
+    async fn observe_decoded_block(
+        &mut self,
+        modification: &mut DatadirModification<'_>,
+        blk: &DecodedBkpBlock,
+        ctx: &RequestContext,
+    ) -> Result<(), PageReconstructError> {
+        let rel = RelTag {
+            spcnode: blk.rnode_spcnode,
+            dbnode: blk.rnode_dbnode,
+            relnode: blk.rnode_relnode,
+            forknum: blk.forknum,
+        };
+        self.handle_rel_extend(modification, rel, blk.blkno, ctx)
+            .await
+    }
+
     async fn ingest_decoded_block(
         &mut self,
         modification: &mut DatadirModification<'_>,
@@ -1440,7 +1495,16 @@ impl<'a> WalIngest<'a> {
         // record.
         // TODO: would be nice if to be more explicit about it
         let last_lsn = modification.lsn;
-        let old_nblocks = if !self
+
+        // Get current size and put rel creation if rel doesn't exist
+        //
+        // NOTE: we check the cache first even though get_rel_exists and get_rel_size would
+        //       check the cache too. This is because eagerly checking the cache results in
+        //       less work overall and 10% better performance. It's more work on cache miss
+        //       but cache miss is rare.
+        let old_nblocks = if let Some(nblocks) = self.timeline.get_cached_rel_size(&rel, last_lsn) {
+            nblocks
+        } else if !self
             .timeline
             .get_rel_exists(rel, last_lsn, true, ctx)
             .await?
@@ -1459,8 +1523,15 @@ impl<'a> WalIngest<'a> {
             //info!("extending {} {} to {}", rel, old_nblocks, new_nblocks);
             modification.put_rel_extend(rel, new_nblocks, ctx).await?;
 
+            let mut key = rel_block_to_key(rel, blknum);
             // fill the gap with zeros
             for gap_blknum in old_nblocks..blknum {
+                key.field6 = gap_blknum;
+
+                if self.shard.get_shard_number(&key) != self.shard.number {
+                    continue;
+                }
+
                 modification.put_rel_page_image(rel, gap_blknum, ZERO_PAGE.clone())?;
             }
         }
@@ -2079,4 +2150,88 @@ mod tests {
 
         Ok(())
     }
+
+    /// Replay a wal segment file taken directly from safekeepers.
+    ///
+    /// This test is useful for benchmarking since it allows us to profile only
+    /// the walingest code in a single-threaded executor, and iterate more quickly
+    /// without waiting for unrelated steps.
+    #[tokio::test]
+    async fn test_ingest_real_wal() {
+        use crate::tenant::harness::*;
+        use postgres_ffi::waldecoder::WalStreamDecoder;
+        use postgres_ffi::WAL_SEGMENT_SIZE;
+
+        // Define test data path and constants.
+        //
+        // Steps to reconstruct the data, if needed:
+        // 1. Run the pgbench python test
+        // 2. Take the first wal segment file from safekeeper
+        // 3. Compress it using `zstd --long input_file`
+        // 4. Copy initdb.tar.zst from local_fs_remote_storage
+        // 5. Grep sk logs for "restart decoder" to get startpoint
+        // 6. Run just the decoder from this test to get the endpoint.
+        //    It's the last LSN the decoder will output.
+        let pg_version = 15; // The test data was generated by pg15
+        let path = "test_data/sk_wal_segment_from_pgbench";
+        let wal_segment_path = format!("{path}/000000010000000000000001.zst");
+        let startpoint = Lsn::from_hex("14AEC08").unwrap();
+        let endpoint = Lsn::from_hex("1FFFF98").unwrap();
+
+        // Bootstrap a real timeline. We can't use create_test_timeline because
+        // it doesn't create a real checkpoint, and Walingest::new tries to parse
+        // the garbage data.
+        //
+        // TODO use the initdb.tar.zst file stored with the test data to avoid
+        //      problems with inconsistent initdb results after pg minor version bumps.
+        let (tenant, ctx) = TenantHarness::create("test_ingest_real_wal")
+            .unwrap()
+            .load()
+            .await;
+        let tline = tenant
+            .bootstrap_timeline(TIMELINE_ID, pg_version, None, &ctx)
+            .await
+            .unwrap();
+
+        // We fully read and decompress this into memory before decoding
+        // to get a more accurate perf profile of the decoder.
+        let bytes = {
+            use async_compression::tokio::bufread::ZstdDecoder;
+            let file = tokio::fs::File::open(wal_segment_path).await.unwrap();
+            let reader = tokio::io::BufReader::new(file);
+            let decoder = ZstdDecoder::new(reader);
+            let mut reader = tokio::io::BufReader::new(decoder);
+            let mut buffer = Vec::new();
+            tokio::io::copy_buf(&mut reader, &mut buffer).await.unwrap();
+            buffer
+        };
+
+        // TODO start a profiler too
+        let started_at = std::time::Instant::now();
+
+        // Initialize walingest
+        let xlogoff: usize = startpoint.segment_offset(WAL_SEGMENT_SIZE);
+        let mut decoder = WalStreamDecoder::new(startpoint, pg_version);
+        let mut walingest = WalIngest::new(tline.as_ref(), startpoint, &ctx)
+            .await
+            .unwrap();
+        let mut modification = tline.begin_modification(endpoint);
+        let mut decoded = DecodedWALRecord::default();
+        println!("decoding {} bytes", bytes.len() - xlogoff);
+
+        // Decode and ingest wal. We process the wal in chunks because
+        // that's what happens when we get bytes from safekeepers.
+        for chunk in bytes[xlogoff..].chunks(50) {
+            decoder.feed_bytes(chunk);
+            while let Some((lsn, recdata)) = decoder.poll_decode().unwrap() {
+                walingest
+                    .ingest_record(recdata, lsn, &mut modification, &mut decoded, &ctx)
+                    .await
+                    .unwrap();
+            }
+        }
+
+        let duration = started_at.elapsed();
+        println!("done in {:?}", duration);
+    }
 }

pageserver/src/walredo.rs

@@ -34,17 +34,20 @@ use std::process::{Child, ChildStdin, ChildStdout, Command};
 use std::sync::{Arc, Mutex, MutexGuard, RwLock};
 use std::time::Duration;
 use std::time::Instant;
-use tokio_util::sync::CancellationToken;
 use tracing::*;
 use utils::{bin_ser::BeSer, id::TenantId, lsn::Lsn, nonblock::set_nonblock};
 
 #[cfg(feature = "testing")]
 use std::sync::atomic::{AtomicUsize, Ordering};
 
+#[cfg(feature = "testing")]
+use pageserver_api::shard::TenantShardId;
+
 use crate::config::PageServerConf;
 use crate::metrics::{
     WalRedoKillCause, WAL_REDO_BYTES_HISTOGRAM, WAL_REDO_PROCESS_COUNTERS,
-    WAL_REDO_RECORDS_HISTOGRAM, WAL_REDO_RECORD_COUNTER, WAL_REDO_TIME,
+    WAL_REDO_PROCESS_LAUNCH_DURATION_HISTOGRAM, WAL_REDO_RECORDS_HISTOGRAM,
+    WAL_REDO_RECORD_COUNTER, WAL_REDO_TIME,
 };
 use crate::pgdatadir_mapping::{key_to_rel_block, key_to_slru_block};
 use crate::repository::Key;
@@ -120,7 +123,9 @@ impl PostgresRedoManager {
     /// The WAL redo is handled by a separate thread, so this just sends a request
     /// to the thread and waits for response.
     ///
-    /// CANCEL SAFETY: NOT CANCEL SAFE.
+    /// # Cancel-Safety
+    ///
+    /// This method is cancellation-safe.
     pub async fn request_redo(
         &self,
         key: Key,
@@ -153,7 +158,6 @@ impl PostgresRedoManager {
                         self.conf.wal_redo_timeout,
                         pg_version,
                     )
-                    .await
                 };
                 img = Some(result?);
 
@@ -174,7 +178,6 @@ impl PostgresRedoManager {
                 self.conf.wal_redo_timeout,
                 pg_version,
             )
-            .await
         }
     }
 }
@@ -212,7 +215,7 @@ impl PostgresRedoManager {
     /// Process one request for WAL redo using wal-redo postgres
     ///
     #[allow(clippy::too_many_arguments)]
-    async fn apply_batch_postgres(
+    fn apply_batch_postgres(
         &self,
         key: Key,
         lsn: Lsn,
@@ -238,10 +241,13 @@ impl PostgresRedoManager {
                         let mut proc_guard = self.redo_process.write().unwrap();
                         match &*proc_guard {
                             None => {
+                                let timer =
+                                    WAL_REDO_PROCESS_LAUNCH_DURATION_HISTOGRAM.start_timer();
                                 let proc = Arc::new(
                                     WalRedoProcess::launch(self.conf, self.tenant_id, pg_version)
                                         .context("launch walredo process")?,
                                 );
+                                timer.observe_duration();
                                 *proc_guard = Some(Arc::clone(&proc));
                                 proc
                             }
@@ -325,12 +331,7 @@ impl PostgresRedoManager {
                 // than we can SIGKILL & `wait` for them to exit. By doing it the way we do here,
                 // we limit this risk of run-away to at most $num_runtimes * $num_executor_threads.
                 // This probably needs revisiting at some later point.
-                let mut wait_done = proc.stderr_logger_task_done.clone();
                 drop(proc);
-                wait_done
-                    .wait_for(|v| *v)
-                    .await
-                    .expect("we use scopeguard to ensure we always send `true` to the channel before dropping the sender");
             } else if n_attempts != 0 {
                 info!(n_attempts, "retried walredo succeeded");
             }
@@ -642,8 +643,6 @@ struct WalRedoProcess {
     child: Option<NoLeakChild>,
     stdout: Mutex<ProcessOutput>,
     stdin: Mutex<ProcessInput>,
-    stderr_logger_cancel: CancellationToken,
-    stderr_logger_task_done: tokio::sync::watch::Receiver<bool>,
     /// Counter to separate same sized walredo inputs failing at the same millisecond.
     #[cfg(feature = "testing")]
     dump_sequence: AtomicUsize,
@@ -692,6 +691,8 @@ impl WalRedoProcess {
         let stdin = child.stdin.take().unwrap();
         let stdout = child.stdout.take().unwrap();
         let stderr = child.stderr.take().unwrap();
+        let stderr = tokio::process::ChildStderr::from_std(stderr)
+            .context("convert to tokio::ChildStderr")?;
         macro_rules! set_nonblock_or_log_err {
             ($file:ident) => {{
                 let res = set_nonblock($file.as_raw_fd());
@@ -703,69 +704,45 @@ impl WalRedoProcess {
         }
         set_nonblock_or_log_err!(stdin)?;
         set_nonblock_or_log_err!(stdout)?;
-        set_nonblock_or_log_err!(stderr)?;
-
-        let mut stderr = tokio::io::unix::AsyncFd::new(stderr).context("AsyncFd::with_interest")?;
 
         // all fallible operations post-spawn are complete, so get rid of the guard
         let child = scopeguard::ScopeGuard::into_inner(child);
 
-        let stderr_logger_cancel = CancellationToken::new();
-        let (stderr_logger_task_done_tx, stderr_logger_task_done_rx) =
-            tokio::sync::watch::channel(false);
-        tokio::spawn({
-            let stderr_logger_cancel = stderr_logger_cancel.clone();
+        tokio::spawn(
             async move {
                 scopeguard::defer! {
                     debug!("wal-redo-postgres stderr_logger_task finished");
-                    let _ = stderr_logger_task_done_tx.send(true);
+                    crate::metrics::WAL_REDO_PROCESS_COUNTERS.active_stderr_logger_tasks_finished.inc();
                 }
                 debug!("wal-redo-postgres stderr_logger_task started");
-                loop {
-                    // NB: we purposefully don't do a select! for the cancellation here.
-                    // The cancellation would likely cause us to miss stderr messages.
-                    // We can rely on this to return from .await because when we SIGKILL
-                    // the child, the writing end of the stderr pipe gets closed.
-                    match stderr.readable_mut().await {
-                        Ok(mut guard) => {
-                            let mut errbuf = [0; 16384];
-                            let res = guard.try_io(|fd| {
-                                use std::io::Read;
-                                fd.get_mut().read(&mut errbuf)
-                            });
-                            match res {
-                                Ok(Ok(0)) => {
-                                    // it closed the stderr pipe
-                                    break;
-                                }
-                                Ok(Ok(n)) => {
-                                    // The message might not be split correctly into lines here. But this is
-                                    // good enough, the important thing is to get the message to the log.
-                                    let output = String::from_utf8_lossy(&errbuf[0..n]).to_string();
-                                    error!(output, "received output");
-                                },
-                                Ok(Err(e)) => {
-                                    error!(error = ?e, "read() error, waiting for cancellation");
-                                    stderr_logger_cancel.cancelled().await;
-                                    error!(error = ?e, "read() error, cancellation complete");
-                                    break;
-                                }
-                                Err(e) => {
-                                    let _e: tokio::io::unix::TryIoError = e;
-                                    // the read() returned WouldBlock, that's expected
-                                }
-                            }
+                crate::metrics::WAL_REDO_PROCESS_COUNTERS.active_stderr_logger_tasks_started.inc();
+
+                use tokio::io::AsyncBufReadExt;
+                let mut stderr_lines = tokio::io::BufReader::new(stderr);
+                let mut buf = Vec::new();
+                let res = loop {
+                    buf.clear();
+                    // TODO we don't trust the process to cap its stderr length.
+                    // Currently it can do unbounded Vec allocation.
+                    match stderr_lines.read_until(b'\n', &mut buf).await {
+                        Ok(0) => break Ok(()), // eof
+                        Ok(num_bytes) => {
+                            let output = String::from_utf8_lossy(&buf[..num_bytes]);
+                            error!(%output, "received output");
                         }
                         Err(e) => {
-                            error!(error = ?e, "read() error, waiting for cancellation");
-                            stderr_logger_cancel.cancelled().await;
-                            error!(error = ?e, "read() error, cancellation complete");
-                            break;
+                            break Err(e);
                         }
                     }
+                };
+                match res {
+                    Ok(()) => (),
+                    Err(e) => {
+                        error!(error=?e, "failed to read from walredo stderr");
+                    }
                 }
             }.instrument(tracing::info_span!(parent: None, "wal-redo-postgres-stderr", pid = child.id(), tenant_id = %tenant_id, %pg_version))
-        });
+        );
 
         Ok(Self {
             conf,
@@ -780,8 +757,6 @@ impl WalRedoProcess {
                 pending_responses: VecDeque::new(),
                 n_processed_responses: 0,
             }),
-            stderr_logger_cancel,
-            stderr_logger_task_done: stderr_logger_task_done_rx,
             #[cfg(feature = "testing")]
             dump_sequence: AtomicUsize::default(),
         })
@@ -991,7 +966,11 @@ impl WalRedoProcess {
         // these files will be collected to an allure report
         let filename = format!("walredo-{millis}-{}-{seq}.walredo", writebuf.len());
 
-        let path = self.conf.tenant_path(&self.tenant_id).join(&filename);
+        // TODO(sharding): update this call when WalRedoProcess gets a TenantShardId.
+        let path = self
+            .conf
+            .tenant_path(&TenantShardId::unsharded(self.tenant_id))
+            .join(&filename);
 
         let res = std::fs::OpenOptions::new()
             .write(true)
@@ -1018,7 +997,6 @@ impl Drop for WalRedoProcess {
             .take()
             .expect("we only do this once")
             .kill_and_wait(WalRedoKillCause::WalRedoProcessDrop);
-        self.stderr_logger_cancel.cancel();
         // no way to wait for stderr_logger_task from Drop because that is async only
     }
 }
@@ -1182,7 +1160,7 @@ mod tests {
 
     #[tokio::test]
     async fn short_v14_redo() {
-        let expected = std::fs::read("fixtures/short_v14_redo.page").unwrap();
+        let expected = std::fs::read("test_data/short_v14_redo.page").unwrap();
 
         let h = RedoHarness::new().unwrap();
 

pgxn/neon/Makefile

@@ -41,6 +41,17 @@ libwalproposer.a: $(WALPROP_OBJS)
 	rm -f $@
 	$(AR) $(AROPT) $@ $^
 
+# needs vars:
+# FIND_TYPEDEF pointing to find_typedef
+# INDENT pointing to pg_bsd_indent
+# PGINDENT_SCRIPT pointing to pgindent (be careful with PGINDENT var name:
+#   pgindent will pick it up as pg_bsd_indent path).
+.PHONY: pgindent
+pgindent:
+	+@ echo top_srcdir=$(top_srcdir) top_builddir=$(top_builddir) srcdir=$(srcdir)
+	$(FIND_TYPEDEF) . > neon.typedefs
+	INDENT=$(INDENT) $(PGINDENT_SCRIPT) --typedefs neon.typedefs $(srcdir)/*.c $(srcdir)/*.h
+
 PG_CONFIG = pg_config
 PGXS := $(shell $(PG_CONFIG) --pgxs)
 include $(PGXS)

pgxn/neon/control_plane_connector.c

@@ -41,7 +41,7 @@ static char *ConsoleURL = NULL;
 static bool ForwardDDL = true;
 
 /* Curl structures for sending the HTTP requests */
-static CURL * CurlHandle;
+static CURL *CurlHandle;
 static struct curl_slist *ContentHeader = NULL;
 
 /*
@@ -54,7 +54,7 @@ typedef enum
 {
 	Op_Set,						/* An upsert: Either a creation or an alter */
 	Op_Delete,
-}			OpType;
+} OpType;
 
 typedef struct
 {
@@ -62,7 +62,7 @@ typedef struct
 	Oid			owner;
 	char		old_name[NAMEDATALEN];
 	OpType		type;
-}			DbEntry;
+} DbEntry;
 
 typedef struct
 {
@@ -70,7 +70,7 @@ typedef struct
 	char		old_name[NAMEDATALEN];
 	const char *password;
 	OpType		type;
-}			RoleEntry;
+} RoleEntry;
 
 /*
  * We keep one of these for each subtransaction in a stack. When a subtransaction
@@ -82,10 +82,10 @@ typedef struct DdlHashTable
 	struct DdlHashTable *prev_table;
 	HTAB	   *db_table;
 	HTAB	   *role_table;
-}			DdlHashTable;
+} DdlHashTable;
 
 static DdlHashTable RootTable;
-static DdlHashTable * CurrentDdlTable = &RootTable;
+static DdlHashTable *CurrentDdlTable = &RootTable;
 
 static void
 PushKeyValue(JsonbParseState **state, char *key, char *value)
@@ -199,7 +199,7 @@ typedef struct
 {
 	char		str[ERROR_SIZE];
 	size_t		size;
-}			ErrorString;
+} ErrorString;
 
 static size_t
 ErrorWriteCallback(char *ptr, size_t size, size_t nmemb, void *userdata)
@@ -478,7 +478,7 @@ NeonXactCallback(XactEvent event, void *arg)
 static bool
 RoleIsNeonSuperuser(const char *role_name)
 {
-    return strcmp(role_name, "neon_superuser") == 0;
+	return strcmp(role_name, "neon_superuser") == 0;
 }
 
 static void
@@ -509,6 +509,7 @@ HandleCreateDb(CreatedbStmt *stmt)
 	if (downer && downer->arg)
 	{
 		const char *owner_name = defGetString(downer);
+
 		if (RoleIsNeonSuperuser(owner_name))
 			elog(ERROR, "can't create a database with owner neon_superuser");
 		entry->owner = get_role_oid(owner_name, false);
@@ -536,6 +537,7 @@ HandleAlterOwner(AlterOwnerStmt *stmt)
 	if (!found)
 		memset(entry->old_name, 0, sizeof(entry->old_name));
 	const char *new_owner = get_rolespec_name(stmt->newowner);
+
 	if (RoleIsNeonSuperuser(new_owner))
 		elog(ERROR, "can't alter owner to neon_superuser");
 	entry->owner = get_role_oid(new_owner, false);
@@ -633,6 +635,7 @@ HandleAlterRole(AlterRoleStmt *stmt)
 	DefElem    *dpass = NULL;
 	ListCell   *option;
 	const char *role_name = stmt->role->rolename;
+
 	if (RoleIsNeonSuperuser(role_name))
 		elog(ERROR, "can't ALTER neon_superuser");
 

pgxn/neon/extension_server.c

@@ -25,79 +25,81 @@
 
 #include <curl/curl.h>
 
-static int extension_server_port = 0;
+static int	extension_server_port = 0;
 
 static download_extension_file_hook_type prev_download_extension_file_hook = NULL;
 
-// to download all SQL (and data) files for an extension:
-// curl -X POST http://localhost:8080/extension_server/postgis
-// it covers two possible extension files layouts:
-// 1. extension_name--version--platform.sql
-// 2. extension_name/extension_name--version.sql
-//    extension_name/extra_files.csv
-//
-// to download specific library file:
-// curl -X POST http://localhost:8080/extension_server/postgis-3.so?is_library=true
+/*
+  * to download all SQL (and data) files for an extension:
+  * curl -X POST http://localhost:8080/extension_server/postgis
+  * it covers two possible extension files layouts:
+  * 1. extension_name--version--platform.sql
+  * 2. extension_name/extension_name--version.sql
+  *    extension_name/extra_files.csv
+  * to download specific library file:
+  * curl -X POST http://localhost:8080/extension_server/postgis-3.so?is_library=true
+  */
 static bool
 neon_download_extension_file_http(const char *filename, bool is_library)
 {
-    CURL *curl;
-    CURLcode res;
-    char *compute_ctl_url;
-    char *postdata;
-    bool ret = false;
+	CURL	   *curl;
+	CURLcode	res;
+	char	   *compute_ctl_url;
+	char	   *postdata;
+	bool		ret = false;
 
-    if ((curl = curl_easy_init()) == NULL)
-    {
-        elog(ERROR, "Failed to initialize curl handle");
-    }
+	if ((curl = curl_easy_init()) == NULL)
+	{
+		elog(ERROR, "Failed to initialize curl handle");
+	}
 
-    compute_ctl_url = psprintf("http://localhost:%d/extension_server/%s%s",
-                               extension_server_port, filename, is_library ? "?is_library=true" : "");
+	compute_ctl_url = psprintf("http://localhost:%d/extension_server/%s%s",
+							   extension_server_port, filename, is_library ? "?is_library=true" : "");
 
-    elog(LOG, "Sending request to compute_ctl: %s", compute_ctl_url);
+	elog(LOG, "Sending request to compute_ctl: %s", compute_ctl_url);
 
-    curl_easy_setopt(curl, CURLOPT_CUSTOMREQUEST, "POST");
-    curl_easy_setopt(curl, CURLOPT_URL, compute_ctl_url);
-    curl_easy_setopt(curl, CURLOPT_TIMEOUT, 3L /* seconds */);
+	curl_easy_setopt(curl, CURLOPT_CUSTOMREQUEST, "POST");
+	curl_easy_setopt(curl, CURLOPT_URL, compute_ctl_url);
+	curl_easy_setopt(curl, CURLOPT_TIMEOUT, 3L /* seconds */ );
 
-    if (curl)
-    {
-        /* Perform the request, res will get the return code */
-        res = curl_easy_perform(curl);
-        /* Check for errors */
-        if (res == CURLE_OK)
-        {
-            ret = true;
-        }
-        else
-        {
-            // Don't error here because postgres will try to find the file
-            // and will fail with some proper error message if it's not found.
-            elog(WARNING, "neon_download_extension_file_http failed: %s\n", curl_easy_strerror(res));
-        }
+	if (curl)
+	{
+		/* Perform the request, res will get the return code */
+		res = curl_easy_perform(curl);
+		/* Check for errors */
+		if (res == CURLE_OK)
+		{
+			ret = true;
+		}
+		else
+		{
+			/* Don't error here because postgres will try to find the file */
+			/* and will fail with some proper error message if it's not found. */
+			elog(WARNING, "neon_download_extension_file_http failed: %s\n", curl_easy_strerror(res));
+		}
 
-        /* always cleanup */
-        curl_easy_cleanup(curl);
-    }
+		/* always cleanup */
+		curl_easy_cleanup(curl);
+	}
 
-    return ret;
+	return ret;
 }
 
-void pg_init_extension_server()
+void
+pg_init_extension_server()
 {
-    // Port to connect to compute_ctl on localhost
-    // to request extension files.
-    DefineCustomIntVariable("neon.extension_server_port",
-                            "connection string to the compute_ctl",
-                            NULL,
-                            &extension_server_port,
-                            0, 0, INT_MAX,
-                            PGC_POSTMASTER,
-                            0, /* no flags required */
-                            NULL, NULL, NULL);
+	/* Port to connect to compute_ctl on localhost */
+	/* to request extension files. */
+	DefineCustomIntVariable("neon.extension_server_port",
+							"connection string to the compute_ctl",
+							NULL,
+							&extension_server_port,
+							0, 0, INT_MAX,
+							PGC_POSTMASTER,
+							0,	/* no flags required */
+							NULL, NULL, NULL);
 
-    // set download_extension_file_hook
-    prev_download_extension_file_hook = download_extension_file_hook;
-    download_extension_file_hook = neon_download_extension_file_http;
+	/* set download_extension_file_hook */
+	prev_download_extension_file_hook = download_extension_file_hook;
+	download_extension_file_hook = neon_download_extension_file_http;
 }

pgxn/neon/file_cache.c

@@ -67,32 +67,34 @@
 typedef struct FileCacheEntry
 {
 	BufferTag	key;
-	uint32      hash;
+	uint32		hash;
 	uint32		offset;
 	uint32		access_count;
-	uint32		bitmap[BLOCKS_PER_CHUNK/32];
-	dlist_node	lru_node; /* LRU list node */
+	uint32		bitmap[BLOCKS_PER_CHUNK / 32];
+	dlist_node	lru_node;		/* LRU list node */
 } FileCacheEntry;
 
 typedef struct FileCacheControl
 {
-	uint64 generation; /* generation is needed to handle correct hash reenabling */
-	uint32 size; /* size of cache file in chunks */
-	uint32 used; /* number of used chunks */
-	uint32 limit; /* shared copy of lfc_size_limit */
-	uint64 hits;
-	uint64 misses;
-	uint64 writes;
-	dlist_head lru; /* double linked list for LRU replacement algorithm */
+	uint64		generation;		/* generation is needed to handle correct hash
+								 * reenabling */
+	uint32		size;			/* size of cache file in chunks */
+	uint32		used;			/* number of used chunks */
+	uint32		limit;			/* shared copy of lfc_size_limit */
+	uint64		hits;
+	uint64		misses;
+	uint64		writes;
+	dlist_head	lru;			/* double linked list for LRU replacement
+								 * algorithm */
 } FileCacheControl;
 
-static HTAB* lfc_hash;
-static int   lfc_desc = 0;
+static HTAB *lfc_hash;
+static int	lfc_desc = 0;
 static LWLockId lfc_lock;
-static int   lfc_max_size;
-static int   lfc_size_limit;
-static char* lfc_path;
-static  FileCacheControl* lfc_ctl;
+static int	lfc_max_size;
+static int	lfc_size_limit;
+static char *lfc_path;
+static FileCacheControl *lfc_ctl;
 static shmem_startup_hook_type prev_shmem_startup_hook;
 #if PG_VERSION_NUM>=150000
 static shmem_request_hook_type prev_shmem_request_hook;
@@ -100,7 +102,7 @@ static shmem_request_hook_type prev_shmem_request_hook;
 
 #define LFC_ENABLED() (lfc_ctl->limit != 0)
 
-void PGDLLEXPORT FileCacheMonitorMain(Datum main_arg);
+void		PGDLLEXPORT FileCacheMonitorMain(Datum main_arg);
 
 /*
  * Local file cache is optional and Neon can work without it.
@@ -109,9 +111,10 @@ void PGDLLEXPORT FileCacheMonitorMain(Datum main_arg);
  * All cache content should be invalidated to avoid reading of stale or corrupted data
  */
 static void
-lfc_disable(char const* op)
+lfc_disable(char const *op)
 {
-	int fd;
+	int			fd;
+
 	elog(WARNING, "Failed to %s local file cache at %s: %m, disabling local file cache", op, lfc_path);
 
 	/* Invalidate hash */
@@ -120,7 +123,7 @@ lfc_disable(char const* op)
 	if (LFC_ENABLED())
 	{
 		HASH_SEQ_STATUS status;
-		FileCacheEntry* entry;
+		FileCacheEntry *entry;
 
 		hash_seq_init(&status, lfc_hash);
 		while ((entry = hash_seq_search(&status)) != NULL)
@@ -135,16 +138,24 @@ lfc_disable(char const* op)
 
 		if (lfc_desc > 0)
 		{
-			/* If the reason of error is ENOSPC, then truncation of file may help to reclaim some space */
-			int rc = ftruncate(lfc_desc, 0);
+			/*
+			 * If the reason of error is ENOSPC, then truncation of file may
+			 * help to reclaim some space
+			 */
+			int			rc = ftruncate(lfc_desc, 0);
+
 			if (rc < 0)
 				elog(WARNING, "Failed to truncate local file cache %s: %m", lfc_path);
 		}
 	}
-	/* We need to use unlink to to avoid races in LFC write, because it is not protectedby */
+
+	/*
+	 * We need to use unlink to to avoid races in LFC write, because it is not
+	 * protectedby
+	 */
 	unlink(lfc_path);
 
-	fd = BasicOpenFile(lfc_path, O_RDWR|O_CREAT|O_TRUNC);
+	fd = BasicOpenFile(lfc_path, O_RDWR | O_CREAT | O_TRUNC);
 	if (fd < 0)
 		elog(WARNING, "Failed to recreate local file cache %s: %m", lfc_path);
 	else
@@ -170,13 +181,15 @@ lfc_maybe_disabled(void)
 static bool
 lfc_ensure_opened(void)
 {
-	bool enabled = !lfc_maybe_disabled();
+	bool		enabled = !lfc_maybe_disabled();
+
 	/* Open cache file if not done yet */
 	if (lfc_desc <= 0 && enabled)
 	{
 		lfc_desc = BasicOpenFile(lfc_path, O_RDWR);
 
-		if (lfc_desc < 0) {
+		if (lfc_desc < 0)
+		{
 			lfc_disable("open");
 			return false;
 		}
@@ -187,7 +200,7 @@ lfc_ensure_opened(void)
 static void
 lfc_shmem_startup(void)
 {
-	bool found;
+	bool		found;
 	static HASHCTL info;
 
 	if (prev_shmem_startup_hook)
@@ -197,17 +210,22 @@ lfc_shmem_startup(void)
 
 	LWLockAcquire(AddinShmemInitLock, LW_EXCLUSIVE);
 
-	lfc_ctl = (FileCacheControl*)ShmemInitStruct("lfc", sizeof(FileCacheControl), &found);
+	lfc_ctl = (FileCacheControl *) ShmemInitStruct("lfc", sizeof(FileCacheControl), &found);
 	if (!found)
 	{
-		int fd;
-		uint32 lfc_size = SIZE_MB_TO_CHUNKS(lfc_max_size);
-		lfc_lock = (LWLockId)GetNamedLWLockTranche("lfc_lock");
+		int			fd;
+		uint32		lfc_size = SIZE_MB_TO_CHUNKS(lfc_max_size);
+
+		lfc_lock = (LWLockId) GetNamedLWLockTranche("lfc_lock");
 		info.keysize = sizeof(BufferTag);
 		info.entrysize = sizeof(FileCacheEntry);
+
+		/*
+		 * lfc_size+1 because we add new element to hash table before eviction
+		 * of victim
+		 */
 		lfc_hash = ShmemInitHash("lfc_hash",
-								 /* lfc_size+1 because we add new element to hash table before eviction of victim */
-								 lfc_size+1, lfc_size+1,
+								 lfc_size + 1, lfc_size + 1,
 								 &info,
 								 HASH_ELEM | HASH_BLOBS);
 		lfc_ctl->generation = 0;
@@ -219,7 +237,7 @@ lfc_shmem_startup(void)
 		dlist_init(&lfc_ctl->lru);
 
 		/* Recreate file cache on restart */
-		fd = BasicOpenFile(lfc_path, O_RDWR|O_CREAT|O_TRUNC);
+		fd = BasicOpenFile(lfc_path, O_RDWR | O_CREAT | O_TRUNC);
 		if (fd < 0)
 		{
 			elog(WARNING, "Failed to create local file cache %s: %m", lfc_path);
@@ -242,7 +260,7 @@ lfc_shmem_request(void)
 		prev_shmem_request_hook();
 #endif
 
-	RequestAddinShmemSpace(sizeof(FileCacheControl) + hash_estimate_size(SIZE_MB_TO_CHUNKS(lfc_max_size)+1, sizeof(FileCacheEntry)));
+	RequestAddinShmemSpace(sizeof(FileCacheControl) + hash_estimate_size(SIZE_MB_TO_CHUNKS(lfc_max_size) + 1, sizeof(FileCacheEntry)));
 	RequestNamedLWLockTranche("lfc_lock", 1);
 }
 
@@ -250,9 +268,11 @@ static bool
 is_normal_backend(void)
 {
 	/*
-	 * Stats collector detach shared memory, so we should not try to access shared memory here.
-	 * Parallel workers first assign default value (0), so not perform truncation in parallel workers.
-	 * The Postmaster can handle SIGHUP and it has access to shared memory (UsedShmemSegAddr != NULL), but has no PGPROC.
+	 * Stats collector detach shared memory, so we should not try to access
+	 * shared memory here. Parallel workers first assign default value (0), so
+	 * not perform truncation in parallel workers. The Postmaster can handle
+	 * SIGHUP and it has access to shared memory (UsedShmemSegAddr != NULL),
+	 * but has no PGPROC.
 	 */
 	return lfc_ctl && MyProc && UsedShmemSegAddr && !IsParallelWorker();
 }
@@ -271,7 +291,7 @@ lfc_check_limit_hook(int *newval, void **extra, GucSource source)
 static void
 lfc_change_limit_hook(int newval, void *extra)
 {
-	uint32 new_size = SIZE_MB_TO_CHUNKS(newval);
+	uint32		new_size = SIZE_MB_TO_CHUNKS(newval);
 
 	if (!is_normal_backend())
 		return;
@@ -283,11 +303,15 @@ lfc_change_limit_hook(int newval, void *extra)
 
 	while (new_size < lfc_ctl->used && !dlist_is_empty(&lfc_ctl->lru))
 	{
-		/* Shrink cache by throwing away least recently accessed chunks and returning their space to file system */
-		FileCacheEntry* victim = dlist_container(FileCacheEntry, lru_node, dlist_pop_head_node(&lfc_ctl->lru));
+		/*
+		 * Shrink cache by throwing away least recently accessed chunks and
+		 * returning their space to file system
+		 */
+		FileCacheEntry *victim = dlist_container(FileCacheEntry, lru_node, dlist_pop_head_node(&lfc_ctl->lru));
+
 		Assert(victim->access_count == 0);
 #ifdef FALLOC_FL_PUNCH_HOLE
-		if (fallocate(lfc_desc, FALLOC_FL_PUNCH_HOLE|FALLOC_FL_KEEP_SIZE, (off_t)victim->offset*BLOCKS_PER_CHUNK*BLCKSZ, BLOCKS_PER_CHUNK*BLCKSZ) < 0)
+		if (fallocate(lfc_desc, FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE, (off_t) victim->offset * BLOCKS_PER_CHUNK * BLCKSZ, BLOCKS_PER_CHUNK * BLCKSZ) < 0)
 			elog(LOG, "Failed to punch hole in file: %m");
 #endif
 		hash_search_with_hash_value(lfc_hash, &victim->key, victim->hash, HASH_REMOVE, NULL);
@@ -314,7 +338,7 @@ lfc_init(void)
 							"Maximal size of Neon local file cache",
 							NULL,
 							&lfc_max_size,
-							0, /* disabled by default */
+							0,	/* disabled by default */
 							0,
 							INT_MAX,
 							PGC_POSTMASTER,
@@ -327,7 +351,7 @@ lfc_init(void)
 							"Current limit for size of Neon local file cache",
 							NULL,
 							&lfc_size_limit,
-							0, /* disabled by default */
+							0,	/* disabled by default */
 							0,
 							INT_MAX,
 							PGC_SIGHUP,
@@ -367,18 +391,18 @@ lfc_init(void)
 bool
 lfc_cache_contains(NRelFileInfo rinfo, ForkNumber forkNum, BlockNumber blkno)
 {
-	BufferTag tag;
-	FileCacheEntry* entry;
-	int chunk_offs = blkno & (BLOCKS_PER_CHUNK-1);
-	bool found = false;
-	uint32 hash;
+	BufferTag	tag;
+	FileCacheEntry *entry;
+	int			chunk_offs = blkno & (BLOCKS_PER_CHUNK - 1);
+	bool		found = false;
+	uint32		hash;
 
-	if (lfc_maybe_disabled()) /* fast exit if file cache is disabled */
+	if (lfc_maybe_disabled())	/* fast exit if file cache is disabled */
 		return false;
 
 	CopyNRelFileInfoToBufTag(tag, rinfo);
 	tag.forkNum = forkNum;
-	tag.blockNum = blkno & ~(BLOCKS_PER_CHUNK-1);
+	tag.blockNum = blkno & ~(BLOCKS_PER_CHUNK - 1);
 	hash = get_hash_value(lfc_hash, &tag);
 
 	LWLockAcquire(lfc_lock, LW_SHARED);
@@ -397,13 +421,13 @@ lfc_cache_contains(NRelFileInfo rinfo, ForkNumber forkNum, BlockNumber blkno)
 void
 lfc_evict(NRelFileInfo rinfo, ForkNumber forkNum, BlockNumber blkno)
 {
-	BufferTag tag;
-	FileCacheEntry* entry;
-	bool found;
-	int chunk_offs = blkno & (BLOCKS_PER_CHUNK-1);
-	uint32 hash;
+	BufferTag	tag;
+	FileCacheEntry *entry;
+	bool		found;
+	int			chunk_offs = blkno & (BLOCKS_PER_CHUNK - 1);
+	uint32		hash;
 
-	if (lfc_maybe_disabled()) /* fast exit if file cache is disabled */
+	if (lfc_maybe_disabled())	/* fast exit if file cache is disabled */
 		return;
 
 	CopyNRelFileInfoToBufTag(tag, rinfo);
@@ -438,9 +462,10 @@ lfc_evict(NRelFileInfo rinfo, ForkNumber forkNum, BlockNumber blkno)
 	 */
 	if (entry->bitmap[chunk_offs >> 5] == 0)
 	{
-		bool has_remaining_pages;
+		bool		has_remaining_pages;
 
-		for (int i = 0; i < (BLOCKS_PER_CHUNK / 32); i++) {
+		for (int i = 0; i < (BLOCKS_PER_CHUNK / 32); i++)
+		{
 			if (entry->bitmap[i] != 0)
 			{
 				has_remaining_pages = true;
@@ -449,8 +474,8 @@ lfc_evict(NRelFileInfo rinfo, ForkNumber forkNum, BlockNumber blkno)
 		}
 
 		/*
-		 * Put the entry at the position that is first to be reclaimed when
-		 * we have no cached pages remaining in the chunk
+		 * Put the entry at the position that is first to be reclaimed when we
+		 * have no cached pages remaining in the chunk
 		 */
 		if (!has_remaining_pages)
 		{
@@ -476,16 +501,16 @@ bool
 lfc_read(NRelFileInfo rinfo, ForkNumber forkNum, BlockNumber blkno,
 		 char *buffer)
 {
-	BufferTag tag;
-	FileCacheEntry* entry;
-	ssize_t rc;
-	int chunk_offs = blkno & (BLOCKS_PER_CHUNK-1);
-	bool result = true;
-	uint32 hash;
-	uint64 generation;
-	uint32 entry_offset;
+	BufferTag	tag;
+	FileCacheEntry *entry;
+	ssize_t		rc;
+	int			chunk_offs = blkno & (BLOCKS_PER_CHUNK - 1);
+	bool		result = true;
+	uint32		hash;
+	uint64		generation;
+	uint32		entry_offset;
 
-	if (lfc_maybe_disabled()) /* fast exit if file cache is disabled */
+	if (lfc_maybe_disabled())	/* fast exit if file cache is disabled */
 		return false;
 
 	if (!lfc_ensure_opened())
@@ -493,7 +518,7 @@ lfc_read(NRelFileInfo rinfo, ForkNumber forkNum, BlockNumber blkno,
 
 	CopyNRelFileInfoToBufTag(tag, rinfo);
 	tag.forkNum = forkNum;
-	tag.blockNum = blkno & ~(BLOCKS_PER_CHUNK-1);
+	tag.blockNum = blkno & ~(BLOCKS_PER_CHUNK - 1);
 	hash = get_hash_value(lfc_hash, &tag);
 
 	LWLockAcquire(lfc_lock, LW_EXCLUSIVE);
@@ -520,7 +545,7 @@ lfc_read(NRelFileInfo rinfo, ForkNumber forkNum, BlockNumber blkno,
 
 	LWLockRelease(lfc_lock);
 
-	rc = pread(lfc_desc, buffer, BLCKSZ, ((off_t)entry_offset*BLOCKS_PER_CHUNK + chunk_offs)*BLCKSZ);
+	rc = pread(lfc_desc, buffer, BLCKSZ, ((off_t) entry_offset * BLOCKS_PER_CHUNK + chunk_offs) * BLCKSZ);
 	if (rc != BLCKSZ)
 	{
 		lfc_disable("read");
@@ -551,30 +576,29 @@ lfc_read(NRelFileInfo rinfo, ForkNumber forkNum, BlockNumber blkno,
  * If cache is full then evict some other page.
  */
 void
-lfc_write(NRelFileInfo rinfo, ForkNumber forkNum, BlockNumber blkno,
 #if PG_MAJORVERSION_NUM < 16
-		  char *buffer)
+lfc_write(NRelFileInfo rinfo, ForkNumber forkNum, BlockNumber blkno, char *buffer)
 #else
-		  const void *buffer)
+lfc_write(NRelFileInfo rinfo, ForkNumber forkNum, BlockNumber blkno, const void *buffer)
 #endif
 {
-	BufferTag tag;
-	FileCacheEntry* entry;
-	ssize_t rc;
-	bool found;
-	int chunk_offs = blkno & (BLOCKS_PER_CHUNK-1);
-	uint32 hash;
-	uint64 generation;
-	uint32 entry_offset;
+	BufferTag	tag;
+	FileCacheEntry *entry;
+	ssize_t		rc;
+	bool		found;
+	int			chunk_offs = blkno & (BLOCKS_PER_CHUNK - 1);
+	uint32		hash;
+	uint64		generation;
+	uint32		entry_offset;
 
-	if (lfc_maybe_disabled()) /* fast exit if file cache is disabled */
+	if (lfc_maybe_disabled())	/* fast exit if file cache is disabled */
 		return;
 
 	if (!lfc_ensure_opened())
 		return;
 
 	tag.forkNum = forkNum;
-	tag.blockNum = blkno & ~(BLOCKS_PER_CHUNK-1);
+	tag.blockNum = blkno & ~(BLOCKS_PER_CHUNK - 1);
 	CopyNRelFileInfoToBufTag(tag, rinfo);
 	hash = get_hash_value(lfc_hash, &tag);
 
@@ -590,24 +614,36 @@ lfc_write(NRelFileInfo rinfo, ForkNumber forkNum, BlockNumber blkno,
 
 	if (found)
 	{
-		/* Unlink entry from LRU list to pin it for the duration of IO operation */
+		/*
+		 * Unlink entry from LRU list to pin it for the duration of IO
+		 * operation
+		 */
 		if (entry->access_count++ == 0)
 			dlist_delete(&entry->lru_node);
 	}
 	else
 	{
 		/*
-		 * We have two choices if all cache pages are pinned (i.e. used in IO operations):
-		 * 1. Wait until some of this operation is completed and pages is unpinned
-		 * 2. Allocate one more chunk, so that specified cache size is more recommendation than hard limit.
-		 * As far as probability of such event (that all pages are pinned) is considered to be very very small:
-		 * there are should be very large number of concurrent IO operations and them are limited by max_connections,
-		 * we prefer not to complicate code and use second approach.
+		 * We have two choices if all cache pages are pinned (i.e. used in IO
+		 * operations):
+		 *
+		 * 1) Wait until some of this operation is completed and pages is
+		 * unpinned.
+		 *
+		 * 2) Allocate one more chunk, so that specified cache size is more
+		 * recommendation than hard limit.
+		 *
+		 * As far as probability of such event (that all pages are pinned) is
+		 * considered to be very very small: there are should be very large
+		 * number of concurrent IO operations and them are limited by
+		 * max_connections, we prefer not to complicate code and use second
+		 * approach.
 		 */
 		if (lfc_ctl->used >= lfc_ctl->limit && !dlist_is_empty(&lfc_ctl->lru))
 		{
 			/* Cache overflow: evict least recently used chunk */
-			FileCacheEntry* victim = dlist_container(FileCacheEntry, lru_node, dlist_pop_head_node(&lfc_ctl->lru));
+			FileCacheEntry *victim = dlist_container(FileCacheEntry, lru_node, dlist_pop_head_node(&lfc_ctl->lru));
+
 			Assert(victim->access_count == 0);
 			entry->offset = victim->offset; /* grab victim's chunk */
 			hash_search_with_hash_value(lfc_hash, &victim->key, victim->hash, HASH_REMOVE, NULL);
@@ -616,7 +652,8 @@ lfc_write(NRelFileInfo rinfo, ForkNumber forkNum, BlockNumber blkno,
 		else
 		{
 			lfc_ctl->used += 1;
-			entry->offset = lfc_ctl->size++; /* allocate new chunk at end of file */
+			entry->offset = lfc_ctl->size++;	/* allocate new chunk at end
+												 * of file */
 		}
 		entry->access_count = 1;
 		entry->hash = hash;
@@ -628,7 +665,7 @@ lfc_write(NRelFileInfo rinfo, ForkNumber forkNum, BlockNumber blkno,
 	lfc_ctl->writes += 1;
 	LWLockRelease(lfc_lock);
 
-	rc = pwrite(lfc_desc, buffer, BLCKSZ, ((off_t)entry_offset*BLOCKS_PER_CHUNK + chunk_offs)*BLCKSZ);
+	rc = pwrite(lfc_desc, buffer, BLCKSZ, ((off_t) entry_offset * BLOCKS_PER_CHUNK + chunk_offs) * BLCKSZ);
 	if (rc != BLCKSZ)
 	{
 		lfc_disable("write");
@@ -665,13 +702,13 @@ Datum
 neon_get_lfc_stats(PG_FUNCTION_ARGS)
 {
 	FuncCallContext *funcctx;
-	NeonGetStatsCtx* fctx;
+	NeonGetStatsCtx *fctx;
 	MemoryContext oldcontext;
 	TupleDesc	tupledesc;
 	Datum		result;
 	HeapTuple	tuple;
-	char const* key;
-	uint64      value;
+	char const *key;
+	uint64		value;
 	Datum		values[NUM_NEON_GET_STATS_COLS];
 	bool		nulls[NUM_NEON_GET_STATS_COLS];
 
@@ -683,7 +720,7 @@ neon_get_lfc_stats(PG_FUNCTION_ARGS)
 		oldcontext = MemoryContextSwitchTo(funcctx->multi_call_memory_ctx);
 
 		/* Create a user function context for cross-call persistence */
-		fctx = (NeonGetStatsCtx*) palloc(sizeof(NeonGetStatsCtx));
+		fctx = (NeonGetStatsCtx *) palloc(sizeof(NeonGetStatsCtx));
 
 		/* Construct a tuple descriptor for the result rows. */
 		tupledesc = CreateTemplateTupleDesc(NUM_NEON_GET_STATS_COLS);
@@ -704,7 +741,7 @@ neon_get_lfc_stats(PG_FUNCTION_ARGS)
 	funcctx = SRF_PERCALL_SETUP();
 
 	/* Get the saved state */
-	fctx = (NeonGetStatsCtx*) funcctx->user_fctx;
+	fctx = (NeonGetStatsCtx *) funcctx->user_fctx;
 
 	switch (funcctx->call_cntr)
 	{
@@ -792,9 +829,9 @@ local_cache_pages(PG_FUNCTION_ARGS)
 
 	if (SRF_IS_FIRSTCALL())
 	{
-        HASH_SEQ_STATUS status;
-		FileCacheEntry* entry;
-		uint32 n_pages = 0;
+		HASH_SEQ_STATUS status;
+		FileCacheEntry *entry;
+		uint32		n_pages = 0;
 
 		funcctx = SRF_FIRSTCALL_INIT();
 
@@ -851,7 +888,7 @@ local_cache_pages(PG_FUNCTION_ARGS)
 				hash_seq_init(&status, lfc_hash);
 				while ((entry = hash_seq_search(&status)) != NULL)
 				{
-					for (int i = 0; i < BLOCKS_PER_CHUNK/32; i++)
+					for (int i = 0; i < BLOCKS_PER_CHUNK / 32; i++)
 						n_pages += pg_popcount32(entry->bitmap[i]);
 				}
 			}
@@ -870,10 +907,11 @@ local_cache_pages(PG_FUNCTION_ARGS)
 		if (n_pages != 0)
 		{
 			/*
-			 * Scan through all the cache entries, saving the relevant fields in the
-			 * fctx->record structure.
+			 * Scan through all the cache entries, saving the relevant fields
+			 * in the fctx->record structure.
 			 */
-			uint32 n = 0;
+			uint32		n = 0;
+
 			hash_seq_init(&status, lfc_hash);
 			while ((entry = hash_seq_search(&status)) != NULL)
 			{
@@ -881,7 +919,7 @@ local_cache_pages(PG_FUNCTION_ARGS)
 				{
 					if (entry->bitmap[i >> 5] & (1 << (i & 31)))
 					{
-						fctx->record[n].pageoffs = entry->offset*BLOCKS_PER_CHUNK + i;
+						fctx->record[n].pageoffs = entry->offset * BLOCKS_PER_CHUNK + i;
 						fctx->record[n].relfilenode = NInfoGetRelNumber(BufTagGetNRelFileInfo(entry->key));
 						fctx->record[n].reltablespace = NInfoGetSpcOid(BufTagGetNRelFileInfo(entry->key));
 						fctx->record[n].reldatabase = NInfoGetDbOid(BufTagGetNRelFileInfo(entry->key));