Add new compaction abstraction, simulator, and implementation.

This consists of three parts:

1. A refactoring and new contract for implementing and testing
compaction.

The logic is now in a separate crate, with no dependency on the
'pageserver' crate. It defines an interface that the real pageserver
must implement, in order to call the compaction algorithm. The
interface models things like delta and image layers, but just the
parts that the compaction algorithm needs to make decisions. That
makes it easier unit test the algorithm and experiment with different
implementations.

I did not convert the current code to the new abstraction, however.
When compaction algorithm is set to "Legacy", we just use the old
code. It might be worthwhile to convert the old code to the new
abstraction, so that we can compare the behavior of the new algorithm
against the old one, using the same simulated cases. If we do that,
have to be careful that the converted code really is equivalent to the
old.

This inclues only trivial changes to the main pageserver code. All the
new code is behind a tenant config option. So this should be pretty
safe to merge, even if the new implementation is buggy, as long as we
don't enable it.

2. A new compaction algorithm, implemented using the new
abstraction.

The new algorithm is tiered compaction.  It is inspired by the PoC at
PR #4539, although I did not use that code directly, as I needed the
new implementation to fit the new abstraction. The algorithm here is
less advanced, I did not implement partial image layers, for example.
I wanted to keep it simple on purpose, so that as we add bells and
whistles, we can see the effects using the included simulator.

One difference to #4539 and your typical LSM tree implementations is
how we keep track of the LSM tree levels. This PR doesn't have a
permanent concept of a level, tier or sorted run at all. There are
just delta and image layers. However, when compaction starts, we look
at the layers that exist, and arrange them into levels, depending on
their shapes. That is ephemeral: when the compaction finishes, we
forget that information. This allows the new algorithm to work without
any extra bookkeeping. That makes it easier to transition from the old
algorithm to new, and back again.

There is just a new tenant config option to choose the compaction
algorithm. The default is "Legacy", meaning the current algorithm in
'main'. If you set it to "Tiered".

3. A simulator, which implements the new abstraction.

The simulator can be used to analyze write and storage amplification,
without running a test with the full pageserver. It can also draw an
SVG animation of the simulation, to visualize how layers are created
and deleted.

To run the simulator:

    ./target/debug/compaction-simulator run-suite

.github/actionlint.yml

@@ -5,4 +5,6 @@ self-hosted-runner:
     - small
     - us-east-2
 config-variables:
+  - REMOTE_STORAGE_AZURE_CONTAINER
+  - REMOTE_STORAGE_AZURE_REGION
   - SLACK_UPCOMING_RELEASE_CHANNEL_ID

.github/actions/allure-report-generate/action.yml

@@ -203,6 +203,10 @@ runs:
         COMMIT_SHA: ${{ github.event.pull_request.head.sha || github.sha }}
         BASE_S3_URL: ${{ steps.generate-report.outputs.base-s3-url }}
       run: |
+        if [ ! -d "${WORKDIR}/report/data/test-cases" ]; then
+          exit 0
+        fi
+
         export DATABASE_URL=${REGRESS_TEST_RESULT_CONNSTR_NEW}
 
         ./scripts/pysync

.github/workflows/build_and_test.yml

@@ -320,6 +320,9 @@ jobs:
       - name: Build neon extensions
         run: mold -run make neon-pg-ext -j$(nproc)
 
+      - name: Build walproposer-lib
+        run: mold -run make walproposer-lib -j$(nproc)
+
       - name: Run cargo build
         run: |
           ${cov_prefix} mold -run cargo build $CARGO_FLAGS $CARGO_FEATURES --bins --tests
@@ -335,6 +338,16 @@ jobs:
           # Avoid `$CARGO_FEATURES` since there's no `testing` feature in the e2e tests now
           ${cov_prefix} cargo test $CARGO_FLAGS --package remote_storage --test test_real_s3
 
+          # Run separate tests for real Azure Blob Storage
+          # XXX: replace region with `eu-central-1`-like region
+          export ENABLE_REAL_AZURE_REMOTE_STORAGE=y
+          export AZURE_STORAGE_ACCOUNT="${{ secrets.AZURE_STORAGE_ACCOUNT_DEV }}"
+          export AZURE_STORAGE_ACCESS_KEY="${{ secrets.AZURE_STORAGE_ACCESS_KEY_DEV }}"
+          export REMOTE_STORAGE_AZURE_CONTAINER="${{ vars.REMOTE_STORAGE_AZURE_CONTAINER }}"
+          export REMOTE_STORAGE_AZURE_REGION="${{ vars.REMOTE_STORAGE_AZURE_REGION }}"
+          # Avoid `$CARGO_FEATURES` since there's no `testing` feature in the e2e tests now
+          ${cov_prefix} cargo test $CARGO_FLAGS --package remote_storage --test test_real_azure
+
       - name: Install rust binaries
         run: |
           # Install target binaries
@@ -420,7 +433,7 @@ jobs:
           rerun_flaky: true
           pg_version: ${{ matrix.pg_version }}
         env:
-          TEST_RESULT_CONNSTR: ${{ secrets.REGRESS_TEST_RESULT_CONNSTR }}
+          TEST_RESULT_CONNSTR: ${{ secrets.REGRESS_TEST_RESULT_CONNSTR_NEW }}
           CHECK_ONDISK_DATA_COMPATIBILITY: nonempty
 
       - name: Merge and upload coverage data
@@ -455,7 +468,7 @@ jobs:
         env:
           VIP_VAP_ACCESS_TOKEN: "${{ secrets.VIP_VAP_ACCESS_TOKEN }}"
           PERF_TEST_RESULT_CONNSTR: "${{ secrets.PERF_TEST_RESULT_CONNSTR }}"
-          TEST_RESULT_CONNSTR: "${{ secrets.REGRESS_TEST_RESULT_CONNSTR }}"
+          TEST_RESULT_CONNSTR: "${{ secrets.REGRESS_TEST_RESULT_CONNSTR_NEW }}"
       # XXX: no coverage data handling here, since benchmarks are run on release builds,
       # while coverage is currently collected for the debug ones
 
@@ -834,7 +847,7 @@ jobs:
       run:
         shell: sh -eu {0}
     env:
-      VM_BUILDER_VERSION: v0.17.12
+      VM_BUILDER_VERSION: v0.18.5
 
     steps:
       - name: Checkout

.github/workflows/neon_extra_builds.yml

@@ -32,7 +32,7 @@ jobs:
 
     steps:
       - name: Checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
         with:
           submodules: true
           fetch-depth: 1
@@ -90,18 +90,21 @@ jobs:
 
       - name: Build postgres v14
         if: steps.cache_pg_14.outputs.cache-hit != 'true'
-        run: make postgres-v14 -j$(nproc)
+        run: make postgres-v14 -j$(sysctl -n hw.ncpu)
 
       - name: Build postgres v15
         if: steps.cache_pg_15.outputs.cache-hit != 'true'
-        run: make postgres-v15 -j$(nproc)
+        run: make postgres-v15 -j$(sysctl -n hw.ncpu)
 
       - name: Build postgres v16
         if: steps.cache_pg_16.outputs.cache-hit != 'true'
-        run: make postgres-v16 -j$(nproc)
+        run: make postgres-v16 -j$(sysctl -n hw.ncpu)
 
       - name: Build neon extensions
-        run: make neon-pg-ext -j$(nproc)
+        run: make neon-pg-ext -j$(sysctl -n hw.ncpu)
+
+      - name: Build walproposer-lib
+        run: make walproposer-lib -j$(sysctl -n hw.ncpu)
 
       - name: Run cargo build
         run: cargo build --all --release
@@ -126,7 +129,7 @@ jobs:
 
     steps:
       - name: Checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
         with:
           submodules: true
           fetch-depth: 1
@@ -135,6 +138,9 @@ jobs:
       - name: Get postgres headers
         run: make postgres-headers -j$(nproc)
 
+      - name: Build walproposer-lib
+        run: make walproposer-lib -j$(nproc)
+
       - name: Produce the build stats
         run: cargo build --all --release --timings
 

.github/workflows/release.yml

@@ -2,7 +2,7 @@ name: Create Release Branch
 
 on:
   schedule:
-    - cron: '0 7 * * 2'
+    - cron: '0 7 * * 5'
   workflow_dispatch:
 
 jobs:

Cargo.toml

@@ -4,6 +4,7 @@ members = [
     "compute_tools",
     "control_plane",
     "pageserver",
+    "pageserver/compaction",
     "pageserver/ctl",
     "proxy",
     "safekeeper",
@@ -26,6 +27,7 @@ members = [
     "libs/tracing-utils",
     "libs/postgres_ffi/wal_craft",
     "libs/vm_monitor",
+    "libs/walproposer",
 ]
 
 [workspace.package]
@@ -36,6 +38,10 @@ license = "Apache-2.0"
 [workspace.dependencies]
 anyhow = { version = "1.0", features = ["backtrace"] }
 async-compression = { version = "0.4.0", features = ["tokio", "gzip"] }
+azure_core = "0.16"
+azure_identity = "0.16"
+azure_storage = "0.16"
+azure_storage_blobs = "0.16"
 flate2 = "1.0.26"
 async-stream = "0.3"
 async-trait = "0.1"
@@ -76,6 +82,7 @@ hex = "0.4"
 hex-literal = "0.4"
 hmac = "0.12.1"
 hostname = "0.3.1"
+http-types = "2"
 humantime = "2.1"
 humantime-serde = "1.1.1"
 hyper = "0.14"
@@ -155,11 +162,11 @@ env_logger = "0.10"
 log = "0.4"
 
 ## Libraries from neondatabase/ git forks, ideally with changes to be upstreamed
-postgres = { git = "https://github.com/neondatabase/rust-postgres.git", rev="9011f7110db12b5e15afaf98f8ac834501d50ddc" }
-postgres-native-tls = { git = "https://github.com/neondatabase/rust-postgres.git", rev="9011f7110db12b5e15afaf98f8ac834501d50ddc" }
-postgres-protocol = { git = "https://github.com/neondatabase/rust-postgres.git", rev="9011f7110db12b5e15afaf98f8ac834501d50ddc" }
-postgres-types = { git = "https://github.com/neondatabase/rust-postgres.git", rev="9011f7110db12b5e15afaf98f8ac834501d50ddc" }
-tokio-postgres = { git = "https://github.com/neondatabase/rust-postgres.git", rev="9011f7110db12b5e15afaf98f8ac834501d50ddc" }
+postgres = { git = "https://github.com/neondatabase/rust-postgres.git", rev="ce7260db5998fe27167da42503905a12e7ad9048" }
+postgres-native-tls = { git = "https://github.com/neondatabase/rust-postgres.git", rev="ce7260db5998fe27167da42503905a12e7ad9048" }
+postgres-protocol = { git = "https://github.com/neondatabase/rust-postgres.git", rev="ce7260db5998fe27167da42503905a12e7ad9048" }
+postgres-types = { git = "https://github.com/neondatabase/rust-postgres.git", rev="ce7260db5998fe27167da42503905a12e7ad9048" }
+tokio-postgres = { git = "https://github.com/neondatabase/rust-postgres.git", rev="ce7260db5998fe27167da42503905a12e7ad9048" }
 
 ## Other git libraries
 heapless = { default-features=false, features=[], git = "https://github.com/japaric/heapless.git", rev = "644653bf3b831c6bb4963be2de24804acf5e5001" } # upstream release pending
@@ -169,6 +176,7 @@ compute_api = { version = "0.1", path = "./libs/compute_api/" }
 consumption_metrics = { version = "0.1", path = "./libs/consumption_metrics/" }
 metrics = { version = "0.1", path = "./libs/metrics/" }
 pageserver_api = { version = "0.1", path = "./libs/pageserver_api/" }
+pageserver_compaction = { version = "0.1", path = "./pageserver/compaction/" }
 postgres_backend = { version = "0.1", path = "./libs/postgres_backend/" }
 postgres_connection = { version = "0.1", path = "./libs/postgres_connection/" }
 postgres_ffi = { version = "0.1", path = "./libs/postgres_ffi/" }
@@ -180,6 +188,7 @@ tenant_size_model = { version = "0.1", path = "./libs/tenant_size_model/" }
 tracing-utils = { version = "0.1", path = "./libs/tracing-utils/" }
 utils = { version = "0.1", path = "./libs/utils/" }
 vm_monitor = { version = "0.1", path = "./libs/vm_monitor/" }
+walproposer = { version = "0.1", path = "./libs/walproposer/" }
 
 ## Common library dependency
 workspace_hack = { version = "0.1", path = "./workspace_hack/" }
@@ -195,7 +204,7 @@ tonic-build = "0.9"
 
 # This is only needed for proxy's tests.
 # TODO: we should probably fork `tokio-postgres-rustls` instead.
-tokio-postgres = { git = "https://github.com/neondatabase/rust-postgres.git", rev="9011f7110db12b5e15afaf98f8ac834501d50ddc" }
+tokio-postgres = { git = "https://github.com/neondatabase/rust-postgres.git", rev="ce7260db5998fe27167da42503905a12e7ad9048" }
 
 ################# Binary contents sections
 

Makefile

@@ -62,7 +62,7 @@ all: neon postgres neon-pg-ext
 #
 # The 'postgres_ffi' depends on the Postgres headers.
 .PHONY: neon
-neon: postgres-headers
+neon: postgres-headers walproposer-lib
 	+@echo "Compiling Neon"
 	$(CARGO_CMD_PREFIX) cargo build $(CARGO_BUILD_FLAGS)
 
@@ -168,6 +168,42 @@ neon-pg-ext-clean-%:
 	-C $(POSTGRES_INSTALL_DIR)/build/neon-utils-$* \
 	-f $(ROOT_PROJECT_DIR)/pgxn/neon_utils/Makefile clean
 
+# Build walproposer as a static library. walproposer source code is located
+# in the pgxn/neon directory.
+# 
+# We also need to include libpgport.a and libpgcommon.a, because walproposer
+# uses some functions from those libraries.
+# 
+# Some object files are removed from libpgport.a and libpgcommon.a because
+# they depend on openssl and other libraries that are not included in our
+# Rust build.
+.PHONY: walproposer-lib
+walproposer-lib: neon-pg-ext-v16
+	+@echo "Compiling walproposer-lib"
+	mkdir -p $(POSTGRES_INSTALL_DIR)/build/walproposer-lib
+	$(MAKE) PG_CONFIG=$(POSTGRES_INSTALL_DIR)/v16/bin/pg_config CFLAGS='$(PG_CFLAGS) $(COPT)' \
+		-C $(POSTGRES_INSTALL_DIR)/build/walproposer-lib \
+		-f $(ROOT_PROJECT_DIR)/pgxn/neon/Makefile walproposer-lib
+	cp $(POSTGRES_INSTALL_DIR)/v16/lib/libpgport.a $(POSTGRES_INSTALL_DIR)/build/walproposer-lib
+	cp $(POSTGRES_INSTALL_DIR)/v16/lib/libpgcommon.a $(POSTGRES_INSTALL_DIR)/build/walproposer-lib
+ifeq ($(UNAME_S),Linux)
+	$(AR) d $(POSTGRES_INSTALL_DIR)/build/walproposer-lib/libpgport.a \
+		pg_strong_random.o
+	$(AR) d $(POSTGRES_INSTALL_DIR)/build/walproposer-lib/libpgcommon.a \
+		pg_crc32c.o \
+		hmac_openssl.o \
+		cryptohash_openssl.o \
+		scram-common.o \
+		md5_common.o \
+		checksum_helper.o
+endif
+
+.PHONY: walproposer-lib-clean
+walproposer-lib-clean:
+	$(MAKE) PG_CONFIG=$(POSTGRES_INSTALL_DIR)/v16/bin/pg_config \
+		-C $(POSTGRES_INSTALL_DIR)/build/walproposer-lib \
+		-f $(ROOT_PROJECT_DIR)/pgxn/neon/Makefile clean
+
 .PHONY: neon-pg-ext
 neon-pg-ext: \
 	neon-pg-ext-v14 \

NOTICE

@@ -1,5 +1,5 @@
 Neon
 Copyright 2022 Neon Inc.
 
-The PostgreSQL submodules in vendor/postgres-v14 and vendor/postgres-v15 are licensed under the
-PostgreSQL license. See vendor/postgres-v14/COPYRIGHT and vendor/postgres-v15/COPYRIGHT.
+The PostgreSQL submodules in vendor/ are licensed under the PostgreSQL license.
+See vendor/postgres-vX/COPYRIGHT for details.

compute_tools/src/bin/compute_ctl.rs

@@ -156,6 +156,7 @@ fn main() -> Result<()> {
                 let path = Path::new(sp);
                 let file = File::open(path)?;
                 spec = Some(serde_json::from_reader(file)?);
+                live_config_allowed = true;
             } else if let Some(id) = compute_id {
                 if let Some(cp_base) = control_plane_uri {
                     live_config_allowed = true;
@@ -277,32 +278,26 @@ fn main() -> Result<()> {
         if #[cfg(target_os = "linux")] {
             use std::env;
             use tokio_util::sync::CancellationToken;
-            use tracing::warn;
-            let vm_monitor_addr = matches.get_one::<String>("vm-monitor-addr");
+            let vm_monitor_addr = matches
+                .get_one::<String>("vm-monitor-addr")
+                .expect("--vm-monitor-addr should always be set because it has a default arg");
             let file_cache_connstr = matches.get_one::<String>("filecache-connstr");
             let cgroup = matches.get_one::<String>("cgroup");
-            let file_cache_on_disk = matches.get_flag("file-cache-on-disk");
 
             // Only make a runtime if we need to.
             // Note: it seems like you can make a runtime in an inner scope and
             // if you start a task in it it won't be dropped. However, make it
             // in the outermost scope just to be safe.
-            let rt = match (env::var_os("AUTOSCALING"), vm_monitor_addr) {
-                (None, None) => None,
-                (None, Some(_)) => {
-                    warn!("--vm-monitor-addr option set but AUTOSCALING env var not present");
-                    None
-                }
-                (Some(_), None) => {
-                    panic!("AUTOSCALING env var present but --vm-monitor-addr option not set")
-                }
-                (Some(_), Some(_)) => Some(
+            let rt = if env::var_os("AUTOSCALING").is_some() {
+                Some(
                     tokio::runtime::Builder::new_multi_thread()
                         .worker_threads(4)
                         .enable_all()
                         .build()
-                        .expect("failed to create tokio runtime for monitor"),
-                ),
+                        .expect("failed to create tokio runtime for monitor")
+                )
+            } else {
+                None
             };
 
             // This token is used internally by the monitor to clean up all threads
@@ -313,8 +308,7 @@ fn main() -> Result<()> {
                     Box::leak(Box::new(vm_monitor::Args {
                         cgroup: cgroup.cloned(),
                         pgconnstr: file_cache_connstr.cloned(),
-                        addr: vm_monitor_addr.cloned().unwrap(),
-                        file_cache_on_disk,
+                        addr: vm_monitor_addr.clone(),
                     })),
                     token.clone(),
                 ))
@@ -486,6 +480,8 @@ fn cli() -> clap::Command {
                 .value_name("FILECACHE_CONNSTR"),
         )
         .arg(
+            // DEPRECATED, NO LONGER DOES ANYTHING.
+            // See https://github.com/neondatabase/cloud/issues/7516
             Arg::new("file-cache-on-disk")
                 .long("file-cache-on-disk")
                 .action(clap::ArgAction::SetTrue),

compute_tools/src/compute.rs

@@ -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 IN ROLE pg_read_all_data, pg_write_all_data;
+                        CREATE ROLE neon_superuser CREATEDB CREATEROLE NOLOGIN REPLICATION 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), ', '));
@@ -692,10 +692,11 @@ impl ComputeNode {
         // Proceed with post-startup configuration. Note, that order of operations is important.
         let spec = &compute_state.pspec.as_ref().expect("spec must be set").spec;
         create_neon_superuser(spec, &mut client)?;
+        cleanup_instance(&mut client)?;
         handle_roles(spec, &mut client)?;
         handle_databases(spec, &mut client)?;
         handle_role_deletions(spec, self.connstr.as_str(), &mut client)?;
-        handle_grants(spec, self.connstr.as_str())?;
+        handle_grants(spec, &mut client, self.connstr.as_str())?;
         handle_extensions(spec, &mut client)?;
         create_availability_check_data(&mut client)?;
 
@@ -731,10 +732,11 @@ impl ComputeNode {
         // Disable DDL forwarding because control plane already knows about these roles/databases.
         if spec.mode == ComputeMode::Primary {
             client.simple_query("SET neon.forward_ddl = false")?;
+            cleanup_instance(&mut client)?;
             handle_roles(&spec, &mut client)?;
             handle_databases(&spec, &mut client)?;
             handle_role_deletions(&spec, self.connstr.as_str(), &mut client)?;
-            handle_grants(&spec, self.connstr.as_str())?;
+            handle_grants(&spec, &mut client, self.connstr.as_str())?;
             handle_extensions(&spec, &mut client)?;
         }
 

compute_tools/src/pg_helpers.rs

@@ -1,3 +1,4 @@
+use std::collections::HashMap;
 use std::fmt::Write;
 use std::fs;
 use std::fs::File;
@@ -192,11 +193,16 @@ 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 FROM pg_catalog.pg_authid", &[])?
+        .query(
+            "SELECT rolname, rolpassword, rolreplication, rolbypassrls 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();
@@ -205,22 +211,37 @@ pub fn get_existing_roles(xact: &mut Transaction<'_>) -> Result<Vec<Role>> {
 }
 
 /// Build a list of existing Postgres databases
-pub fn get_existing_dbs(client: &mut Client) -> Result<Vec<Database>> {
-    let postgres_dbs = client
+pub fn get_existing_dbs(client: &mut Client) -> Result<HashMap<String, Database>> {
+    // `pg_database.datconnlimit = -2` means that the database is in the
+    // invalid state. See:
+    //   https://github.com/postgres/postgres/commit/a4b4cc1d60f7e8ccfcc8ff8cb80c28ee411ad9a9
+    let postgres_dbs: Vec<Database> = client
         .query(
-            "SELECT datname, datdba::regrole::text as owner
-               FROM pg_catalog.pg_database;",
+            "SELECT
+                datname AS name,
+                datdba::regrole::text AS owner,
+                NOT datallowconn AS restrict_conn,
+                datconnlimit = - 2 AS invalid
+            FROM
+                pg_catalog.pg_database;",
             &[],
         )?
         .iter()
         .map(|row| Database {
-            name: row.get("datname"),
+            name: row.get("name"),
             owner: row.get("owner"),
+            restrict_conn: row.get("restrict_conn"),
+            invalid: row.get("invalid"),
             options: None,
         })
         .collect();
 
-    Ok(postgres_dbs)
+    let dbs_map = postgres_dbs
+        .iter()
+        .map(|db| (db.name.clone(), db.clone()))
+        .collect::<HashMap<_, _>>();
+
+    Ok(dbs_map)
 }
 
 /// Wait for Postgres to become ready to accept connections. It's ready to

compute_tools/src/spec.rs

@@ -13,7 +13,7 @@ use crate::params::PG_HBA_ALL_MD5;
 use crate::pg_helpers::*;
 
 use compute_api::responses::{ControlPlaneComputeStatus, ControlPlaneSpecResponse};
-use compute_api::spec::{ComputeSpec, Database, PgIdent, Role};
+use compute_api::spec::{ComputeSpec, PgIdent, Role};
 
 // Do control plane request and return response if any. In case of error it
 // returns a bool flag indicating whether it makes sense to retry the request
@@ -24,7 +24,7 @@ fn do_control_plane_request(
 ) -> Result<ControlPlaneSpecResponse, (bool, String)> {
     let resp = reqwest::blocking::Client::new()
         .get(uri)
-        .header("Authorization", jwt)
+        .header("Authorization", format!("Bearer {}", jwt))
         .send()
         .map_err(|e| {
             (
@@ -68,7 +68,7 @@ pub fn get_spec_from_control_plane(
     base_uri: &str,
     compute_id: &str,
 ) -> Result<Option<ComputeSpec>> {
-    let cp_uri = format!("{base_uri}/management/api/v2/computes/{compute_id}/spec");
+    let cp_uri = format!("{base_uri}/compute/api/v2/computes/{compute_id}/spec");
     let jwt: String = match std::env::var("NEON_CONTROL_PLANE_TOKEN") {
         Ok(v) => v,
         Err(_) => "".to_string(),
@@ -161,6 +161,38 @@ pub fn add_standby_signal(pgdata_path: &Path) -> Result<()> {
     Ok(())
 }
 
+/// Compute could be unexpectedly shut down, for example, during the
+/// database dropping. This leaves the database in the invalid state,
+/// which prevents new db creation with the same name. This function
+/// will clean it up before proceeding with catalog updates. All
+/// possible future cleanup operations may go here too.
+#[instrument(skip_all)]
+pub fn cleanup_instance(client: &mut Client) -> Result<()> {
+    let existing_dbs = get_existing_dbs(client)?;
+
+    for (_, db) in existing_dbs {
+        if db.invalid {
+            // After recent commit in Postgres, interrupted DROP DATABASE
+            // leaves the database in the invalid state. According to the
+            // commit message, the only option for user is to drop it again.
+            // See:
+            //   https://github.com/postgres/postgres/commit/a4b4cc1d60f7e8ccfcc8ff8cb80c28ee411ad9a9
+            //
+            // Postgres Neon extension is done the way, that db is de-registered
+            // in the control plane metadata only after it is dropped. So there is
+            // a chance that it still thinks that db should exist. This means
+            // that it will be re-created by `handle_databases()`. Yet, it's fine
+            // as user can just repeat drop (in vanilla Postgres they would need
+            // to do the same, btw).
+            let query = format!("DROP DATABASE IF EXISTS {}", db.name.pg_quote());
+            info!("dropping invalid database {}", db.name);
+            client.execute(query.as_str(), &[])?;
+        }
+    }
+
+    Ok(())
+}
+
 /// Given a cluster spec json and open transaction it handles roles creation,
 /// deletion and update.
 #[instrument(skip_all)]
@@ -233,6 +265,8 @@ 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 {
@@ -264,13 +298,14 @@ pub fn handle_roles(spec: &ComputeSpec, client: &mut Client) -> Result<()> {
         match action {
             RoleAction::None => {}
             RoleAction::Update => {
-                let mut query: String = format!("ALTER ROLE {} ", name.pg_quote());
+                let mut query: String =
+                    format!("ALTER ROLE {} BYPASSRLS REPLICATION", name.pg_quote());
                 query.push_str(&role.to_pg_options());
                 xact.execute(query.as_str(), &[])?;
             }
             RoleAction::Create => {
                 let mut query: String = format!(
-                    "CREATE ROLE {} CREATEROLE CREATEDB BYPASSRLS IN ROLE neon_superuser",
+                    "CREATE ROLE {} CREATEROLE CREATEDB BYPASSRLS REPLICATION IN ROLE neon_superuser",
                     name.pg_quote()
                 );
                 info!("role create query: '{}'", &query);
@@ -379,13 +414,13 @@ fn reassign_owned_objects(spec: &ComputeSpec, connstr: &str, role_name: &PgIdent
 /// which together provide us idempotency.
 #[instrument(skip_all)]
 pub fn handle_databases(spec: &ComputeSpec, client: &mut Client) -> Result<()> {
-    let existing_dbs: Vec<Database> = get_existing_dbs(client)?;
+    let existing_dbs = get_existing_dbs(client)?;
 
     // Print a list of existing Postgres databases (only in debug mode)
     if span_enabled!(Level::INFO) {
         info!("postgres databases:");
-        for r in &existing_dbs {
-            info!("    {}:{}", r.name, r.owner);
+        for (dbname, db) in &existing_dbs {
+            info!("    {}:{}", dbname, db.owner);
         }
     }
 
@@ -439,8 +474,7 @@ pub fn handle_databases(spec: &ComputeSpec, client: &mut Client) -> Result<()> {
                 "rename_db" => {
                     let new_name = op.new_name.as_ref().unwrap();
 
-                    // XXX: with a limited number of roles it is fine, but consider making it a HashMap
-                    if existing_dbs.iter().any(|r| r.name == op.name) {
+                    if existing_dbs.get(&op.name).is_some() {
                         let query: String = format!(
                             "ALTER DATABASE {} RENAME TO {}",
                             op.name.pg_quote(),
@@ -457,14 +491,12 @@ pub fn handle_databases(spec: &ComputeSpec, client: &mut Client) -> Result<()> {
     }
 
     // Refresh Postgres databases info to handle possible renames
-    let existing_dbs: Vec<Database> = get_existing_dbs(client)?;
+    let existing_dbs = get_existing_dbs(client)?;
 
     info!("cluster spec databases:");
     for db in &spec.cluster.databases {
         let name = &db.name;
-
-        // XXX: with a limited number of databases it is fine, but consider making it a HashMap
-        let pg_db = existing_dbs.iter().find(|r| r.name == *name);
+        let pg_db = existing_dbs.get(name);
 
         enum DatabaseAction {
             None,
@@ -530,13 +562,32 @@ pub fn handle_databases(spec: &ComputeSpec, client: &mut Client) -> Result<()> {
 /// Grant CREATE ON DATABASE to the database owner and do some other alters and grants
 /// to allow users creating trusted extensions and re-creating `public` schema, for example.
 #[instrument(skip_all)]
-pub fn handle_grants(spec: &ComputeSpec, connstr: &str) -> Result<()> {
-    info!("cluster spec grants:");
+pub fn handle_grants(spec: &ComputeSpec, client: &mut Client, connstr: &str) -> Result<()> {
+    info!("modifying database permissions");
+    let existing_dbs = get_existing_dbs(client)?;
 
     // Do some per-database access adjustments. We'd better do this at db creation time,
     // but CREATE DATABASE isn't transactional. So we cannot create db + do some grants
     // atomically.
     for db in &spec.cluster.databases {
+        match existing_dbs.get(&db.name) {
+            Some(pg_db) => {
+                if pg_db.restrict_conn || pg_db.invalid {
+                    info!(
+                        "skipping grants for db {} (invalid: {}, connections not allowed: {})",
+                        db.name, pg_db.invalid, pg_db.restrict_conn
+                    );
+                    continue;
+                }
+            }
+            None => {
+                bail!(
+                    "database {} doesn't exist in Postgres after handle_databases()",
+                    db.name
+                );
+            }
+        }
+
         let mut conf = Config::from_str(connstr)?;
         conf.dbname(&db.name);
 
@@ -575,6 +626,11 @@ pub fn handle_grants(spec: &ComputeSpec, connstr: &str) -> Result<()> {
 
         // Explicitly grant CREATE ON SCHEMA PUBLIC to the web_access user.
         // This is needed because since postgres 15 this privilege is removed by default.
+        // TODO: web_access isn't created for almost 1 year. It could be that we have
+        // active users of 1 year old projects, but hopefully not, so check it and
+        // remove this code if possible. The worst thing that could happen is that
+        // user won't be able to use public schema in NEW databases created in the
+        // very OLD project.
         let grant_query = "DO $$\n\
                 BEGIN\n\
                     IF EXISTS(\n\

compute_tools/tests/pg_helpers_tests.rs

@@ -28,7 +28,7 @@ mod pg_helpers_tests {
         assert_eq!(
             spec.cluster.settings.as_pg_settings(),
             r#"fsync = off
-wal_level = replica
+wal_level = logical
 hot_standby = on
 neon.safekeepers = '127.0.0.1:6502,127.0.0.1:6503,127.0.0.1:6501'
 wal_log_hints = on

control_plane/src/attachment_service.rs

@@ -19,7 +19,7 @@ const COMMAND: &str = "attachment_service";
 pub struct AttachHookRequest {
     #[serde_as(as = "DisplayFromStr")]
     pub tenant_id: TenantId,
-    pub pageserver_id: Option<NodeId>,
+    pub node_id: Option<NodeId>,
 }
 
 #[derive(Serialize, Deserialize)]
@@ -85,7 +85,7 @@ impl AttachmentService {
             .control_plane_api
             .clone()
             .unwrap()
-            .join("attach_hook")
+            .join("attach-hook")
             .unwrap();
         let client = reqwest::blocking::ClientBuilder::new()
             .build()
@@ -93,7 +93,7 @@ impl AttachmentService {
 
         let request = AttachHookRequest {
             tenant_id,
-            pageserver_id: Some(pageserver_id),
+            node_id: Some(pageserver_id),
         };
 
         let response = client.post(url).json(&request).send()?;

control_plane/src/background_process.rs

@@ -86,7 +86,7 @@ where
         .stdout(process_log_file)
         .stderr(same_file_for_stderr)
         .args(args);
-    let filled_cmd = fill_aws_secrets_vars(fill_rust_env_vars(background_command));
+    let filled_cmd = fill_remote_storage_secrets_vars(fill_rust_env_vars(background_command));
     filled_cmd.envs(envs);
 
     let pid_file_to_check = match initial_pid_file {
@@ -238,11 +238,13 @@ fn fill_rust_env_vars(cmd: &mut Command) -> &mut Command {
     filled_cmd
 }
 
-fn fill_aws_secrets_vars(mut cmd: &mut Command) -> &mut Command {
+fn fill_remote_storage_secrets_vars(mut cmd: &mut Command) -> &mut Command {
     for env_key in [
         "AWS_ACCESS_KEY_ID",
         "AWS_SECRET_ACCESS_KEY",
         "AWS_SESSION_TOKEN",
+        "AZURE_STORAGE_ACCOUNT",
+        "AZURE_STORAGE_ACCESS_KEY",
     ] {
         if let Ok(value) = std::env::var(env_key) {
             cmd = cmd.env(env_key, value);

control_plane/src/bin/attachment_service.rs

@@ -12,7 +12,9 @@ use hyper::{Body, Request, Response};
 use serde::{Deserialize, Serialize};
 use std::path::{Path, PathBuf};
 use std::{collections::HashMap, sync::Arc};
+use utils::http::endpoint::request_span;
 use utils::logging::{self, LogFormat};
+use utils::signals::{ShutdownSignals, Signal};
 
 use utils::{
     http::{
@@ -170,7 +172,7 @@ async fn handle_re_attach(mut req: Request<Body>) -> Result<Response<Body>, ApiE
             state.generation += 1;
             response.tenants.push(ReAttachResponseTenant {
                 id: *t,
-                generation: state.generation,
+                gen: state.generation,
             });
         }
     }
@@ -216,14 +218,31 @@ async fn handle_attach_hook(mut req: Request<Body>) -> Result<Response<Body>, Ap
         .tenants
         .entry(attach_req.tenant_id)
         .or_insert_with(|| TenantState {
-            pageserver: attach_req.pageserver_id,
+            pageserver: attach_req.node_id,
             generation: 0,
         });
 
-    if attach_req.pageserver_id.is_some() {
+    if let Some(attaching_pageserver) = attach_req.node_id.as_ref() {
         tenant_state.generation += 1;
+        tracing::info!(
+            tenant_id = %attach_req.tenant_id,
+            ps_id = %attaching_pageserver,
+            generation = %tenant_state.generation,
+            "issuing",
+        );
+    } else if let Some(ps_id) = tenant_state.pageserver {
+        tracing::info!(
+            tenant_id = %attach_req.tenant_id,
+            %ps_id,
+            generation = %tenant_state.generation,
+            "dropping",
+        );
+    } else {
+        tracing::info!(
+            tenant_id = %attach_req.tenant_id,
+            "no-op: tenant already has no pageserver");
     }
-    tenant_state.pageserver = attach_req.pageserver_id;
+    tenant_state.pageserver = attach_req.node_id;
     let generation = tenant_state.generation;
 
     locked.save().await.map_err(ApiError::InternalServerError)?;
@@ -231,7 +250,7 @@ async fn handle_attach_hook(mut req: Request<Body>) -> Result<Response<Body>, Ap
     json_response(
         StatusCode::OK,
         AttachHookResponse {
-            gen: attach_req.pageserver_id.map(|_| generation),
+            gen: attach_req.node_id.map(|_| generation),
         },
     )
 }
@@ -239,9 +258,9 @@ async fn handle_attach_hook(mut req: Request<Body>) -> Result<Response<Body>, Ap
 fn make_router(persistent_state: PersistentState) -> RouterBuilder<hyper::Body, ApiError> {
     endpoint::make_router()
         .data(Arc::new(State::new(persistent_state)))
-        .post("/re-attach", handle_re_attach)
-        .post("/validate", handle_validate)
-        .post("/attach_hook", handle_attach_hook)
+        .post("/re-attach", |r| request_span(r, handle_re_attach))
+        .post("/validate", |r| request_span(r, handle_validate))
+        .post("/attach-hook", |r| request_span(r, handle_attach_hook))
 }
 
 #[tokio::main]
@@ -268,7 +287,16 @@ async fn main() -> anyhow::Result<()> {
     let server = hyper::Server::from_tcp(http_listener)?.serve(service);
 
     tracing::info!("Serving on {0}", args.listen);
-    server.await?;
+
+    tokio::task::spawn(server);
+
+    ShutdownSignals::handle(|signal| match signal {
+        Signal::Interrupt | Signal::Terminate | Signal::Quit => {
+            tracing::info!("Got {}. Terminating", signal.name());
+            // We're just a test helper: no graceful shutdown.
+            std::process::exit(0);
+        }
+    })?;
 
     Ok(())
 }

control_plane/src/bin/neon_local.rs

@@ -798,6 +798,24 @@ fn handle_endpoint(ep_match: &ArgMatches, env: &local_env::LocalEnv) -> Result<(
                 ep.start(&auth_token, safekeepers, remote_ext_config)?;
             }
         }
+        "reconfigure" => {
+            let endpoint_id = sub_args
+                .get_one::<String>("endpoint_id")
+                .ok_or_else(|| anyhow!("No endpoint ID provided to reconfigure"))?;
+            let endpoint = cplane
+                .endpoints
+                .get(endpoint_id.as_str())
+                .with_context(|| format!("postgres endpoint {endpoint_id} is not found"))?;
+            let pageserver_id =
+                if let Some(id_str) = sub_args.get_one::<String>("endpoint-pageserver-id") {
+                    Some(NodeId(
+                        id_str.parse().context("while parsing pageserver id")?,
+                    ))
+                } else {
+                    None
+                };
+            endpoint.reconfigure(pageserver_id)?;
+        }
         "stop" => {
             let endpoint_id = sub_args
                 .get_one::<String>("endpoint_id")
@@ -1369,6 +1387,12 @@ fn cli() -> Command {
                     .arg(safekeepers_arg)
                     .arg(remote_ext_config_args)
                 )
+                .subcommand(Command::new("reconfigure")
+                            .about("Reconfigure the endpoint")
+                            .arg(endpoint_pageserver_id_arg)
+                            .arg(endpoint_id_arg.clone())
+                            .arg(tenant_id_arg.clone())
+                )
                 .subcommand(
                     Command::new("stop")
                     .arg(endpoint_id_arg)

control_plane/src/endpoint.rs

@@ -253,7 +253,7 @@ impl Endpoint {
         conf.append("shared_buffers", "1MB");
         conf.append("fsync", "off");
         conf.append("max_connections", "100");
-        conf.append("wal_level", "replica");
+        conf.append("wal_level", "logical");
         // wal_sender_timeout is the maximum time to wait for WAL replication.
         // It also defines how often the walreciever will send a feedback message to the wal sender.
         conf.append("wal_sender_timeout", "5s");
@@ -414,18 +414,34 @@ impl Endpoint {
             );
         }
 
-        // Also wait for the compute_ctl process to die. It might have some cleanup
-        // work to do after postgres stops, like syncing safekeepers, etc.
-        //
+        Ok(())
+    }
+
+    fn wait_for_compute_ctl_to_exit(&self) -> Result<()> {
         // TODO use background_process::stop_process instead
         let pidfile_path = self.endpoint_path().join("compute_ctl.pid");
         let pid: u32 = std::fs::read_to_string(pidfile_path)?.parse()?;
         let pid = nix::unistd::Pid::from_raw(pid as i32);
         crate::background_process::wait_until_stopped("compute_ctl", pid)?;
-
         Ok(())
     }
 
+    fn read_postgresql_conf(&self) -> Result<String> {
+        // Slurp the endpoints/<endpoint id>/postgresql.conf file into
+        // memory. We will include it in the spec file that we pass to
+        // `compute_ctl`, and `compute_ctl` will write it to the postgresql.conf
+        // in the data directory.
+        let postgresql_conf_path = self.endpoint_path().join("postgresql.conf");
+        match std::fs::read(&postgresql_conf_path) {
+            Ok(content) => Ok(String::from_utf8(content)?),
+            Err(e) if e.kind() == std::io::ErrorKind::NotFound => Ok("".to_string()),
+            Err(e) => Err(anyhow::Error::new(e).context(format!(
+                "failed to read config file in {}",
+                postgresql_conf_path.to_str().unwrap()
+            ))),
+        }
+    }
+
     pub fn start(
         &self,
         auth_token: &Option<String>,
@@ -436,21 +452,7 @@ impl Endpoint {
             anyhow::bail!("The endpoint is already running");
         }
 
-        // Slurp the endpoints/<endpoint id>/postgresql.conf file into
-        // memory. We will include it in the spec file that we pass to
-        // `compute_ctl`, and `compute_ctl` will write it to the postgresql.conf
-        // in the data directory.
-        let postgresql_conf_path = self.endpoint_path().join("postgresql.conf");
-        let postgresql_conf = match std::fs::read(&postgresql_conf_path) {
-            Ok(content) => String::from_utf8(content)?,
-            Err(e) if e.kind() == std::io::ErrorKind::NotFound => "".to_string(),
-            Err(e) => {
-                return Err(anyhow::Error::new(e).context(format!(
-                    "failed to read config file in {}",
-                    postgresql_conf_path.to_str().unwrap()
-                )))
-            }
-        };
+        let postgresql_conf = self.read_postgresql_conf()?;
 
         // We always start the compute node from scratch, so if the Postgres
         // data dir exists from a previous launch, remove it first.
@@ -621,6 +623,61 @@ impl Endpoint {
         }
     }
 
+    pub fn reconfigure(&self, pageserver_id: Option<NodeId>) -> Result<()> {
+        let mut spec: ComputeSpec = {
+            let spec_path = self.endpoint_path().join("spec.json");
+            let file = std::fs::File::open(spec_path)?;
+            serde_json::from_reader(file)?
+        };
+
+        let postgresql_conf = self.read_postgresql_conf()?;
+        spec.cluster.postgresql_conf = Some(postgresql_conf);
+
+        if let Some(pageserver_id) = pageserver_id {
+            let endpoint_config_path = self.endpoint_path().join("endpoint.json");
+            let mut endpoint_conf: EndpointConf = {
+                let file = std::fs::File::open(&endpoint_config_path)?;
+                serde_json::from_reader(file)?
+            };
+            endpoint_conf.pageserver_id = pageserver_id;
+            std::fs::write(
+                endpoint_config_path,
+                serde_json::to_string_pretty(&endpoint_conf)?,
+            )?;
+
+            let pageserver =
+                PageServerNode::from_env(&self.env, self.env.get_pageserver_conf(pageserver_id)?);
+            let ps_http_conf = &pageserver.pg_connection_config;
+            let (host, port) = (ps_http_conf.host(), ps_http_conf.port());
+            spec.pageserver_connstring = Some(format!("postgresql://no_user@{host}:{port}"));
+        }
+
+        let client = reqwest::blocking::Client::new();
+        let response = client
+            .post(format!(
+                "http://{}:{}/configure",
+                self.http_address.ip(),
+                self.http_address.port()
+            ))
+            .body(format!(
+                "{{\"spec\":{}}}",
+                serde_json::to_string_pretty(&spec)?
+            ))
+            .send()?;
+
+        let status = response.status();
+        if !(status.is_client_error() || status.is_server_error()) {
+            Ok(())
+        } else {
+            let url = response.url().to_owned();
+            let msg = match response.text() {
+                Ok(err_body) => format!("Error: {}", err_body),
+                Err(_) => format!("Http error ({}) at {}.", status.as_u16(), url),
+            };
+            Err(anyhow::anyhow!(msg))
+        }
+    }
+
     pub fn stop(&self, destroy: bool) -> Result<()> {
         // If we are going to destroy data directory,
         // use immediate shutdown mode, otherwise,
@@ -629,15 +686,25 @@ impl Endpoint {
         // Postgres is always started from scratch, so stop
         // without destroy only used for testing and debugging.
         //
+        self.pg_ctl(
+            if destroy {
+                &["-m", "immediate", "stop"]
+            } else {
+                &["stop"]
+            },
+            &None,
+        )?;
+
+        // Also wait for the compute_ctl process to die. It might have some cleanup
+        // work to do after postgres stops, like syncing safekeepers, etc.
+        //
+        self.wait_for_compute_ctl_to_exit()?;
         if destroy {
-            self.pg_ctl(&["-m", "immediate", "stop"], &None)?;
             println!(
                 "Destroying postgres data directory '{}'",
                 self.pgdata().to_str().unwrap()
             );
             std::fs::remove_dir_all(self.endpoint_path())?;
-        } else {
-            self.pg_ctl(&["stop"], &None)?;
         }
         Ok(())
     }

control_plane/src/pageserver.rs

@@ -345,6 +345,11 @@ impl PageServerNode {
                 .remove("compaction_threshold")
                 .map(|x| x.parse::<usize>())
                 .transpose()?,
+            compaction_algorithm: settings
+                .remove("compaction_algorithm")
+                .map(serde_json::from_str)
+                .transpose()
+                .context("Failed to parse 'compaction_algorithm' json")?,
             gc_horizon: settings
                 .remove("gc_horizon")
                 .map(|x| x.parse::<u64>())
@@ -440,6 +445,11 @@ impl PageServerNode {
                     .map(|x| x.parse::<usize>())
                     .transpose()
                     .context("Failed to parse 'compaction_threshold' as an integer")?,
+                compaction_algorithm: settings
+                    .remove("compactin_algorithm")
+                    .map(serde_json::from_str)
+                    .transpose()
+                    .context("Failed to parse 'compaction_algorithm' json")?,
                 gc_horizon: settings
                     .remove("gc_horizon")
                     .map(|x| x.parse::<u64>())

docker-compose/compute_wrapper/var/db/postgres/specs/spec.json

@@ -25,7 +25,7 @@
             },
             {
                 "name": "wal_level",
-                "value": "replica",
+                "value": "logical",
                 "vartype": "enum"
             },
             {

docs/error-handling.md

@@ -188,11 +188,60 @@ that.
 
 ## Error message style
 
+### PostgreSQL extensions
+
 PostgreSQL has a style guide for writing error messages:
 
 https://www.postgresql.org/docs/current/error-style-guide.html
 
 Follow that guide when writing error messages in the PostgreSQL
-extension. We don't follow it strictly in the pageserver and
-safekeeper, but the advice in the PostgreSQL style guide is generally
-good, and you can't go wrong by following it.
+extensions.
+
+### Neon Rust code
+
+#### Anyhow Context
+
+When adding anyhow `context()`, use form `present-tense-verb+action`.
+
+Example:
+- Bad: `file.metadata().context("could not get file metadata")?;`
+- Good: `file.metadata().context("get file metadata")?;`
+
+#### Logging Errors
+
+When logging any error `e`, use `could not {e:#}` or `failed to {e:#}`.
+
+If `e` is an `anyhow` error and you want to log the backtrace that it contains,
+use `{e:?}` instead of `{e:#}`.
+
+#### Rationale
+
+The `{:#}` ("alternate Display") of an `anyhow` error chain is concatenation fo the contexts, using `: `.
+
+For example, the following Rust code will result in output
+```
+ERROR  failed to list users: load users from server: parse response: invalid json
+```
+
+This is more concise / less noisy than what happens if you do `.context("could not ...")?` at each level, i.e.:
+
+```
+ERROR  could not list users: could not load users from server: could not parse response: invalid json
+```
+
+
+```rust
+fn main() {
+  match list_users().context("list users") else {
+    Ok(_) => ...,
+    Err(e) => tracing::error!("failed to {e:#}"),
+  }
+}
+fn list_users() {
+  http_get_users().context("load users from server")?;
+}
+fn http_get_users() {
+  let response = client....?;
+  response.parse().context("parse response")?; // fails with serde error "invalid json"
+}
+```

docs/pageserver-services.md

@@ -96,6 +96,16 @@ prefix_in_bucket = '/test_prefix/'
 
 `AWS_SECRET_ACCESS_KEY` and `AWS_ACCESS_KEY_ID` env variables can be used to specify the S3 credentials if needed.
 
+or
+
+```toml
+[remote_storage]
+container_name = 'some-container-name'
+container_region = 'us-east'
+prefix_in_container = '/test-prefix/'
+```
+
+`AZURE_STORAGE_ACCOUNT` and `AZURE_STORAGE_ACCESS_KEY` env variables can be used to specify the azure credentials if needed.
 
 ## Repository background tasks
 

docs/updating-postgres.md

@@ -0,0 +1,108 @@
+# Updating Postgres
+
+## Minor Versions
+
+When upgrading to a new minor version of Postgres, please follow these steps:
+
+_Example: 15.4 is the new minor version to upgrade to from 15.3._
+
+1. Clone the Neon Postgres repository if you have not done so already.
+
+    ```shell
+    git clone git@github.com:neondatabase/postgres.git
+    ```
+
+1. Add the Postgres upstream remote.
+
+    ```shell
+    git remote add upstream https://git.postgresql.org/git/postgresql.git
+    ```
+
+1. Create a new branch based on the stable branch you are updating.
+
+    ```shell
+    git checkout -b my-branch REL_15_STABLE_neon
+    ```
+
+1. Tag the last commit on the stable branch you are updating.
+
+    ```shell
+    git tag REL_15_3_neon
+    ```
+
+1. Push the new tag to the Neon Postgres repository.
+
+    ```shell
+    git push origin REL_15_3_neon
+    ```
+
+1. Find the release tags you're looking for. They are of the form `REL_X_Y`.
+
+1. Rebase the branch you created on the tag and resolve any conflicts.
+
+    ```shell
+    git fetch upstream REL_15_4
+    git rebase REL_15_4
+    ```
+
+1. Run the Postgres test suite to make sure our commits have not affected
+Postgres in a negative way.
+
+    ```shell
+    make check
+    # OR
+    meson test -C builddir
+    ```
+
+1. Push your branch to the Neon Postgres repository.
+
+    ```shell
+    git push origin my-branch
+    ```
+
+1. Clone the Neon repository if you have not done so already.
+
+    ```shell
+    git clone git@github.com:neondatabase/neon.git
+    ```
+
+1. Create a new branch.
+
+1. Change the `revisions.json` file to point at the HEAD of your Postgres
+branch.
+
+1. Update the Git submodule.
+
+    ```shell
+    git submodule set-branch --branch my-branch vendor/postgres-v15
+    git submodule update --remote vendor/postgres-v15
+    ```
+
+1. Run the Neon test suite to make sure that Neon is still good to go on this
+minor Postgres release.
+
+    ```shell
+    ./scripts/poetry -k pg15
+    ```
+
+1. Commit your changes.
+
+1. Create a pull request, and wait for CI to go green.
+
+1. Force push the rebased Postgres branches into the Neon Postgres repository.
+
+    ```shell
+    git push --force origin my-branch:REL_15_STABLE_neon
+    ```
+
+    It may require disabling various branch protections.
+
+1. Update your Neon PR to point at the branches.
+
+    ```shell
+    git submodule set-branch --branch REL_15_STABLE_neon vendor/postgres-v15
+    git commit --amend --no-edit
+    git push --force origin
+    ```
+
+1. Merge the pull request after getting approval(s) and CI completion.

libs/compute_api/src/spec.rs

@@ -190,6 +190,8 @@ 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,
 }
 
@@ -200,6 +202,12 @@ pub struct Database {
     pub name: PgIdent,
     pub owner: PgIdent,
     pub options: GenericOptions,
+    // These are derived flags, not present in the spec file.
+    // They are never set by the control plane.
+    #[serde(skip_deserializing, default)]
+    pub restrict_conn: bool,
+    #[serde(skip_deserializing, default)]
+    pub invalid: bool,
 }
 
 /// Common type representing both SQL statement params with or without value,

libs/compute_api/tests/cluster_spec.json

@@ -76,7 +76,7 @@
             },
             {
                 "name": "wal_level",
-                "value": "replica",
+                "value": "logical",
                 "vartype": "enum"
             },
             {

libs/metrics/src/lib.rs

@@ -89,14 +89,14 @@ pub const DISK_WRITE_SECONDS_BUCKETS: &[f64] = &[
     0.000_050, 0.000_100, 0.000_500, 0.001, 0.003, 0.005, 0.01, 0.05, 0.1, 0.3, 0.5,
 ];
 
-pub fn set_build_info_metric(revision: &str) {
+pub fn set_build_info_metric(revision: &str, build_tag: &str) {
     let metric = register_int_gauge_vec!(
         "libmetrics_build_info",
         "Build/version information",
-        &["revision"]
+        &["revision", "build_tag"]
     )
     .expect("Failed to register build info metric");
-    metric.with_label_values(&[revision]).set(1);
+    metric.with_label_values(&[revision, build_tag]).set(1);
 }
 
 // Records I/O stats in a "cross-platform" way.

libs/metrics/src/wrappers.rs

@@ -1,6 +1,6 @@
 use std::io::{Read, Result, Write};
 
-/// A wrapper for an object implementing [Read](std::io::Read)
+/// A wrapper for an object implementing [Read]
 /// which allows a closure to observe the amount of bytes read.
 /// This is useful in conjunction with metrics (e.g. [IntCounter](crate::IntCounter)).
 ///
@@ -51,17 +51,17 @@ impl<'a, T> CountedReader<'a, T> {
         }
     }
 
-    /// Get an immutable reference to the underlying [Read](std::io::Read) implementor
+    /// Get an immutable reference to the underlying [Read] implementor
     pub fn inner(&self) -> &T {
         &self.reader
     }
 
-    /// Get a mutable reference to the underlying [Read](std::io::Read) implementor
+    /// Get a mutable reference to the underlying [Read] implementor
     pub fn inner_mut(&mut self) -> &mut T {
         &mut self.reader
     }
 
-    /// Consume the wrapper and return the underlying [Read](std::io::Read) implementor
+    /// Consume the wrapper and return the underlying [Read] implementor
     pub fn into_inner(self) -> T {
         self.reader
     }
@@ -75,7 +75,7 @@ impl<T: Read> Read for CountedReader<'_, T> {
     }
 }
 
-/// A wrapper for an object implementing [Write](std::io::Write)
+/// A wrapper for an object implementing [Write]
 /// which allows a closure to observe the amount of bytes written.
 /// This is useful in conjunction with metrics (e.g. [IntCounter](crate::IntCounter)).
 ///
@@ -122,17 +122,17 @@ impl<'a, T> CountedWriter<'a, T> {
         }
     }
 
-    /// Get an immutable reference to the underlying [Write](std::io::Write) implementor
+    /// Get an immutable reference to the underlying [Write] implementor
     pub fn inner(&self) -> &T {
         &self.writer
     }
 
-    /// Get a mutable reference to the underlying [Write](std::io::Write) implementor
+    /// Get a mutable reference to the underlying [Write] implementor
     pub fn inner_mut(&mut self) -> &mut T {
         &mut self.writer
     }
 
-    /// Consume the wrapper and return the underlying [Write](std::io::Write) implementor
+    /// Consume the wrapper and return the underlying [Write] implementor
     pub fn into_inner(self) -> T {
         self.writer
     }

libs/pageserver_api/src/control_api.rs

@@ -17,7 +17,7 @@ pub struct ReAttachRequest {
 pub struct ReAttachResponseTenant {
     #[serde_as(as = "DisplayFromStr")]
     pub id: TenantId,
-    pub generation: u32,
+    pub gen: u32,
 }
 
 #[derive(Serialize, Deserialize)]

libs/pageserver_api/src/models.rs

@@ -110,7 +110,6 @@ impl TenantState {
             // So, return `Maybe` while Attaching, making Console wait for the attach task to finish.
             Self::Attaching | Self::Activating(ActivatingFrom::Attaching) => Maybe,
             // tenant mgr startup distinguishes attaching from loading via marker file.
-            // If it's loading, there is no attach marker file, i.e., attach had finished in the past.
             Self::Loading | Self::Activating(ActivatingFrom::Loading) => Attached,
             // We only reach Active after successful load / attach.
             // So, call atttachment status Attached.
@@ -228,6 +227,8 @@ pub struct TenantConfig {
     pub compaction_target_size: Option<u64>,
     pub compaction_period: Option<String>,
     pub compaction_threshold: Option<usize>,
+    // defer parsing compaction_algorithm, like eviction_policy
+    pub compaction_algorithm: Option<serde_json::Value>,
     pub gc_horizon: Option<u64>,
     pub gc_period: Option<String>,
     pub image_creation_threshold: Option<usize>,
@@ -325,6 +326,7 @@ impl TenantConfigRequest {
             compaction_target_size: None,
             compaction_period: None,
             compaction_threshold: None,
+            compaction_algorithm: None,
             gc_horizon: None,
             gc_period: None,
             image_creation_threshold: None,

libs/pageserver_api/src/reltag.rs

@@ -22,9 +22,9 @@ use postgres_ffi::Oid;
 /// [See more related comments here](https:///github.com/postgres/postgres/blob/99c5852e20a0987eca1c38ba0c09329d4076b6a0/src/include/storage/relfilenode.h#L57).
 ///
 // FIXME: should move 'forknum' as last field to keep this consistent with Postgres.
-// Then we could replace the custo Ord and PartialOrd implementations below with
-// deriving them.
-#[derive(Debug, PartialEq, Eq, Hash, Clone, Copy, Serialize, Deserialize)]
+// Then we could replace the custom Ord and PartialOrd implementations below with
+// deriving them. This will require changes in walredoproc.c.
+#[derive(Debug, PartialEq, Eq, Hash, Clone, Copy, Serialize)]
 pub struct RelTag {
     pub forknum: u8,
     pub spcnode: Oid,
@@ -40,21 +40,9 @@ impl PartialOrd for RelTag {
 
 impl Ord for RelTag {
     fn cmp(&self, other: &Self) -> Ordering {
-        let mut cmp = self.spcnode.cmp(&other.spcnode);
-        if cmp != Ordering::Equal {
-            return cmp;
-        }
-        cmp = self.dbnode.cmp(&other.dbnode);
-        if cmp != Ordering::Equal {
-            return cmp;
-        }
-        cmp = self.relnode.cmp(&other.relnode);
-        if cmp != Ordering::Equal {
-            return cmp;
-        }
-        cmp = self.forknum.cmp(&other.forknum);
-
-        cmp
+        // Custom ordering where we put forknum to the end of the list
+        let other_tup = (other.spcnode, other.dbnode, other.relnode, other.forknum);
+        (self.spcnode, self.dbnode, self.relnode, self.forknum).cmp(&other_tup)
     }
 }
 

libs/postgres_backend/src/lib.rs

@@ -19,8 +19,8 @@ use tracing::{debug, error, info, trace};
 
 use pq_proto::framed::{ConnectionError, Framed, FramedReader, FramedWriter};
 use pq_proto::{
-    BeMessage, FeMessage, FeStartupPacket, ProtocolError, SQLSTATE_INTERNAL_ERROR,
-    SQLSTATE_SUCCESSFUL_COMPLETION,
+    BeMessage, FeMessage, FeStartupPacket, ProtocolError, SQLSTATE_ADMIN_SHUTDOWN,
+    SQLSTATE_INTERNAL_ERROR, SQLSTATE_SUCCESSFUL_COMPLETION,
 };
 
 /// An error, occurred during query processing:
@@ -30,6 +30,9 @@ pub enum QueryError {
     /// The connection was lost while processing the query.
     #[error(transparent)]
     Disconnected(#[from] ConnectionError),
+    /// We were instructed to shutdown while processing the query
+    #[error("Shutting down")]
+    Shutdown,
     /// Some other error
     #[error(transparent)]
     Other(#[from] anyhow::Error),
@@ -44,7 +47,8 @@ impl From<io::Error> for QueryError {
 impl QueryError {
     pub fn pg_error_code(&self) -> &'static [u8; 5] {
         match self {
-            Self::Disconnected(_) => b"08006",         // connection failure
+            Self::Disconnected(_) => b"08006", // connection failure
+            Self::Shutdown => SQLSTATE_ADMIN_SHUTDOWN,
             Self::Other(_) => SQLSTATE_INTERNAL_ERROR, // internal error
         }
     }
@@ -238,6 +242,7 @@ impl<IO: AsyncRead + AsyncWrite + Unpin> MaybeWriteOnly<IO> {
         }
     }
 
+    /// Cancellation safe as long as the underlying IO is cancellation safe.
     async fn shutdown(&mut self) -> io::Result<()> {
         match self {
             MaybeWriteOnly::Full(framed) => framed.shutdown().await,
@@ -389,14 +394,37 @@ impl<IO: AsyncRead + AsyncWrite + Unpin> PostgresBackend<IO> {
         shutdown_watcher: F,
     ) -> Result<(), QueryError>
     where
-        F: Fn() -> S,
+        F: Fn() -> S + Clone,
         S: Future,
     {
-        let ret = self.run_message_loop(handler, shutdown_watcher).await;
-        // socket might be already closed, e.g. if previously received error,
-        // so ignore result.
-        self.framed.shutdown().await.ok();
-        ret
+        let ret = self
+            .run_message_loop(handler, shutdown_watcher.clone())
+            .await;
+
+        tokio::select! {
+            _ = shutdown_watcher() => {
+                // do nothing; we most likely got already stopped by shutdown and will log it next.
+            }
+            _ = self.framed.shutdown() => {
+                // socket might be already closed, e.g. if previously received error,
+                // so ignore result.
+            },
+        }
+
+        match ret {
+            Ok(()) => Ok(()),
+            Err(QueryError::Shutdown) => {
+                info!("Stopped due to shutdown");
+                Ok(())
+            }
+            Err(QueryError::Disconnected(e)) => {
+                info!("Disconnected ({e:#})");
+                // Disconnection is not an error: we just use it that way internally to drop
+                // out of loops.
+                Ok(())
+            }
+            e => e,
+        }
     }
 
     async fn run_message_loop<F, S>(
@@ -416,15 +444,11 @@ impl<IO: AsyncRead + AsyncWrite + Unpin> PostgresBackend<IO> {
             _ = shutdown_watcher() => {
                 // We were requested to shut down.
                 tracing::info!("shutdown request received during handshake");
-                return Ok(())
+                return Err(QueryError::Shutdown)
             },
 
-            result = self.handshake(handler) => {
-                // Handshake complete.
-                result?;
-                if self.state == ProtoState::Closed {
-                    return Ok(()); // EOF during handshake
-                }
+            handshake_r = self.handshake(handler) => {
+                handshake_r?;
             }
         );
 
@@ -435,7 +459,7 @@ impl<IO: AsyncRead + AsyncWrite + Unpin> PostgresBackend<IO> {
             _ = shutdown_watcher() => {
                 // We were requested to shut down.
                 tracing::info!("shutdown request received in run_message_loop");
-                Ok(None)
+                return Err(QueryError::Shutdown)
             },
             msg = self.read_message() => { msg },
         )? {
@@ -447,7 +471,14 @@ impl<IO: AsyncRead + AsyncWrite + Unpin> PostgresBackend<IO> {
                 _ = shutdown_watcher() => {
                     // We were requested to shut down.
                     tracing::info!("shutdown request received during response flush");
-                    return Ok(())
+
+                    // If we exited process_message with a shutdown error, there may be
+                    // some valid response content on in our transmit buffer: permit sending
+                    // this within a short timeout.  This is a best effort thing so we don't
+                    // care about the result.
+                    tokio::time::timeout(std::time::Duration::from_millis(500), self.flush()).await.ok();
+
+                    return Err(QueryError::Shutdown)
                 },
                 flush_r = self.flush() => {
                     flush_r?;
@@ -560,7 +591,9 @@ impl<IO: AsyncRead + AsyncWrite + Unpin> PostgresBackend<IO> {
                         self.peer_addr
                     );
                     self.state = ProtoState::Closed;
-                    return Ok(());
+                    return Err(QueryError::Disconnected(ConnectionError::Protocol(
+                        ProtocolError::Protocol("EOF during handshake".to_string()),
+                    )));
                 }
             }
         }
@@ -599,7 +632,9 @@ impl<IO: AsyncRead + AsyncWrite + Unpin> PostgresBackend<IO> {
                         self.peer_addr
                     );
                     self.state = ProtoState::Closed;
-                    return Ok(());
+                    return Err(QueryError::Disconnected(ConnectionError::Protocol(
+                        ProtocolError::Protocol("EOF during auth".to_string()),
+                    )));
                 }
             }
         }
@@ -923,6 +958,7 @@ impl<'a, IO: AsyncRead + AsyncWrite + Unpin> AsyncWrite for CopyDataWriter<'a, I
 pub fn short_error(e: &QueryError) -> String {
     match e {
         QueryError::Disconnected(connection_error) => connection_error.to_string(),
+        QueryError::Shutdown => "shutdown".to_string(),
         QueryError::Other(e) => format!("{e:#}"),
     }
 }
@@ -939,6 +975,9 @@ fn log_query_error(query: &str, e: &QueryError) {
         QueryError::Disconnected(other_connection_error) => {
             error!("query handler for '{query}' failed with connection error: {other_connection_error:?}")
         }
+        QueryError::Shutdown => {
+            info!("query handler for '{query}' cancelled during tenant shutdown")
+        }
         QueryError::Other(e) => {
             error!("query handler for '{query}' failed: {e:?}");
         }

libs/postgres_ffi/src/lib.rs

@@ -131,6 +131,7 @@ pub const MAX_SEND_SIZE: usize = XLOG_BLCKSZ * 16;
 
 // Export some version independent functions that are used outside of this mod
 pub use v14::xlog_utils::encode_logical_message;
+pub use v14::xlog_utils::from_pg_timestamp;
 pub use v14::xlog_utils::get_current_timestamp;
 pub use v14::xlog_utils::to_pg_timestamp;
 pub use v14::xlog_utils::XLogFileName;

libs/postgres_ffi/src/pg_constants.rs

@@ -220,6 +220,10 @@ pub const XLOG_CHECKPOINT_ONLINE: u8 = 0x10;
 pub const XLP_FIRST_IS_CONTRECORD: u16 = 0x0001;
 pub const XLP_LONG_HEADER: u16 = 0x0002;
 
+/* From replication/slot.h */
+pub const REPL_SLOT_ON_DISK_OFFSETOF_RESTART_LSN: usize = 4*4  /* offset of `slotdata` in ReplicationSlotOnDisk  */
+   + 64 /* NameData */  + 4*4;
+
 /* From fsm_internals.h */
 const FSM_NODES_PER_PAGE: usize = BLCKSZ as usize - SIZEOF_PAGE_HEADER_DATA - 4;
 const FSM_NON_LEAF_NODES_PER_PAGE: usize = BLCKSZ as usize / 2 - 1;

libs/postgres_ffi/src/xlog_utils.rs

@@ -136,21 +136,42 @@ pub fn get_current_timestamp() -> TimestampTz {
     to_pg_timestamp(SystemTime::now())
 }
 
-pub fn to_pg_timestamp(time: SystemTime) -> TimestampTz {
-    const UNIX_EPOCH_JDATE: u64 = 2440588; /* == date2j(1970, 1, 1) */
-    const POSTGRES_EPOCH_JDATE: u64 = 2451545; /* == date2j(2000, 1, 1) */
+// Module to reduce the scope of the constants
+mod timestamp_conversions {
+    use std::time::Duration;
+
+    use super::*;
+
+    const UNIX_EPOCH_JDATE: u64 = 2440588; // == date2j(1970, 1, 1)
+    const POSTGRES_EPOCH_JDATE: u64 = 2451545; // == date2j(2000, 1, 1)
     const SECS_PER_DAY: u64 = 86400;
     const USECS_PER_SEC: u64 = 1000000;
-    match time.duration_since(SystemTime::UNIX_EPOCH) {
-        Ok(n) => {
-            ((n.as_secs() - ((POSTGRES_EPOCH_JDATE - UNIX_EPOCH_JDATE) * SECS_PER_DAY))
-                * USECS_PER_SEC
-                + n.subsec_micros() as u64) as i64
+    const SECS_DIFF_UNIX_TO_POSTGRES_EPOCH: u64 =
+        (POSTGRES_EPOCH_JDATE - UNIX_EPOCH_JDATE) * SECS_PER_DAY;
+
+    pub fn to_pg_timestamp(time: SystemTime) -> TimestampTz {
+        match time.duration_since(SystemTime::UNIX_EPOCH) {
+            Ok(n) => {
+                ((n.as_secs() - SECS_DIFF_UNIX_TO_POSTGRES_EPOCH) * USECS_PER_SEC
+                    + n.subsec_micros() as u64) as i64
+            }
+            Err(_) => panic!("SystemTime before UNIX EPOCH!"),
         }
-        Err(_) => panic!("SystemTime before UNIX EPOCH!"),
+    }
+
+    pub fn from_pg_timestamp(time: TimestampTz) -> SystemTime {
+        let time: u64 = time
+            .try_into()
+            .expect("timestamp before millenium (postgres epoch)");
+        let since_unix_epoch = time + SECS_DIFF_UNIX_TO_POSTGRES_EPOCH * USECS_PER_SEC;
+        SystemTime::UNIX_EPOCH
+            .checked_add(Duration::from_micros(since_unix_epoch))
+            .expect("SystemTime overflow")
     }
 }
 
+pub use timestamp_conversions::{from_pg_timestamp, to_pg_timestamp};
+
 // Returns (aligned) end_lsn of the last record in data_dir with WAL segments.
 // start_lsn must point to some previously known record boundary (beginning of
 // the next record). If no valid record after is found, start_lsn is returned
@@ -481,4 +502,24 @@ pub fn encode_logical_message(prefix: &str, message: &str) -> Vec<u8> {
     wal
 }
 
-// If you need to craft WAL and write tests for this module, put it at wal_craft crate.
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn test_ts_conversion() {
+        let now = SystemTime::now();
+        let round_trip = from_pg_timestamp(to_pg_timestamp(now));
+
+        let now_since = now.duration_since(SystemTime::UNIX_EPOCH).unwrap();
+        let round_trip_since = round_trip.duration_since(SystemTime::UNIX_EPOCH).unwrap();
+        assert_eq!(now_since.as_micros(), round_trip_since.as_micros());
+
+        let now_pg = get_current_timestamp();
+        let round_trip_pg = to_pg_timestamp(from_pg_timestamp(now_pg));
+
+        assert_eq!(now_pg, round_trip_pg);
+    }
+
+    // If you need to craft WAL and write tests for this module, put it at wal_craft crate.
+}

libs/postgres_ffi/wal_craft/src/lib.rs

@@ -14,6 +14,7 @@ macro_rules! xlog_utils_test {
     ($version:ident) => {
         #[path = "."]
         mod $version {
+            #[allow(unused_imports)]
             pub use postgres_ffi::$version::wal_craft_test_export::*;
             #[allow(clippy::duplicate_mod)]
             #[cfg(test)]

libs/pq_proto/src/framed.rs

@@ -214,27 +214,24 @@ where
     }
 }
 
+/// Cancellation safe as long as the AsyncWrite is cancellation safe.
 async fn flush<S: AsyncWrite + Unpin>(
     stream: &mut S,
     write_buf: &mut BytesMut,
 ) -> Result<(), io::Error> {
     while write_buf.has_remaining() {
-        let bytes_written = stream.write(write_buf.chunk()).await?;
+        let bytes_written = stream.write_buf(write_buf).await?;
         if bytes_written == 0 {
             return Err(io::Error::new(
                 ErrorKind::WriteZero,
                 "failed to write message",
             ));
         }
-        // The advanced part will be garbage collected, likely during shifting
-        // data left on next attempt to write to buffer when free space is not
-        // enough.
-        write_buf.advance(bytes_written);
     }
-    write_buf.clear();
     stream.flush().await
 }
 
+/// Cancellation safe as long as the AsyncWrite is cancellation safe.
 async fn shutdown<S: AsyncWrite + Unpin>(
     stream: &mut S,
     write_buf: &mut BytesMut,

libs/pq_proto/src/lib.rs

@@ -670,6 +670,7 @@ pub fn read_cstr(buf: &mut Bytes) -> Result<Bytes, ProtocolError> {
 }
 
 pub const SQLSTATE_INTERNAL_ERROR: &[u8; 5] = b"XX000";
+pub const SQLSTATE_ADMIN_SHUTDOWN: &[u8; 5] = b"57P01";
 pub const SQLSTATE_SUCCESSFUL_COMPLETION: &[u8; 5] = b"00000";
 
 impl<'a> BeMessage<'a> {

libs/remote_storage/Cargo.toml

@@ -13,6 +13,7 @@ aws-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"] }
 serde.workspace = true
@@ -26,6 +27,13 @@ metrics.workspace = true
 utils.workspace = true
 pin-project-lite.workspace = true
 workspace_hack.workspace = true
+azure_core.workspace = true
+azure_identity.workspace = true
+azure_storage.workspace = true
+azure_storage_blobs.workspace = true
+futures-util.workspace = true
+http-types.workspace = true
+itertools.workspace = true
 
 [dev-dependencies]
 camino-tempfile.workspace = true

libs/remote_storage/src/azure_blob.rs

@@ -0,0 +1,337 @@
+//! Azure Blob Storage wrapper
+
+use std::env;
+use std::num::NonZeroU32;
+use std::sync::Arc;
+use std::{borrow::Cow, collections::HashMap, io::Cursor};
+
+use super::REMOTE_STORAGE_PREFIX_SEPARATOR;
+use anyhow::Result;
+use azure_core::request_options::{MaxResults, Metadata, Range};
+use azure_core::Header;
+use azure_identity::DefaultAzureCredential;
+use azure_storage::StorageCredentials;
+use azure_storage_blobs::prelude::ClientBuilder;
+use azure_storage_blobs::{
+    blob::operations::GetBlobBuilder,
+    prelude::{BlobClient, ContainerClient},
+};
+use futures_util::StreamExt;
+use http_types::StatusCode;
+use tokio::io::AsyncRead;
+use tracing::debug;
+
+use crate::s3_bucket::RequestKind;
+use crate::{
+    AzureConfig, ConcurrencyLimiter, Download, DownloadError, Listing, ListingMode, RemotePath,
+    RemoteStorage, StorageMetadata,
+};
+
+pub struct AzureBlobStorage {
+    client: ContainerClient,
+    prefix_in_container: Option<String>,
+    max_keys_per_list_response: Option<NonZeroU32>,
+    concurrency_limiter: ConcurrencyLimiter,
+}
+
+impl AzureBlobStorage {
+    pub fn new(azure_config: &AzureConfig) -> Result<Self> {
+        debug!(
+            "Creating azure remote storage for azure container {}",
+            azure_config.container_name
+        );
+
+        let account = env::var("AZURE_STORAGE_ACCOUNT").expect("missing AZURE_STORAGE_ACCOUNT");
+
+        // If the `AZURE_STORAGE_ACCESS_KEY` env var has an access key, use that,
+        // otherwise try the token based credentials.
+        let credentials = if let Ok(access_key) = env::var("AZURE_STORAGE_ACCESS_KEY") {
+            StorageCredentials::access_key(account.clone(), access_key)
+        } else {
+            let token_credential = DefaultAzureCredential::default();
+            StorageCredentials::token_credential(Arc::new(token_credential))
+        };
+
+        let builder = ClientBuilder::new(account, credentials);
+
+        let client = builder.container_client(azure_config.container_name.to_owned());
+
+        let max_keys_per_list_response =
+            if let Some(limit) = azure_config.max_keys_per_list_response {
+                Some(
+                    NonZeroU32::new(limit as u32)
+                        .ok_or_else(|| anyhow::anyhow!("max_keys_per_list_response can't be 0"))?,
+                )
+            } else {
+                None
+            };
+
+        Ok(AzureBlobStorage {
+            client,
+            prefix_in_container: azure_config.prefix_in_container.to_owned(),
+            max_keys_per_list_response,
+            concurrency_limiter: ConcurrencyLimiter::new(azure_config.concurrency_limit.get()),
+        })
+    }
+
+    pub fn relative_path_to_name(&self, path: &RemotePath) -> String {
+        assert_eq!(std::path::MAIN_SEPARATOR, REMOTE_STORAGE_PREFIX_SEPARATOR);
+        let path_string = path
+            .get_path()
+            .as_str()
+            .trim_end_matches(REMOTE_STORAGE_PREFIX_SEPARATOR);
+        match &self.prefix_in_container {
+            Some(prefix) => {
+                if prefix.ends_with(REMOTE_STORAGE_PREFIX_SEPARATOR) {
+                    prefix.clone() + path_string
+                } else {
+                    format!("{prefix}{REMOTE_STORAGE_PREFIX_SEPARATOR}{path_string}")
+                }
+            }
+            None => path_string.to_string(),
+        }
+    }
+
+    fn name_to_relative_path(&self, key: &str) -> RemotePath {
+        let relative_path =
+            match key.strip_prefix(self.prefix_in_container.as_deref().unwrap_or_default()) {
+                Some(stripped) => stripped,
+                // we rely on Azure to return properly prefixed paths
+                // for requests with a certain prefix
+                None => panic!(
+                    "Key {key} does not start with container prefix {:?}",
+                    self.prefix_in_container
+                ),
+            };
+        RemotePath(
+            relative_path
+                .split(REMOTE_STORAGE_PREFIX_SEPARATOR)
+                .collect(),
+        )
+    }
+
+    async fn download_for_builder(
+        &self,
+        metadata: StorageMetadata,
+        builder: GetBlobBuilder,
+    ) -> Result<Download, DownloadError> {
+        let mut response = builder.into_stream();
+
+        // TODO give proper streaming response instead of buffering into RAM
+        // https://github.com/neondatabase/neon/issues/5563
+        let mut buf = Vec::new();
+        while let Some(part) = response.next().await {
+            let part = part.map_err(to_download_error)?;
+            let data = part
+                .data
+                .collect()
+                .await
+                .map_err(|e| DownloadError::Other(e.into()))?;
+            buf.extend_from_slice(&data.slice(..));
+        }
+        Ok(Download {
+            download_stream: Box::pin(Cursor::new(buf)),
+            metadata: Some(metadata),
+        })
+    }
+    // TODO get rid of this function once we have metadata included in the response
+    // https://github.com/Azure/azure-sdk-for-rust/issues/1439
+    async fn get_metadata(
+        &self,
+        blob_client: &BlobClient,
+    ) -> Result<StorageMetadata, DownloadError> {
+        let builder = blob_client.get_metadata();
+
+        let response = builder.into_future().await.map_err(to_download_error)?;
+        let mut map = HashMap::new();
+
+        for md in response.metadata.iter() {
+            map.insert(
+                md.name().as_str().to_string(),
+                md.value().as_str().to_string(),
+            );
+        }
+        Ok(StorageMetadata(map))
+    }
+
+    async fn permit(&self, kind: RequestKind) -> tokio::sync::SemaphorePermit<'_> {
+        self.concurrency_limiter
+            .acquire(kind)
+            .await
+            .expect("semaphore is never closed")
+    }
+}
+
+fn to_azure_metadata(metadata: StorageMetadata) -> Metadata {
+    let mut res = Metadata::new();
+    for (k, v) in metadata.0.into_iter() {
+        res.insert(k, v);
+    }
+    res
+}
+
+fn to_download_error(error: azure_core::Error) -> DownloadError {
+    if let Some(http_err) = error.as_http_error() {
+        match http_err.status() {
+            StatusCode::NotFound => DownloadError::NotFound,
+            StatusCode::BadRequest => DownloadError::BadInput(anyhow::Error::new(error)),
+            _ => DownloadError::Other(anyhow::Error::new(error)),
+        }
+    } else {
+        DownloadError::Other(error.into())
+    }
+}
+
+#[async_trait::async_trait]
+impl RemoteStorage for AzureBlobStorage {
+    async fn list(
+        &self,
+        prefix: Option<&RemotePath>,
+        mode: ListingMode,
+    ) -> anyhow::Result<Listing, DownloadError> {
+        // get the passed prefix or if it is not set use prefix_in_bucket value
+        let list_prefix = prefix
+            .map(|p| self.relative_path_to_name(p))
+            .or_else(|| self.prefix_in_container.clone())
+            .map(|mut p| {
+                // required to end with a separator
+                // otherwise request will return only the entry of a prefix
+                if matches!(mode, ListingMode::WithDelimiter)
+                    && !p.ends_with(REMOTE_STORAGE_PREFIX_SEPARATOR)
+                {
+                    p.push(REMOTE_STORAGE_PREFIX_SEPARATOR);
+                }
+                p
+            });
+
+        let mut builder = self.client.list_blobs();
+
+        if let ListingMode::WithDelimiter = mode {
+            builder = builder.delimiter(REMOTE_STORAGE_PREFIX_SEPARATOR.to_string());
+        }
+
+        if let Some(prefix) = list_prefix {
+            builder = builder.prefix(Cow::from(prefix.to_owned()));
+        }
+
+        if let Some(limit) = self.max_keys_per_list_response {
+            builder = builder.max_results(MaxResults::new(limit));
+        }
+
+        let mut response = builder.into_stream();
+        let mut res = Listing::default();
+        while let Some(l) = response.next().await {
+            let entry = l.map_err(to_download_error)?;
+            let prefix_iter = entry
+                .blobs
+                .prefixes()
+                .map(|prefix| self.name_to_relative_path(&prefix.name));
+            res.prefixes.extend(prefix_iter);
+
+            let blob_iter = entry
+                .blobs
+                .blobs()
+                .map(|k| self.name_to_relative_path(&k.name));
+            res.keys.extend(blob_iter);
+        }
+        Ok(res)
+    }
+    async fn upload(
+        &self,
+        mut from: impl AsyncRead + Unpin + Send + Sync + 'static,
+        data_size_bytes: usize,
+        to: &RemotePath,
+        metadata: Option<StorageMetadata>,
+    ) -> anyhow::Result<()> {
+        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 mut builder = blob_client.put_block_blob(body);
+
+        if let Some(metadata) = metadata {
+            builder = builder.metadata(to_azure_metadata(metadata));
+        }
+
+        let _response = builder.into_future().await?;
+
+        Ok(())
+    }
+
+    async fn download(&self, from: &RemotePath) -> Result<Download, DownloadError> {
+        let _permit = self.permit(RequestKind::Get).await;
+        let blob_client = self.client.blob_client(self.relative_path_to_name(from));
+
+        let metadata = self.get_metadata(&blob_client).await?;
+
+        let builder = blob_client.get();
+
+        self.download_for_builder(metadata, builder).await
+    }
+
+    async fn download_byte_range(
+        &self,
+        from: &RemotePath,
+        start_inclusive: u64,
+        end_exclusive: Option<u64>,
+    ) -> Result<Download, DownloadError> {
+        let _permit = self.permit(RequestKind::Get).await;
+        let blob_client = self.client.blob_client(self.relative_path_to_name(from));
+
+        let metadata = self.get_metadata(&blob_client).await?;
+
+        let mut builder = blob_client.get();
+
+        if let Some(end_exclusive) = end_exclusive {
+            builder = builder.range(Range::new(start_inclusive, end_exclusive));
+        } else {
+            // Open ranges are not supported by the SDK so we work around
+            // by setting the upper limit extremely high (but high enough
+            // to still be representable by signed 64 bit integers).
+            // TODO remove workaround once the SDK adds open range support
+            // https://github.com/Azure/azure-sdk-for-rust/issues/1438
+            let end_exclusive = u64::MAX / 4;
+            builder = builder.range(Range::new(start_inclusive, end_exclusive));
+        }
+
+        self.download_for_builder(metadata, builder).await
+    }
+
+    async fn delete(&self, path: &RemotePath) -> anyhow::Result<()> {
+        let _permit = self.permit(RequestKind::Delete).await;
+        let blob_client = self.client.blob_client(self.relative_path_to_name(path));
+
+        let builder = blob_client.delete();
+
+        match builder.into_future().await {
+            Ok(_response) => Ok(()),
+            Err(e) => {
+                if let Some(http_err) = e.as_http_error() {
+                    if http_err.status() == StatusCode::NotFound {
+                        return Ok(());
+                    }
+                }
+                Err(anyhow::Error::new(e))
+            }
+        }
+    }
+
+    async fn delete_objects<'a>(&self, paths: &'a [RemotePath]) -> anyhow::Result<()> {
+        // Permit is already obtained by inner delete function
+
+        // TODO batch requests are also not supported by the SDK
+        // https://github.com/Azure/azure-sdk-for-rust/issues/1068
+        // https://github.com/Azure/azure-sdk-for-rust/issues/1249
+        for path in paths {
+            self.delete(path).await?;
+        }
+        Ok(())
+    }
+}

libs/remote_storage/src/lib.rs

@@ -4,7 +4,10 @@
 //! [`RemoteStorage`] trait a CRUD-like generic abstraction to use for adapting external storages with a few implementations:
 //!   * [`local_fs`] allows to use local file system as an external storage
 //!   * [`s3_bucket`] uses AWS S3 bucket as an external storage
+//!   * [`azure_blob`] allows to use Azure Blob storage as an external storage
 //!
+
+mod azure_blob;
 mod local_fs;
 mod s3_bucket;
 mod simulate_failures;
@@ -21,11 +24,15 @@ use anyhow::{bail, Context};
 use camino::{Utf8Path, Utf8PathBuf};
 
 use serde::{Deserialize, Serialize};
-use tokio::io;
+use tokio::{io, sync::Semaphore};
 use toml_edit::Item;
 use tracing::info;
 
-pub use self::{local_fs::LocalFs, s3_bucket::S3Bucket, simulate_failures::UnreliableWrapper};
+pub use self::{
+    azure_blob::AzureBlobStorage, local_fs::LocalFs, s3_bucket::S3Bucket,
+    simulate_failures::UnreliableWrapper,
+};
+use s3_bucket::RequestKind;
 
 /// How many different timelines can be processed simultaneously when synchronizing layers with the remote storage.
 /// During regular work, pageserver produces one layer file per timeline checkpoint, with bursts of concurrency
@@ -39,6 +46,11 @@ pub const DEFAULT_REMOTE_STORAGE_MAX_SYNC_ERRORS: u32 = 10;
 /// ~3500 PUT/COPY/POST/DELETE or 5500 GET/HEAD S3 requests
 /// <https://aws.amazon.com/premiumsupport/knowledge-center/s3-request-limit-avoid-throttling/>
 pub const DEFAULT_REMOTE_STORAGE_S3_CONCURRENCY_LIMIT: usize = 100;
+/// We set this a little bit low as we currently buffer the entire file into RAM
+///
+/// Here, a limit of max 20k concurrent connections was noted.
+/// <https://learn.microsoft.com/en-us/answers/questions/1301863/is-there-any-limitation-to-concurrent-connections>
+pub const DEFAULT_REMOTE_STORAGE_AZURE_CONCURRENCY_LIMIT: usize = 30;
 /// No limits on the client side, which currenltly means 1000 for AWS S3.
 /// <https://docs.aws.amazon.com/AmazonS3/latest/API/API_ListObjectsV2.html#API_ListObjectsV2_RequestSyntax>
 pub const DEFAULT_MAX_KEYS_PER_LIST_RESPONSE: Option<i32> = None;
@@ -117,6 +129,22 @@ impl RemotePath {
     }
 }
 
+/// We don't need callers to be able to pass arbitrary delimiters: just control
+/// whether listings will use a '/' separator or not.
+///
+/// The WithDelimiter mode will populate `prefixes` and `keys` in the result.  The
+/// NoDelimiter mode will only populate `keys`.
+pub enum ListingMode {
+    WithDelimiter,
+    NoDelimiter,
+}
+
+#[derive(Default)]
+pub struct Listing {
+    pub prefixes: Vec<RemotePath>,
+    pub keys: Vec<RemotePath>,
+}
+
 /// Storage (potentially remote) API to manage its state.
 /// This storage tries to be unaware of any layered repository context,
 /// providing basic CRUD operations for storage files.
@@ -129,8 +157,13 @@ pub trait RemoteStorage: Send + Sync + 'static {
     async fn list_prefixes(
         &self,
         prefix: Option<&RemotePath>,
-    ) -> Result<Vec<RemotePath>, DownloadError>;
-
+    ) -> Result<Vec<RemotePath>, DownloadError> {
+        let result = self
+            .list(prefix, ListingMode::WithDelimiter)
+            .await?
+            .prefixes;
+        Ok(result)
+    }
     /// Lists all files in directory "recursively"
     /// (not really recursively, because AWS has a flat namespace)
     /// Note: This is subtely different than list_prefixes,
@@ -142,7 +175,16 @@ pub trait RemoteStorage: Send + Sync + 'static {
     /// whereas,
     /// list_prefixes("foo/bar/") = ["cat", "dog"]
     /// See `test_real_s3.rs` for more details.
-    async fn list_files(&self, folder: Option<&RemotePath>) -> anyhow::Result<Vec<RemotePath>>;
+    async fn list_files(&self, prefix: Option<&RemotePath>) -> anyhow::Result<Vec<RemotePath>> {
+        let result = self.list(prefix, ListingMode::NoDelimiter).await?.keys;
+        Ok(result)
+    }
+
+    async fn list(
+        &self,
+        prefix: Option<&RemotePath>,
+        _mode: ListingMode,
+    ) -> anyhow::Result<Listing, DownloadError>;
 
     /// Streams the local file contents into remote into the remote storage entry.
     async fn upload(
@@ -193,6 +235,9 @@ pub enum DownloadError {
     BadInput(anyhow::Error),
     /// The file was not found in the remote storage.
     NotFound,
+    /// A cancellation token aborted the download, typically during
+    /// tenant detach or process shutdown.
+    Cancelled,
     /// The file was found in the remote storage, but the download failed.
     Other(anyhow::Error),
 }
@@ -203,6 +248,7 @@ impl std::fmt::Display for DownloadError {
             DownloadError::BadInput(e) => {
                 write!(f, "Failed to download a remote file due to user input: {e}")
             }
+            DownloadError::Cancelled => write!(f, "Cancelled, shutting down"),
             DownloadError::NotFound => write!(f, "No file found for the remote object id given"),
             DownloadError::Other(e) => write!(f, "Failed to download a remote file: {e:?}"),
         }
@@ -217,10 +263,24 @@ impl std::error::Error for DownloadError {}
 pub enum GenericRemoteStorage {
     LocalFs(LocalFs),
     AwsS3(Arc<S3Bucket>),
+    AzureBlob(Arc<AzureBlobStorage>),
     Unreliable(Arc<UnreliableWrapper>),
 }
 
 impl GenericRemoteStorage {
+    pub async fn list(
+        &self,
+        prefix: Option<&RemotePath>,
+        mode: ListingMode,
+    ) -> anyhow::Result<Listing, DownloadError> {
+        match self {
+            Self::LocalFs(s) => s.list(prefix, mode).await,
+            Self::AwsS3(s) => s.list(prefix, mode).await,
+            Self::AzureBlob(s) => s.list(prefix, mode).await,
+            Self::Unreliable(s) => s.list(prefix, mode).await,
+        }
+    }
+
     // A function for listing all the files in a "directory"
     // Example:
     // list_files("foo/bar") = ["foo/bar/a.txt", "foo/bar/b.txt"]
@@ -228,6 +288,7 @@ impl GenericRemoteStorage {
         match self {
             Self::LocalFs(s) => s.list_files(folder).await,
             Self::AwsS3(s) => s.list_files(folder).await,
+            Self::AzureBlob(s) => s.list_files(folder).await,
             Self::Unreliable(s) => s.list_files(folder).await,
         }
     }
@@ -242,6 +303,7 @@ impl GenericRemoteStorage {
         match self {
             Self::LocalFs(s) => s.list_prefixes(prefix).await,
             Self::AwsS3(s) => s.list_prefixes(prefix).await,
+            Self::AzureBlob(s) => s.list_prefixes(prefix).await,
             Self::Unreliable(s) => s.list_prefixes(prefix).await,
         }
     }
@@ -256,6 +318,7 @@ impl GenericRemoteStorage {
         match self {
             Self::LocalFs(s) => s.upload(from, data_size_bytes, to, metadata).await,
             Self::AwsS3(s) => s.upload(from, data_size_bytes, to, metadata).await,
+            Self::AzureBlob(s) => s.upload(from, data_size_bytes, to, metadata).await,
             Self::Unreliable(s) => s.upload(from, data_size_bytes, to, metadata).await,
         }
     }
@@ -264,6 +327,7 @@ impl GenericRemoteStorage {
         match self {
             Self::LocalFs(s) => s.download(from).await,
             Self::AwsS3(s) => s.download(from).await,
+            Self::AzureBlob(s) => s.download(from).await,
             Self::Unreliable(s) => s.download(from).await,
         }
     }
@@ -283,6 +347,10 @@ impl GenericRemoteStorage {
                 s.download_byte_range(from, start_inclusive, end_exclusive)
                     .await
             }
+            Self::AzureBlob(s) => {
+                s.download_byte_range(from, start_inclusive, end_exclusive)
+                    .await
+            }
             Self::Unreliable(s) => {
                 s.download_byte_range(from, start_inclusive, end_exclusive)
                     .await
@@ -294,6 +362,7 @@ impl GenericRemoteStorage {
         match self {
             Self::LocalFs(s) => s.delete(path).await,
             Self::AwsS3(s) => s.delete(path).await,
+            Self::AzureBlob(s) => s.delete(path).await,
             Self::Unreliable(s) => s.delete(path).await,
         }
     }
@@ -302,6 +371,7 @@ impl GenericRemoteStorage {
         match self {
             Self::LocalFs(s) => s.delete_objects(paths).await,
             Self::AwsS3(s) => s.delete_objects(paths).await,
+            Self::AzureBlob(s) => s.delete_objects(paths).await,
             Self::Unreliable(s) => s.delete_objects(paths).await,
         }
     }
@@ -319,6 +389,11 @@ impl GenericRemoteStorage {
                       s3_config.bucket_name, s3_config.bucket_region, s3_config.prefix_in_bucket, s3_config.endpoint);
                 Self::AwsS3(Arc::new(S3Bucket::new(s3_config)?))
             }
+            RemoteStorageKind::AzureContainer(azure_config) => {
+                info!("Using azure container '{}' in region '{}' as a remote storage, prefix in container: '{:?}'",
+                      azure_config.container_name, azure_config.container_region, azure_config.prefix_in_container);
+                Self::AzureBlob(Arc::new(AzureBlobStorage::new(azure_config)?))
+            }
         })
     }
 
@@ -383,6 +458,9 @@ pub enum RemoteStorageKind {
     /// AWS S3 based storage, storing all files in the S3 bucket
     /// specified by the config
     AwsS3(S3Config),
+    /// Azure Blob based storage, storing all files in the container
+    /// specified by the config
+    AzureContainer(AzureConfig),
 }
 
 /// AWS S3 bucket coordinates and access credentials to manage the bucket contents (read and write).
@@ -422,11 +500,45 @@ impl Debug for S3Config {
     }
 }
 
+/// Azure  bucket coordinates and access credentials to manage the bucket contents (read and write).
+#[derive(Clone, PartialEq, Eq)]
+pub struct AzureConfig {
+    /// Name of the container to connect to.
+    pub container_name: String,
+    /// The region where the bucket is located at.
+    pub container_region: String,
+    /// A "subfolder" in the container, to use the same container separately by multiple remote storage users at once.
+    pub prefix_in_container: Option<String>,
+    /// Azure has various limits on its API calls, we need not to exceed those.
+    /// See [`DEFAULT_REMOTE_STORAGE_AZURE_CONCURRENCY_LIMIT`] for more details.
+    pub concurrency_limit: NonZeroUsize,
+    pub max_keys_per_list_response: Option<i32>,
+}
+
+impl Debug for AzureConfig {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        f.debug_struct("AzureConfig")
+            .field("bucket_name", &self.container_name)
+            .field("bucket_region", &self.container_region)
+            .field("prefix_in_bucket", &self.prefix_in_container)
+            .field("concurrency_limit", &self.concurrency_limit)
+            .field(
+                "max_keys_per_list_response",
+                &self.max_keys_per_list_response,
+            )
+            .finish()
+    }
+}
+
 impl RemoteStorageConfig {
     pub fn from_toml(toml: &toml_edit::Item) -> anyhow::Result<Option<RemoteStorageConfig>> {
         let local_path = toml.get("local_path");
         let bucket_name = toml.get("bucket_name");
         let bucket_region = toml.get("bucket_region");
+        let container_name = toml.get("container_name");
+        let container_region = toml.get("container_region");
+
+        let use_azure = container_name.is_some() && container_region.is_some();
 
         let max_concurrent_syncs = NonZeroUsize::new(
             parse_optional_integer("max_concurrent_syncs", toml)?
@@ -440,9 +552,13 @@ impl RemoteStorageConfig {
         )
         .context("Failed to parse 'max_sync_errors' as a positive integer")?;
 
+        let default_concurrency_limit = if use_azure {
+            DEFAULT_REMOTE_STORAGE_AZURE_CONCURRENCY_LIMIT
+        } else {
+            DEFAULT_REMOTE_STORAGE_S3_CONCURRENCY_LIMIT
+        };
         let concurrency_limit = NonZeroUsize::new(
-            parse_optional_integer("concurrency_limit", toml)?
-                .unwrap_or(DEFAULT_REMOTE_STORAGE_S3_CONCURRENCY_LIMIT),
+            parse_optional_integer("concurrency_limit", toml)?.unwrap_or(default_concurrency_limit),
         )
         .context("Failed to parse 'concurrency_limit' as a positive integer")?;
 
@@ -451,33 +567,70 @@ impl RemoteStorageConfig {
                 .context("Failed to parse 'max_keys_per_list_response' as a positive integer")?
                 .or(DEFAULT_MAX_KEYS_PER_LIST_RESPONSE);
 
-        let storage = match (local_path, bucket_name, bucket_region) {
+        let endpoint = toml
+            .get("endpoint")
+            .map(|endpoint| parse_toml_string("endpoint", endpoint))
+            .transpose()?;
+
+        let storage = match (
+            local_path,
+            bucket_name,
+            bucket_region,
+            container_name,
+            container_region,
+        ) {
             // no 'local_path' nor 'bucket_name' options are provided, consider this remote storage disabled
-            (None, None, None) => return Ok(None),
-            (_, Some(_), None) => {
+            (None, None, None, None, None) => return Ok(None),
+            (_, Some(_), None, ..) => {
                 bail!("'bucket_region' option is mandatory if 'bucket_name' is given ")
             }
-            (_, None, Some(_)) => {
+            (_, None, Some(_), ..) => {
                 bail!("'bucket_name' option is mandatory if 'bucket_region' is given ")
             }
-            (None, Some(bucket_name), Some(bucket_region)) => RemoteStorageKind::AwsS3(S3Config {
-                bucket_name: parse_toml_string("bucket_name", bucket_name)?,
-                bucket_region: parse_toml_string("bucket_region", bucket_region)?,
-                prefix_in_bucket: toml
-                    .get("prefix_in_bucket")
-                    .map(|prefix_in_bucket| parse_toml_string("prefix_in_bucket", prefix_in_bucket))
-                    .transpose()?,
-                endpoint: toml
-                    .get("endpoint")
-                    .map(|endpoint| parse_toml_string("endpoint", endpoint))
-                    .transpose()?,
-                concurrency_limit,
-                max_keys_per_list_response,
-            }),
-            (Some(local_path), None, None) => RemoteStorageKind::LocalFs(Utf8PathBuf::from(
-                parse_toml_string("local_path", local_path)?,
-            )),
-            (Some(_), Some(_), _) => bail!("local_path and bucket_name are mutually exclusive"),
+            (None, Some(bucket_name), Some(bucket_region), ..) => {
+                RemoteStorageKind::AwsS3(S3Config {
+                    bucket_name: parse_toml_string("bucket_name", bucket_name)?,
+                    bucket_region: parse_toml_string("bucket_region", bucket_region)?,
+                    prefix_in_bucket: toml
+                        .get("prefix_in_bucket")
+                        .map(|prefix_in_bucket| {
+                            parse_toml_string("prefix_in_bucket", prefix_in_bucket)
+                        })
+                        .transpose()?,
+                    endpoint,
+                    concurrency_limit,
+                    max_keys_per_list_response,
+                })
+            }
+            (_, _, _, Some(_), None) => {
+                bail!("'container_name' option is mandatory if 'container_region' is given ")
+            }
+            (_, _, _, None, Some(_)) => {
+                bail!("'container_name' option is mandatory if 'container_region' is given ")
+            }
+            (None, None, None, Some(container_name), Some(container_region)) => {
+                RemoteStorageKind::AzureContainer(AzureConfig {
+                    container_name: parse_toml_string("container_name", container_name)?,
+                    container_region: parse_toml_string("container_region", container_region)?,
+                    prefix_in_container: toml
+                        .get("prefix_in_container")
+                        .map(|prefix_in_container| {
+                            parse_toml_string("prefix_in_container", prefix_in_container)
+                        })
+                        .transpose()?,
+                    concurrency_limit,
+                    max_keys_per_list_response,
+                })
+            }
+            (Some(local_path), None, None, None, None) => RemoteStorageKind::LocalFs(
+                Utf8PathBuf::from(parse_toml_string("local_path", local_path)?),
+            ),
+            (Some(_), Some(_), ..) => {
+                bail!("'local_path' and 'bucket_name' are mutually exclusive")
+            }
+            (Some(_), _, _, Some(_), Some(_)) => {
+                bail!("local_path and 'container_name' are mutually exclusive")
+            }
         };
 
         Ok(Some(RemoteStorageConfig {
@@ -513,6 +666,46 @@ fn parse_toml_string(name: &str, item: &Item) -> anyhow::Result<String> {
     Ok(s.to_string())
 }
 
+struct ConcurrencyLimiter {
+    // Every request to S3 can be throttled or cancelled, if a certain number of requests per second is exceeded.
+    // Same goes to IAM, which is queried before every S3 request, if enabled. IAM has even lower RPS threshold.
+    // The helps to ensure we don't exceed the thresholds.
+    write: Arc<Semaphore>,
+    read: Arc<Semaphore>,
+}
+
+impl ConcurrencyLimiter {
+    fn for_kind(&self, kind: RequestKind) -> &Arc<Semaphore> {
+        match kind {
+            RequestKind::Get => &self.read,
+            RequestKind::Put => &self.write,
+            RequestKind::List => &self.read,
+            RequestKind::Delete => &self.write,
+        }
+    }
+
+    async fn acquire(
+        &self,
+        kind: RequestKind,
+    ) -> Result<tokio::sync::SemaphorePermit<'_>, tokio::sync::AcquireError> {
+        self.for_kind(kind).acquire().await
+    }
+
+    async fn acquire_owned(
+        &self,
+        kind: RequestKind,
+    ) -> Result<tokio::sync::OwnedSemaphorePermit, tokio::sync::AcquireError> {
+        Arc::clone(self.for_kind(kind)).acquire_owned().await
+    }
+
+    fn new(limit: usize) -> ConcurrencyLimiter {
+        Self {
+            read: Arc::new(Semaphore::new(limit)),
+            write: Arc::new(Semaphore::new(limit)),
+        }
+    }
+}
+
 #[cfg(test)]
 mod tests {
     use super::*;

libs/remote_storage/src/local_fs.rs

@@ -15,7 +15,7 @@ use tokio::{
 use tracing::*;
 use utils::{crashsafe::path_with_suffix_extension, fs_ext::is_directory_empty};
 
-use crate::{Download, DownloadError, RemotePath};
+use crate::{Download, DownloadError, Listing, ListingMode, RemotePath};
 
 use super::{RemoteStorage, StorageMetadata};
 
@@ -75,7 +75,7 @@ impl LocalFs {
     }
 
     #[cfg(test)]
-    async fn list(&self) -> anyhow::Result<Vec<RemotePath>> {
+    async fn list_all(&self) -> anyhow::Result<Vec<RemotePath>> {
         Ok(get_all_files(&self.storage_root, true)
             .await?
             .into_iter()
@@ -89,52 +89,10 @@ impl LocalFs {
             })
             .collect())
     }
-}
-
-#[async_trait::async_trait]
-impl RemoteStorage for LocalFs {
-    async fn list_prefixes(
-        &self,
-        prefix: Option<&RemotePath>,
-    ) -> Result<Vec<RemotePath>, DownloadError> {
-        let path = match prefix {
-            Some(prefix) => Cow::Owned(prefix.with_base(&self.storage_root)),
-            None => Cow::Borrowed(&self.storage_root),
-        };
-
-        let prefixes_to_filter = get_all_files(path.as_ref(), false)
-            .await
-            .map_err(DownloadError::Other)?;
-
-        let mut prefixes = Vec::with_capacity(prefixes_to_filter.len());
-
-        // filter out empty directories to mirror s3 behavior.
-        for prefix in prefixes_to_filter {
-            if prefix.is_dir()
-                && is_directory_empty(&prefix)
-                    .await
-                    .map_err(DownloadError::Other)?
-            {
-                continue;
-            }
-
-            prefixes.push(
-                prefix
-                    .strip_prefix(&self.storage_root)
-                    .context("Failed to strip prefix")
-                    .and_then(RemotePath::new)
-                    .expect(
-                        "We list files for storage root, hence should be able to remote the prefix",
-                    ),
-            )
-        }
-
-        Ok(prefixes)
-    }
 
     // recursively lists all files in a directory,
     // mirroring the `list_files` for `s3_bucket`
-    async fn list_files(&self, folder: Option<&RemotePath>) -> anyhow::Result<Vec<RemotePath>> {
+    async fn list_recursive(&self, folder: Option<&RemotePath>) -> anyhow::Result<Vec<RemotePath>> {
         let full_path = match folder {
             Some(folder) => folder.with_base(&self.storage_root),
             None => self.storage_root.clone(),
@@ -186,6 +144,70 @@ impl RemoteStorage for LocalFs {
 
         Ok(files)
     }
+}
+
+#[async_trait::async_trait]
+impl RemoteStorage for LocalFs {
+    async fn list(
+        &self,
+        prefix: Option<&RemotePath>,
+        mode: ListingMode,
+    ) -> Result<Listing, DownloadError> {
+        let mut result = Listing::default();
+
+        if let ListingMode::NoDelimiter = mode {
+            let keys = self
+                .list_recursive(prefix)
+                .await
+                .map_err(DownloadError::Other)?;
+
+            result.keys = keys
+                .into_iter()
+                .filter(|k| {
+                    let path = k.with_base(&self.storage_root);
+                    !path.is_dir()
+                })
+                .collect();
+
+            return Ok(result);
+        }
+
+        let path = match prefix {
+            Some(prefix) => Cow::Owned(prefix.with_base(&self.storage_root)),
+            None => Cow::Borrowed(&self.storage_root),
+        };
+
+        let prefixes_to_filter = get_all_files(path.as_ref(), false)
+            .await
+            .map_err(DownloadError::Other)?;
+
+        // filter out empty directories to mirror s3 behavior.
+        for prefix in prefixes_to_filter {
+            if prefix.is_dir()
+                && is_directory_empty(&prefix)
+                    .await
+                    .map_err(DownloadError::Other)?
+            {
+                continue;
+            }
+
+            let stripped = prefix
+                .strip_prefix(&self.storage_root)
+                .context("Failed to strip prefix")
+                .and_then(RemotePath::new)
+                .expect(
+                    "We list files for storage root, hence should be able to remote the prefix",
+                );
+
+            if prefix.is_dir() {
+                result.prefixes.push(stripped);
+            } else {
+                result.keys.push(stripped);
+            }
+        }
+
+        Ok(result)
+    }
 
     async fn upload(
         &self,
@@ -479,7 +501,7 @@ mod fs_tests {
 
         let target_path_1 = upload_dummy_file(&storage, "upload_1", None).await?;
         assert_eq!(
-            storage.list().await?,
+            storage.list_all().await?,
             vec![target_path_1.clone()],
             "Should list a single file after first upload"
         );
@@ -667,7 +689,7 @@ mod fs_tests {
         let upload_target = upload_dummy_file(&storage, upload_name, None).await?;
 
         storage.delete(&upload_target).await?;
-        assert!(storage.list().await?.is_empty());
+        assert!(storage.list_all().await?.is_empty());
 
         storage
             .delete(&upload_target)
@@ -725,6 +747,43 @@ mod fs_tests {
         Ok(())
     }
 
+    #[tokio::test]
+    async fn list() -> anyhow::Result<()> {
+        // No delimiter: should recursively list everything
+        let storage = create_storage()?;
+        let child = upload_dummy_file(&storage, "grandparent/parent/child", None).await?;
+        let uncle = upload_dummy_file(&storage, "grandparent/uncle", None).await?;
+
+        let listing = storage.list(None, ListingMode::NoDelimiter).await?;
+        assert!(listing.prefixes.is_empty());
+        assert_eq!(listing.keys, [uncle.clone(), child.clone()].to_vec());
+
+        // Delimiter: should only go one deep
+        let listing = storage.list(None, ListingMode::WithDelimiter).await?;
+
+        assert_eq!(
+            listing.prefixes,
+            [RemotePath::from_string("timelines").unwrap()].to_vec()
+        );
+        assert!(listing.keys.is_empty());
+
+        // Delimiter & prefix
+        let listing = storage
+            .list(
+                Some(&RemotePath::from_string("timelines/some_timeline/grandparent").unwrap()),
+                ListingMode::WithDelimiter,
+            )
+            .await?;
+        assert_eq!(
+            listing.prefixes,
+            [RemotePath::from_string("timelines/some_timeline/grandparent/parent").unwrap()]
+                .to_vec()
+        );
+        assert_eq!(listing.keys, [uncle.clone()].to_vec());
+
+        Ok(())
+    }
+
     async fn upload_dummy_file(
         storage: &LocalFs,
         name: &str,
@@ -777,7 +836,7 @@ mod fs_tests {
     }
 
     async fn list_files_sorted(storage: &LocalFs) -> anyhow::Result<Vec<RemotePath>> {
-        let mut files = storage.list().await?;
+        let mut files = storage.list_all().await?;
         files.sort_by(|a, b| a.0.cmp(&b.0));
         Ok(files)
     }

libs/remote_storage/src/s3_bucket.rs

@@ -4,7 +4,7 @@
 //! 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;
 
 use anyhow::Context;
 use aws_config::{
@@ -24,22 +24,20 @@ use aws_sdk_s3::{
 use aws_smithy_http::body::SdkBody;
 use hyper::Body;
 use scopeguard::ScopeGuard;
-use tokio::{
-    io::{self, AsyncRead},
-    sync::Semaphore,
-};
+use tokio::io::{self, AsyncRead};
 use tokio_util::io::ReaderStream;
 use tracing::debug;
 
 use super::StorageMetadata;
 use crate::{
-    Download, DownloadError, RemotePath, RemoteStorage, S3Config, MAX_KEYS_PER_DELETE,
-    REMOTE_STORAGE_PREFIX_SEPARATOR,
+    ConcurrencyLimiter, Download, DownloadError, Listing, ListingMode, RemotePath, RemoteStorage,
+    S3Config, MAX_KEYS_PER_DELETE, REMOTE_STORAGE_PREFIX_SEPARATOR,
 };
 
 pub(super) mod metrics;
 
-use self::metrics::{AttemptOutcome, RequestKind};
+use self::metrics::AttemptOutcome;
+pub(super) use self::metrics::RequestKind;
 
 /// AWS S3 storage.
 pub struct S3Bucket {
@@ -50,46 +48,6 @@ pub struct S3Bucket {
     concurrency_limiter: ConcurrencyLimiter,
 }
 
-struct ConcurrencyLimiter {
-    // Every request to S3 can be throttled or cancelled, if a certain number of requests per second is exceeded.
-    // Same goes to IAM, which is queried before every S3 request, if enabled. IAM has even lower RPS threshold.
-    // The helps to ensure we don't exceed the thresholds.
-    write: Arc<Semaphore>,
-    read: Arc<Semaphore>,
-}
-
-impl ConcurrencyLimiter {
-    fn for_kind(&self, kind: RequestKind) -> &Arc<Semaphore> {
-        match kind {
-            RequestKind::Get => &self.read,
-            RequestKind::Put => &self.write,
-            RequestKind::List => &self.read,
-            RequestKind::Delete => &self.write,
-        }
-    }
-
-    async fn acquire(
-        &self,
-        kind: RequestKind,
-    ) -> Result<tokio::sync::SemaphorePermit<'_>, tokio::sync::AcquireError> {
-        self.for_kind(kind).acquire().await
-    }
-
-    async fn acquire_owned(
-        &self,
-        kind: RequestKind,
-    ) -> Result<tokio::sync::OwnedSemaphorePermit, tokio::sync::AcquireError> {
-        Arc::clone(self.for_kind(kind)).acquire_owned().await
-    }
-
-    fn new(limit: usize) -> ConcurrencyLimiter {
-        Self {
-            read: Arc::new(Semaphore::new(limit)),
-            write: Arc::new(Semaphore::new(limit)),
-        }
-    }
-}
-
 #[derive(Default)]
 struct GetObjectRequest {
     bucket: String,
@@ -341,13 +299,13 @@ impl<S: AsyncRead> AsyncRead for TimedDownload<S> {
 
 #[async_trait::async_trait]
 impl RemoteStorage for S3Bucket {
-    /// See the doc for `RemoteStorage::list_prefixes`
-    /// Note: it wont include empty "directories"
-    async fn list_prefixes(
+    async fn list(
         &self,
         prefix: Option<&RemotePath>,
-    ) -> Result<Vec<RemotePath>, DownloadError> {
+        mode: ListingMode,
+    ) -> Result<Listing, DownloadError> {
         let kind = RequestKind::List;
+        let mut result = Listing::default();
 
         // get the passed prefix or if it is not set use prefix_in_bucket value
         let list_prefix = prefix
@@ -356,28 +314,33 @@ impl RemoteStorage for S3Bucket {
             .map(|mut p| {
                 // required to end with a separator
                 // otherwise request will return only the entry of a prefix
-                if !p.ends_with(REMOTE_STORAGE_PREFIX_SEPARATOR) {
+                if matches!(mode, ListingMode::WithDelimiter)
+                    && !p.ends_with(REMOTE_STORAGE_PREFIX_SEPARATOR)
+                {
                     p.push(REMOTE_STORAGE_PREFIX_SEPARATOR);
                 }
                 p
             });
 
-        let mut document_keys = Vec::new();
-
         let mut continuation_token = None;
 
         loop {
             let _guard = self.permit(kind).await;
             let started_at = start_measuring_requests(kind);
 
-            let fetch_response = self
+            let mut request = self
                 .client
                 .list_objects_v2()
                 .bucket(self.bucket_name.clone())
                 .set_prefix(list_prefix.clone())
                 .set_continuation_token(continuation_token)
-                .delimiter(REMOTE_STORAGE_PREFIX_SEPARATOR.to_string())
-                .set_max_keys(self.max_keys_per_list_response)
+                .set_max_keys(self.max_keys_per_list_response);
+
+            if let ListingMode::WithDelimiter = mode {
+                request = request.delimiter(REMOTE_STORAGE_PREFIX_SEPARATOR.to_string());
+            }
+
+            let response = request
                 .send()
                 .await
                 .context("Failed to list S3 prefixes")
@@ -387,71 +350,35 @@ impl RemoteStorage for S3Bucket {
 
             metrics::BUCKET_METRICS
                 .req_seconds
-                .observe_elapsed(kind, &fetch_response, started_at);
+                .observe_elapsed(kind, &response, started_at);
 
-            let fetch_response = fetch_response?;
+            let response = response?;
 
-            document_keys.extend(
-                fetch_response
-                    .common_prefixes
-                    .unwrap_or_default()
-                    .into_iter()
+            let keys = response.contents().unwrap_or_default();
+            let empty = Vec::new();
+            let prefixes = response.common_prefixes.as_ref().unwrap_or(&empty);
+
+            tracing::info!("list: {} prefixes, {} keys", prefixes.len(), keys.len());
+
+            for object in keys {
+                let object_path = object.key().expect("response does not contain a key");
+                let remote_path = self.s3_object_to_relative_path(object_path);
+                result.keys.push(remote_path);
+            }
+
+            result.prefixes.extend(
+                prefixes
+                    .iter()
                     .filter_map(|o| Some(self.s3_object_to_relative_path(o.prefix()?))),
             );
 
-            continuation_token = match fetch_response.next_continuation_token {
+            continuation_token = match response.next_continuation_token {
                 Some(new_token) => Some(new_token),
                 None => break,
             };
         }
 
-        Ok(document_keys)
-    }
-
-    /// See the doc for `RemoteStorage::list_files`
-    async fn list_files(&self, folder: Option<&RemotePath>) -> anyhow::Result<Vec<RemotePath>> {
-        let kind = RequestKind::List;
-
-        let folder_name = folder
-            .map(|p| self.relative_path_to_s3_object(p))
-            .or_else(|| self.prefix_in_bucket.clone());
-
-        // AWS may need to break the response into several parts
-        let mut continuation_token = None;
-        let mut all_files = vec![];
-        loop {
-            let _guard = self.permit(kind).await;
-            let started_at = start_measuring_requests(kind);
-
-            let response = self
-                .client
-                .list_objects_v2()
-                .bucket(self.bucket_name.clone())
-                .set_prefix(folder_name.clone())
-                .set_continuation_token(continuation_token)
-                .set_max_keys(self.max_keys_per_list_response)
-                .send()
-                .await
-                .context("Failed to list files in S3 bucket");
-
-            let started_at = ScopeGuard::into_inner(started_at);
-            metrics::BUCKET_METRICS
-                .req_seconds
-                .observe_elapsed(kind, &response, started_at);
-
-            let response = response?;
-
-            for object in response.contents().unwrap_or_default() {
-                let object_path = object.key().expect("response does not contain a key");
-                let remote_path = self.s3_object_to_relative_path(object_path);
-                all_files.push(remote_path);
-            }
-            match response.next_continuation_token {
-                Some(new_token) => continuation_token = Some(new_token),
-                None => break,
-            }
-        }
-        Ok(all_files)
+        Ok(result)
     }
 
     async fn upload(

libs/remote_storage/src/s3_bucket/metrics.rs

@@ -6,7 +6,7 @@ use once_cell::sync::Lazy;
 pub(super) static BUCKET_METRICS: Lazy<BucketMetrics> = Lazy::new(Default::default);
 
 #[derive(Clone, Copy, Debug)]
-pub(super) enum RequestKind {
+pub(crate) enum RequestKind {
     Get = 0,
     Put = 1,
     Delete = 2,

libs/remote_storage/src/simulate_failures.rs

@@ -5,7 +5,9 @@ use std::collections::hash_map::Entry;
 use std::collections::HashMap;
 use std::sync::Mutex;
 
-use crate::{Download, DownloadError, RemotePath, RemoteStorage, StorageMetadata};
+use crate::{
+    Download, DownloadError, Listing, ListingMode, RemotePath, RemoteStorage, StorageMetadata,
+};
 
 pub struct UnreliableWrapper {
     inner: crate::GenericRemoteStorage,
@@ -95,6 +97,15 @@ impl RemoteStorage for UnreliableWrapper {
         self.inner.list_files(folder).await
     }
 
+    async fn list(
+        &self,
+        prefix: Option<&RemotePath>,
+        mode: ListingMode,
+    ) -> Result<Listing, DownloadError> {
+        self.attempt(RemoteOp::ListPrefixes(prefix.cloned()))?;
+        self.inner.list(prefix, mode).await
+    }
+
     async fn upload(
         &self,
         data: impl tokio::io::AsyncRead + Unpin + Send + Sync + 'static,

libs/remote_storage/tests/test_real_azure.rs

@@ -0,0 +1,625 @@
+use std::collections::HashSet;
+use std::env;
+use std::num::{NonZeroU32, NonZeroUsize};
+use std::ops::ControlFlow;
+use std::path::PathBuf;
+use std::sync::Arc;
+use std::time::UNIX_EPOCH;
+
+use anyhow::Context;
+use camino::Utf8Path;
+use once_cell::sync::OnceCell;
+use remote_storage::{
+    AzureConfig, Download, GenericRemoteStorage, RemotePath, RemoteStorageConfig, RemoteStorageKind,
+};
+use test_context::{test_context, AsyncTestContext};
+use tokio::task::JoinSet;
+use tracing::{debug, error, info};
+
+static LOGGING_DONE: OnceCell<()> = OnceCell::new();
+
+const ENABLE_REAL_AZURE_REMOTE_STORAGE_ENV_VAR_NAME: &str = "ENABLE_REAL_AZURE_REMOTE_STORAGE";
+
+const BASE_PREFIX: &str = "test";
+
+/// Tests that the Azure client can list all prefixes, even if the response comes paginated and requires multiple HTTP queries.
+/// Uses real Azure and requires [`ENABLE_REAL_AZURE_REMOTE_STORAGE_ENV_VAR_NAME`] and related Azure cred env vars specified.
+/// See the client creation in [`create_azure_client`] for details on the required env vars.
+/// If real Azure tests are disabled, the test passes, skipping any real test run: currently, there's no way to mark the test ignored in runtime with the
+/// deafult test framework, see https://github.com/rust-lang/rust/issues/68007 for details.
+///
+/// First, the test creates a set of Azure blobs with keys `/${random_prefix_part}/${base_prefix_str}/sub_prefix_${i}/blob_${i}` in [`upload_azure_data`]
+/// where
+/// * `random_prefix_part` is set for the entire Azure client during the Azure client creation in [`create_azure_client`], to avoid multiple test runs interference
+/// * `base_prefix_str` is a common prefix to use in the client requests: we would want to ensure that the client is able to list nested prefixes inside the bucket
+///
+/// Then, verifies that the client does return correct prefixes when queried:
+/// * with no prefix, it lists everything after its `${random_prefix_part}/` — that should be `${base_prefix_str}` value only
+/// * with `${base_prefix_str}/` prefix, it lists every `sub_prefix_${i}`
+///
+/// With the real Azure enabled and `#[cfg(test)]` Rust configuration used, the Azure client test adds a `max-keys` param to limit the response keys.
+/// This way, we are able to test the pagination implicitly, by ensuring all results are returned from the remote storage and avoid uploading too many blobs to Azure.
+///
+/// Lastly, the test attempts to clean up and remove all uploaded Azure files.
+/// If any errors appear during the clean up, they get logged, but the test is not failed or stopped until clean up is finished.
+#[test_context(MaybeEnabledAzureWithTestBlobs)]
+#[tokio::test]
+async fn azure_pagination_should_work(
+    ctx: &mut MaybeEnabledAzureWithTestBlobs,
+) -> anyhow::Result<()> {
+    let ctx = match ctx {
+        MaybeEnabledAzureWithTestBlobs::Enabled(ctx) => ctx,
+        MaybeEnabledAzureWithTestBlobs::Disabled => return Ok(()),
+        MaybeEnabledAzureWithTestBlobs::UploadsFailed(e, _) => {
+            anyhow::bail!("Azure init failed: {e:?}")
+        }
+    };
+
+    let test_client = Arc::clone(&ctx.enabled.client);
+    let expected_remote_prefixes = ctx.remote_prefixes.clone();
+
+    let base_prefix = RemotePath::new(Utf8Path::new(ctx.enabled.base_prefix))
+        .context("common_prefix construction")?;
+    let root_remote_prefixes = test_client
+        .list_prefixes(None)
+        .await
+        .context("client list root prefixes failure")?
+        .into_iter()
+        .collect::<HashSet<_>>();
+    assert_eq!(
+        root_remote_prefixes, HashSet::from([base_prefix.clone()]),
+        "remote storage root prefixes list mismatches with the uploads. Returned prefixes: {root_remote_prefixes:?}"
+    );
+
+    let nested_remote_prefixes = test_client
+        .list_prefixes(Some(&base_prefix))
+        .await
+        .context("client list nested prefixes failure")?
+        .into_iter()
+        .collect::<HashSet<_>>();
+    let remote_only_prefixes = nested_remote_prefixes
+        .difference(&expected_remote_prefixes)
+        .collect::<HashSet<_>>();
+    let missing_uploaded_prefixes = expected_remote_prefixes
+        .difference(&nested_remote_prefixes)
+        .collect::<HashSet<_>>();
+    assert_eq!(
+        remote_only_prefixes.len() + missing_uploaded_prefixes.len(), 0,
+        "remote storage nested prefixes list mismatches with the uploads. Remote only prefixes: {remote_only_prefixes:?}, missing uploaded prefixes: {missing_uploaded_prefixes:?}",
+    );
+
+    Ok(())
+}
+
+/// Tests that Azure client can list all files in a folder, even if the response comes paginated and requirees multiple Azure queries.
+/// Uses real Azure and requires [`ENABLE_REAL_AZURE_REMOTE_STORAGE_ENV_VAR_NAME`] and related Azure cred env vars specified. Test will skip real code and pass if env vars not set.
+/// See `Azure_pagination_should_work` for more information.
+///
+/// First, create a set of Azure objects with keys `random_prefix/folder{j}/blob_{i}.txt` in [`upload_azure_data`]
+/// Then performs the following queries:
+///    1. `list_files(None)`. This should return all files `random_prefix/folder{j}/blob_{i}.txt`
+///    2. `list_files("folder1")`.  This  should return all files `random_prefix/folder1/blob_{i}.txt`
+#[test_context(MaybeEnabledAzureWithSimpleTestBlobs)]
+#[tokio::test]
+async fn azure_list_files_works(
+    ctx: &mut MaybeEnabledAzureWithSimpleTestBlobs,
+) -> anyhow::Result<()> {
+    let ctx = match ctx {
+        MaybeEnabledAzureWithSimpleTestBlobs::Enabled(ctx) => ctx,
+        MaybeEnabledAzureWithSimpleTestBlobs::Disabled => return Ok(()),
+        MaybeEnabledAzureWithSimpleTestBlobs::UploadsFailed(e, _) => {
+            anyhow::bail!("Azure init failed: {e:?}")
+        }
+    };
+    let test_client = Arc::clone(&ctx.enabled.client);
+    let base_prefix =
+        RemotePath::new(Utf8Path::new("folder1")).context("common_prefix construction")?;
+    let root_files = test_client
+        .list_files(None)
+        .await
+        .context("client list root files failure")?
+        .into_iter()
+        .collect::<HashSet<_>>();
+    assert_eq!(
+        root_files,
+        ctx.remote_blobs.clone(),
+        "remote storage list_files on root mismatches with the uploads."
+    );
+    let nested_remote_files = test_client
+        .list_files(Some(&base_prefix))
+        .await
+        .context("client list nested files failure")?
+        .into_iter()
+        .collect::<HashSet<_>>();
+    let trim_remote_blobs: HashSet<_> = ctx
+        .remote_blobs
+        .iter()
+        .map(|x| x.get_path())
+        .filter(|x| x.starts_with("folder1"))
+        .map(|x| RemotePath::new(x).expect("must be valid path"))
+        .collect();
+    assert_eq!(
+        nested_remote_files, trim_remote_blobs,
+        "remote storage list_files on subdirrectory mismatches with the uploads."
+    );
+    Ok(())
+}
+
+#[test_context(MaybeEnabledAzure)]
+#[tokio::test]
+async fn azure_delete_non_exising_works(ctx: &mut MaybeEnabledAzure) -> anyhow::Result<()> {
+    let ctx = match ctx {
+        MaybeEnabledAzure::Enabled(ctx) => ctx,
+        MaybeEnabledAzure::Disabled => return Ok(()),
+    };
+
+    let path = RemotePath::new(Utf8Path::new(
+        format!("{}/for_sure_there_is_nothing_there_really", ctx.base_prefix).as_str(),
+    ))
+    .with_context(|| "RemotePath conversion")?;
+
+    ctx.client.delete(&path).await.expect("should succeed");
+
+    Ok(())
+}
+
+#[test_context(MaybeEnabledAzure)]
+#[tokio::test]
+async fn azure_delete_objects_works(ctx: &mut MaybeEnabledAzure) -> anyhow::Result<()> {
+    let ctx = match ctx {
+        MaybeEnabledAzure::Enabled(ctx) => ctx,
+        MaybeEnabledAzure::Disabled => return Ok(()),
+    };
+
+    let path1 = RemotePath::new(Utf8Path::new(format!("{}/path1", ctx.base_prefix).as_str()))
+        .with_context(|| "RemotePath conversion")?;
+
+    let path2 = RemotePath::new(Utf8Path::new(format!("{}/path2", ctx.base_prefix).as_str()))
+        .with_context(|| "RemotePath conversion")?;
+
+    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?;
+
+    ctx.client
+        .upload(std::io::Cursor::new(data2), data2_len, &path2, None)
+        .await?;
+
+    ctx.client
+        .upload(std::io::Cursor::new(data3), data3_len, &path3, None)
+        .await?;
+
+    ctx.client.delete_objects(&[path1, path2]).await?;
+
+    let prefixes = ctx.client.list_prefixes(None).await?;
+
+    assert_eq!(prefixes.len(), 1);
+
+    ctx.client.delete_objects(&[path3]).await?;
+
+    Ok(())
+}
+
+#[test_context(MaybeEnabledAzure)]
+#[tokio::test]
+async fn azure_upload_download_works(ctx: &mut MaybeEnabledAzure) -> anyhow::Result<()> {
+    let MaybeEnabledAzure::Enabled(ctx) = ctx else {
+        return Ok(());
+    };
+
+    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;
+
+    ctx.client
+        .upload(std::io::Cursor::new(data), data.len(), &path, None)
+        .await?;
+
+    async fn download_and_compare(mut dl: Download) -> anyhow::Result<Vec<u8>> {
+        let mut buf = Vec::new();
+        tokio::io::copy(&mut 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);
+
+    // Full range (end specified)
+    let dl = ctx
+        .client
+        .download_byte_range(&path, 0, Some(data_len))
+        .await?;
+    let buf = download_and_compare(dl).await?;
+    assert_eq!(buf, data);
+
+    // 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]);
+
+    // partial range (end beyond real end)
+    let dl = ctx
+        .client
+        .download_byte_range(&path, 8, Some(data_len * 100))
+        .await?;
+    let buf = download_and_compare(dl).await?;
+    assert_eq!(buf, data[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..]);
+
+    // 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);
+
+    debug!("Cleanup: deleting file at path {path:?}");
+    ctx.client
+        .delete(&path)
+        .await
+        .with_context(|| format!("{path:?} removal"))?;
+
+    Ok(())
+}
+
+fn ensure_logging_ready() {
+    LOGGING_DONE.get_or_init(|| {
+        utils::logging::init(
+            utils::logging::LogFormat::Test,
+            utils::logging::TracingErrorLayerEnablement::Disabled,
+        )
+        .expect("logging init failed");
+    });
+}
+
+struct EnabledAzure {
+    client: Arc<GenericRemoteStorage>,
+    base_prefix: &'static str,
+}
+
+impl EnabledAzure {
+    async fn setup(max_keys_in_list_response: Option<i32>) -> Self {
+        let client = create_azure_client(max_keys_in_list_response)
+            .context("Azure client creation")
+            .expect("Azure client creation failed");
+
+        EnabledAzure {
+            client,
+            base_prefix: BASE_PREFIX,
+        }
+    }
+}
+
+enum MaybeEnabledAzure {
+    Enabled(EnabledAzure),
+    Disabled,
+}
+
+#[async_trait::async_trait]
+impl AsyncTestContext for MaybeEnabledAzure {
+    async fn setup() -> Self {
+        ensure_logging_ready();
+
+        if env::var(ENABLE_REAL_AZURE_REMOTE_STORAGE_ENV_VAR_NAME).is_err() {
+            info!(
+                "`{}` env variable is not set, skipping the test",
+                ENABLE_REAL_AZURE_REMOTE_STORAGE_ENV_VAR_NAME
+            );
+            return Self::Disabled;
+        }
+
+        Self::Enabled(EnabledAzure::setup(None).await)
+    }
+}
+
+enum MaybeEnabledAzureWithTestBlobs {
+    Enabled(AzureWithTestBlobs),
+    Disabled,
+    UploadsFailed(anyhow::Error, AzureWithTestBlobs),
+}
+
+struct AzureWithTestBlobs {
+    enabled: EnabledAzure,
+    remote_prefixes: HashSet<RemotePath>,
+    remote_blobs: HashSet<RemotePath>,
+}
+
+#[async_trait::async_trait]
+impl AsyncTestContext for MaybeEnabledAzureWithTestBlobs {
+    async fn setup() -> Self {
+        ensure_logging_ready();
+        if env::var(ENABLE_REAL_AZURE_REMOTE_STORAGE_ENV_VAR_NAME).is_err() {
+            info!(
+                "`{}` env variable is not set, skipping the test",
+                ENABLE_REAL_AZURE_REMOTE_STORAGE_ENV_VAR_NAME
+            );
+            return Self::Disabled;
+        }
+
+        let max_keys_in_list_response = 10;
+        let upload_tasks_count = 1 + (2 * usize::try_from(max_keys_in_list_response).unwrap());
+
+        let enabled = EnabledAzure::setup(Some(max_keys_in_list_response)).await;
+
+        match upload_azure_data(&enabled.client, enabled.base_prefix, upload_tasks_count).await {
+            ControlFlow::Continue(uploads) => {
+                info!("Remote objects created successfully");
+
+                Self::Enabled(AzureWithTestBlobs {
+                    enabled,
+                    remote_prefixes: uploads.prefixes,
+                    remote_blobs: uploads.blobs,
+                })
+            }
+            ControlFlow::Break(uploads) => Self::UploadsFailed(
+                anyhow::anyhow!("One or multiple blobs failed to upload to Azure"),
+                AzureWithTestBlobs {
+                    enabled,
+                    remote_prefixes: uploads.prefixes,
+                    remote_blobs: uploads.blobs,
+                },
+            ),
+        }
+    }
+
+    async fn teardown(self) {
+        match self {
+            Self::Disabled => {}
+            Self::Enabled(ctx) | Self::UploadsFailed(_, ctx) => {
+                cleanup(&ctx.enabled.client, ctx.remote_blobs).await;
+            }
+        }
+    }
+}
+
+// NOTE: the setups for the list_prefixes test and the list_files test are very similar
+// However, they are not idential. The list_prefixes function is concerned with listing prefixes,
+// whereas the list_files function is concerned with listing files.
+// See `RemoteStorage::list_files` documentation for more details
+enum MaybeEnabledAzureWithSimpleTestBlobs {
+    Enabled(AzureWithSimpleTestBlobs),
+    Disabled,
+    UploadsFailed(anyhow::Error, AzureWithSimpleTestBlobs),
+}
+struct AzureWithSimpleTestBlobs {
+    enabled: EnabledAzure,
+    remote_blobs: HashSet<RemotePath>,
+}
+
+#[async_trait::async_trait]
+impl AsyncTestContext for MaybeEnabledAzureWithSimpleTestBlobs {
+    async fn setup() -> Self {
+        ensure_logging_ready();
+        if env::var(ENABLE_REAL_AZURE_REMOTE_STORAGE_ENV_VAR_NAME).is_err() {
+            info!(
+                "`{}` env variable is not set, skipping the test",
+                ENABLE_REAL_AZURE_REMOTE_STORAGE_ENV_VAR_NAME
+            );
+            return Self::Disabled;
+        }
+
+        let max_keys_in_list_response = 10;
+        let upload_tasks_count = 1 + (2 * usize::try_from(max_keys_in_list_response).unwrap());
+
+        let enabled = EnabledAzure::setup(Some(max_keys_in_list_response)).await;
+
+        match upload_simple_azure_data(&enabled.client, upload_tasks_count).await {
+            ControlFlow::Continue(uploads) => {
+                info!("Remote objects created successfully");
+
+                Self::Enabled(AzureWithSimpleTestBlobs {
+                    enabled,
+                    remote_blobs: uploads,
+                })
+            }
+            ControlFlow::Break(uploads) => Self::UploadsFailed(
+                anyhow::anyhow!("One or multiple blobs failed to upload to Azure"),
+                AzureWithSimpleTestBlobs {
+                    enabled,
+                    remote_blobs: uploads,
+                },
+            ),
+        }
+    }
+
+    async fn teardown(self) {
+        match self {
+            Self::Disabled => {}
+            Self::Enabled(ctx) | Self::UploadsFailed(_, ctx) => {
+                cleanup(&ctx.enabled.client, ctx.remote_blobs).await;
+            }
+        }
+    }
+}
+
+fn create_azure_client(
+    max_keys_per_list_response: Option<i32>,
+) -> anyhow::Result<Arc<GenericRemoteStorage>> {
+    use rand::Rng;
+
+    let remote_storage_azure_container = env::var("REMOTE_STORAGE_AZURE_CONTAINER").context(
+        "`REMOTE_STORAGE_AZURE_CONTAINER` env var is not set, but real Azure tests are enabled",
+    )?;
+    let remote_storage_azure_region = env::var("REMOTE_STORAGE_AZURE_REGION").context(
+        "`REMOTE_STORAGE_AZURE_REGION` env var is not set, but real Azure tests are enabled",
+    )?;
+
+    // due to how time works, we've had test runners use the same nanos as bucket prefixes.
+    // millis is just a debugging aid for easier finding the prefix later.
+    let millis = std::time::SystemTime::now()
+        .duration_since(UNIX_EPOCH)
+        .context("random Azure test prefix part calculation")?
+        .as_millis();
+
+    // because nanos can be the same for two threads so can millis, add randomness
+    let random = rand::thread_rng().gen::<u32>();
+
+    let remote_storage_config = RemoteStorageConfig {
+        max_concurrent_syncs: NonZeroUsize::new(100).unwrap(),
+        max_sync_errors: NonZeroU32::new(5).unwrap(),
+        storage: RemoteStorageKind::AzureContainer(AzureConfig {
+            container_name: remote_storage_azure_container,
+            container_region: remote_storage_azure_region,
+            prefix_in_container: Some(format!("test_{millis}_{random:08x}/")),
+            concurrency_limit: NonZeroUsize::new(100).unwrap(),
+            max_keys_per_list_response,
+        }),
+    };
+    Ok(Arc::new(
+        GenericRemoteStorage::from_config(&remote_storage_config).context("remote storage init")?,
+    ))
+}
+
+struct Uploads {
+    prefixes: HashSet<RemotePath>,
+    blobs: HashSet<RemotePath>,
+}
+
+async fn upload_azure_data(
+    client: &Arc<GenericRemoteStorage>,
+    base_prefix_str: &'static str,
+    upload_tasks_count: usize,
+) -> ControlFlow<Uploads, Uploads> {
+    info!("Creating {upload_tasks_count} Azure files");
+    let mut upload_tasks = JoinSet::new();
+    for i in 1..upload_tasks_count + 1 {
+        let task_client = Arc::clone(client);
+        upload_tasks.spawn(async move {
+            let prefix = format!("{base_prefix_str}/sub_prefix_{i}/");
+            let blob_prefix = RemotePath::new(Utf8Path::new(&prefix))
+                .with_context(|| format!("{prefix:?} to RemotePath conversion"))?;
+            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?;
+
+            Ok::<_, anyhow::Error>((blob_prefix, blob_path))
+        });
+    }
+
+    let mut upload_tasks_failed = false;
+    let mut uploaded_prefixes = HashSet::with_capacity(upload_tasks_count);
+    let mut uploaded_blobs = HashSet::with_capacity(upload_tasks_count);
+    while let Some(task_run_result) = upload_tasks.join_next().await {
+        match task_run_result
+            .context("task join failed")
+            .and_then(|task_result| task_result.context("upload task failed"))
+        {
+            Ok((upload_prefix, upload_path)) => {
+                uploaded_prefixes.insert(upload_prefix);
+                uploaded_blobs.insert(upload_path);
+            }
+            Err(e) => {
+                error!("Upload task failed: {e:?}");
+                upload_tasks_failed = true;
+            }
+        }
+    }
+
+    let uploads = Uploads {
+        prefixes: uploaded_prefixes,
+        blobs: uploaded_blobs,
+    };
+    if upload_tasks_failed {
+        ControlFlow::Break(uploads)
+    } else {
+        ControlFlow::Continue(uploads)
+    }
+}
+
+async fn cleanup(client: &Arc<GenericRemoteStorage>, objects_to_delete: HashSet<RemotePath>) {
+    info!(
+        "Removing {} objects from the remote storage during cleanup",
+        objects_to_delete.len()
+    );
+    let mut delete_tasks = JoinSet::new();
+    for object_to_delete in objects_to_delete {
+        let task_client = Arc::clone(client);
+        delete_tasks.spawn(async move {
+            debug!("Deleting remote item at path {object_to_delete:?}");
+            task_client
+                .delete(&object_to_delete)
+                .await
+                .with_context(|| format!("{object_to_delete:?} removal"))
+        });
+    }
+
+    while let Some(task_run_result) = delete_tasks.join_next().await {
+        match task_run_result {
+            Ok(task_result) => match task_result {
+                Ok(()) => {}
+                Err(e) => error!("Delete task failed: {e:?}"),
+            },
+            Err(join_err) => error!("Delete task did not finish correctly: {join_err}"),
+        }
+    }
+}
+
+// Uploads files `folder{j}/blob{i}.txt`. See test description for more details.
+async fn upload_simple_azure_data(
+    client: &Arc<GenericRemoteStorage>,
+    upload_tasks_count: usize,
+) -> ControlFlow<HashSet<RemotePath>, HashSet<RemotePath>> {
+    info!("Creating {upload_tasks_count} Azure files");
+    let mut upload_tasks = JoinSet::new();
+    for i in 1..upload_tasks_count + 1 {
+        let task_client = Arc::clone(client);
+        upload_tasks.spawn(async move {
+            let blob_path = PathBuf::from(format!("folder{}/blob_{}.txt", i / 7, i));
+            let blob_path = RemotePath::new(
+                Utf8Path::from_path(blob_path.as_path()).expect("must be valid blob path"),
+            )
+            .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?;
+
+            Ok::<_, anyhow::Error>(blob_path)
+        });
+    }
+
+    let mut upload_tasks_failed = false;
+    let mut uploaded_blobs = HashSet::with_capacity(upload_tasks_count);
+    while let Some(task_run_result) = upload_tasks.join_next().await {
+        match task_run_result
+            .context("task join failed")
+            .and_then(|task_result| task_result.context("upload task failed"))
+        {
+            Ok(upload_path) => {
+                uploaded_blobs.insert(upload_path);
+            }
+            Err(e) => {
+                error!("Upload task failed: {e:?}");
+                upload_tasks_failed = true;
+            }
+        }
+    }
+
+    if upload_tasks_failed {
+        ControlFlow::Break(uploaded_blobs)
+    } else {
+        ControlFlow::Continue(uploaded_blobs)
+    }
+}

libs/utils/src/generation.rs

@@ -7,7 +7,7 @@ use serde::{Deserialize, Serialize};
 ///
 /// See docs/rfcs/025-generation-numbers.md for detail on how generation
 /// numbers are used.
-#[derive(Copy, Clone, Eq, PartialEq, PartialOrd, Ord)]
+#[derive(Copy, Clone, Eq, PartialEq, PartialOrd, Ord, Hash)]
 pub enum Generation {
     // Generations with this magic value will not add a suffix to S3 keys, and will not
     // be included in persisted index_part.json.  This value is only to be used

libs/utils/src/http/endpoint.rs

@@ -14,6 +14,11 @@ use tracing::{self, debug, info, info_span, warn, Instrument};
 use std::future::Future;
 use std::str::FromStr;
 
+use bytes::{Bytes, BytesMut};
+use std::io::Write as _;
+use tokio::sync::mpsc;
+use tokio_stream::wrappers::ReceiverStream;
+
 static SERVE_METRICS_COUNT: Lazy<IntCounter> = Lazy::new(|| {
     register_int_counter!(
         "libmetrics_metric_handler_requests_total",
@@ -146,94 +151,89 @@ impl Drop for RequestCancelled {
     }
 }
 
+/// An [`std::io::Write`] implementation on top of a channel sending [`bytes::Bytes`] chunks.
+pub struct ChannelWriter {
+    buffer: BytesMut,
+    pub tx: mpsc::Sender<std::io::Result<Bytes>>,
+    written: usize,
+}
+
+impl ChannelWriter {
+    pub fn new(buf_len: usize, tx: mpsc::Sender<std::io::Result<Bytes>>) -> Self {
+        assert_ne!(buf_len, 0);
+        ChannelWriter {
+            // split about half off the buffer from the start, because we flush depending on
+            // capacity. first flush will come sooner than without this, but now resizes will
+            // have better chance of picking up the "other" half. not guaranteed of course.
+            buffer: BytesMut::with_capacity(buf_len).split_off(buf_len / 2),
+            tx,
+            written: 0,
+        }
+    }
+
+    pub fn flush0(&mut self) -> std::io::Result<usize> {
+        let n = self.buffer.len();
+        if n == 0 {
+            return Ok(0);
+        }
+
+        tracing::trace!(n, "flushing");
+        let ready = self.buffer.split().freeze();
+
+        // not ideal to call from blocking code to block_on, but we are sure that this
+        // operation does not spawn_blocking other tasks
+        let res: Result<(), ()> = tokio::runtime::Handle::current().block_on(async {
+            self.tx.send(Ok(ready)).await.map_err(|_| ())?;
+
+            // throttle sending to allow reuse of our buffer in `write`.
+            self.tx.reserve().await.map_err(|_| ())?;
+
+            // now the response task has picked up the buffer and hopefully started
+            // sending it to the client.
+            Ok(())
+        });
+        if res.is_err() {
+            return Err(std::io::ErrorKind::BrokenPipe.into());
+        }
+        self.written += n;
+        Ok(n)
+    }
+
+    pub fn flushed_bytes(&self) -> usize {
+        self.written
+    }
+}
+
+impl std::io::Write for ChannelWriter {
+    fn write(&mut self, mut buf: &[u8]) -> std::io::Result<usize> {
+        let remaining = self.buffer.capacity() - self.buffer.len();
+
+        let out_of_space = remaining < buf.len();
+
+        let original_len = buf.len();
+
+        if out_of_space {
+            let can_still_fit = buf.len() - remaining;
+            self.buffer.extend_from_slice(&buf[..can_still_fit]);
+            buf = &buf[can_still_fit..];
+            self.flush0()?;
+        }
+
+        // assume that this will often under normal operation just move the pointer back to the
+        // beginning of allocation, because previous split off parts are already sent and
+        // dropped.
+        self.buffer.extend_from_slice(buf);
+        Ok(original_len)
+    }
+
+    fn flush(&mut self) -> std::io::Result<()> {
+        self.flush0().map(|_| ())
+    }
+}
+
 async fn prometheus_metrics_handler(_req: Request<Body>) -> Result<Response<Body>, ApiError> {
-    use bytes::{Bytes, BytesMut};
-    use std::io::Write as _;
-    use tokio::sync::mpsc;
-    use tokio_stream::wrappers::ReceiverStream;
-
     SERVE_METRICS_COUNT.inc();
 
-    /// An [`std::io::Write`] implementation on top of a channel sending [`bytes::Bytes`] chunks.
-    struct ChannelWriter {
-        buffer: BytesMut,
-        tx: mpsc::Sender<std::io::Result<Bytes>>,
-        written: usize,
-    }
-
-    impl ChannelWriter {
-        fn new(buf_len: usize, tx: mpsc::Sender<std::io::Result<Bytes>>) -> Self {
-            assert_ne!(buf_len, 0);
-            ChannelWriter {
-                // split about half off the buffer from the start, because we flush depending on
-                // capacity. first flush will come sooner than without this, but now resizes will
-                // have better chance of picking up the "other" half. not guaranteed of course.
-                buffer: BytesMut::with_capacity(buf_len).split_off(buf_len / 2),
-                tx,
-                written: 0,
-            }
-        }
-
-        fn flush0(&mut self) -> std::io::Result<usize> {
-            let n = self.buffer.len();
-            if n == 0 {
-                return Ok(0);
-            }
-
-            tracing::trace!(n, "flushing");
-            let ready = self.buffer.split().freeze();
-
-            // not ideal to call from blocking code to block_on, but we are sure that this
-            // operation does not spawn_blocking other tasks
-            let res: Result<(), ()> = tokio::runtime::Handle::current().block_on(async {
-                self.tx.send(Ok(ready)).await.map_err(|_| ())?;
-
-                // throttle sending to allow reuse of our buffer in `write`.
-                self.tx.reserve().await.map_err(|_| ())?;
-
-                // now the response task has picked up the buffer and hopefully started
-                // sending it to the client.
-                Ok(())
-            });
-            if res.is_err() {
-                return Err(std::io::ErrorKind::BrokenPipe.into());
-            }
-            self.written += n;
-            Ok(n)
-        }
-
-        fn flushed_bytes(&self) -> usize {
-            self.written
-        }
-    }
-
-    impl std::io::Write for ChannelWriter {
-        fn write(&mut self, mut buf: &[u8]) -> std::io::Result<usize> {
-            let remaining = self.buffer.capacity() - self.buffer.len();
-
-            let out_of_space = remaining < buf.len();
-
-            let original_len = buf.len();
-
-            if out_of_space {
-                let can_still_fit = buf.len() - remaining;
-                self.buffer.extend_from_slice(&buf[..can_still_fit]);
-                buf = &buf[can_still_fit..];
-                self.flush0()?;
-            }
-
-            // assume that this will often under normal operation just move the pointer back to the
-            // beginning of allocation, because previous split off parts are already sent and
-            // dropped.
-            self.buffer.extend_from_slice(buf);
-            Ok(original_len)
-        }
-
-        fn flush(&mut self) -> std::io::Result<()> {
-            self.flush0().map(|_| ())
-        }
-    }
-
     let started_at = std::time::Instant::now();
 
     let (tx, rx) = mpsc::channel(1);

libs/utils/src/lib.rs

@@ -73,6 +73,8 @@ pub mod completion;
 /// Reporting utilities
 pub mod error;
 
+pub mod sync;
+
 /// This is a shortcut to embed git sha into binaries and avoid copying the same build script to all packages
 ///
 /// we have several cases:
@@ -128,6 +130,21 @@ macro_rules! project_git_version {
     };
 }
 
+/// This is a shortcut to embed build tag into binaries and avoid copying the same build script to all packages
+#[macro_export]
+macro_rules! project_build_tag {
+    ($const_identifier:ident) => {
+        const $const_identifier: &::core::primitive::str = {
+            const __ARG: &[&::core::primitive::str; 2] = &match ::core::option_env!("BUILD_TAG") {
+                ::core::option::Option::Some(x) => ["build_tag-env:", x],
+                ::core::option::Option::None => ["build_tag:", ""],
+            };
+
+            $crate::__const_format::concatcp!(__ARG[0], __ARG[1])
+        };
+    };
+}
+
 /// Re-export for `project_git_version` macro
 #[doc(hidden)]
 pub use const_format as __const_format;

libs/utils/src/sync.rs

@@ -0,0 +1 @@
+pub mod heavier_once_cell;

libs/utils/src/sync/heavier_once_cell.rs

@@ -0,0 +1,383 @@
+use std::sync::{
+    atomic::{AtomicUsize, Ordering},
+    Arc, Mutex, MutexGuard,
+};
+use tokio::sync::Semaphore;
+
+/// Custom design like [`tokio::sync::OnceCell`] but using [`OwnedSemaphorePermit`] instead of
+/// `SemaphorePermit`, allowing use of `take` which does not require holding an outer mutex guard
+/// for the duration of initialization.
+///
+/// Has no unsafe, builds upon [`tokio::sync::Semaphore`] and [`std::sync::Mutex`].
+///
+/// [`OwnedSemaphorePermit`]: tokio::sync::OwnedSemaphorePermit
+pub struct OnceCell<T> {
+    inner: Mutex<Inner<T>>,
+    initializers: AtomicUsize,
+}
+
+impl<T> Default for OnceCell<T> {
+    /// Create new uninitialized [`OnceCell`].
+    fn default() -> Self {
+        Self {
+            inner: Default::default(),
+            initializers: AtomicUsize::new(0),
+        }
+    }
+}
+
+/// Semaphore is the current state:
+/// - open semaphore means the value is `None`, not yet initialized
+/// - closed semaphore means the value has been initialized
+#[derive(Debug)]
+struct Inner<T> {
+    init_semaphore: Arc<Semaphore>,
+    value: Option<T>,
+}
+
+impl<T> Default for Inner<T> {
+    fn default() -> Self {
+        Self {
+            init_semaphore: Arc::new(Semaphore::new(1)),
+            value: None,
+        }
+    }
+}
+
+impl<T> OnceCell<T> {
+    /// Creates an already initialized `OnceCell` with the given value.
+    pub fn new(value: T) -> Self {
+        let sem = Semaphore::new(1);
+        sem.close();
+        Self {
+            inner: Mutex::new(Inner {
+                init_semaphore: Arc::new(sem),
+                value: Some(value),
+            }),
+            initializers: AtomicUsize::new(0),
+        }
+    }
+
+    /// Returns a guard to an existing initialized value, or uniquely initializes the value before
+    /// returning the guard.
+    ///
+    /// Initializing might wait on any existing [`Guard::take_and_deinit`] deinitialization.
+    ///
+    /// Initialization is panic-safe and cancellation-safe.
+    pub async fn get_or_init<F, Fut, E>(&self, factory: F) -> Result<Guard<'_, T>, E>
+    where
+        F: FnOnce(InitPermit) -> Fut,
+        Fut: std::future::Future<Output = Result<(T, InitPermit), E>>,
+    {
+        let sem = {
+            let guard = self.inner.lock().unwrap();
+            if guard.value.is_some() {
+                return Ok(Guard(guard));
+            }
+            guard.init_semaphore.clone()
+        };
+
+        let permit = {
+            // increment the count for the duration of queued
+            let _guard = CountWaitingInitializers::start(self);
+            sem.acquire_owned().await
+        };
+
+        match permit {
+            Ok(permit) => {
+                let permit = InitPermit(permit);
+                let (value, _permit) = factory(permit).await?;
+
+                let guard = self.inner.lock().unwrap();
+
+                Ok(Self::set0(value, guard))
+            }
+            Err(_closed) => {
+                let guard = self.inner.lock().unwrap();
+                assert!(
+                    guard.value.is_some(),
+                    "semaphore got closed, must be initialized"
+                );
+                return Ok(Guard(guard));
+            }
+        }
+    }
+
+    /// Assuming a permit is held after previous call to [`Guard::take_and_deinit`], it can be used
+    /// to complete initializing the inner value.
+    ///
+    /// # Panics
+    ///
+    /// If the inner has already been initialized.
+    pub fn set(&self, value: T, _permit: InitPermit) -> Guard<'_, T> {
+        let guard = self.inner.lock().unwrap();
+
+        // cannot assert that this permit is for self.inner.semaphore, but we can assert it cannot
+        // give more permits right now.
+        if guard.init_semaphore.try_acquire().is_ok() {
+            drop(guard);
+            panic!("permit is of wrong origin");
+        }
+
+        Self::set0(value, guard)
+    }
+
+    fn set0(value: T, mut guard: std::sync::MutexGuard<'_, Inner<T>>) -> Guard<'_, T> {
+        if guard.value.is_some() {
+            drop(guard);
+            unreachable!("we won permit, must not be initialized");
+        }
+        guard.value = Some(value);
+        guard.init_semaphore.close();
+        Guard(guard)
+    }
+
+    /// Returns a guard to an existing initialized value, if any.
+    pub fn get(&self) -> Option<Guard<'_, T>> {
+        let guard = self.inner.lock().unwrap();
+        if guard.value.is_some() {
+            Some(Guard(guard))
+        } else {
+            None
+        }
+    }
+
+    /// Return the number of [`Self::get_or_init`] calls waiting for initialization to complete.
+    pub fn initializer_count(&self) -> usize {
+        self.initializers.load(Ordering::Relaxed)
+    }
+}
+
+/// DropGuard counter for queued tasks waiting to initialize, mainly accessible for the
+/// initializing task for example at the end of initialization.
+struct CountWaitingInitializers<'a, T>(&'a OnceCell<T>);
+
+impl<'a, T> CountWaitingInitializers<'a, T> {
+    fn start(target: &'a OnceCell<T>) -> Self {
+        target.initializers.fetch_add(1, Ordering::Relaxed);
+        CountWaitingInitializers(target)
+    }
+}
+
+impl<'a, T> Drop for CountWaitingInitializers<'a, T> {
+    fn drop(&mut self) {
+        self.0.initializers.fetch_sub(1, Ordering::Relaxed);
+    }
+}
+
+/// Uninteresting guard object to allow short-lived access to inspect or clone the held,
+/// initialized value.
+#[derive(Debug)]
+pub struct Guard<'a, T>(MutexGuard<'a, Inner<T>>);
+
+impl<T> std::ops::Deref for Guard<'_, T> {
+    type Target = T;
+
+    fn deref(&self) -> &Self::Target {
+        self.0
+            .value
+            .as_ref()
+            .expect("guard is not created unless value has been initialized")
+    }
+}
+
+impl<T> std::ops::DerefMut for Guard<'_, T> {
+    fn deref_mut(&mut self) -> &mut Self::Target {
+        self.0
+            .value
+            .as_mut()
+            .expect("guard is not created unless value has been initialized")
+    }
+}
+
+impl<'a, T> Guard<'a, T> {
+    /// Take the current value, and a new permit for it's deinitialization.
+    ///
+    /// The permit will be on a semaphore part of the new internal value, and any following
+    /// [`OnceCell::get_or_init`] will wait on it to complete.
+    pub fn take_and_deinit(&mut self) -> (T, InitPermit) {
+        let mut swapped = Inner::default();
+        let permit = swapped
+            .init_semaphore
+            .clone()
+            .try_acquire_owned()
+            .expect("we just created this");
+        std::mem::swap(&mut *self.0, &mut swapped);
+        swapped
+            .value
+            .map(|v| (v, InitPermit(permit)))
+            .expect("guard is not created unless value has been initialized")
+    }
+}
+
+/// Type held by OnceCell (de)initializing task.
+pub struct InitPermit(tokio::sync::OwnedSemaphorePermit);
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use std::{
+        convert::Infallible,
+        sync::atomic::{AtomicUsize, Ordering},
+        time::Duration,
+    };
+
+    #[tokio::test]
+    async fn many_initializers() {
+        #[derive(Default, Debug)]
+        struct Counters {
+            factory_got_to_run: AtomicUsize,
+            future_polled: AtomicUsize,
+            winners: AtomicUsize,
+        }
+
+        let initializers = 100;
+
+        let cell = Arc::new(OnceCell::default());
+        let counters = Arc::new(Counters::default());
+        let barrier = Arc::new(tokio::sync::Barrier::new(initializers + 1));
+
+        let mut js = tokio::task::JoinSet::new();
+        for i in 0..initializers {
+            js.spawn({
+                let cell = cell.clone();
+                let counters = counters.clone();
+                let barrier = barrier.clone();
+
+                async move {
+                    barrier.wait().await;
+                    let won = {
+                        let g = cell
+                            .get_or_init(|permit| {
+                                counters.factory_got_to_run.fetch_add(1, Ordering::Relaxed);
+                                async {
+                                    counters.future_polled.fetch_add(1, Ordering::Relaxed);
+                                    Ok::<_, Infallible>((i, permit))
+                                }
+                            })
+                            .await
+                            .unwrap();
+
+                        *g == i
+                    };
+
+                    if won {
+                        counters.winners.fetch_add(1, Ordering::Relaxed);
+                    }
+                }
+            });
+        }
+
+        barrier.wait().await;
+
+        while let Some(next) = js.join_next().await {
+            next.expect("no panics expected");
+        }
+
+        let mut counters = Arc::try_unwrap(counters).unwrap();
+
+        assert_eq!(*counters.factory_got_to_run.get_mut(), 1);
+        assert_eq!(*counters.future_polled.get_mut(), 1);
+        assert_eq!(*counters.winners.get_mut(), 1);
+    }
+
+    #[tokio::test(start_paused = true)]
+    async fn reinit_waits_for_deinit() {
+        // with the tokio::time paused, we will "sleep" for 1s while holding the reinitialization
+        let sleep_for = Duration::from_secs(1);
+        let initial = 42;
+        let reinit = 1;
+        let cell = Arc::new(OnceCell::new(initial));
+
+        let deinitialization_started = Arc::new(tokio::sync::Barrier::new(2));
+
+        let jh = tokio::spawn({
+            let cell = cell.clone();
+            let deinitialization_started = deinitialization_started.clone();
+            async move {
+                let (answer, _permit) = cell.get().expect("initialized to value").take_and_deinit();
+                assert_eq!(answer, initial);
+
+                deinitialization_started.wait().await;
+                tokio::time::sleep(sleep_for).await;
+            }
+        });
+
+        deinitialization_started.wait().await;
+
+        let started_at = tokio::time::Instant::now();
+        cell.get_or_init(|permit| async { Ok::<_, Infallible>((reinit, permit)) })
+            .await
+            .unwrap();
+
+        let elapsed = started_at.elapsed();
+        assert!(
+            elapsed >= sleep_for,
+            "initialization should had taken at least the time time slept with permit"
+        );
+
+        jh.await.unwrap();
+
+        assert_eq!(*cell.get().unwrap(), reinit);
+    }
+
+    #[test]
+    fn reinit_with_deinit_permit() {
+        let cell = Arc::new(OnceCell::new(42));
+
+        let (mol, permit) = cell.get().unwrap().take_and_deinit();
+        cell.set(5, permit);
+        assert_eq!(*cell.get().unwrap(), 5);
+
+        let (five, permit) = cell.get().unwrap().take_and_deinit();
+        assert_eq!(5, five);
+        cell.set(mol, permit);
+        assert_eq!(*cell.get().unwrap(), 42);
+    }
+
+    #[tokio::test]
+    async fn initialization_attemptable_until_ok() {
+        let cell = OnceCell::default();
+
+        for _ in 0..10 {
+            cell.get_or_init(|_permit| async { Err("whatever error") })
+                .await
+                .unwrap_err();
+        }
+
+        let g = cell
+            .get_or_init(|permit| async { Ok::<_, Infallible>(("finally success", permit)) })
+            .await
+            .unwrap();
+        assert_eq!(*g, "finally success");
+    }
+
+    #[tokio::test]
+    async fn initialization_is_cancellation_safe() {
+        let cell = OnceCell::default();
+
+        let barrier = tokio::sync::Barrier::new(2);
+
+        let initializer = cell.get_or_init(|permit| async {
+            barrier.wait().await;
+            futures::future::pending::<()>().await;
+
+            Ok::<_, Infallible>(("never reached", permit))
+        });
+
+        tokio::select! {
+            _ = initializer => { unreachable!("cannot complete; stuck in pending().await") },
+            _ = barrier.wait() => {}
+        };
+
+        // now initializer is dropped
+
+        assert!(cell.get().is_none());
+
+        let g = cell
+            .get_or_init(|permit| async { Ok::<_, Infallible>(("now initialized", permit)) })
+            .await
+            .unwrap();
+        assert_eq!(*g, "now initialized");
+    }
+}

libs/vm_monitor/README.md

@@ -27,8 +27,8 @@ and old one if it exists.
 * the filecache: a struct that allows communication with the Postgres file cache.
 On startup, we connect to the filecache and hold on to the connection for the
 entire monitor lifetime.
-* the cgroup watcher: the `CgroupWatcher` manages the `neon-postgres` cgroup by
-listening for `memory.high` events and setting its `memory.{high,max}` values.
+* the cgroup watcher: the `CgroupWatcher` polls the `neon-postgres` cgroup's memory
+usage and sends rolling aggregates to the runner.
 * the runner: the runner marries the filecache and cgroup watcher together,
 communicating with the agent throught the `Dispatcher`, and then calling filecache
 and cgroup watcher functions as needed to upscale and downscale

libs/vm_monitor/src/cgroup.rs

@@ -1,161 +1,38 @@
-use std::{
-    fmt::{Debug, Display},
-    fs,
-    pin::pin,
-    sync::atomic::{AtomicU64, Ordering},
-};
+use std::fmt::{self, Debug, Formatter};
+use std::time::{Duration, Instant};
 
-use anyhow::{anyhow, bail, Context};
+use anyhow::{anyhow, Context};
 use cgroups_rs::{
-    freezer::FreezerController,
-    hierarchies::{self, is_cgroup2_unified_mode, UNIFIED_MOUNTPOINT},
+    hierarchies::{self, is_cgroup2_unified_mode},
     memory::MemController,
-    MaxValue,
-    Subsystem::{Freezer, Mem},
+    Subsystem,
 };
-use inotify::{EventStream, Inotify, WatchMask};
-use tokio::sync::mpsc::{self, error::TryRecvError};
-use tokio::time::{Duration, Instant};
-use tokio_stream::{Stream, StreamExt};
+use tokio::sync::watch;
 use tracing::{info, warn};
 
-use crate::protocol::Resources;
-use crate::MiB;
-
-/// Monotonically increasing counter of the number of memory.high events
-/// the cgroup has experienced.
-///
-/// We use this to determine if a modification to the `memory.events` file actually
-/// changed the `high` field. If not, we don't care about the change. When we
-/// read the file, we check the `high` field in the file against `MEMORY_EVENT_COUNT`
-/// to see if it changed since last time.
-pub static MEMORY_EVENT_COUNT: AtomicU64 = AtomicU64::new(0);
-
-/// Monotonically increasing counter that gives each cgroup event a unique id.
-///
-/// This allows us to answer questions like "did this upscale arrive before this
-/// memory.high?". This static is also used by the `Sequenced` type to "tag" values
-/// with a sequence number. As such, prefer to used the `Sequenced` type rather
-/// than this static directly.
-static EVENT_SEQUENCE_NUMBER: AtomicU64 = AtomicU64::new(0);
-
-/// A memory event type reported in memory.events.
-#[derive(Debug, Eq, PartialEq, Copy, Clone)]
-pub enum MemoryEvent {
-    Low,
-    High,
-    Max,
-    Oom,
-    OomKill,
-    OomGroupKill,
-}
-
-impl MemoryEvent {
-    fn as_str(&self) -> &str {
-        match self {
-            MemoryEvent::Low => "low",
-            MemoryEvent::High => "high",
-            MemoryEvent::Max => "max",
-            MemoryEvent::Oom => "oom",
-            MemoryEvent::OomKill => "oom_kill",
-            MemoryEvent::OomGroupKill => "oom_group_kill",
-        }
-    }
-}
-
-impl Display for MemoryEvent {
-    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
-        f.write_str(self.as_str())
-    }
-}
-
 /// Configuration for a `CgroupWatcher`
 #[derive(Debug, Clone)]
 pub struct Config {
-    // The target difference between the total memory reserved for the cgroup
-    // and the value of the cgroup's memory.high.
-    //
-    // In other words, memory.high + oom_buffer_bytes will equal the total memory that the cgroup may
-    // use (equal to system memory, minus whatever's taken out for the file cache).
-    oom_buffer_bytes: u64,
+    /// Interval at which we should be fetching memory statistics
+    memory_poll_interval: Duration,
 
-    // The amount of memory, in bytes, below a proposed new value for
-    // memory.high that the cgroup's memory usage must be for us to downscale
-    //
-    // In other words, we can downscale only when:
-    //
-    //   memory.current + memory_high_buffer_bytes < (proposed) memory.high
-    //
-    // TODO: there's some minor issues with this approach -- in particular, that we might have
-    // memory in use by the kernel's page cache that we're actually ok with getting rid of.
-    pub(crate) memory_high_buffer_bytes: u64,
-
-    // The maximum duration, in milliseconds, that we're allowed to pause
-    // the cgroup for while waiting for the autoscaler-agent to upscale us
-    max_upscale_wait: Duration,
-
-    // The required minimum time, in milliseconds, that we must wait before re-freezing
-    // the cgroup while waiting for the autoscaler-agent to upscale us.
-    do_not_freeze_more_often_than: Duration,
-
-    // The amount of memory, in bytes, that we should periodically increase memory.high
-    // by while waiting for the autoscaler-agent to upscale us.
-    //
-    // This exists to avoid the excessive throttling that happens when a cgroup is above its
-    // memory.high for too long. See more here:
-    // https://github.com/neondatabase/autoscaling/issues/44#issuecomment-1522487217
-    memory_high_increase_by_bytes: u64,
-
-    // The period, in milliseconds, at which we should repeatedly increase the value
-    // of the cgroup's memory.high while we're waiting on upscaling and memory.high
-    // is still being hit.
-    //
-    // Technically speaking, this actually serves as a rate limit to moderate responding to
-    // memory.high events, but these are roughly equivalent if the process is still allocating
-    // memory.
-    memory_high_increase_every: Duration,
-}
-
-impl Config {
-    /// Calculate the new value for the cgroups memory.high based on system memory
-    pub fn calculate_memory_high_value(&self, total_system_mem: u64) -> u64 {
-        total_system_mem.saturating_sub(self.oom_buffer_bytes)
-    }
+    /// The number of samples used in constructing aggregated memory statistics
+    memory_history_len: usize,
+    /// The number of most recent samples that will be periodically logged.
+    ///
+    /// Each sample is logged exactly once. Increasing this value means that recent samples will be
+    /// logged less frequently, and vice versa.
+    ///
+    /// For simplicity, this value must be greater than or equal to `memory_history_len`.
+    memory_history_log_interval: usize,
 }
 
 impl Default for Config {
     fn default() -> Self {
         Self {
-            oom_buffer_bytes: 100 * MiB,
-            memory_high_buffer_bytes: 100 * MiB,
-            // while waiting for upscale, don't freeze for more than 20ms every 1s
-            max_upscale_wait: Duration::from_millis(20),
-            do_not_freeze_more_often_than: Duration::from_millis(1000),
-            // while waiting for upscale, increase memory.high by 10MiB every 25ms
-            memory_high_increase_by_bytes: 10 * MiB,
-            memory_high_increase_every: Duration::from_millis(25),
-        }
-    }
-}
-
-/// Used to represent data that is associated with a certain point in time, such
-/// as an upscale request or memory.high event.
-///
-/// Internally, creating a `Sequenced` uses a static atomic counter to obtain
-/// a unique sequence number. Sequence numbers are monotonically increasing,
-/// allowing us to answer questions like "did this upscale happen after this
-/// memory.high event?" by comparing the sequence numbers of the two events.
-#[derive(Debug, Clone)]
-pub struct Sequenced<T> {
-    seqnum: u64,
-    data: T,
-}
-
-impl<T> Sequenced<T> {
-    pub fn new(data: T) -> Self {
-        Self {
-            seqnum: EVENT_SEQUENCE_NUMBER.fetch_add(1, Ordering::AcqRel),
-            data,
+            memory_poll_interval: Duration::from_millis(100),
+            memory_history_len: 5, // use 500ms of history for decision-making
+            memory_history_log_interval: 20, // but only log every ~2s (otherwise it's spammy)
         }
     }
 }
@@ -170,74 +47,14 @@ impl<T> Sequenced<T> {
 pub struct CgroupWatcher {
     pub config: Config,
 
-    /// The sequence number of the last upscale.
-    ///
-    /// If we receive a memory.high event that has a _lower_ sequence number than
-    /// `last_upscale_seqnum`, then we know it occured before the upscale, and we
-    /// can safely ignore it.
-    ///
-    /// Note: Like the `events` field, this doesn't _need_ interior mutability but we
-    /// use it anyways so that methods take `&self`, not `&mut self`.
-    last_upscale_seqnum: AtomicU64,
-
-    /// A channel on which we send messages to request upscale from the dispatcher.
-    upscale_requester: mpsc::Sender<()>,
-
     /// The actual cgroup we are watching and managing.
     cgroup: cgroups_rs::Cgroup,
 }
 
-/// Read memory.events for the desired event type.
-///
-/// `path` specifies the path to the desired `memory.events` file.
-/// For more info, see the `memory.events` section of the [kernel docs]
-/// <https://docs.kernel.org/admin-guide/cgroup-v2.html#memory-interface-files>
-fn get_event_count(path: &str, event: MemoryEvent) -> anyhow::Result<u64> {
-    let contents = fs::read_to_string(path)
-        .with_context(|| format!("failed to read memory.events from {path}"))?;
-
-    // Then contents of the file look like:
-    // low 42
-    // high 101
-    // ...
-    contents
-        .lines()
-        .filter_map(|s| s.split_once(' '))
-        .find(|(e, _)| *e == event.as_str())
-        .ok_or_else(|| anyhow!("failed to find entry for memory.{event} events in {path}"))
-        .and_then(|(_, count)| {
-            count
-                .parse::<u64>()
-                .with_context(|| format!("failed to parse memory.{event} as u64"))
-        })
-}
-
-/// Create an event stream that produces events whenever the file at the provided
-/// path is modified.
-fn create_file_watcher(path: &str) -> anyhow::Result<EventStream<[u8; 1024]>> {
-    info!("creating file watcher for {path}");
-    let inotify = Inotify::init().context("failed to initialize file watcher")?;
-    inotify
-        .watches()
-        .add(path, WatchMask::MODIFY)
-        .with_context(|| format!("failed to start watching {path}"))?;
-    inotify
-        // The inotify docs use [0u8; 1024] so we'll just copy them. We only need
-        // to store one event at a time - if the event gets written over, that's
-        // ok. We still see that there is an event. For more information, see:
-        // https://man7.org/linux/man-pages/man7/inotify.7.html
-        .into_event_stream([0u8; 1024])
-        .context("failed to start inotify event stream")
-}
-
 impl CgroupWatcher {
     /// Create a new `CgroupWatcher`.
     #[tracing::instrument(skip_all, fields(%name))]
-    pub fn new(
-        name: String,
-        // A channel on which to send upscale requests
-        upscale_requester: mpsc::Sender<()>,
-    ) -> anyhow::Result<(Self, impl Stream<Item = Sequenced<u64>>)> {
+    pub fn new(name: String) -> anyhow::Result<Self> {
         // TODO: clarify exactly why we need v2
         // Make sure cgroups v2 (aka unified) are supported
         if !is_cgroup2_unified_mode() {
@@ -245,410 +62,203 @@ impl CgroupWatcher {
         }
         let cgroup = cgroups_rs::Cgroup::load(hierarchies::auto(), &name);
 
-        // Start monitoring the cgroup for memory events. In general, for
-        // cgroups v2 (aka unified), metrics are reported in files like
-        // > `/sys/fs/cgroup/{name}/{metric}`
-        // We are looking for `memory.high` events, which are stored in the
-        // file `memory.events`. For more info, see the `memory.events` section
-        // of https://docs.kernel.org/admin-guide/cgroup-v2.html#memory-interface-files
-        let path = format!("{}/{}/memory.events", UNIFIED_MOUNTPOINT, &name);
-        let memory_events = create_file_watcher(&path)
-            .with_context(|| format!("failed to create event watcher for {path}"))?
-            // This would be nice with with .inspect_err followed by .ok
-            .filter_map(move |_| match get_event_count(&path, MemoryEvent::High) {
-                Ok(high) => Some(high),
-                Err(error) => {
-                    // TODO: Might want to just panic here
-                    warn!(?error, "failed to read high events count from {}", &path);
-                    None
-                }
-            })
-            // Only report the event if the memory.high count increased
-            .filter_map(|high| {
-                if MEMORY_EVENT_COUNT.fetch_max(high, Ordering::AcqRel) < high {
-                    Some(high)
-                } else {
-                    None
-                }
-            })
-            .map(Sequenced::new);
-
-        let initial_count = get_event_count(
-            &format!("{}/{}/memory.events", UNIFIED_MOUNTPOINT, &name),
-            MemoryEvent::High,
-        )?;
-
-        info!(initial_count, "initial memory.high event count");
-
-        // Hard update `MEMORY_EVENT_COUNT` since there could have been processes
-        // running in the cgroup before that caused it to be non-zero.
-        MEMORY_EVENT_COUNT.fetch_max(initial_count, Ordering::AcqRel);
-
-        Ok((
-            Self {
-                cgroup,
-                upscale_requester,
-                last_upscale_seqnum: AtomicU64::new(0),
-                config: Default::default(),
-            },
-            memory_events,
-        ))
+        Ok(Self {
+            cgroup,
+            config: Default::default(),
+        })
     }
 
     /// The entrypoint for the `CgroupWatcher`.
     #[tracing::instrument(skip_all)]
-    pub async fn watch<E>(
+    pub async fn watch(
         &self,
-        // These are ~dependency injected~ (fancy, I know) because this function
-        // should never return.
-        // -> therefore: when we tokio::spawn it, we don't await the JoinHandle.
-        // -> therefore: if we want to stick it in an Arc so many threads can access
-        //    it, methods can never take mutable access.
-        //     - note: we use the Arc strategy so that a) we can call this function
-        //             right here and b) the runner can call the set/get_memory methods
-        // -> since calling recv() on a tokio::sync::mpsc::Receiver takes &mut self,
-        //    we just pass them in here instead of holding them in fields, as that
-        //    would require this method to take &mut self.
-        mut upscales: mpsc::Receiver<Sequenced<Resources>>,
-        events: E,
-    ) -> anyhow::Result<()>
-    where
-        E: Stream<Item = Sequenced<u64>>,
-    {
-        let mut wait_to_freeze = pin!(tokio::time::sleep(Duration::ZERO));
-        let mut last_memory_high_increase_at: Option<Instant> = None;
-        let mut events = pin!(events);
-
-        // Are we waiting to be upscaled? Could be true if we request upscale due
-        // to a memory.high event and it does not arrive in time.
-        let mut waiting_on_upscale = false;
-
-        loop {
-            tokio::select! {
-                upscale = upscales.recv() => {
-                    let Sequenced { seqnum, data } = upscale
-                        .context("failed to listen on upscale notification channel")?;
-                    waiting_on_upscale = false;
-                    last_memory_high_increase_at = None;
-                    self.last_upscale_seqnum.store(seqnum, Ordering::Release);
-                    info!(cpu = data.cpu, mem_bytes = data.mem, "received upscale");
-                }
-                event = events.next() => {
-                    let Some(Sequenced { seqnum, .. }) = event else {
-                        bail!("failed to listen for memory.high events")
-                    };
-                    // The memory.high came before our last upscale, so we consider
-                    // it resolved
-                    if self.last_upscale_seqnum.fetch_max(seqnum, Ordering::AcqRel) > seqnum {
-                        info!(
-                            "received memory.high event, but it came before our last upscale -> ignoring it"
-                        );
-                        continue;
-                    }
-
-                    // The memory.high came after our latest upscale. We don't
-                    // want to do anything yet, so peek the next event in hopes
-                    // that it's an upscale.
-                    if let Some(upscale_num) = self
-                        .upscaled(&mut upscales)
-                        .context("failed to check if we were upscaled")?
-                    {
-                        if upscale_num > seqnum {
-                            info!(
-                                "received memory.high event, but it came before our last upscale -> ignoring it"
-                            );
-                            continue;
-                        }
-                    }
-
-                    // If it's been long enough since we last froze, freeze the
-                    // cgroup and request upscale
-                    if wait_to_freeze.is_elapsed() {
-                        info!("received memory.high event -> requesting upscale");
-                        waiting_on_upscale = self
-                            .handle_memory_high_event(&mut upscales)
-                            .await
-                            .context("failed to handle upscale")?;
-                        wait_to_freeze
-                            .as_mut()
-                            .reset(Instant::now() + self.config.do_not_freeze_more_often_than);
-                        continue;
-                    }
-
-                    // Ok, we can't freeze, just request upscale
-                    if !waiting_on_upscale {
-                        info!("received memory.high event, but too soon to refreeze -> requesting upscale");
-
-                        // Make check to make sure we haven't been upscaled in the
-                        // meantine (can happen if the agent independently decides
-                        // to upscale us again)
-                        if self
-                            .upscaled(&mut upscales)
-                            .context("failed to check if we were upscaled")?
-                            .is_some()
-                        {
-                            info!("no need to request upscaling because we got upscaled");
-                            continue;
-                        }
-                        self.upscale_requester
-                            .send(())
-                            .await
-                            .context("failed to request upscale")?;
-                        waiting_on_upscale = true;
-                        continue;
-                    }
-
-                    // Shoot, we can't freeze or and we're still waiting on upscale,
-                    // increase memory.high to reduce throttling
-                    let can_increase_memory_high = match last_memory_high_increase_at {
-                        None => true,
-                        Some(t) => t.elapsed() > self.config.memory_high_increase_every,
-                    };
-                    if can_increase_memory_high {
-                        info!(
-                            "received memory.high event, \
-                            but too soon to refreeze and already requested upscale \
-                            -> increasing memory.high"
-                        );
-
-                        // Make check to make sure we haven't been upscaled in the
-                        // meantine (can happen if the agent independently decides
-                        // to upscale us again)
-                        if self
-                            .upscaled(&mut upscales)
-                            .context("failed to check if we were upscaled")?
-                            .is_some()
-                        {
-                            info!("no need to increase memory.high because got upscaled");
-                            continue;
-                        }
-
-                        // Request upscale anyways (the agent will handle deduplicating
-                        // requests)
-                        self.upscale_requester
-                            .send(())
-                            .await
-                            .context("failed to request upscale")?;
-
-                        let memory_high =
-                            self.get_memory_high_bytes().context("failed to get memory.high")?;
-                        let new_high = memory_high + self.config.memory_high_increase_by_bytes;
-                        info!(
-                            current_high_bytes = memory_high,
-                            new_high_bytes = new_high,
-                            "updating memory.high"
-                        );
-                        self.set_memory_high_bytes(new_high)
-                            .context("failed to set memory.high")?;
-                        last_memory_high_increase_at = Some(Instant::now());
-                        continue;
-                    }
-
-                    info!("received memory.high event, but can't do anything");
-                }
-            };
-        }
-    }
-
-    /// Handle a `memory.high`, returning whether we are still waiting on upscale
-    /// by the time the function returns.
-    ///
-    /// The general plan for handling a `memory.high` event is as follows:
-    /// 1. Freeze the cgroup
-    /// 2. Start a timer for `self.config.max_upscale_wait`
-    /// 3. Request upscale
-    /// 4. After the timer elapses or we receive upscale, thaw the cgroup.
-    /// 5. Return whether or not we are still waiting for upscale. If we are,
-    ///    we'll increase the cgroups memory.high to avoid getting oom killed
-    #[tracing::instrument(skip_all)]
-    async fn handle_memory_high_event(
-        &self,
-        upscales: &mut mpsc::Receiver<Sequenced<Resources>>,
-    ) -> anyhow::Result<bool> {
-        // Immediately freeze the cgroup before doing anything else.
-        info!("received memory.high event -> freezing cgroup");
-        self.freeze().context("failed to freeze cgroup")?;
-
-        // We'll use this for logging durations
-        let start_time = Instant::now();
-
-        // Await the upscale until we have to unfreeze
-        let timed =
-            tokio::time::timeout(self.config.max_upscale_wait, self.await_upscale(upscales));
-
-        // Request the upscale
-        info!(
-            wait = ?self.config.max_upscale_wait,
-            "sending request for immediate upscaling",
-        );
-        self.upscale_requester
-            .send(())
-            .await
-            .context("failed to request upscale")?;
-
-        let waiting_on_upscale = match timed.await {
-            Ok(Ok(())) => {
-                info!(elapsed = ?start_time.elapsed(), "received upscale in time");
-                false
-            }
-            // **important**: unfreeze the cgroup before ?-reporting the error
-            Ok(Err(e)) => {
-                info!("error waiting for upscale -> thawing cgroup");
-                self.thaw()
-                    .context("failed to thaw cgroup after errored waiting for upscale")?;
-                Err(e.context("failed to await upscale"))?
-            }
-            Err(_) => {
-                info!(elapsed = ?self.config.max_upscale_wait, "timed out waiting for upscale");
-                true
-            }
-        };
-
-        info!("thawing cgroup");
-        self.thaw().context("failed to thaw cgroup")?;
-
-        Ok(waiting_on_upscale)
-    }
-
-    /// Checks whether we were just upscaled, returning the upscale's sequence
-    /// number if so.
-    #[tracing::instrument(skip_all)]
-    fn upscaled(
-        &self,
-        upscales: &mut mpsc::Receiver<Sequenced<Resources>>,
-    ) -> anyhow::Result<Option<u64>> {
-        let Sequenced { seqnum, data } = match upscales.try_recv() {
-            Ok(upscale) => upscale,
-            Err(TryRecvError::Empty) => return Ok(None),
-            Err(TryRecvError::Disconnected) => {
-                bail!("upscale notification channel was disconnected")
-            }
-        };
-
-        // Make sure to update the last upscale sequence number
-        self.last_upscale_seqnum.store(seqnum, Ordering::Release);
-        info!(cpu = data.cpu, mem_bytes = data.mem, "received upscale");
-        Ok(Some(seqnum))
-    }
-
-    /// Await an upscale event, discarding any `memory.high` events received in
-    /// the process.
-    ///
-    /// This is used in `handle_memory_high_event`, where we need to listen
-    /// for upscales in particular so we know if we can thaw the cgroup early.
-    #[tracing::instrument(skip_all)]
-    async fn await_upscale(
-        &self,
-        upscales: &mut mpsc::Receiver<Sequenced<Resources>>,
+        updates: watch::Sender<(Instant, MemoryHistory)>,
     ) -> anyhow::Result<()> {
-        let Sequenced { seqnum, .. } = upscales
-            .recv()
-            .await
-            .context("error listening for upscales")?;
+        // this requirement makes the code a bit easier to work with; see the config for more.
+        assert!(self.config.memory_history_len <= self.config.memory_history_log_interval);
 
-        self.last_upscale_seqnum.store(seqnum, Ordering::Release);
-        Ok(())
-    }
+        let mut ticker = tokio::time::interval(self.config.memory_poll_interval);
+        ticker.set_missed_tick_behavior(tokio::time::MissedTickBehavior::Skip);
+        // ticker.reset_immediately(); // FIXME: enable this once updating to tokio >= 1.30.0
 
-    /// Get the cgroup's name.
-    pub fn path(&self) -> &str {
-        self.cgroup.path()
-    }
-}
+        let mem_controller = self.memory()?;
 
-// Methods for manipulating the actual cgroup
-impl CgroupWatcher {
-    /// Get a handle on the freezer subsystem.
-    fn freezer(&self) -> anyhow::Result<&FreezerController> {
-        if let Some(Freezer(freezer)) = self
-            .cgroup
-            .subsystems()
-            .iter()
-            .find(|sub| matches!(sub, Freezer(_)))
-        {
-            Ok(freezer)
-        } else {
-            anyhow::bail!("could not find freezer subsystem")
+        // buffer for samples that will be logged. once full, it remains so.
+        let history_log_len = self.config.memory_history_log_interval;
+        let mut history_log_buf = vec![MemoryStatus::zeroed(); history_log_len];
+
+        for t in 0_u64.. {
+            ticker.tick().await;
+
+            let now = Instant::now();
+            let mem = Self::memory_usage(mem_controller);
+
+            let i = t as usize % history_log_len;
+            history_log_buf[i] = mem;
+
+            // We're taking *at most* memory_history_len values; we may be bounded by the total
+            // number of samples that have come in so far.
+            let samples_count = (t + 1).min(self.config.memory_history_len as u64) as usize;
+            // NB: in `ring_buf_recent_values_iter`, `i` is *inclusive*, which matches the fact
+            // that we just inserted a value there, so the end of the iterator will *include* the
+            // value at i, rather than stopping just short of it.
+            let samples = ring_buf_recent_values_iter(&history_log_buf, i, samples_count);
+
+            let summary = MemoryHistory {
+                avg_non_reclaimable: samples.map(|h| h.non_reclaimable).sum::<u64>()
+                    / samples_count as u64,
+                samples_count,
+                samples_span: self.config.memory_poll_interval * (samples_count - 1) as u32,
+            };
+
+            // Log the current history if it's time to do so. Because `history_log_buf` has length
+            // equal to the logging interval, we can just log the entire buffer every time we set
+            // the last entry, which also means that for this log line, we can ignore that it's a
+            // ring buffer (because all the entries are in order of increasing time).
+            if i == history_log_len - 1 {
+                info!(
+                    history = ?MemoryStatus::debug_slice(&history_log_buf),
+                    summary = ?summary,
+                    "Recent cgroup memory statistics history"
+                );
+            }
+
+            updates
+                .send((now, summary))
+                .context("failed to send MemoryHistory")?;
         }
-    }
 
-    /// Attempt to freeze the cgroup.
-    pub fn freeze(&self) -> anyhow::Result<()> {
-        self.freezer()
-            .context("failed to get freezer subsystem")?
-            .freeze()
-            .context("failed to freeze")
-    }
-
-    /// Attempt to thaw the cgroup.
-    pub fn thaw(&self) -> anyhow::Result<()> {
-        self.freezer()
-            .context("failed to get freezer subsystem")?
-            .thaw()
-            .context("failed to thaw")
+        unreachable!()
     }
 
     /// Get a handle on the memory subsystem.
-    ///
-    /// Note: this method does not require `self.memory_update_lock` because
-    /// getting a handle to the subsystem does not access any of the files we
-    /// care about, such as memory.high and memory.events
     fn memory(&self) -> anyhow::Result<&MemController> {
-        if let Some(Mem(memory)) = self
-            .cgroup
+        self.cgroup
             .subsystems()
             .iter()
-            .find(|sub| matches!(sub, Mem(_)))
-        {
-            Ok(memory)
-        } else {
-            anyhow::bail!("could not find memory subsystem")
-        }
-    }
-
-    /// Get cgroup current memory usage.
-    pub fn current_memory_usage(&self) -> anyhow::Result<u64> {
-        Ok(self
-            .memory()
-            .context("failed to get memory subsystem")?
-            .memory_stat()
-            .usage_in_bytes)
-    }
-
-    /// Set cgroup memory.high threshold.
-    pub fn set_memory_high_bytes(&self, bytes: u64) -> anyhow::Result<()> {
-        self.set_memory_high_internal(MaxValue::Value(u64::min(bytes, i64::MAX as u64) as i64))
-    }
-
-    /// Set the cgroup's memory.high to 'max', disabling it.
-    pub fn unset_memory_high(&self) -> anyhow::Result<()> {
-        self.set_memory_high_internal(MaxValue::Max)
-    }
-
-    fn set_memory_high_internal(&self, value: MaxValue) -> anyhow::Result<()> {
-        self.memory()
-            .context("failed to get memory subsystem")?
-            .set_mem(cgroups_rs::memory::SetMemory {
-                low: None,
-                high: Some(value),
-                min: None,
-                max: None,
+            .find_map(|sub| match sub {
+                Subsystem::Mem(c) => Some(c),
+                _ => None,
             })
-            .map_err(anyhow::Error::from)
+            .ok_or_else(|| anyhow!("could not find memory subsystem"))
     }
 
-    /// Get memory.high threshold.
-    pub fn get_memory_high_bytes(&self) -> anyhow::Result<u64> {
-        let high = self
-            .memory()
-            .context("failed to get memory subsystem while getting memory statistics")?
-            .get_mem()
-            .map(|mem| mem.high)
-            .context("failed to get memory statistics from subsystem")?;
-        match high {
-            Some(MaxValue::Max) => Ok(i64::MAX as u64),
-            Some(MaxValue::Value(high)) => Ok(high as u64),
-            None => anyhow::bail!("failed to read memory.high from memory subsystem"),
+    /// Given a handle on the memory subsystem, returns the current memory information
+    fn memory_usage(mem_controller: &MemController) -> MemoryStatus {
+        let stat = mem_controller.memory_stat().stat;
+        MemoryStatus {
+            non_reclaimable: stat.active_anon + stat.inactive_anon,
         }
     }
 }
+
+// Helper function for `CgroupWatcher::watch`
+fn ring_buf_recent_values_iter<T>(
+    buf: &[T],
+    last_value_idx: usize,
+    count: usize,
+) -> impl '_ + Iterator<Item = &T> {
+    // Assertion carried over from `CgroupWatcher::watch`, to make the logic in this function
+    // easier (we only have to add `buf.len()` once, rather than a dynamic number of times).
+    assert!(count <= buf.len());
+
+    buf.iter()
+        // 'cycle' because the values could wrap around
+        .cycle()
+        // with 'cycle', this skip is more like 'offset', and functionally this is
+        // offsettting by 'last_value_idx - count (mod buf.len())', but we have to be
+        // careful to avoid underflow, so we pre-add buf.len().
+        // The '+ 1' is because `last_value_idx` is inclusive, rather than exclusive.
+        .skip((buf.len() + last_value_idx + 1 - count) % buf.len())
+        .take(count)
+}
+
+/// Summary of recent memory usage
+#[derive(Debug, Copy, Clone)]
+pub struct MemoryHistory {
+    /// Rolling average of non-reclaimable memory usage samples over the last `history_period`
+    pub avg_non_reclaimable: u64,
+
+    /// The number of samples used to construct this summary
+    pub samples_count: usize,
+    /// Total timespan between the first and last sample used for this summary
+    pub samples_span: Duration,
+}
+
+#[derive(Debug, Copy, Clone)]
+pub struct MemoryStatus {
+    non_reclaimable: u64,
+}
+
+impl MemoryStatus {
+    fn zeroed() -> Self {
+        MemoryStatus { non_reclaimable: 0 }
+    }
+
+    fn debug_slice(slice: &[Self]) -> impl '_ + Debug {
+        struct DS<'a>(&'a [MemoryStatus]);
+
+        impl<'a> Debug for DS<'a> {
+            fn fmt(&self, f: &mut Formatter) -> fmt::Result {
+                f.debug_struct("[MemoryStatus]")
+                    .field(
+                        "non_reclaimable[..]",
+                        &Fields(self.0, |stat: &MemoryStatus| {
+                            BytesToGB(stat.non_reclaimable)
+                        }),
+                    )
+                    .finish()
+            }
+        }
+
+        struct Fields<'a, F>(&'a [MemoryStatus], F);
+
+        impl<'a, F: Fn(&MemoryStatus) -> T, T: Debug> Debug for Fields<'a, F> {
+            fn fmt(&self, f: &mut Formatter) -> fmt::Result {
+                f.debug_list().entries(self.0.iter().map(&self.1)).finish()
+            }
+        }
+
+        struct BytesToGB(u64);
+
+        impl Debug for BytesToGB {
+            fn fmt(&self, f: &mut Formatter) -> fmt::Result {
+                f.write_fmt(format_args!(
+                    "{:.3}Gi",
+                    self.0 as f64 / (1_u64 << 30) as f64
+                ))
+            }
+        }
+
+        DS(slice)
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    #[test]
+    fn ring_buf_iter() {
+        let buf = vec![0_i32, 1, 2, 3, 4, 5, 6, 7, 8, 9];
+
+        let values = |offset, count| {
+            super::ring_buf_recent_values_iter(&buf, offset, count)
+                .copied()
+                .collect::<Vec<i32>>()
+        };
+
+        // Boundary conditions: start, end, and entire thing:
+        assert_eq!(values(0, 1), [0]);
+        assert_eq!(values(3, 4), [0, 1, 2, 3]);
+        assert_eq!(values(9, 4), [6, 7, 8, 9]);
+        assert_eq!(values(9, 10), [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]);
+
+        // "normal" operation: no wraparound
+        assert_eq!(values(7, 4), [4, 5, 6, 7]);
+
+        // wraparound:
+        assert_eq!(values(0, 4), [7, 8, 9, 0]);
+        assert_eq!(values(1, 4), [8, 9, 0, 1]);
+        assert_eq!(values(2, 4), [9, 0, 1, 2]);
+        assert_eq!(values(2, 10), [3, 4, 5, 6, 7, 8, 9, 0, 1, 2]);
+    }
+}

libs/vm_monitor/src/dispatcher.rs

@@ -12,12 +12,10 @@ use futures::{
     stream::{SplitSink, SplitStream},
     SinkExt, StreamExt,
 };
-use tokio::sync::mpsc;
 use tracing::info;
 
-use crate::cgroup::Sequenced;
 use crate::protocol::{
-    OutboundMsg, ProtocolRange, ProtocolResponse, ProtocolVersion, Resources, PROTOCOL_MAX_VERSION,
+    OutboundMsg, ProtocolRange, ProtocolResponse, ProtocolVersion, PROTOCOL_MAX_VERSION,
     PROTOCOL_MIN_VERSION,
 };
 
@@ -36,13 +34,6 @@ pub struct Dispatcher {
     /// We send messages to the agent through `sink`
     sink: SplitSink<WebSocket, Message>,
 
-    /// Used to notify the cgroup when we are upscaled.
-    pub(crate) notify_upscale_events: mpsc::Sender<Sequenced<Resources>>,
-
-    /// When the cgroup requests upscale it will send on this channel. In response
-    /// we send an `UpscaleRequst` to the agent.
-    pub(crate) request_upscale_events: mpsc::Receiver<()>,
-
     /// The protocol version we have agreed to use with the agent. This is negotiated
     /// during the creation of the dispatcher, and should be the highest shared protocol
     /// version.
@@ -61,11 +52,7 @@ impl Dispatcher {
     /// 1. Wait for the agent to sent the range of protocols it supports.
     /// 2. Send a protocol version that works for us as well, or an error if there
     ///    is no compatible version.
-    pub async fn new(
-        stream: WebSocket,
-        notify_upscale_events: mpsc::Sender<Sequenced<Resources>>,
-        request_upscale_events: mpsc::Receiver<()>,
-    ) -> anyhow::Result<Self> {
+    pub async fn new(stream: WebSocket) -> anyhow::Result<Self> {
         let (mut sink, mut source) = stream.split();
 
         // Figure out the highest protocol version we both support
@@ -119,22 +106,10 @@ impl Dispatcher {
         Ok(Self {
             sink,
             source,
-            notify_upscale_events,
-            request_upscale_events,
             proto_version: highest_shared_version,
         })
     }
 
-    /// Notify the cgroup manager that we have received upscale and wait for
-    /// the acknowledgement.
-    #[tracing::instrument(skip_all, fields(?resources))]
-    pub async fn notify_upscale(&self, resources: Sequenced<Resources>) -> anyhow::Result<()> {
-        self.notify_upscale_events
-            .send(resources)
-            .await
-            .context("failed to send resources and oneshot sender across channel")
-    }
-
     /// Send a message to the agent.
     ///
     /// Although this function is small, it has one major benefit: it is the only

libs/vm_monitor/src/filecache.rs

@@ -21,11 +21,6 @@ pub struct FileCacheState {
 
 #[derive(Debug)]
 pub struct FileCacheConfig {
-    /// Whether the file cache is *actually* stored in memory (e.g. by writing to
-    /// a tmpfs or shmem file). If true, the size of the file cache will be counted against the
-    /// memory available for the cgroup.
-    pub(crate) in_memory: bool,
-
     /// The size of the file cache, in terms of the size of the resource it consumes
     /// (currently: only memory)
     ///
@@ -59,22 +54,9 @@ pub struct FileCacheConfig {
     spread_factor: f64,
 }
 
-impl FileCacheConfig {
-    pub fn default_in_memory() -> Self {
+impl Default for FileCacheConfig {
+    fn default() -> Self {
         Self {
-            in_memory: true,
-            // 75 %
-            resource_multiplier: 0.75,
-            // 640 MiB; (512 + 128)
-            min_remaining_after_cache: NonZeroU64::new(640 * MiB).unwrap(),
-            // ensure any increase in file cache size is split 90-10 with 10% to other memory
-            spread_factor: 0.1,
-        }
-    }
-
-    pub fn default_on_disk() -> Self {
-        Self {
-            in_memory: false,
             resource_multiplier: 0.75,
             // 256 MiB - lower than when in memory because overcommitting is safe; if we don't have
             // memory, the kernel will just evict from its page cache, rather than e.g. killing
@@ -83,7 +65,9 @@ impl FileCacheConfig {
             spread_factor: 0.1,
         }
     }
+}
 
+impl FileCacheConfig {
     /// Make sure fields of the config are consistent.
     pub fn validate(&self) -> anyhow::Result<()> {
         // Single field validity

libs/vm_monitor/src/lib.rs

@@ -39,16 +39,6 @@ pub struct Args {
     #[arg(short, long)]
     pub pgconnstr: Option<String>,
 
-    /// Flag to signal that the Postgres file cache is on disk (i.e. not in memory aside from the
-    /// kernel's page cache), and therefore should not count against available memory.
-    //
-    // NB: Ideally this flag would directly refer to whether the file cache is in memory (rather
-    // than a roundabout way, via whether it's on disk), but in order to be backwards compatible
-    // during the switch away from an in-memory file cache, we had to default to the previous
-    // behavior.
-    #[arg(long)]
-    pub file_cache_on_disk: bool,
-
     /// The address we should listen on for connection requests. For the
     /// agent, this is 0.0.0.0:10301. For the informant, this is 127.0.0.1:10369.
     #[arg(short, long)]

libs/vm_monitor/src/runner.rs

@@ -5,18 +5,16 @@
 //! all functionality.
 
 use std::fmt::Debug;
-use std::sync::Arc;
 use std::time::{Duration, Instant};
 
 use anyhow::{bail, Context};
 use axum::extract::ws::{Message, WebSocket};
 use futures::StreamExt;
-use tokio::sync::broadcast;
-use tokio::sync::mpsc;
+use tokio::sync::{broadcast, watch};
 use tokio_util::sync::CancellationToken;
 use tracing::{error, info, warn};
 
-use crate::cgroup::{CgroupWatcher, Sequenced};
+use crate::cgroup::{self, CgroupWatcher};
 use crate::dispatcher::Dispatcher;
 use crate::filecache::{FileCacheConfig, FileCacheState};
 use crate::protocol::{InboundMsg, InboundMsgKind, OutboundMsg, OutboundMsgKind, Resources};
@@ -28,7 +26,7 @@ use crate::{bytes_to_mebibytes, get_total_system_memory, spawn_with_cancel, Args
 pub struct Runner {
     config: Config,
     filecache: Option<FileCacheState>,
-    cgroup: Option<Arc<CgroupWatcher>>,
+    cgroup: Option<CgroupState>,
     dispatcher: Dispatcher,
 
     /// We "mint" new message ids by incrementing this counter and taking the value.
@@ -45,6 +43,14 @@ pub struct Runner {
     kill: broadcast::Receiver<()>,
 }
 
+#[derive(Debug)]
+struct CgroupState {
+    watcher: watch::Receiver<(Instant, cgroup::MemoryHistory)>,
+    /// If [`cgroup::MemoryHistory::avg_non_reclaimable`] exceeds `threshold`, we send upscale
+    /// requests.
+    threshold: u64,
+}
+
 /// Configuration for a `Runner`
 #[derive(Debug)]
 pub struct Config {
@@ -62,16 +68,56 @@ pub struct Config {
     /// upscale resource amounts (because we might not *actually* have been upscaled yet). This field
     /// should be removed once we have a better solution there.
     sys_buffer_bytes: u64,
+
+    /// Minimum fraction of total system memory reserved *before* the the cgroup threshold; in
+    /// other words, providing a ceiling for the highest value of the threshold by enforcing that
+    /// there's at least `cgroup_min_overhead_fraction` of the total memory remaining beyond the
+    /// threshold.
+    ///
+    /// For example, a value of `0.1` means that 10% of total memory must remain after exceeding
+    /// the threshold, so the value of the cgroup threshold would always be capped at 90% of total
+    /// memory.
+    ///
+    /// The default value of `0.15` means that we *guarantee* sending upscale requests if the
+    /// cgroup is using more than 85% of total memory (even if we're *not* separately reserving
+    /// memory for the file cache).
+    cgroup_min_overhead_fraction: f64,
+
+    cgroup_downscale_threshold_buffer_bytes: u64,
 }
 
 impl Default for Config {
     fn default() -> Self {
         Self {
             sys_buffer_bytes: 100 * MiB,
+            cgroup_min_overhead_fraction: 0.15,
+            cgroup_downscale_threshold_buffer_bytes: 100 * MiB,
         }
     }
 }
 
+impl Config {
+    fn cgroup_threshold(&self, total_mem: u64, file_cache_disk_size: u64) -> u64 {
+        // If the file cache is in tmpfs, then it will count towards shmem usage of the cgroup,
+        // and thus be non-reclaimable, so we should allow for additional memory usage.
+        //
+        // If the file cache sits on disk, our desired stable system state is for it to be fully
+        // page cached (its contents should only be paged to/from disk in situations where we can't
+        // upscale fast enough). Page-cached memory is reclaimable, so we need to lower the
+        // threshold for non-reclaimable memory so we scale up *before* the kernel starts paging
+        // out the file cache.
+        let memory_remaining_for_cgroup = total_mem.saturating_sub(file_cache_disk_size);
+
+        // Even if we're not separately making room for the file cache (if it's in tmpfs), we still
+        // want our threshold to be met gracefully instead of letting postgres get OOM-killed.
+        // So we guarantee that there's at least `cgroup_min_overhead_fraction` of total memory
+        // remaining above the threshold.
+        let max_threshold = (total_mem as f64 * (1.0 - self.cgroup_min_overhead_fraction)) as u64;
+
+        memory_remaining_for_cgroup.min(max_threshold)
+    }
+}
+
 impl Runner {
     /// Create a new monitor.
     #[tracing::instrument(skip_all, fields(?config, ?args))]
@@ -87,12 +133,7 @@ impl Runner {
             "invalid monitor Config: sys_buffer_bytes cannot be 0"
         );
 
-        // *NOTE*: the dispatcher and cgroup manager talk through these channels
-        // so make sure they each get the correct half, nothing is droppped, etc.
-        let (notified_send, notified_recv) = mpsc::channel(1);
-        let (requesting_send, requesting_recv) = mpsc::channel(1);
-
-        let dispatcher = Dispatcher::new(ws, notified_send, requesting_recv)
+        let dispatcher = Dispatcher::new(ws)
             .await
             .context("error creating new dispatcher")?;
 
@@ -106,57 +147,18 @@ impl Runner {
             kill,
         };
 
-        // If we have both the cgroup and file cache integrations enabled, it's possible for
-        // temporary failures to result in cgroup throttling (from memory.high), that in turn makes
-        // it near-impossible to connect to the file cache (because it times out). Unfortunately,
-        // we *do* still want to determine the file cache size before setting the cgroup's
-        // memory.high, so it's not as simple as just swapping the order.
-        //
-        // Instead, the resolution here is that on vm-monitor startup (note: happens on each
-        // connection from autoscaler-agent, possibly multiple times per compute_ctl lifecycle), we
-        // temporarily unset memory.high, to allow any existing throttling to dissipate. It's a bit
-        // of a hacky solution, but helps with reliability.
-        if let Some(name) = &args.cgroup {
-            // Best not to set up cgroup stuff more than once, so we'll initialize cgroup state
-            // now, and then set limits later.
-            info!("initializing cgroup");
-
-            let (cgroup, cgroup_event_stream) = CgroupWatcher::new(name.clone(), requesting_send)
-                .context("failed to create cgroup manager")?;
-
-            info!("temporarily unsetting memory.high");
-
-            // Temporarily un-set cgroup memory.high; see above.
-            cgroup
-                .unset_memory_high()
-                .context("failed to unset memory.high")?;
-
-            let cgroup = Arc::new(cgroup);
-
-            let cgroup_clone = Arc::clone(&cgroup);
-            spawn_with_cancel(
-                token.clone(),
-                |_| error!("cgroup watcher terminated"),
-                async move { cgroup_clone.watch(notified_recv, cgroup_event_stream).await },
-            );
-
-            state.cgroup = Some(cgroup);
-        }
-
-        let mut file_cache_reserved_bytes = 0;
         let mem = get_total_system_memory();
 
+        let mut file_cache_disk_size = 0;
+
         // We need to process file cache initialization before cgroup initialization, so that the memory
         // allocated to the file cache is appropriately taken into account when we decide the cgroup's
         // memory limits.
         if let Some(connstr) = &args.pgconnstr {
             info!("initializing file cache");
-            let config = match args.file_cache_on_disk {
-                true => FileCacheConfig::default_on_disk(),
-                false => FileCacheConfig::default_in_memory(),
-            };
+            let config = FileCacheConfig::default();
 
-            let mut file_cache = FileCacheState::new(connstr, config, token)
+            let mut file_cache = FileCacheState::new(connstr, config, token.clone())
                 .await
                 .context("failed to create file cache")?;
 
@@ -181,23 +183,37 @@ impl Runner {
             if actual_size != new_size {
                 info!("file cache size actually got set to {actual_size}")
             }
-            // Mark the resources given to the file cache as reserved, but only if it's in memory.
-            if !args.file_cache_on_disk {
-                file_cache_reserved_bytes = actual_size;
-            }
 
+            file_cache_disk_size = actual_size;
             state.filecache = Some(file_cache);
         }
 
-        if let Some(cgroup) = &state.cgroup {
-            let available = mem - file_cache_reserved_bytes;
-            let value = cgroup.config.calculate_memory_high_value(available);
+        if let Some(name) = &args.cgroup {
+            // Best not to set up cgroup stuff more than once, so we'll initialize cgroup state
+            // now, and then set limits later.
+            info!("initializing cgroup");
 
-            info!(value, "setting memory.high");
+            let cgroup =
+                CgroupWatcher::new(name.clone()).context("failed to create cgroup manager")?;
 
-            cgroup
-                .set_memory_high_bytes(value)
-                .context("failed to set cgroup memory.high")?;
+            let init_value = cgroup::MemoryHistory {
+                avg_non_reclaimable: 0,
+                samples_count: 0,
+                samples_span: Duration::ZERO,
+            };
+            let (hist_tx, hist_rx) = watch::channel((Instant::now(), init_value));
+
+            spawn_with_cancel(token, |_| error!("cgroup watcher terminated"), async move {
+                cgroup.watch(hist_tx).await
+            });
+
+            let threshold = state.config.cgroup_threshold(mem, file_cache_disk_size);
+            info!(threshold, "set initial cgroup threshold",);
+
+            state.cgroup = Some(CgroupState {
+                watcher: hist_rx,
+                threshold,
+            });
         }
 
         Ok(state)
@@ -217,28 +233,45 @@ impl Runner {
 
         let requested_mem = target.mem;
         let usable_system_memory = requested_mem.saturating_sub(self.config.sys_buffer_bytes);
-        let expected_file_cache_mem_usage = self
+        let expected_file_cache_size = self
             .filecache
             .as_ref()
             .map(|file_cache| file_cache.config.calculate_cache_size(usable_system_memory))
             .unwrap_or(0);
-        let mut new_cgroup_mem_high = 0;
         if let Some(cgroup) = &self.cgroup {
-            new_cgroup_mem_high = cgroup
+            let (last_time, last_history) = *cgroup.watcher.borrow();
+
+            // NB: The ordering of these conditions is intentional. During startup, we should deny
+            // downscaling until we have enough information to determine that it's safe to do so
+            // (i.e. enough samples have come in). But if it's been a while and we *still* haven't
+            // received any information, we should *fail* instead of just denying downscaling.
+            //
+            // `last_time` is set to `Instant::now()` on startup, so checking `last_time.elapsed()`
+            // serves double-duty: it trips if we haven't received *any* metrics for long enough,
+            // OR if we haven't received metrics *recently enough*.
+            //
+            // TODO: make the duration here configurable.
+            if last_time.elapsed() > Duration::from_secs(5) {
+                bail!("haven't gotten cgroup memory stats recently enough to determine downscaling information");
+            } else if last_history.samples_count <= 1 {
+                let status = "haven't received enough cgroup memory stats yet";
+                info!(status, "discontinuing downscale");
+                return Ok((false, status.to_owned()));
+            }
+
+            let new_threshold = self
                 .config
-                .calculate_memory_high_value(usable_system_memory - expected_file_cache_mem_usage);
+                .cgroup_threshold(usable_system_memory, expected_file_cache_size);
 
-            let current = cgroup
-                .current_memory_usage()
-                .context("failed to fetch cgroup memory")?;
+            let current = last_history.avg_non_reclaimable;
 
-            if new_cgroup_mem_high < current + cgroup.config.memory_high_buffer_bytes {
+            if new_threshold < current + self.config.cgroup_downscale_threshold_buffer_bytes {
                 let status = format!(
-                    "{}: {} MiB (new high) < {} (current usage) + {} (buffer)",
-                    "calculated memory.high too low",
-                    bytes_to_mebibytes(new_cgroup_mem_high),
+                    "{}: {} MiB (new threshold) < {} (current usage) + {} (downscale buffer)",
+                    "calculated memory threshold too low",
+                    bytes_to_mebibytes(new_threshold),
                     bytes_to_mebibytes(current),
-                    bytes_to_mebibytes(cgroup.config.memory_high_buffer_bytes)
+                    bytes_to_mebibytes(self.config.cgroup_downscale_threshold_buffer_bytes)
                 );
 
                 info!(status, "discontinuing downscale");
@@ -249,42 +282,33 @@ impl Runner {
 
         // The downscaling has been approved. Downscale the file cache, then the cgroup.
         let mut status = vec![];
-        let mut file_cache_mem_usage = 0;
+        let mut file_cache_disk_size = 0;
         if let Some(file_cache) = &mut self.filecache {
             let actual_usage = file_cache
-                .set_file_cache_size(expected_file_cache_mem_usage)
+                .set_file_cache_size(expected_file_cache_size)
                 .await
                 .context("failed to set file cache size")?;
-            if file_cache.config.in_memory {
-                file_cache_mem_usage = actual_usage;
-            }
+            file_cache_disk_size = actual_usage;
             let message = format!(
-                "set file cache size to {} MiB (in memory = {})",
+                "set file cache size to {} MiB",
                 bytes_to_mebibytes(actual_usage),
-                file_cache.config.in_memory,
             );
             info!("downscale: {message}");
             status.push(message);
         }
 
-        if let Some(cgroup) = &self.cgroup {
-            let available_memory = usable_system_memory - file_cache_mem_usage;
-
-            if file_cache_mem_usage != expected_file_cache_mem_usage {
-                new_cgroup_mem_high = cgroup.config.calculate_memory_high_value(available_memory);
-            }
-
-            // new_cgroup_mem_high is initialized to 0 but it is guaranteed to not be here
-            // since it is properly initialized in the previous cgroup if let block
-            cgroup
-                .set_memory_high_bytes(new_cgroup_mem_high)
-                .context("failed to set cgroup memory.high")?;
+        if let Some(cgroup) = &mut self.cgroup {
+            let new_threshold = self
+                .config
+                .cgroup_threshold(usable_system_memory, file_cache_disk_size);
 
             let message = format!(
-                "set cgroup memory.high to {} MiB, of new max {} MiB",
-                bytes_to_mebibytes(new_cgroup_mem_high),
-                bytes_to_mebibytes(available_memory)
+                "set cgroup memory threshold from {} MiB to {} MiB, of new total {} MiB",
+                bytes_to_mebibytes(cgroup.threshold),
+                bytes_to_mebibytes(new_threshold),
+                bytes_to_mebibytes(usable_system_memory)
             );
+            cgroup.threshold = new_threshold;
             info!("downscale: {message}");
             status.push(message);
         }
@@ -305,8 +329,7 @@ impl Runner {
         let new_mem = resources.mem;
         let usable_system_memory = new_mem.saturating_sub(self.config.sys_buffer_bytes);
 
-        // Get the file cache's expected contribution to the memory usage
-        let mut file_cache_mem_usage = 0;
+        let mut file_cache_disk_size = 0;
         if let Some(file_cache) = &mut self.filecache {
             let expected_usage = file_cache.config.calculate_cache_size(usable_system_memory);
             info!(
@@ -319,9 +342,7 @@ impl Runner {
                 .set_file_cache_size(expected_usage)
                 .await
                 .context("failed to set file cache size")?;
-            if file_cache.config.in_memory {
-                file_cache_mem_usage = actual_usage;
-            }
+            file_cache_disk_size = actual_usage;
 
             if actual_usage != expected_usage {
                 warn!(
@@ -332,18 +353,18 @@ impl Runner {
             }
         }
 
-        if let Some(cgroup) = &self.cgroup {
-            let available_memory = usable_system_memory - file_cache_mem_usage;
-            let new_cgroup_mem_high = cgroup.config.calculate_memory_high_value(available_memory);
+        if let Some(cgroup) = &mut self.cgroup {
+            let new_threshold = self
+                .config
+                .cgroup_threshold(usable_system_memory, file_cache_disk_size);
+
             info!(
-                target = bytes_to_mebibytes(new_cgroup_mem_high),
-                total = bytes_to_mebibytes(new_mem),
-                name = cgroup.path(),
-                "updating cgroup memory.high",
+                "set cgroup memory threshold from {} MiB to {} MiB of new total {} MiB",
+                bytes_to_mebibytes(cgroup.threshold),
+                bytes_to_mebibytes(new_threshold),
+                bytes_to_mebibytes(usable_system_memory)
             );
-            cgroup
-                .set_memory_high_bytes(new_cgroup_mem_high)
-                .context("failed to set cgroup memory.high")?;
+            cgroup.threshold = new_threshold;
         }
 
         Ok(())
@@ -361,10 +382,6 @@ impl Runner {
                 self.handle_upscale(granted)
                     .await
                     .context("failed to handle upscale")?;
-                self.dispatcher
-                    .notify_upscale(Sequenced::new(granted))
-                    .await
-                    .context("failed to notify notify cgroup of upscale")?;
                 Ok(Some(OutboundMsg::new(
                     OutboundMsgKind::UpscaleConfirmation {},
                     id,
@@ -408,33 +425,53 @@ impl Runner {
                         Err(e) => bail!("failed to receive kill signal: {e}")
                     }
                 }
-                // we need to propagate an upscale request
-                request = self.dispatcher.request_upscale_events.recv(), if self.cgroup.is_some() => {
-                    if request.is_none() {
-                        bail!("failed to listen for upscale event from cgroup")
+
+                // New memory stats from the cgroup, *may* need to request upscaling, if we've
+                // exceeded the threshold
+                result = self.cgroup.as_mut().unwrap().watcher.changed(), if self.cgroup.is_some() => {
+                    result.context("failed to receive from cgroup memory stats watcher")?;
+
+                    let cgroup = self.cgroup.as_ref().unwrap();
+
+                    let (_time, cgroup_mem_stat) = *cgroup.watcher.borrow();
+
+                    // If we haven't exceeded the threshold, then we're all ok
+                    if cgroup_mem_stat.avg_non_reclaimable < cgroup.threshold {
+                        continue;
                     }
 
-                    // If it's been less than 1 second since the last time we requested upscaling,
-                    // ignore the event, to avoid spamming the agent (otherwise, this can happen
-                    // ~1k times per second).
+                    // Otherwise, we generally want upscaling. But, if it's been less than 1 second
+                    // since the last time we requested upscaling, ignore the event, to avoid
+                    // spamming the agent.
                     if let Some(t) = self.last_upscale_request_at {
                         let elapsed = t.elapsed();
                         if elapsed < Duration::from_secs(1) {
-                            info!(elapsed_millis = elapsed.as_millis(), "cgroup asked for upscale but too soon to forward the request, ignoring");
+                            info!(
+                                elapsed_millis = elapsed.as_millis(),
+                                avg_non_reclaimable = bytes_to_mebibytes(cgroup_mem_stat.avg_non_reclaimable),
+                                threshold = bytes_to_mebibytes(cgroup.threshold),
+                                "cgroup memory stats are high enough to upscale but too soon to forward the request, ignoring",
+                            );
                             continue;
                         }
                     }
 
                     self.last_upscale_request_at = Some(Instant::now());
 
-                    info!("cgroup asking for upscale; forwarding request");
+                    info!(
+                        avg_non_reclaimable = bytes_to_mebibytes(cgroup_mem_stat.avg_non_reclaimable),
+                        threshold = bytes_to_mebibytes(cgroup.threshold),
+                        "cgroup memory stats are high enough to upscale, requesting upscale",
+                    );
+
                     self.counter += 2; // Increment, preserving parity (i.e. keep the
                                        // counter odd). See the field comment for more.
                     self.dispatcher
                         .send(OutboundMsg::new(OutboundMsgKind::UpscaleRequest {}, self.counter))
                         .await
                         .context("failed to send message")?;
-                }
+                },
+
                 // there is a message from the agent
                 msg = self.dispatcher.source.next() => {
                     if let Some(msg) = msg {
@@ -462,11 +499,14 @@ impl Runner {
                                     Ok(Some(out)) => out,
                                     Ok(None) => continue,
                                     Err(e) => {
-                                        let error = e.to_string();
-                                        warn!(?error, "error handling message");
+                                        // use {:#} for our logging because the display impl only
+                                        // gives the outermost cause, and the debug impl
+                                        // pretty-prints the error, whereas {:#} contains all the
+                                        // causes, but is compact (no newlines).
+                                        warn!(error = format!("{e:#}"), "error handling message");
                                         OutboundMsg::new(
                                             OutboundMsgKind::InternalError {
-                                                error
+                                                error: e.to_string(),
                                             },
                                             message.id
                                         )

libs/walproposer/Cargo.toml

@@ -0,0 +1,16 @@
+[package]
+name = "walproposer"
+version = "0.1.0"
+edition.workspace = true
+license.workspace = true
+
+[dependencies]
+anyhow.workspace = true
+utils.workspace = true
+postgres_ffi.workspace = true
+
+workspace_hack.workspace = true
+
+[build-dependencies]
+anyhow.workspace = true
+bindgen.workspace = true

libs/walproposer/bindgen_deps.h

@@ -0,0 +1 @@
+#include "walproposer.h"

libs/walproposer/build.rs

@@ -0,0 +1,113 @@
+use std::{env, path::PathBuf, process::Command};
+
+use anyhow::{anyhow, Context};
+use bindgen::CargoCallbacks;
+
+fn main() -> anyhow::Result<()> {
+    // Tell cargo to invalidate the built crate whenever the wrapper changes
+    println!("cargo:rerun-if-changed=bindgen_deps.h");
+
+    // Finding the location of built libraries and Postgres C headers:
+    // - if POSTGRES_INSTALL_DIR is set look into it, otherwise look into `<project_root>/pg_install`
+    // - if there's a `bin/pg_config` file use it for getting include server, otherwise use `<project_root>/pg_install/{PG_MAJORVERSION}/include/postgresql/server`
+    let pg_install_dir = if let Some(postgres_install_dir) = env::var_os("POSTGRES_INSTALL_DIR") {
+        postgres_install_dir.into()
+    } else {
+        PathBuf::from(env!("CARGO_MANIFEST_DIR")).join("../../pg_install")
+    };
+
+    let pg_install_abs = std::fs::canonicalize(pg_install_dir)?;
+    let walproposer_lib_dir = pg_install_abs.join("build/walproposer-lib");
+    let walproposer_lib_search_str = walproposer_lib_dir
+        .to_str()
+        .ok_or(anyhow!("Bad non-UTF path"))?;
+
+    let pgxn_neon = PathBuf::from(env!("CARGO_MANIFEST_DIR")).join("../../pgxn/neon");
+    let pgxn_neon = std::fs::canonicalize(pgxn_neon)?;
+    let pgxn_neon = pgxn_neon.to_str().ok_or(anyhow!("Bad non-UTF path"))?;
+
+    println!("cargo:rustc-link-lib=static=pgport");
+    println!("cargo:rustc-link-lib=static=pgcommon");
+    println!("cargo:rustc-link-lib=static=walproposer");
+    println!("cargo:rustc-link-search={walproposer_lib_search_str}");
+
+    let pg_config_bin = pg_install_abs.join("v16").join("bin").join("pg_config");
+    let inc_server_path: String = if pg_config_bin.exists() {
+        let output = Command::new(pg_config_bin)
+            .arg("--includedir-server")
+            .output()
+            .context("failed to execute `pg_config --includedir-server`")?;
+
+        if !output.status.success() {
+            panic!("`pg_config --includedir-server` failed")
+        }
+
+        String::from_utf8(output.stdout)
+            .context("pg_config output is not UTF-8")?
+            .trim_end()
+            .into()
+    } else {
+        let server_path = pg_install_abs
+            .join("v16")
+            .join("include")
+            .join("postgresql")
+            .join("server")
+            .into_os_string();
+        server_path
+            .into_string()
+            .map_err(|s| anyhow!("Bad postgres server path {s:?}"))?
+    };
+
+    // The bindgen::Builder is the main entry point
+    // to bindgen, and lets you build up options for
+    // the resulting bindings.
+    let bindings = bindgen::Builder::default()
+        // The input header we would like to generate
+        // bindings for.
+        .header("bindgen_deps.h")
+        // Tell cargo to invalidate the built crate whenever any of the
+        // included header files changed.
+        .parse_callbacks(Box::new(CargoCallbacks))
+        .allowlist_type("WalProposer")
+        .allowlist_type("WalProposerConfig")
+        .allowlist_type("walproposer_api")
+        .allowlist_function("WalProposerCreate")
+        .allowlist_function("WalProposerStart")
+        .allowlist_function("WalProposerBroadcast")
+        .allowlist_function("WalProposerPoll")
+        .allowlist_function("WalProposerFree")
+        .allowlist_var("DEBUG5")
+        .allowlist_var("DEBUG4")
+        .allowlist_var("DEBUG3")
+        .allowlist_var("DEBUG2")
+        .allowlist_var("DEBUG1")
+        .allowlist_var("LOG")
+        .allowlist_var("INFO")
+        .allowlist_var("NOTICE")
+        .allowlist_var("WARNING")
+        .allowlist_var("ERROR")
+        .allowlist_var("FATAL")
+        .allowlist_var("PANIC")
+        .allowlist_var("WPEVENT")
+        .allowlist_var("WL_LATCH_SET")
+        .allowlist_var("WL_SOCKET_READABLE")
+        .allowlist_var("WL_SOCKET_WRITEABLE")
+        .allowlist_var("WL_TIMEOUT")
+        .allowlist_var("WL_SOCKET_CLOSED")
+        .allowlist_var("WL_SOCKET_MASK")
+        .clang_arg("-DWALPROPOSER_LIB")
+        .clang_arg(format!("-I{pgxn_neon}"))
+        .clang_arg(format!("-I{inc_server_path}"))
+        // Finish the builder and generate the bindings.
+        .generate()
+        // Unwrap the Result and panic on failure.
+        .expect("Unable to generate bindings");
+
+    // Write the bindings to the $OUT_DIR/bindings.rs file.
+    let out_path = PathBuf::from(env::var("OUT_DIR").unwrap()).join("bindings.rs");
+    bindings
+        .write_to_file(out_path)
+        .expect("Couldn't write bindings!");
+
+    Ok(())
+}

libs/walproposer/src/api_bindings.rs

@@ -0,0 +1,455 @@
+#![allow(dead_code)]
+
+use std::ffi::CStr;
+use std::ffi::CString;
+
+use crate::bindings::uint32;
+use crate::bindings::walproposer_api;
+use crate::bindings::PGAsyncReadResult;
+use crate::bindings::PGAsyncWriteResult;
+use crate::bindings::Safekeeper;
+use crate::bindings::Size;
+use crate::bindings::StringInfoData;
+use crate::bindings::TimeLineID;
+use crate::bindings::TimestampTz;
+use crate::bindings::WalProposer;
+use crate::bindings::WalProposerConnStatusType;
+use crate::bindings::WalProposerConnectPollStatusType;
+use crate::bindings::WalProposerExecStatusType;
+use crate::bindings::WalproposerShmemState;
+use crate::bindings::XLogRecPtr;
+use crate::walproposer::ApiImpl;
+use crate::walproposer::WaitResult;
+
+extern "C" fn get_shmem_state(wp: *mut WalProposer) -> *mut WalproposerShmemState {
+    unsafe {
+        let callback_data = (*(*wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).get_shmem_state()
+    }
+}
+
+extern "C" fn start_streaming(wp: *mut WalProposer, startpos: XLogRecPtr) {
+    unsafe {
+        let callback_data = (*(*wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).start_streaming(startpos)
+    }
+}
+
+extern "C" fn get_flush_rec_ptr(wp: *mut WalProposer) -> XLogRecPtr {
+    unsafe {
+        let callback_data = (*(*wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).get_flush_rec_ptr()
+    }
+}
+
+extern "C" fn get_current_timestamp(wp: *mut WalProposer) -> TimestampTz {
+    unsafe {
+        let callback_data = (*(*wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).get_current_timestamp()
+    }
+}
+
+extern "C" fn conn_error_message(sk: *mut Safekeeper) -> *mut ::std::os::raw::c_char {
+    unsafe {
+        let callback_data = (*(*(*sk).wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        let msg = (*api).conn_error_message(&mut (*sk));
+        let msg = CString::new(msg).unwrap();
+        // TODO: fix leaking error message
+        msg.into_raw()
+    }
+}
+
+extern "C" fn conn_status(sk: *mut Safekeeper) -> WalProposerConnStatusType {
+    unsafe {
+        let callback_data = (*(*(*sk).wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).conn_status(&mut (*sk))
+    }
+}
+
+extern "C" fn conn_connect_start(sk: *mut Safekeeper) {
+    unsafe {
+        let callback_data = (*(*(*sk).wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).conn_connect_start(&mut (*sk))
+    }
+}
+
+extern "C" fn conn_connect_poll(sk: *mut Safekeeper) -> WalProposerConnectPollStatusType {
+    unsafe {
+        let callback_data = (*(*(*sk).wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).conn_connect_poll(&mut (*sk))
+    }
+}
+
+extern "C" fn conn_send_query(sk: *mut Safekeeper, query: *mut ::std::os::raw::c_char) -> bool {
+    let query = unsafe { CStr::from_ptr(query) };
+    let query = query.to_str().unwrap();
+
+    unsafe {
+        let callback_data = (*(*(*sk).wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).conn_send_query(&mut (*sk), query)
+    }
+}
+
+extern "C" fn conn_get_query_result(sk: *mut Safekeeper) -> WalProposerExecStatusType {
+    unsafe {
+        let callback_data = (*(*(*sk).wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).conn_get_query_result(&mut (*sk))
+    }
+}
+
+extern "C" fn conn_flush(sk: *mut Safekeeper) -> ::std::os::raw::c_int {
+    unsafe {
+        let callback_data = (*(*(*sk).wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).conn_flush(&mut (*sk))
+    }
+}
+
+extern "C" fn conn_finish(sk: *mut Safekeeper) {
+    unsafe {
+        let callback_data = (*(*(*sk).wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).conn_finish(&mut (*sk))
+    }
+}
+
+extern "C" fn conn_async_read(
+    sk: *mut Safekeeper,
+    buf: *mut *mut ::std::os::raw::c_char,
+    amount: *mut ::std::os::raw::c_int,
+) -> PGAsyncReadResult {
+    unsafe {
+        let callback_data = (*(*(*sk).wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        let (res, result) = (*api).conn_async_read(&mut (*sk));
+
+        // This function has guarantee that returned buf will be valid until
+        // the next call. So we can store a Vec in each Safekeeper and reuse
+        // it on the next call.
+        let mut inbuf = take_vec_u8(&mut (*sk).inbuf).unwrap_or_default();
+
+        inbuf.clear();
+        inbuf.extend_from_slice(res);
+
+        // Put a Vec back to sk->inbuf and return data ptr.
+        *buf = store_vec_u8(&mut (*sk).inbuf, inbuf);
+        *amount = res.len() as i32;
+
+        result
+    }
+}
+
+extern "C" fn conn_async_write(
+    sk: *mut Safekeeper,
+    buf: *const ::std::os::raw::c_void,
+    size: usize,
+) -> PGAsyncWriteResult {
+    unsafe {
+        let buf = std::slice::from_raw_parts(buf as *const u8, size);
+        let callback_data = (*(*(*sk).wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).conn_async_write(&mut (*sk), buf)
+    }
+}
+
+extern "C" fn conn_blocking_write(
+    sk: *mut Safekeeper,
+    buf: *const ::std::os::raw::c_void,
+    size: usize,
+) -> bool {
+    unsafe {
+        let buf = std::slice::from_raw_parts(buf as *const u8, size);
+        let callback_data = (*(*(*sk).wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).conn_blocking_write(&mut (*sk), buf)
+    }
+}
+
+extern "C" fn recovery_download(
+    sk: *mut Safekeeper,
+    _timeline: TimeLineID,
+    startpos: XLogRecPtr,
+    endpos: XLogRecPtr,
+) -> bool {
+    unsafe {
+        let callback_data = (*(*(*sk).wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).recovery_download(&mut (*sk), startpos, endpos)
+    }
+}
+
+extern "C" fn wal_read(
+    sk: *mut Safekeeper,
+    buf: *mut ::std::os::raw::c_char,
+    startptr: XLogRecPtr,
+    count: Size,
+) {
+    unsafe {
+        let buf = std::slice::from_raw_parts_mut(buf as *mut u8, count);
+        let callback_data = (*(*(*sk).wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).wal_read(&mut (*sk), buf, startptr)
+    }
+}
+
+extern "C" fn wal_reader_allocate(sk: *mut Safekeeper) {
+    unsafe {
+        let callback_data = (*(*(*sk).wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).wal_reader_allocate(&mut (*sk));
+    }
+}
+
+extern "C" fn free_event_set(wp: *mut WalProposer) {
+    unsafe {
+        let callback_data = (*(*wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).free_event_set(&mut (*wp));
+    }
+}
+
+extern "C" fn init_event_set(wp: *mut WalProposer) {
+    unsafe {
+        let callback_data = (*(*wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).init_event_set(&mut (*wp));
+    }
+}
+
+extern "C" fn update_event_set(sk: *mut Safekeeper, events: uint32) {
+    unsafe {
+        let callback_data = (*(*(*sk).wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).update_event_set(&mut (*sk), events);
+    }
+}
+
+extern "C" fn add_safekeeper_event_set(sk: *mut Safekeeper, events: uint32) {
+    unsafe {
+        let callback_data = (*(*(*sk).wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).add_safekeeper_event_set(&mut (*sk), events);
+    }
+}
+
+extern "C" fn wait_event_set(
+    wp: *mut WalProposer,
+    timeout: ::std::os::raw::c_long,
+    event_sk: *mut *mut Safekeeper,
+    events: *mut uint32,
+) -> ::std::os::raw::c_int {
+    unsafe {
+        let callback_data = (*(*wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        let result = (*api).wait_event_set(&mut (*wp), timeout);
+        match result {
+            WaitResult::Latch => {
+                *event_sk = std::ptr::null_mut();
+                *events = crate::bindings::WL_LATCH_SET;
+                1
+            }
+            WaitResult::Timeout => {
+                *event_sk = std::ptr::null_mut();
+                *events = crate::bindings::WL_TIMEOUT;
+                0
+            }
+            WaitResult::Network(sk, event_mask) => {
+                *event_sk = sk;
+                *events = event_mask;
+                1
+            }
+        }
+    }
+}
+
+extern "C" fn strong_random(
+    wp: *mut WalProposer,
+    buf: *mut ::std::os::raw::c_void,
+    len: usize,
+) -> bool {
+    unsafe {
+        let buf = std::slice::from_raw_parts_mut(buf as *mut u8, len);
+        let callback_data = (*(*wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).strong_random(buf)
+    }
+}
+
+extern "C" fn get_redo_start_lsn(wp: *mut WalProposer) -> XLogRecPtr {
+    unsafe {
+        let callback_data = (*(*wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).get_redo_start_lsn()
+    }
+}
+
+extern "C" fn finish_sync_safekeepers(wp: *mut WalProposer, lsn: XLogRecPtr) {
+    unsafe {
+        let callback_data = (*(*wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).finish_sync_safekeepers(lsn)
+    }
+}
+
+extern "C" fn process_safekeeper_feedback(wp: *mut WalProposer, commit_lsn: XLogRecPtr) {
+    unsafe {
+        let callback_data = (*(*wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).process_safekeeper_feedback(&mut (*wp), commit_lsn)
+    }
+}
+
+extern "C" fn confirm_wal_streamed(wp: *mut WalProposer, lsn: XLogRecPtr) {
+    unsafe {
+        let callback_data = (*(*wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).confirm_wal_streamed(&mut (*wp), lsn)
+    }
+}
+
+extern "C" fn log_internal(
+    wp: *mut WalProposer,
+    level: ::std::os::raw::c_int,
+    line: *const ::std::os::raw::c_char,
+) {
+    unsafe {
+        let callback_data = (*(*wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        let line = CStr::from_ptr(line);
+        let line = line.to_str().unwrap();
+        (*api).log_internal(&mut (*wp), Level::from(level as u32), line)
+    }
+}
+
+extern "C" fn after_election(wp: *mut WalProposer) {
+    unsafe {
+        let callback_data = (*(*wp).config).callback_data;
+        let api = callback_data as *mut Box<dyn ApiImpl>;
+        (*api).after_election(&mut (*wp))
+    }
+}
+
+#[derive(Debug)]
+pub enum Level {
+    Debug5,
+    Debug4,
+    Debug3,
+    Debug2,
+    Debug1,
+    Log,
+    Info,
+    Notice,
+    Warning,
+    Error,
+    Fatal,
+    Panic,
+    WPEvent,
+}
+
+impl Level {
+    pub fn from(elevel: u32) -> Level {
+        use crate::bindings::*;
+
+        match elevel {
+            DEBUG5 => Level::Debug5,
+            DEBUG4 => Level::Debug4,
+            DEBUG3 => Level::Debug3,
+            DEBUG2 => Level::Debug2,
+            DEBUG1 => Level::Debug1,
+            LOG => Level::Log,
+            INFO => Level::Info,
+            NOTICE => Level::Notice,
+            WARNING => Level::Warning,
+            ERROR => Level::Error,
+            FATAL => Level::Fatal,
+            PANIC => Level::Panic,
+            WPEVENT => Level::WPEvent,
+            _ => panic!("unknown log level {}", elevel),
+        }
+    }
+}
+
+pub(crate) fn create_api() -> walproposer_api {
+    walproposer_api {
+        get_shmem_state: Some(get_shmem_state),
+        start_streaming: Some(start_streaming),
+        get_flush_rec_ptr: Some(get_flush_rec_ptr),
+        get_current_timestamp: Some(get_current_timestamp),
+        conn_error_message: Some(conn_error_message),
+        conn_status: Some(conn_status),
+        conn_connect_start: Some(conn_connect_start),
+        conn_connect_poll: Some(conn_connect_poll),
+        conn_send_query: Some(conn_send_query),
+        conn_get_query_result: Some(conn_get_query_result),
+        conn_flush: Some(conn_flush),
+        conn_finish: Some(conn_finish),
+        conn_async_read: Some(conn_async_read),
+        conn_async_write: Some(conn_async_write),
+        conn_blocking_write: Some(conn_blocking_write),
+        recovery_download: Some(recovery_download),
+        wal_read: Some(wal_read),
+        wal_reader_allocate: Some(wal_reader_allocate),
+        free_event_set: Some(free_event_set),
+        init_event_set: Some(init_event_set),
+        update_event_set: Some(update_event_set),
+        add_safekeeper_event_set: Some(add_safekeeper_event_set),
+        wait_event_set: Some(wait_event_set),
+        strong_random: Some(strong_random),
+        get_redo_start_lsn: Some(get_redo_start_lsn),
+        finish_sync_safekeepers: Some(finish_sync_safekeepers),
+        process_safekeeper_feedback: Some(process_safekeeper_feedback),
+        confirm_wal_streamed: Some(confirm_wal_streamed),
+        log_internal: Some(log_internal),
+        after_election: Some(after_election),
+    }
+}
+
+impl std::fmt::Display for Level {
+    fn fmt(&self, f: &mut std::fmt::Formatter) -> std::fmt::Result {
+        write!(f, "{:?}", self)
+    }
+}
+
+/// Take ownership of `Vec<u8>` from StringInfoData.
+pub(crate) fn take_vec_u8(pg: &mut StringInfoData) -> Option<Vec<u8>> {
+    if pg.data.is_null() {
+        return None;
+    }
+
+    let ptr = pg.data as *mut u8;
+    let length = pg.len as usize;
+    let capacity = pg.maxlen as usize;
+
+    pg.data = std::ptr::null_mut();
+    pg.len = 0;
+    pg.maxlen = 0;
+
+    unsafe { Some(Vec::from_raw_parts(ptr, length, capacity)) }
+}
+
+/// Store `Vec<u8>` in StringInfoData.
+fn store_vec_u8(pg: &mut StringInfoData, vec: Vec<u8>) -> *mut ::std::os::raw::c_char {
+    let ptr = vec.as_ptr() as *mut ::std::os::raw::c_char;
+    let length = vec.len();
+    let capacity = vec.capacity();
+
+    assert!(pg.data.is_null());
+
+    pg.data = ptr;
+    pg.len = length as i32;
+    pg.maxlen = capacity as i32;
+
+    std::mem::forget(vec);
+
+    ptr
+}

libs/walproposer/src/lib.rs

@@ -0,0 +1,14 @@
+pub mod bindings {
+    #![allow(non_upper_case_globals)]
+    #![allow(non_camel_case_types)]
+    #![allow(non_snake_case)]
+    // bindgen creates some unsafe code with no doc comments.
+    #![allow(clippy::missing_safety_doc)]
+    // noted at 1.63 that in many cases there's a u32 -> u32 transmutes in bindgen code.
+    #![allow(clippy::useless_transmute)]
+
+    include!(concat!(env!("OUT_DIR"), "/bindings.rs"));
+}
+
+pub mod api_bindings;
+pub mod walproposer;

libs/walproposer/src/walproposer.rs

@@ -0,0 +1,485 @@
+use std::ffi::CString;
+
+use postgres_ffi::WAL_SEGMENT_SIZE;
+use utils::id::TenantTimelineId;
+
+use crate::{
+    api_bindings::{create_api, take_vec_u8, Level},
+    bindings::{
+        Safekeeper, WalProposer, WalProposerConfig, WalProposerCreate, WalProposerFree,
+        WalProposerStart,
+    },
+};
+
+/// Rust high-level wrapper for C walproposer API. Many methods are not required
+/// for simple cases, hence todo!() in default implementations.
+///
+/// Refer to `pgxn/neon/walproposer.h` for documentation.
+pub trait ApiImpl {
+    fn get_shmem_state(&self) -> &mut crate::bindings::WalproposerShmemState {
+        todo!()
+    }
+
+    fn start_streaming(&self, _startpos: u64) {
+        todo!()
+    }
+
+    fn get_flush_rec_ptr(&self) -> u64 {
+        todo!()
+    }
+
+    fn get_current_timestamp(&self) -> i64 {
+        todo!()
+    }
+
+    fn conn_error_message(&self, _sk: &mut Safekeeper) -> String {
+        todo!()
+    }
+
+    fn conn_status(&self, _sk: &mut Safekeeper) -> crate::bindings::WalProposerConnStatusType {
+        todo!()
+    }
+
+    fn conn_connect_start(&self, _sk: &mut Safekeeper) {
+        todo!()
+    }
+
+    fn conn_connect_poll(
+        &self,
+        _sk: &mut Safekeeper,
+    ) -> crate::bindings::WalProposerConnectPollStatusType {
+        todo!()
+    }
+
+    fn conn_send_query(&self, _sk: &mut Safekeeper, _query: &str) -> bool {
+        todo!()
+    }
+
+    fn conn_get_query_result(
+        &self,
+        _sk: &mut Safekeeper,
+    ) -> crate::bindings::WalProposerExecStatusType {
+        todo!()
+    }
+
+    fn conn_flush(&self, _sk: &mut Safekeeper) -> i32 {
+        todo!()
+    }
+
+    fn conn_finish(&self, _sk: &mut Safekeeper) {
+        todo!()
+    }
+
+    fn conn_async_read(&self, _sk: &mut Safekeeper) -> (&[u8], crate::bindings::PGAsyncReadResult) {
+        todo!()
+    }
+
+    fn conn_async_write(
+        &self,
+        _sk: &mut Safekeeper,
+        _buf: &[u8],
+    ) -> crate::bindings::PGAsyncWriteResult {
+        todo!()
+    }
+
+    fn conn_blocking_write(&self, _sk: &mut Safekeeper, _buf: &[u8]) -> bool {
+        todo!()
+    }
+
+    fn recovery_download(&self, _sk: &mut Safekeeper, _startpos: u64, _endpos: u64) -> bool {
+        todo!()
+    }
+
+    fn wal_read(&self, _sk: &mut Safekeeper, _buf: &mut [u8], _startpos: u64) {
+        todo!()
+    }
+
+    fn wal_reader_allocate(&self, _sk: &mut Safekeeper) {
+        todo!()
+    }
+
+    fn free_event_set(&self, _wp: &mut WalProposer) {
+        todo!()
+    }
+
+    fn init_event_set(&self, _wp: &mut WalProposer) {
+        todo!()
+    }
+
+    fn update_event_set(&self, _sk: &mut Safekeeper, _events_mask: u32) {
+        todo!()
+    }
+
+    fn add_safekeeper_event_set(&self, _sk: &mut Safekeeper, _events_mask: u32) {
+        todo!()
+    }
+
+    fn wait_event_set(&self, _wp: &mut WalProposer, _timeout_millis: i64) -> WaitResult {
+        todo!()
+    }
+
+    fn strong_random(&self, _buf: &mut [u8]) -> bool {
+        todo!()
+    }
+
+    fn get_redo_start_lsn(&self) -> u64 {
+        todo!()
+    }
+
+    fn finish_sync_safekeepers(&self, _lsn: u64) {
+        todo!()
+    }
+
+    fn process_safekeeper_feedback(&self, _wp: &mut WalProposer, _commit_lsn: u64) {
+        todo!()
+    }
+
+    fn confirm_wal_streamed(&self, _wp: &mut WalProposer, _lsn: u64) {
+        todo!()
+    }
+
+    fn log_internal(&self, _wp: &mut WalProposer, _level: Level, _msg: &str) {
+        todo!()
+    }
+
+    fn after_election(&self, _wp: &mut WalProposer) {
+        todo!()
+    }
+}
+
+pub enum WaitResult {
+    Latch,
+    Timeout,
+    Network(*mut Safekeeper, u32),
+}
+
+pub struct Config {
+    /// Tenant and timeline id
+    pub ttid: TenantTimelineId,
+    /// List of safekeepers in format `host:port`
+    pub safekeepers_list: Vec<String>,
+    /// Safekeeper reconnect timeout in milliseconds
+    pub safekeeper_reconnect_timeout: i32,
+    /// Safekeeper connection timeout in milliseconds
+    pub safekeeper_connection_timeout: i32,
+    /// walproposer mode, finish when all safekeepers are synced or subscribe
+    /// to WAL streaming
+    pub sync_safekeepers: bool,
+}
+
+/// WalProposer main struct. C methods are reexported as Rust functions.
+pub struct Wrapper {
+    wp: *mut WalProposer,
+    _safekeepers_list_vec: Vec<u8>,
+}
+
+impl Wrapper {
+    pub fn new(api: Box<dyn ApiImpl>, config: Config) -> Wrapper {
+        let neon_tenant = CString::new(config.ttid.tenant_id.to_string())
+            .unwrap()
+            .into_raw();
+        let neon_timeline = CString::new(config.ttid.timeline_id.to_string())
+            .unwrap()
+            .into_raw();
+
+        let mut safekeepers_list_vec = CString::new(config.safekeepers_list.join(","))
+            .unwrap()
+            .into_bytes_with_nul();
+        assert!(safekeepers_list_vec.len() == safekeepers_list_vec.capacity());
+        let safekeepers_list = safekeepers_list_vec.as_mut_ptr() as *mut i8;
+
+        let callback_data = Box::into_raw(Box::new(api)) as *mut ::std::os::raw::c_void;
+
+        let c_config = WalProposerConfig {
+            neon_tenant,
+            neon_timeline,
+            safekeepers_list,
+            safekeeper_reconnect_timeout: config.safekeeper_reconnect_timeout,
+            safekeeper_connection_timeout: config.safekeeper_connection_timeout,
+            wal_segment_size: WAL_SEGMENT_SIZE as i32, // default 16MB
+            syncSafekeepers: config.sync_safekeepers,
+            systemId: 0,
+            pgTimeline: 1,
+            callback_data,
+        };
+        let c_config = Box::into_raw(Box::new(c_config));
+
+        let api = create_api();
+        let wp = unsafe { WalProposerCreate(c_config, api) };
+        Wrapper {
+            wp,
+            _safekeepers_list_vec: safekeepers_list_vec,
+        }
+    }
+
+    pub fn start(&self) {
+        unsafe { WalProposerStart(self.wp) }
+    }
+}
+
+impl Drop for Wrapper {
+    fn drop(&mut self) {
+        unsafe {
+            let config = (*self.wp).config;
+            drop(Box::from_raw(
+                (*config).callback_data as *mut Box<dyn ApiImpl>,
+            ));
+            drop(CString::from_raw((*config).neon_tenant));
+            drop(CString::from_raw((*config).neon_timeline));
+            drop(Box::from_raw(config));
+
+            for i in 0..(*self.wp).n_safekeepers {
+                let sk = &mut (*self.wp).safekeeper[i as usize];
+                take_vec_u8(&mut sk.inbuf);
+            }
+
+            WalProposerFree(self.wp);
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::{
+        cell::Cell,
+        sync::{atomic::AtomicUsize, mpsc::sync_channel},
+    };
+
+    use utils::id::TenantTimelineId;
+
+    use crate::{api_bindings::Level, walproposer::Wrapper};
+
+    use super::ApiImpl;
+
+    #[derive(Clone, Copy, Debug)]
+    struct WaitEventsData {
+        sk: *mut crate::bindings::Safekeeper,
+        event_mask: u32,
+    }
+
+    struct MockImpl {
+        // data to return from wait_event_set
+        wait_events: Cell<WaitEventsData>,
+        // walproposer->safekeeper messages
+        expected_messages: Vec<Vec<u8>>,
+        expected_ptr: AtomicUsize,
+        // safekeeper->walproposer messages
+        safekeeper_replies: Vec<Vec<u8>>,
+        replies_ptr: AtomicUsize,
+        // channel to send LSN to the main thread
+        sync_channel: std::sync::mpsc::SyncSender<u64>,
+    }
+
+    impl MockImpl {
+        fn check_walproposer_msg(&self, msg: &[u8]) {
+            let ptr = self
+                .expected_ptr
+                .fetch_add(1, std::sync::atomic::Ordering::SeqCst);
+
+            if ptr >= self.expected_messages.len() {
+                panic!("unexpected message from walproposer");
+            }
+
+            let expected_msg = &self.expected_messages[ptr];
+            assert_eq!(msg, expected_msg.as_slice());
+        }
+
+        fn next_safekeeper_reply(&self) -> &[u8] {
+            let ptr = self
+                .replies_ptr
+                .fetch_add(1, std::sync::atomic::Ordering::SeqCst);
+
+            if ptr >= self.safekeeper_replies.len() {
+                panic!("no more safekeeper replies");
+            }
+
+            &self.safekeeper_replies[ptr]
+        }
+    }
+
+    impl ApiImpl for MockImpl {
+        fn get_current_timestamp(&self) -> i64 {
+            println!("get_current_timestamp");
+            0
+        }
+
+        fn conn_status(
+            &self,
+            _: &mut crate::bindings::Safekeeper,
+        ) -> crate::bindings::WalProposerConnStatusType {
+            println!("conn_status");
+            crate::bindings::WalProposerConnStatusType_WP_CONNECTION_OK
+        }
+
+        fn conn_connect_start(&self, _: &mut crate::bindings::Safekeeper) {
+            println!("conn_connect_start");
+        }
+
+        fn conn_connect_poll(
+            &self,
+            _: &mut crate::bindings::Safekeeper,
+        ) -> crate::bindings::WalProposerConnectPollStatusType {
+            println!("conn_connect_poll");
+            crate::bindings::WalProposerConnectPollStatusType_WP_CONN_POLLING_OK
+        }
+
+        fn conn_send_query(&self, _: &mut crate::bindings::Safekeeper, query: &str) -> bool {
+            println!("conn_send_query: {}", query);
+            true
+        }
+
+        fn conn_get_query_result(
+            &self,
+            _: &mut crate::bindings::Safekeeper,
+        ) -> crate::bindings::WalProposerExecStatusType {
+            println!("conn_get_query_result");
+            crate::bindings::WalProposerExecStatusType_WP_EXEC_SUCCESS_COPYBOTH
+        }
+
+        fn conn_async_read(
+            &self,
+            _: &mut crate::bindings::Safekeeper,
+        ) -> (&[u8], crate::bindings::PGAsyncReadResult) {
+            println!("conn_async_read");
+            let reply = self.next_safekeeper_reply();
+            println!("conn_async_read result: {:?}", reply);
+            (
+                reply,
+                crate::bindings::PGAsyncReadResult_PG_ASYNC_READ_SUCCESS,
+            )
+        }
+
+        fn conn_blocking_write(&self, _: &mut crate::bindings::Safekeeper, buf: &[u8]) -> bool {
+            println!("conn_blocking_write: {:?}", buf);
+            self.check_walproposer_msg(buf);
+            true
+        }
+
+        fn wal_reader_allocate(&self, _: &mut crate::bindings::Safekeeper) {
+            println!("wal_reader_allocate")
+        }
+
+        fn free_event_set(&self, _: &mut crate::bindings::WalProposer) {
+            println!("free_event_set")
+        }
+
+        fn init_event_set(&self, _: &mut crate::bindings::WalProposer) {
+            println!("init_event_set")
+        }
+
+        fn update_event_set(&self, sk: &mut crate::bindings::Safekeeper, event_mask: u32) {
+            println!(
+                "update_event_set, sk={:?}, events_mask={:#b}",
+                sk as *mut crate::bindings::Safekeeper, event_mask
+            );
+            self.wait_events.set(WaitEventsData { sk, event_mask });
+        }
+
+        fn add_safekeeper_event_set(&self, sk: &mut crate::bindings::Safekeeper, event_mask: u32) {
+            println!(
+                "add_safekeeper_event_set, sk={:?}, events_mask={:#b}",
+                sk as *mut crate::bindings::Safekeeper, event_mask
+            );
+            self.wait_events.set(WaitEventsData { sk, event_mask });
+        }
+
+        fn wait_event_set(
+            &self,
+            _: &mut crate::bindings::WalProposer,
+            timeout_millis: i64,
+        ) -> super::WaitResult {
+            let data = self.wait_events.get();
+            println!(
+                "wait_event_set, timeout_millis={}, res={:?}",
+                timeout_millis, data
+            );
+            super::WaitResult::Network(data.sk, data.event_mask)
+        }
+
+        fn strong_random(&self, buf: &mut [u8]) -> bool {
+            println!("strong_random");
+            buf.fill(0);
+            true
+        }
+
+        fn finish_sync_safekeepers(&self, lsn: u64) {
+            self.sync_channel.send(lsn).unwrap();
+            panic!("sync safekeepers finished at lsn={}", lsn);
+        }
+
+        fn log_internal(&self, _wp: &mut crate::bindings::WalProposer, level: Level, msg: &str) {
+            println!("walprop_log[{}] {}", level, msg);
+        }
+
+        fn after_election(&self, _wp: &mut crate::bindings::WalProposer) {
+            println!("after_election");
+        }
+    }
+
+    /// Test that walproposer can successfully connect to safekeeper and finish
+    /// sync_safekeepers. API is mocked in MockImpl.
+    ///
+    /// Run this test with valgrind to detect leaks:
+    /// `valgrind --leak-check=full target/debug/deps/walproposer-<build>`
+    #[test]
+    fn test_simple_sync_safekeepers() -> anyhow::Result<()> {
+        let ttid = TenantTimelineId::new(
+            "9e4c8f36063c6c6e93bc20d65a820f3d".parse()?,
+            "9e4c8f36063c6c6e93bc20d65a820f3d".parse()?,
+        );
+
+        let (sender, receiver) = sync_channel(1);
+
+        let my_impl: Box<dyn ApiImpl> = Box::new(MockImpl {
+            wait_events: Cell::new(WaitEventsData {
+                sk: std::ptr::null_mut(),
+                event_mask: 0,
+            }),
+            expected_messages: vec![
+                // Greeting(ProposerGreeting { protocol_version: 2, pg_version: 160000, proposer_id: [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], system_id: 0, timeline_id: 9e4c8f36063c6c6e93bc20d65a820f3d, tenant_id: 9e4c8f36063c6c6e93bc20d65a820f3d, tli: 1, wal_seg_size: 16777216 })
+                vec![
+                    103, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 113, 2, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+                    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 158, 76, 143, 54, 6, 60, 108, 110,
+                    147, 188, 32, 214, 90, 130, 15, 61, 158, 76, 143, 54, 6, 60, 108, 110, 147,
+                    188, 32, 214, 90, 130, 15, 61, 1, 0, 0, 0, 0, 0, 0, 1,
+                ],
+                // VoteRequest(VoteRequest { term: 3 })
+                vec![
+                    118, 0, 0, 0, 0, 0, 0, 0, 3, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+                    0, 0, 0, 0, 0, 0,
+                ],
+            ],
+            expected_ptr: AtomicUsize::new(0),
+            safekeeper_replies: vec![
+                // Greeting(AcceptorGreeting { term: 2, node_id: NodeId(1) })
+                vec![
+                    103, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0,
+                ],
+                // VoteResponse(VoteResponse { term: 3, vote_given: 1, flush_lsn: 0/539, truncate_lsn: 0/539, term_history: [(2, 0/539)], timeline_start_lsn: 0/539 })
+                vec![
+                    118, 0, 0, 0, 0, 0, 0, 0, 3, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 57,
+                    5, 0, 0, 0, 0, 0, 0, 57, 5, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 2, 0, 0, 0, 0, 0, 0,
+                    0, 57, 5, 0, 0, 0, 0, 0, 0, 57, 5, 0, 0, 0, 0, 0, 0,
+                ],
+            ],
+            replies_ptr: AtomicUsize::new(0),
+            sync_channel: sender,
+        });
+        let config = crate::walproposer::Config {
+            ttid,
+            safekeepers_list: vec!["localhost:5000".to_string()],
+            safekeeper_reconnect_timeout: 1000,
+            safekeeper_connection_timeout: 10000,
+            sync_safekeepers: true,
+        };
+
+        let wp = Wrapper::new(my_impl, config);
+
+        // walproposer will panic when it finishes sync_safekeepers
+        std::panic::catch_unwind(|| wp.start()).unwrap_err();
+        // validate the resulting LSN
+        assert_eq!(receiver.recv()?, 1337);
+        Ok(())
+        // drop() will free up resources here
+    }
+}

pageserver/Cargo.toml

@@ -68,6 +68,7 @@ url.workspace = true
 walkdir.workspace = true
 metrics.workspace = true
 pageserver_api.workspace = true
+pageserver_compaction.workspace = true
 postgres_connection.workspace = true
 postgres_ffi.workspace = true
 pq_proto.workspace = true

pageserver/benches/bench_walredo.rs

@@ -11,10 +11,7 @@ use std::sync::{Arc, Barrier};
 
 use bytes::{Buf, Bytes};
 use pageserver::{
-    config::PageServerConf,
-    repository::Key,
-    walrecord::NeonWalRecord,
-    walredo::{PostgresRedoManager, WalRedoError},
+    config::PageServerConf, repository::Key, walrecord::NeonWalRecord, walredo::PostgresRedoManager,
 };
 use utils::{id::TenantId, lsn::Lsn};
 
@@ -35,9 +32,15 @@ fn redo_scenarios(c: &mut Criterion) {
 
     let manager = Arc::new(manager);
 
-    tracing::info!("executing first");
-    short().execute(&manager).unwrap();
-    tracing::info!("first executed");
+    {
+        let rt = tokio::runtime::Builder::new_current_thread()
+            .enable_all()
+            .build()
+            .unwrap();
+        tracing::info!("executing first");
+        short().execute(rt.handle(), &manager).unwrap();
+        tracing::info!("first executed");
+    }
 
     let thread_counts = [1, 2, 4, 8, 16];
 
@@ -80,9 +83,14 @@ fn add_multithreaded_walredo_requesters(
     assert_ne!(threads, 0);
 
     if threads == 1 {
+        let rt = tokio::runtime::Builder::new_current_thread()
+            .enable_all()
+            .build()
+            .unwrap();
+        let handle = rt.handle();
         b.iter_batched_ref(
             || Some(input_factory()),
-            |input| execute_all(input.take(), manager),
+            |input| execute_all(input.take(), handle, manager),
             criterion::BatchSize::PerIteration,
         );
     } else {
@@ -98,19 +106,26 @@ fn add_multithreaded_walredo_requesters(
                     let manager = manager.clone();
                     let barrier = barrier.clone();
                     let work_rx = work_rx.clone();
-                    move || loop {
-                        // queue up and wait if we want to go another round
-                        if work_rx.lock().unwrap().recv().is_err() {
-                            break;
+                    move || {
+                        let rt = tokio::runtime::Builder::new_current_thread()
+                            .enable_all()
+                            .build()
+                            .unwrap();
+                        let handle = rt.handle();
+                        loop {
+                            // queue up and wait if we want to go another round
+                            if work_rx.lock().unwrap().recv().is_err() {
+                                break;
+                            }
+
+                            let input = Some(input_factory());
+
+                            barrier.wait();
+
+                            execute_all(input, handle, &manager).unwrap();
+
+                            barrier.wait();
                         }
-
-                        let input = Some(input_factory());
-
-                        barrier.wait();
-
-                        execute_all(input, &manager).unwrap();
-
-                        barrier.wait();
                     }
                 })
             })
@@ -152,15 +167,19 @@ impl Drop for JoinOnDrop {
     }
 }
 
-fn execute_all<I>(input: I, manager: &PostgresRedoManager) -> Result<(), WalRedoError>
+fn execute_all<I>(
+    input: I,
+    handle: &tokio::runtime::Handle,
+    manager: &PostgresRedoManager,
+) -> anyhow::Result<()>
 where
     I: IntoIterator<Item = Request>,
 {
     // just fire all requests as fast as possible
     input.into_iter().try_for_each(|req| {
-        let page = req.execute(manager)?;
+        let page = req.execute(handle, manager)?;
         assert_eq!(page.remaining(), 8192);
-        Ok::<_, WalRedoError>(())
+        anyhow::Ok(())
     })
 }
 
@@ -473,9 +492,11 @@ struct Request {
 }
 
 impl Request {
-    fn execute(self, manager: &PostgresRedoManager) -> Result<Bytes, WalRedoError> {
-        use pageserver::walredo::WalRedoManager;
-
+    fn execute(
+        self,
+        rt: &tokio::runtime::Handle,
+        manager: &PostgresRedoManager,
+    ) -> anyhow::Result<Bytes> {
         let Request {
             key,
             lsn,
@@ -484,6 +505,6 @@ impl Request {
             pg_version,
         } = self;
 
-        manager.request_redo(key, lsn, base_img, records, pg_version)
+        rt.block_on(manager.request_redo(key, lsn, base_img, records, pg_version))
     }
 }

pageserver/compaction/Cargo.toml

@@ -0,0 +1,53 @@
+[package]
+name = "pageserver_compaction"
+version = "0.1.0"
+edition.workspace = true
+license.workspace = true
+
+[features]
+default = []
+
+[dependencies]
+anyhow.workspace = true
+async-compression.workspace = true
+async-stream.workspace = true
+async-trait.workspace = true
+byteorder.workspace = true
+bytes.workspace = true
+chrono = { workspace = true, features = ["serde"] }
+clap = { workspace = true, features = ["string"] }
+const_format.workspace = true
+consumption_metrics.workspace = true
+crossbeam-utils.workspace = true
+either.workspace = true
+flate2.workspace = true
+fail.workspace = true
+futures.workspace = true
+git-version.workspace = true
+hex.workspace = true
+humantime.workspace = true
+humantime-serde.workspace = true
+itertools.workspace = true
+once_cell.workspace = true
+pin-project-lite.workspace = true
+rand.workspace = true
+smallvec = { workspace = true, features = ["write"] }
+svg_fmt.workspace = true
+sync_wrapper.workspace = true
+thiserror.workspace = true
+tokio = { workspace = true, features = ["process", "sync", "fs", "rt", "io-util", "time"] }
+tokio-io-timeout.workspace = true
+tokio-util.workspace = true
+tracing.workspace = true
+tracing-error.workspace = true
+tracing-subscriber.workspace = true
+url.workspace = true
+walkdir.workspace = true
+metrics.workspace = true
+utils.workspace = true
+workspace_hack.workspace = true
+
+[dev-dependencies]
+criterion.workspace = true
+hex-literal.workspace = true
+tokio = { workspace = true, features = ["process", "sync", "fs", "rt", "io-util", "time", "test-util"] }

pageserver/compaction/TODO.md

@@ -0,0 +1,49 @@
+# TODO
+
+- If the key space can be perfectly partitioned at some key, perform planning on each
+  partition separately. For example, if we are compacting a level with layers like this:
+
+              :
+  +--+ +----+ :  +------+
+  |  | |    | :  |      |
+  +--+ +----+ :  +------+
+              :
+  +-----+ +-+ : +--------+
+  |     | | | : |        |
+  +-----+ +-+ : +--------+
+              :
+
+  At the dotted line, there is a natural split in the key space, such that all
+  layers are either on the left or the right of it. We can compact the
+  partitions separately.  We could choose to create image layers for one
+  partition but not the other one, for example.
+
+- All the layers don't have to be exactly the same size, we can choose to cut a
+  layer short or stretch it a little larger than the target size, if it helps
+  the overall system. We can help perfect partitions (see previous bullet point)
+  to happen more frequently, by choosing the cut points wisely. For example, try
+  to cut layers at boundaries of underlying image layers. And "snap to grid",
+  i.e. don't cut layers at any key, but e.g. only when key % 10000 = 0.
+
+- Avoid rewriting layers when we'd just create an identical layer to an input
+  layer.
+
+- Parallelism. The code is already split up into planning and execution, so that
+  we first split up the compaction work into "Jobs", and then execute them.
+  It would be straightforward to execute multiple jobs in parallel.
+
+- Materialize extra pages in delta layers during compaction. This would reduce
+  read amplification. There has been the idea of partial image layers. Materializing
+  extra pages in the delta layers achieve the same goal, without introducing a new
+  concept.
+
+## Simulator
+
+- Expand the simulator for more workloads
+- Automate a test suite that runs the simluator with different workloads and
+  spits out a table of results
+- Model read amplification
+- More sanity checking. One idea is to keep a reference count of each
+  MockRecord, i.e. use Arc<MockRecord> instead of plain MockRecord, and panic if
+  a MockRecord that is newer than PITR horizon is completely dropped. That would
+  indicate that the record was lost.

pageserver/compaction/src/bin/compaction-simulator.rs

@@ -0,0 +1,214 @@
+use clap::{Parser, Subcommand};
+use pageserver_compaction::simulator::MockTimeline;
+use rand::Rng;
+use std::io::Write;
+use std::path::{Path, PathBuf};
+use std::sync::OnceLock;
+
+use utils::project_git_version;
+
+project_git_version!(GIT_VERSION);
+
+#[derive(Parser)]
+#[command(
+    version = GIT_VERSION,
+    about = "Neon Pageserver compaction simulator",
+    long_about = "A developer tool to visualize and test compaction"
+)]
+#[command(propagate_version = true)]
+struct CliOpts {
+    #[command(subcommand)]
+    command: Commands,
+}
+
+#[derive(Subcommand)]
+enum Commands {
+    RunSuite,
+    Simulate(SimulateCmd),
+}
+
+#[derive(Clone, clap::ValueEnum)]
+enum Distribution {
+    Uniform,
+    HotCold,
+}
+
+/// Read and update pageserver metadata file
+#[derive(Parser)]
+struct SimulateCmd {
+    distribution: Distribution,
+
+    /// Number of records to digest
+    num_records: u64,
+    /// Record length
+    record_len: u64,
+
+    // Logical database size in MB
+    logical_size: u64,
+}
+
+async fn simulate(cmd: &SimulateCmd, results_path: &Path) -> anyhow::Result<()> {
+    let mut executor = MockTimeline::new();
+
+    // Convert the logical size in MB into a key range.
+    let key_range = 0..((cmd.logical_size * 1024 * 1024) / 8192);
+    //let key_range = u64::MIN..u64::MAX;
+    println!(
+        "starting simulation with key range {:016X}-{:016X}",
+        key_range.start, key_range.end
+    );
+
+    // helper function to print progress indicator
+    let print_progress = |i| -> anyhow::Result<()> {
+        if i == 0 || (i + 1) % 10000 == 0 || i == cmd.num_records - 1 {
+            print!(
+                "\ringested {} / {} records, {} MiB / {} MiB...",
+                i + 1,
+                cmd.num_records,
+                (i + 1) * cmd.record_len / (1_000_000),
+                cmd.num_records * cmd.record_len / (1_000_000),
+            );
+            std::io::stdout().flush()?;
+        }
+        Ok(())
+    };
+
+    match cmd.distribution {
+        Distribution::Uniform => {
+            for i in 0..cmd.num_records {
+                executor.ingest_uniform(1, cmd.record_len, &key_range)?;
+                executor.compact_if_needed().await?;
+
+                print_progress(i)?;
+            }
+        }
+        Distribution::HotCold => {
+            let splitpoint = key_range.end / 10;
+            let hot_key_range = 0..splitpoint;
+            let cold_key_range = splitpoint..key_range.end;
+
+            for i in 0..cmd.num_records {
+                let chosen_range = if rand::thread_rng().gen_bool(0.9) {
+                    &hot_key_range
+                } else {
+                    &cold_key_range
+                };
+                executor.ingest_uniform(1, cmd.record_len, chosen_range)?;
+                executor.compact_if_needed().await?;
+
+                print_progress(i)?;
+            }
+        }
+    }
+    println!("done!");
+    executor.flush_l0();
+    executor.compact_if_needed().await?;
+    let stats = executor.print_stats()?;
+
+    // Print the stats to stdout, and also to a file
+    print!("{}", stats);
+    std::fs::write(results_path.join("stats.txt"), stats)?;
+
+    let animation_path = results_path.join("compaction-animation.html");
+    executor.draw_history(std::fs::File::create(&animation_path)?)?;
+    println!(
+        "animation: file://{}",
+        animation_path.canonicalize()?.display()
+    );
+
+    Ok(())
+}
+
+async fn run_suite_cmd(results_path: &Path, workload: &SimulateCmd) -> anyhow::Result<()> {
+    std::fs::create_dir(results_path)?;
+
+    set_log_file(File::create(results_path.join("log"))?);
+    let result = simulate(workload, results_path).await;
+    set_log_stdout();
+    result
+}
+
+async fn run_suite() -> anyhow::Result<()> {
+    let top_results_path = PathBuf::from(format!(
+        "compaction-suite-results.{}",
+        std::time::SystemTime::UNIX_EPOCH.elapsed()?.as_secs()
+    ));
+    std::fs::create_dir(&top_results_path)?;
+
+    let workload = SimulateCmd {
+        distribution: Distribution::Uniform,
+        // Generate 20 GB of WAL
+        record_len: 1_000,
+        num_records: 20_000_000,
+        // Logical size 5 GB
+        logical_size: 5_000,
+    };
+
+    run_suite_cmd(&top_results_path.join("uniform-20GB-5GB"), &workload).await?;
+
+    println!(
+        "All tests finished. Results in {}",
+        top_results_path.display()
+    );
+    Ok(())
+}
+
+use std::fs::File;
+use std::io::Stdout;
+use std::sync::Mutex;
+use tracing_subscriber::fmt::writer::EitherWriter;
+use tracing_subscriber::fmt::MakeWriter;
+
+static LOG_FILE: OnceLock<Mutex<EitherWriter<File, Stdout>>> = OnceLock::new();
+fn get_log_output() -> &'static Mutex<EitherWriter<File, Stdout>> {
+    LOG_FILE.get_or_init(|| std::sync::Mutex::new(EitherWriter::B(std::io::stdout())))
+}
+
+fn set_log_file(f: File) {
+    *get_log_output().lock().unwrap() = EitherWriter::A(f);
+}
+
+fn set_log_stdout() {
+    *get_log_output().lock().unwrap() = EitherWriter::B(std::io::stdout());
+}
+
+fn init_logging() -> anyhow::Result<()> {
+    // We fall back to printing all spans at info-level or above if
+    // the RUST_LOG environment variable is not set.
+    let rust_log_env_filter = || {
+        tracing_subscriber::EnvFilter::try_from_default_env()
+            .unwrap_or_else(|_| tracing_subscriber::EnvFilter::new("info"))
+    };
+
+    // NB: the order of the with() calls does not matter.
+    // See https://docs.rs/tracing-subscriber/0.3.16/tracing_subscriber/layer/index.html#per-layer-filtering
+    use tracing_subscriber::prelude::*;
+    tracing_subscriber::registry()
+        .with({
+            let log_layer = tracing_subscriber::fmt::layer()
+                .with_target(false)
+                .with_ansi(false)
+                .with_writer(|| get_log_output().make_writer());
+            log_layer.with_filter(rust_log_env_filter())
+        })
+        .init();
+
+    Ok(())
+}
+
+#[tokio::main]
+async fn main() -> anyhow::Result<()> {
+    let cli = CliOpts::parse();
+
+    init_logging()?;
+
+    match cli.command {
+        Commands::Simulate(cmd) => {
+            simulate(&cmd, &PathBuf::from("/tmp/compactions.html")).await?;
+        }
+        Commands::RunSuite => {
+            run_suite().await?;
+        }
+    };
+    Ok(())
+}

pageserver/compaction/src/compact_tiered.rs

@@ -0,0 +1,870 @@
+//! # Tiered compaction algorithm.
+//!
+//! Read all the input delta files, and write a new set of delta files that
+//! include all the input WAL records. See retile_deltas().
+//!
+//! In a "normal" LSM tree, you get to remove any values that are overwritten by
+//! later values, but in our system, we keep all the history. So the reshuffling
+//! doesn't remove any garbage, it just reshuffles the records to reduce read
+//! amplification, i.e. the number of files that you need to access to find the
+//! WAL records for a given key.
+//!
+//! If the new delta files would be very "narrow", i.e. each file would cover
+//! only a narrow key range, then we create a new set of image files
+//! instead. The current threshold is that if the estimated total size of the
+//! image layers is smaller than the size of the deltas, then we create image
+//! layers. That amounts to 2x storage amplification, and it means that the
+//! distance of image layers in LSN dimension is roughly equal to the logical
+//! database size. For example, if the logical database size is 10 GB, we would
+//! generate new image layers every 10 GB of WAL.
+//!
+use futures::StreamExt;
+use tracing::{debug, info};
+
+use std::collections::{HashSet, VecDeque};
+use std::ops::Range;
+
+use crate::helpers::{accum_key_values, keyspace_total_size, merge_delta_keys, overlaps_with};
+use crate::interface::*;
+use utils::lsn::Lsn;
+
+use crate::identify_levels::identify_level;
+
+/// Main entry point to compaction.
+///
+/// The starting point is a cutoff LSN (`end_lsn`). The compaction is run on
+/// everything below that point, that needs compaction. The cutoff LSN must
+/// partition the layers so that there are no layers that span across that
+/// LSN. To start compaction at the top of the tree, pass the end LSN of the
+/// written last L0 layer.
+pub async fn compact_tiered<E: CompactionJobExecutor>(
+    executor: &mut E,
+    end_lsn: Lsn,
+    target_file_size: u64,
+    fanout: u64,
+    ctx: &E::RequestContext,
+) -> anyhow::Result<()> {
+    assert!(fanout >= 2);
+    // Start at L0
+    let mut current_level_no = 0;
+    let mut current_level_target_height = target_file_size;
+    loop {
+        // end LSN +1 to include possible image layers exactly at 'end_lsn'.
+        let all_layers = executor
+            .get_layers(
+                &(E::Key::MIN..E::Key::MAX),
+                &(Lsn(u64::MIN)..end_lsn + 1),
+                ctx,
+            )
+            .await?;
+        info!(
+            "Compacting L{}, total # of layers: {}",
+            current_level_no,
+            all_layers.len()
+        );
+
+        // Identify the range of LSNs that belong to this level. We assume that
+        // each file in this level span an LSN range up to 1.75x target file
+        // size. That should give us enough slop that if we created a slightly
+        // oversized L0 layer, e.g. because flushing the in-memory layer was
+        // delayed for some reason, we don't consider the oversized layer to
+        // belong to L1. But not too much slop, that we don't accidentally
+        // "skip" levels.
+        let max_height = (current_level_target_height as f64 * 1.75) as u64;
+        let Some(level) = identify_level(all_layers, end_lsn, max_height).await? else {
+            break;
+        };
+
+        // Calculate the height of this level. If the # of tiers exceeds the
+        // fanout parameter, it's time to compact it.
+        let depth = level.depth();
+        info!(
+            "Level {} identified as LSN range {}-{}: depth {}",
+            current_level_no, level.lsn_range.start, level.lsn_range.end, depth
+        );
+        for l in &level.layers {
+            debug!("LEVEL {} layer: {}", current_level_no, l.short_id());
+        }
+        if depth < fanout {
+            debug!(
+                level = current_level_no,
+                depth = depth,
+                fanout,
+                "too few deltas to compact"
+            );
+            break;
+        }
+
+        compact_level(
+            &level.lsn_range,
+            &level.layers,
+            executor,
+            target_file_size,
+            ctx,
+        )
+        .await?;
+        if target_file_size == u64::MAX {
+            break;
+        }
+        current_level_no += 1;
+        current_level_target_height = current_level_target_height.saturating_mul(fanout);
+    }
+    Ok(())
+}
+
+async fn compact_level<E: CompactionJobExecutor>(
+    lsn_range: &Range<Lsn>,
+    layers: &[E::Layer],
+    executor: &mut E,
+    target_file_size: u64,
+    ctx: &E::RequestContext,
+) -> anyhow::Result<bool> {
+    let mut layer_fragments = Vec::new();
+    for l in layers {
+        layer_fragments.push(LayerFragment::new(l.clone()));
+    }
+
+    let mut state = LevelCompactionState {
+        target_file_size,
+        _lsn_range: lsn_range.clone(),
+        layers: layer_fragments,
+        jobs: Vec::new(),
+        job_queue: Vec::new(),
+        next_level: false,
+        executor,
+    };
+
+    let first_job = CompactionJob {
+        key_range: E::Key::MIN..E::Key::MAX,
+        lsn_range: lsn_range.clone(),
+        strategy: CompactionStrategy::Divide,
+        input_layers: state
+            .layers
+            .iter()
+            .enumerate()
+            .map(|i| LayerId(i.0))
+            .collect(),
+        completed: false,
+    };
+
+    state.jobs.push(first_job);
+    state.job_queue.push(JobId(0));
+    state.execute(ctx).await?;
+
+    info!(
+        "compaction completed! Need to process next level: {}",
+        state.next_level
+    );
+
+    Ok(state.next_level)
+}
+
+/// Blackboard that keeps track of the state of all the jobs and work remaining
+struct LevelCompactionState<'a, E>
+where
+    E: CompactionJobExecutor,
+{
+    // parameters
+    target_file_size: u64,
+
+    _lsn_range: Range<Lsn>,
+    layers: Vec<LayerFragment<E>>,
+
+    // job queue
+    jobs: Vec<CompactionJob<E>>,
+    job_queue: Vec<JobId>,
+
+    /// If false, no need to compact levels below this
+    next_level: bool,
+
+    /// Interface to the outside world
+    executor: &'a mut E,
+}
+
+#[derive(Debug, Clone, Copy, Hash, PartialEq, Eq)]
+struct LayerId(usize);
+#[derive(Debug, Clone, Copy, Hash, PartialEq, Eq)]
+struct JobId(usize);
+
+struct PendingJobSet {
+    pending: HashSet<JobId>,
+    completed: HashSet<JobId>,
+}
+
+impl PendingJobSet {
+    fn new() -> Self {
+        PendingJobSet {
+            pending: HashSet::new(),
+            completed: HashSet::new(),
+        }
+    }
+
+    fn complete_job(&mut self, job_id: JobId) {
+        self.pending.remove(&job_id);
+        self.completed.insert(job_id);
+    }
+
+    fn all_completed(&self) -> bool {
+        self.pending.is_empty()
+    }
+}
+
+// When we decide to rewrite a set of layers, LayerFragment is used to keep
+// track which new layers supersede an old layer. When all the stakeholder jobs
+// have completed, this layer can be deleted.
+struct LayerFragment<E>
+where
+    E: CompactionJobExecutor,
+{
+    layer: E::Layer,
+
+    // If we will write new layers to replace this one, this keeps track of the
+    // jobs that need to complete before this layer can be deleted. As the jobs
+    // complete, they are moved from 'pending' to 'completed' set. Once the
+    // 'pending' set becomes empty, the layer can be deleted.
+    //
+    // If None, this layer is not rewritten and must not be deleted.
+    deletable_after: Option<PendingJobSet>,
+
+    deleted: bool,
+}
+
+impl<E> LayerFragment<E>
+where
+    E: CompactionJobExecutor,
+{
+    fn new(layer: E::Layer) -> Self {
+        LayerFragment {
+            layer,
+            deletable_after: None,
+            deleted: false,
+        }
+    }
+}
+
+#[derive(PartialEq)]
+enum CompactionStrategy {
+    Divide,
+    CreateDelta,
+    CreateImage,
+}
+
+#[allow(dead_code)] // Todo
+struct CompactionJob<E: CompactionJobExecutor> {
+    key_range: Range<E::Key>,
+    lsn_range: Range<Lsn>,
+
+    strategy: CompactionStrategy,
+
+    input_layers: Vec<LayerId>,
+
+    completed: bool,
+}
+
+impl<'a, E> LevelCompactionState<'a, E>
+where
+    E: CompactionJobExecutor,
+{
+    /// Main loop of the executor.
+    ///
+    /// In each iteration, we take the next job from the queue, and execute it.
+    /// The execution might add new jobs to the queue. Keep going until the
+    /// queue is empty.
+    ///
+    /// Initially, the job queue consists of one Divide job over the whole
+    /// level. On first call, it is divided into smaller jobs.
+    ///
+    async fn execute(&mut self, ctx: &E::RequestContext) -> anyhow::Result<()> {
+        // TODO: this would be pretty straightforward to parallelize with FuturesUnordered
+        while let Some(next_job_id) = self.job_queue.pop() {
+            info!("executing job {}", next_job_id.0);
+            self.execute_job(next_job_id, ctx).await?;
+        }
+
+        // all done!
+        Ok(())
+    }
+
+    async fn execute_job(&mut self, job_id: JobId, ctx: &E::RequestContext) -> anyhow::Result<()> {
+        let job = &self.jobs[job_id.0];
+        match job.strategy {
+            CompactionStrategy::Divide => {
+                self.divide_job(job_id, ctx).await?;
+                Ok(())
+            }
+            CompactionStrategy::CreateDelta => {
+                let mut deltas: Vec<E::DeltaLayer> = Vec::new();
+                let mut layer_ids: Vec<LayerId> = Vec::new();
+                for layer_id in &job.input_layers {
+                    let layer = &self.layers[layer_id.0].layer;
+                    if let Some(dl) = self.executor.downcast_delta_layer(layer).await? {
+                        deltas.push(dl.clone());
+                        layer_ids.push(*layer_id);
+                    }
+                }
+
+                self.executor
+                    .create_delta(&job.lsn_range, &job.key_range, &deltas, ctx)
+                    .await?;
+                self.jobs[job_id.0].completed = true;
+
+                // did we complete any fragments?
+                for layer_id in layer_ids {
+                    let l = &mut self.layers[layer_id.0];
+                    if let Some(deletable_after) = l.deletable_after.as_mut() {
+                        deletable_after.complete_job(job_id);
+                        if deletable_after.all_completed() {
+                            self.executor.delete_layer(&l.layer, ctx).await?;
+                            l.deleted = true;
+                        }
+                    }
+                }
+
+                self.next_level = true;
+
+                Ok(())
+            }
+            CompactionStrategy::CreateImage => {
+                self.executor
+                    .create_image(job.lsn_range.end, &job.key_range, ctx)
+                    .await?;
+                self.jobs[job_id.0].completed = true;
+
+                // TODO: we could check if any layers < PITR horizon became deletable
+                Ok(())
+            }
+        }
+    }
+
+    fn push_job(&mut self, job: CompactionJob<E>) -> JobId {
+        let job_id = JobId(self.jobs.len());
+        self.jobs.push(job);
+        self.job_queue.push(job_id);
+        job_id
+    }
+
+    ///
+    /// Take a partition of the key space, and decide how to compact it.
+    ///
+    /// TODO: Currently, this is called exactly once for the level, and we
+    /// decide whether to create new image layers to cover the whole level, or
+    /// write a new set of delta. In the future, this should try to partition
+    /// the key space, and make the decision separately for each partition.
+    ///
+    async fn divide_job(&mut self, job_id: JobId, ctx: &E::RequestContext) -> anyhow::Result<()> {
+        let job = &self.jobs[job_id.0];
+        assert!(job.strategy == CompactionStrategy::Divide);
+
+        // Check for dummy cases
+        if job.input_layers.is_empty() {
+            return Ok(());
+        }
+
+        let job = &self.jobs[job_id.0];
+        assert!(job.strategy == CompactionStrategy::Divide);
+
+        // Would it be better to create images for this partition?
+        // Decide based on the average density of the level
+        let keyspace_size = keyspace_total_size(
+            &self
+                .executor
+                .get_keyspace(&job.key_range, job.lsn_range.end, ctx)
+                .await?,
+        ) * 8192;
+
+        let wal_size = job
+            .input_layers
+            .iter()
+            .filter(|layer_id| self.layers[layer_id.0].layer.is_delta())
+            .map(|layer_id| self.layers[layer_id.0].layer.file_size())
+            .sum::<u64>();
+        if keyspace_size < wal_size {
+            // seems worth it
+            info!(
+                "covering with images, because keyspace_size is {}, size of deltas between {}-{} is {}",
+                keyspace_size, job.lsn_range.start, job.lsn_range.end, wal_size
+            );
+            self.cover_with_images(job_id, ctx).await
+        } else {
+            // do deltas
+            info!(
+                "coverage not worth it, keyspace_size {}, wal_size {}",
+                keyspace_size, wal_size
+            );
+            self.retile_deltas(job_id, ctx).await
+        }
+    }
+
+    // LSN
+    //  ^
+    //  |
+    //  |                          ###|###|#####
+    //  | +--+-----+--+            +--+-----+--+
+    //  | |  |     |  |            |  |     |  |
+    //  | +--+--+--+--+            +--+--+--+--+
+    //  | |     |     |            |     |     |
+    //  | +---+-+-+---+     ==>    +---+-+-+---+
+    //  | |   |   |   |            |   |   |   |
+    //  | +---+-+-++--+            +---+-+-++--+
+    //  | |     |  |  |            |     |  |  |
+    //  | +-----+--+--+            +-----+--+--+
+    //  |
+    //  +--------------> key
+    //
+    async fn cover_with_images(
+        &mut self,
+        job_id: JobId,
+        ctx: &E::RequestContext,
+    ) -> anyhow::Result<()> {
+        let job = &self.jobs[job_id.0];
+        assert!(job.strategy == CompactionStrategy::Divide);
+
+        // XXX: do we still need the "holes" stuff?
+
+        let mut new_jobs = Vec::new();
+
+        // Slide a window through the keyspace
+        let keyspace = self
+            .executor
+            .get_keyspace(&job.key_range, job.lsn_range.end, ctx)
+            .await?;
+
+        let mut window = KeyspaceWindow::new(
+            E::Key::MIN..E::Key::MAX,
+            keyspace,
+            self.target_file_size / 8192,
+        );
+        while let Some(key_range) = window.choose_next_image() {
+            new_jobs.push(CompactionJob::<E> {
+                key_range,
+                lsn_range: job.lsn_range.clone(),
+                strategy: CompactionStrategy::CreateImage,
+                input_layers: Vec::new(), // XXX: Is it OK for  this to be empty for image layer?
+                completed: false,
+            });
+        }
+
+        for j in new_jobs.into_iter().rev() {
+            let _job_id = self.push_job(j);
+
+            // TODO: image layers don't let us delete anything. unless < PITR horizon
+            //let j = &self.jobs[job_id.0];
+            // for layer_id in j.input_layers.iter() {
+            //    self.layers[layer_id.0].pending_stakeholders.insert(job_id);
+            //}
+        }
+
+        Ok(())
+    }
+
+    // Merge the contents of all the input delta layers into a new set
+    // of delta layers, based on the current partitioning.
+    //
+    // We split the new delta layers on the key dimension. We iterate through
+    // the key space, and for each key, check if including the next key to the
+    // current output layer we're building would cause the layer to become too
+    // large. If so, dump the current output layer and start new one.  It's
+    // possible that there is a single key with so many page versions that
+    // storing all of them in a single layer file would be too large. In that
+    // case, we also split on the LSN dimension.
+    //
+    // LSN
+    //  ^
+    //  |
+    //  | +-----------+            +--+--+--+--+
+    //  | |           |            |  |  |  |  |
+    //  | +-----------+            |  |  |  |  |
+    //  | |           |            |  |  |  |  |
+    //  | +-----------+     ==>    |  |  |  |  |
+    //  | |           |            |  |  |  |  |
+    //  | +-----------+            |  |  |  |  |
+    //  | |           |            |  |  |  |  |
+    //  | +-----------+            +--+--+--+--+
+    //  |
+    //  +--------------> key
+    //
+    //
+    // If one key (X) has a lot of page versions:
+    //
+    // LSN
+    //  ^
+    //  |                                 (X)
+    //  | +-----------+            +--+--+--+--+
+    //  | |           |            |  |  |  |  |
+    //  | +-----------+            |  |  +--+  |
+    //  | |           |            |  |  |  |  |
+    //  | +-----------+     ==>    |  |  |  |  |
+    //  | |           |            |  |  +--+  |
+    //  | +-----------+            |  |  |  |  |
+    //  | |           |            |  |  |  |  |
+    //  | +-----------+            +--+--+--+--+
+    //  |
+    //  +--------------> key
+    //
+    // TODO: this actually divides the layers into fixed-size chunks, not
+    // based on the partitioning.
+    //
+    // TODO: we should also opportunistically materialize and
+    // garbage collect what we can.
+    async fn retile_deltas(
+        &mut self,
+        job_id: JobId,
+        ctx: &E::RequestContext,
+    ) -> anyhow::Result<()> {
+        let job = &self.jobs[job_id.0];
+        assert!(job.strategy == CompactionStrategy::Divide);
+
+        // Sweep the key space left to right, running an estimate of how much
+        // disk size and keyspace we have accumulated
+        //
+        // Once the disk size reaches the target threshold, stop and think.
+        // If we have accumulated only a narrow band of keyspace, create an
+        // image layer. Otherwise write a delta layer.
+
+        // FIXME: deal with the case of lots of values for same key
+
+        // FIXME: we are ignoring images here. Did we already divide the work
+        // so that we won't encounter them here?
+
+        let mut deltas: Vec<E::DeltaLayer> = Vec::new();
+        for layer_id in &job.input_layers {
+            let l = &self.layers[layer_id.0];
+            if let Some(dl) = self.executor.downcast_delta_layer(&l.layer).await? {
+                deltas.push(dl.clone());
+            }
+        }
+        // Open stream
+        let key_value_stream = std::pin::pin!(merge_delta_keys::<E>(deltas.as_slice(), ctx));
+        let mut new_jobs = Vec::new();
+
+        // Slide a window through the keyspace
+        let mut key_accum = std::pin::pin!(accum_key_values(key_value_stream));
+        let mut all_in_window: bool = false;
+        let mut window = Window::new();
+        loop {
+            if all_in_window && window.elems.is_empty() {
+                // All done!
+                break;
+            }
+            if let Some(key_range) = window.choose_next_delta(self.target_file_size, !all_in_window)
+            {
+                let batch_layers: Vec<LayerId> = job
+                    .input_layers
+                    .iter()
+                    .filter(|layer_id| {
+                        overlaps_with(self.layers[layer_id.0].layer.key_range(), &key_range)
+                    })
+                    .cloned()
+                    .collect();
+                assert!(!batch_layers.is_empty());
+                new_jobs.push(CompactionJob {
+                    key_range,
+                    lsn_range: job.lsn_range.clone(),
+                    strategy: CompactionStrategy::CreateDelta,
+                    input_layers: batch_layers,
+                    completed: false,
+                });
+            } else {
+                assert!(!all_in_window);
+                if let Some(next_key) = key_accum.next().await.transpose()? {
+                    window.feed(next_key.key, next_key.size);
+                } else {
+                    all_in_window = true;
+                }
+            }
+        }
+
+        // All the input files are rewritten. Set up the tracking for when they can
+        // be deleted.
+        for layer_id in job.input_layers.iter() {
+            let l = &mut self.layers[layer_id.0];
+            assert!(l.deletable_after.is_none());
+            l.deletable_after = Some(PendingJobSet::new());
+        }
+        for j in new_jobs.into_iter().rev() {
+            let job_id = self.push_job(j);
+            let j = &self.jobs[job_id.0];
+            for layer_id in j.input_layers.iter() {
+                self.layers[layer_id.0]
+                    .deletable_after
+                    .as_mut()
+                    .unwrap()
+                    .pending
+                    .insert(job_id);
+            }
+        }
+
+        Ok(())
+    }
+}
+
+// Sliding window through keyspace and values
+// This is used by over_with_images to decide on good split points
+struct KeyspaceWindow<K> {
+    head: KeyspaceWindowHead<K>,
+
+    start_pos: KeyspaceWindowPos<K>,
+}
+struct KeyspaceWindowHead<K> {
+    // overall key range to cover
+    key_range: Range<K>,
+
+    keyspace: Vec<Range<K>>,
+    target_keysize: u64,
+}
+
+#[derive(Clone)]
+struct KeyspaceWindowPos<K> {
+    end_key: K,
+
+    keyspace_idx: usize,
+
+    accum_keysize: u64,
+}
+impl<K: CompactionKey> KeyspaceWindowPos<K> {
+    fn reached_end(&self, w: &KeyspaceWindowHead<K>) -> bool {
+        self.keyspace_idx == w.keyspace.len()
+    }
+
+    // Advance the cursor until it reaches 'target_keysize'.
+    fn advance_until_size(&mut self, w: &KeyspaceWindowHead<K>, max_size: u64) {
+        while self.accum_keysize < max_size && !self.reached_end(w) {
+            let curr_range = &w.keyspace[self.keyspace_idx];
+            if self.end_key < curr_range.start {
+                // skip over any unused space
+                self.end_key = curr_range.start;
+            }
+
+            // We're now within 'curr_range'. Can we advance past it completely?
+            let distance = K::key_range_size(&(self.end_key..curr_range.end));
+            if (self.accum_keysize + distance as u64) < max_size {
+                // oh yeah, it fits
+                self.end_key = curr_range.end;
+                self.keyspace_idx += 1;
+                self.accum_keysize += distance as u64;
+            } else {
+                // advance within the range
+                let skip_key = self.end_key.skip_some();
+                let distance = K::key_range_size(&(self.end_key..skip_key));
+                if (self.accum_keysize + distance as u64) < max_size {
+                    self.end_key = skip_key;
+                    self.accum_keysize += distance as u64;
+                } else {
+                    self.end_key = self.end_key.next();
+                    self.accum_keysize += 1;
+                }
+            }
+        }
+    }
+}
+
+impl<K> KeyspaceWindow<K>
+where
+    K: CompactionKey,
+{
+    fn new(key_range: Range<K>, keyspace: CompactionKeySpace<K>, target_keysize: u64) -> Self {
+        assert!(keyspace.first().unwrap().start >= key_range.start);
+
+        let start_key = key_range.start;
+        let start_pos = KeyspaceWindowPos::<K> {
+            end_key: start_key,
+            keyspace_idx: 0,
+            accum_keysize: 0,
+        };
+        Self {
+            head: KeyspaceWindowHead::<K> {
+                key_range,
+                keyspace,
+                target_keysize,
+            },
+            start_pos,
+        }
+    }
+
+    fn choose_next_image(&mut self) -> Option<Range<K>> {
+        if self.start_pos.keyspace_idx == self.head.keyspace.len() {
+            // we've reached the end
+            return None;
+        }
+
+        let mut next_pos = self.start_pos.clone();
+        next_pos.advance_until_size(
+            &self.head,
+            self.start_pos.accum_keysize + self.head.target_keysize,
+        );
+
+        // See if we can gobble up the rest of the keyspace if we stretch out the layer, up to
+        // 1.25x target size
+        let mut end_pos = next_pos.clone();
+        end_pos.advance_until_size(
+            &self.head,
+            self.start_pos.accum_keysize + (self.head.target_keysize * 5 / 4),
+        );
+        if end_pos.reached_end(&self.head) {
+            // gobble up any unused keyspace between the last used key and end of the range
+            assert!(end_pos.end_key <= self.head.key_range.end);
+            end_pos.end_key = self.head.key_range.end;
+            next_pos = end_pos;
+        }
+
+        let start_key = self.start_pos.end_key;
+        self.start_pos = next_pos;
+        Some(start_key..self.start_pos.end_key)
+    }
+}
+
+// Sliding window through keyspace and values
+//
+// This is used to decide what layer to write next, from the beginning of the window.
+//
+// Candidates:
+//
+// 1. Create an image layer, snapping to previous images
+// 2. Create a delta layer, snapping to previous images
+// 3. Create an image layer, snapping to
+//
+//
+
+// Take previous partitioning, based on the image layers below.
+//
+// Candidate is at the front:
+//
+// Consider stretching an image layer to next divider? If it's close enough,
+// that's the image candidate
+//
+// If it's too far, consider splitting at a reasonable point
+//
+// Is the image candidate smaller than the equivalent delta? If so,
+// split off the image. Otherwise, split off one delta.
+// Try to snap off the delta at a reasonable point
+
+struct WindowElement<K> {
+    start_key: K, // inclusive
+    last_key: K,  // inclusive
+    accum_size: u64,
+}
+struct Window<K> {
+    elems: VecDeque<WindowElement<K>>,
+
+    // last key that was split off, inclusive
+    splitoff_key: Option<K>,
+    splitoff_size: u64,
+}
+
+impl<K> Window<K>
+where
+    K: CompactionKey,
+{
+    fn new() -> Self {
+        Self {
+            elems: VecDeque::new(),
+            splitoff_key: None,
+            splitoff_size: 0,
+        }
+    }
+
+    fn feed(&mut self, key: K, size: u64) {
+        let last_size;
+        if let Some(last) = self.elems.back_mut() {
+            assert!(last.last_key <= key);
+            if key == last.last_key {
+                last.accum_size += size;
+                return;
+            }
+            last_size = last.accum_size;
+        } else {
+            last_size = 0;
+        }
+        // This is a new key.
+        let elem = WindowElement {
+            start_key: key,
+            last_key: key,
+            accum_size: last_size + size,
+        };
+        self.elems.push_back(elem);
+    }
+
+    fn remain_size(&self) -> u64 {
+        self.elems.back().unwrap().accum_size - self.splitoff_size
+    }
+
+    fn peek_size(&self) -> u64 {
+        self.elems.front().unwrap().accum_size - self.splitoff_size
+    }
+
+    fn commit_upto(&mut self, mut upto: usize) {
+        while upto > 1 {
+            let popped = self.elems.pop_front().unwrap();
+            self.elems.front_mut().unwrap().start_key = popped.start_key;
+            upto -= 1;
+        }
+    }
+
+    fn find_size_split(&self, target_size: u64) -> usize {
+        self.elems
+            .partition_point(|elem| elem.accum_size - self.splitoff_size < target_size)
+    }
+
+    fn pop(&mut self) {
+        let first = self.elems.pop_front().unwrap();
+        self.splitoff_size = first.accum_size;
+
+        self.splitoff_key = Some(first.last_key);
+    }
+
+    // the difference between delta and image is that an image covers
+    // any unused keyspace before and after, while a delta tries to
+    // minimize that. TODO: difference not implemented
+    fn pop_delta(&mut self) -> Range<K> {
+        let first = self.elems.front().unwrap();
+        let key_range = first.start_key..first.last_key.next();
+
+        self.pop();
+        key_range
+    }
+
+    // Prerequisite: we have enough input in the window
+    //
+    // On return None, the caller should feed more data and call again
+    fn choose_next_delta(&mut self, target_size: u64, has_more: bool) -> Option<Range<K>> {
+        if has_more && self.elems.is_empty() {
+            // Starting up
+            return None;
+        }
+
+        // If we still have an undersized candidate, just keep going
+        while self.peek_size() < target_size {
+            if self.elems.len() > 1 {
+                self.commit_upto(2);
+            } else if has_more {
+                return None;
+            } else {
+                break;
+            }
+        }
+
+        // Ensure we have enough input in the window to make a good decision
+        if has_more && self.remain_size() < target_size * 5 / 4 {
+            return None;
+        }
+
+        // The candidate on the front is now large enough, for a delta.
+        // And we have enough data in the window to decide.
+
+        // If we're willing to stretch it up to 1.25 target size, could we
+        // gobble up the rest of the work? This avoids creating very small
+        // "tail" layers at the end of the keyspace
+        if !has_more && self.remain_size() < target_size * 5 / 3 {
+            self.commit_upto(self.elems.len());
+        } else {
+            let delta_split_at = self.find_size_split(target_size);
+            self.commit_upto(delta_split_at);
+
+            // If it's still not large enough, request the caller to fill the window
+            if self.elems.len() == 1 && has_more {
+                return None;
+            }
+        }
+        Some(self.pop_delta())
+    }
+}

pageserver/compaction/src/helpers.rs

@@ -0,0 +1,251 @@
+//! This file contains generic utility functions over the interface types,
+//! which could be handy for any compaction implementation.
+use crate::interface::*;
+
+use futures::future::BoxFuture;
+use futures::{Stream, StreamExt};
+use itertools::Itertools;
+use pin_project_lite::pin_project;
+use std::cmp::Ord;
+use std::collections::BinaryHeap;
+use std::collections::VecDeque;
+use std::future::Future;
+use std::ops::{DerefMut, Range};
+use std::pin::Pin;
+use std::task::Poll;
+
+pub fn keyspace_total_size<K>(keyspace: &CompactionKeySpace<K>) -> u64
+where
+    K: CompactionKey,
+{
+    let mut total = 0;
+    for r in keyspace.iter() {
+        total += K::key_range_size(r) as u64;
+    }
+    total
+}
+
+pub fn overlaps_with<T: Ord>(a: &Range<T>, b: &Range<T>) -> bool {
+    !(a.end <= b.start || b.end <= a.start)
+}
+
+pub fn union_to_keyspace<K: Ord>(a: &mut CompactionKeySpace<K>, b: CompactionKeySpace<K>) {
+    let x = std::mem::take(a);
+    let mut all_ranges_iter = [x.into_iter(), b.into_iter()]
+        .into_iter()
+        .kmerge_by(|a, b| a.start < b.start);
+    let mut ranges = Vec::new();
+    if let Some(first) = all_ranges_iter.next() {
+        let (mut start, mut end) = (first.start, first.end);
+
+        for r in all_ranges_iter {
+            assert!(r.start >= start);
+            if r.start > end {
+                ranges.push(start..end);
+                start = r.start;
+                end = r.end;
+            } else if r.end > end {
+                end = r.end;
+            }
+        }
+        ranges.push(start..end);
+    }
+    *a = ranges
+}
+
+pub fn intersect_keyspace<K: Ord + Clone + Copy>(
+    a: &CompactionKeySpace<K>,
+    r: &Range<K>,
+) -> CompactionKeySpace<K> {
+    let mut ranges: Vec<Range<K>> = Vec::new();
+
+    for x in a.iter() {
+        if x.end <= r.start {
+            continue;
+        }
+        if x.start >= r.end {
+            break;
+        }
+        ranges.push(x.clone())
+    }
+
+    // trim the ends
+    if let Some(first) = ranges.first_mut() {
+        first.start = std::cmp::max(first.start, r.start);
+    }
+    if let Some(last) = ranges.last_mut() {
+        last.end = std::cmp::min(last.end, r.end);
+    }
+    ranges
+}
+
+/// Create a stream that iterates through all DeltaEntrys among all input
+/// layers, in key-lsn order.
+///
+/// This is public because the create_delta() implementation likely wants to use this too
+/// TODO: move to a more shared place
+pub fn merge_delta_keys<'a, E: CompactionJobExecutor>(
+    layers: &'a [E::DeltaLayer],
+    ctx: &'a E::RequestContext,
+) -> MergeDeltaKeys<'a, E> {
+    // Use a binary heap to merge the layers. Each input layer is initially
+    // represented by a LazyLoadLayer::Unloaded element, which uses the start of
+    // the layer's key range as the key. The first time a layer reaches the top
+    // of the heap, all the keys of the layer are loaded into a sorted vector.
+    //
+    // This helps to keep the memory usage reasonable: we only need to hold in
+    // memory the DeltaEntrys of the layers that overlap with the "current" key.
+    let mut heap: BinaryHeap<LazyLoadLayer<'a, E>> = BinaryHeap::new();
+    for l in layers {
+        heap.push(LazyLoadLayer::Unloaded(l));
+    }
+    MergeDeltaKeys {
+        heap,
+        ctx,
+        load_future: None,
+    }
+}
+
+enum LazyLoadLayer<'a, E: CompactionJobExecutor> {
+    Loaded(VecDeque<<E::DeltaLayer as CompactionDeltaLayer<E>>::DeltaEntry<'a>>),
+    Unloaded(&'a E::DeltaLayer),
+}
+impl<'a, E: CompactionJobExecutor> LazyLoadLayer<'a, E> {
+    fn key(&self) -> E::Key {
+        match self {
+            Self::Loaded(entries) => entries.front().unwrap().key(),
+            Self::Unloaded(dl) => dl.key_range().start,
+        }
+    }
+}
+impl<'a, E: CompactionJobExecutor> PartialOrd for LazyLoadLayer<'a, E> {
+    fn partial_cmp(&self, other: &Self) -> Option<std::cmp::Ordering> {
+        // reverse order so that we get a min-heap
+        other.key().partial_cmp(&self.key())
+    }
+}
+impl<'a, E: CompactionJobExecutor> Ord for LazyLoadLayer<'a, E> {
+    fn cmp(&self, other: &Self) -> std::cmp::Ordering {
+        // reverse order so that we get a min-heap
+        other.key().cmp(&self.key())
+    }
+}
+impl<'a, E: CompactionJobExecutor> PartialEq for LazyLoadLayer<'a, E> {
+    fn eq(&self, other: &Self) -> bool {
+        self.key().eq(&other.key())
+    }
+}
+impl<'a, E: CompactionJobExecutor> Eq for LazyLoadLayer<'a, E> {}
+
+type LoadFuture<'a, E> = BoxFuture<'a, anyhow::Result<Vec<E>>>;
+
+// Stream returned by `merge_delta_keys`
+pin_project! {
+#[allow(clippy::type_complexity)]
+pub struct MergeDeltaKeys<'a, E: CompactionJobExecutor> {
+    heap: BinaryHeap<LazyLoadLayer<'a, E>>,
+
+    #[pin]
+    load_future: Option<LoadFuture<'a, <E::DeltaLayer as CompactionDeltaLayer<E>>::DeltaEntry<'a>>>,
+
+    ctx: &'a E::RequestContext,
+}
+}
+
+impl<'a, E> Stream for MergeDeltaKeys<'a, E>
+where
+    E: CompactionJobExecutor + 'a,
+{
+    type Item = anyhow::Result<<E::DeltaLayer as CompactionDeltaLayer<E>>::DeltaEntry<'a>>;
+
+    fn poll_next(
+        self: Pin<&mut Self>,
+        cx: &mut std::task::Context<'_>,
+    ) -> Poll<std::option::Option<<Self as futures::Stream>::Item>> {
+        let mut this = self.project();
+        loop {
+            if let Some(mut load_future) = this.load_future.as_mut().as_pin_mut() {
+                // We are waiting for loading the keys to finish
+                match load_future.as_mut().poll(cx) {
+                    Poll::Ready(Ok(entries)) => {
+                        this.load_future.set(None);
+                        *this.heap.peek_mut().unwrap() =
+                            LazyLoadLayer::Loaded(VecDeque::from(entries));
+                    }
+                    Poll::Ready(Err(e)) => {
+                        return Poll::Ready(Some(Err(e)));
+                    }
+                    Poll::Pending => {
+                        return Poll::Pending;
+                    }
+                }
+            }
+
+            // If the topmost layer in the heap hasn't been loaded yet, start
+            // loading it. Otherwise return the next entry from it and update
+            // the layer's position in the heap (this decreaseKey operation is
+            // performed implicitly when `top` is dropped).
+            if let Some(mut top) = this.heap.peek_mut() {
+                match top.deref_mut() {
+                    LazyLoadLayer::Unloaded(ref mut l) => {
+                        let fut = l.load_keys(this.ctx);
+                        this.load_future.set(Some(Box::pin(fut)));
+                        continue;
+                    }
+                    LazyLoadLayer::Loaded(ref mut entries) => {
+                        let result = entries.pop_front().unwrap();
+                        if entries.is_empty() {
+                            std::collections::binary_heap::PeekMut::pop(top);
+                        }
+                        return Poll::Ready(Some(Ok(result)));
+                    }
+                }
+            } else {
+                return Poll::Ready(None);
+            }
+        }
+    }
+}
+
+// Accumulate values at key boundaries
+pub struct KeySize<K> {
+    pub key: K,
+    pub num_values: u64,
+    pub size: u64,
+}
+
+pub fn accum_key_values<'a, I, K, D, E>(input: I) -> impl Stream<Item = Result<KeySize<K>, E>>
+where
+    K: Eq,
+    I: Stream<Item = Result<D, E>>,
+    D: CompactionDeltaEntry<'a, K>,
+{
+    async_stream::try_stream! {
+        // Initialize the state from the first value
+        let mut input = std::pin::pin!(input);
+
+        if let Some(first) = input.next().await {
+            let first = first?;
+            let mut accum: KeySize<K> = KeySize {
+                key: first.key(),
+                num_values: 1,
+                size: first.size(),
+            };
+            while let Some(this) = input.next().await {
+                let this = this?;
+                if this.key() == accum.key {
+                    accum.size += this.size();
+                    accum.num_values += 1;
+                } else {
+                    yield accum;
+                    accum = KeySize {
+                        key: this.key(),
+                        num_values: 1,
+                        size: this.size(),
+                    };
+                }
+            }
+            yield accum;
+        }
+    }
+}

pageserver/compaction/src/identify_levels.rs

@@ -0,0 +1,376 @@
+//! An LSM tree consists of multiple levels, each exponential larger than the
+//! previous level. And each level consists of be multiple "tiers". With tiered
+//! compaction, a level is compacted when it has accumulated more than N tiers,
+//! forming one tier on the next level.
+//!
+//! In the pageserver, we don't explicitly track the levels and tiers. Instead,
+//! we identify them by looking at the shapes of the layers. It's an easy task
+//! for a human, but it's not straightforward to come up with the exact
+//! rules. Especially if there are cases like interrupted, half-finished
+//! compactions, or highly skewed data distributions that have let us "skip"
+//! some levels. It's not critical to classify all cases correctly; at worst we
+//! delay some compaction work, and suffer from more read amplification, or we
+//! perform some unnecessary compaction work.
+//!
+//! `identify_level` performs that shape-matching.
+//!
+//! It returns a Level struct, which has `depth()` function to count the number
+//! of "tiers" in the level. The tier count is the max depth of stacked layers
+//! within the level. That's a good measure, because the point of compacting is
+//! to reduce read amplification, and the depth is what determines that.
+//!
+//! One interesting effect of this is that if we generate very small delta
+//! layers at L0, e.g. because the L0 layers are flushed by timeout rather than
+//! because they reach the target size, the L0 compaction will combine them to
+//! one larger file. But if the combined file is still smaller than the target
+//! file size, the file will still be considered to be part of L0 at the next
+//! iteration.
+
+use anyhow::bail;
+use std::collections::BTreeSet;
+use std::ops::Range;
+use utils::lsn::Lsn;
+
+use crate::interface::*;
+
+use tracing::{info, trace};
+
+pub struct Level<L> {
+    pub lsn_range: Range<Lsn>,
+    pub layers: Vec<L>,
+}
+
+/// Identify an LSN > `end_lsn` that partitions the LSN space, so that there are
+/// no layers that cross the boundary LSN.
+///
+/// A further restriction is that all layers in the returned partition cover at
+/// most 'lsn_max_size' LSN bytes.
+pub async fn identify_level<K, L>(
+    all_layers: Vec<L>,
+    end_lsn: Lsn,
+    lsn_max_size: u64,
+) -> anyhow::Result<Option<Level<L>>>
+where
+    K: CompactionKey,
+    L: CompactionLayer<K> + Clone,
+{
+    // filter out layers that are above the `end_lsn`, they are completely irrelevant.
+    let mut layers = Vec::new();
+    for l in all_layers {
+        if l.lsn_range().start < end_lsn && l.lsn_range().end > end_lsn {
+            // shouldn't happen. Indicates that the caller passed a bogus
+            // end_lsn.
+            bail!("identify_level() called with end_lsn that does not partition the LSN space: end_lsn {} intersects with layer {}", end_lsn, l.short_id());
+        }
+        // include image layers sitting exacty at `end_lsn`.
+        let is_image = !l.is_delta();
+        if (is_image && l.lsn_range().start > end_lsn)
+            || (!is_image && l.lsn_range().start >= end_lsn)
+        {
+            continue;
+        }
+        layers.push(l);
+    }
+    // All the remaining layers either belong to this level, or are below it.
+    info!(
+        "identify level at {}, size {}, num layers below: {}",
+        end_lsn,
+        lsn_max_size,
+        layers.len()
+    );
+    if layers.is_empty() {
+        return Ok(None);
+    }
+
+    // Walk the ranges in LSN order.
+    //
+    // ----- end_lsn
+    //  |
+    //  |
+    //  v
+    //
+    layers.sort_by_key(|l| l.lsn_range().end);
+    let mut candidate_start_lsn = end_lsn;
+    let mut candidate_layers: Vec<L> = Vec::new();
+    let mut current_best_start_lsn = end_lsn;
+    let mut current_best_layers: Vec<L> = Vec::new();
+    let mut iter = layers.into_iter();
+    loop {
+        let Some(l) = iter.next_back() else {
+            // Reached end. Accept the last candidate
+            current_best_start_lsn = candidate_start_lsn;
+            current_best_layers.extend_from_slice(&std::mem::take(&mut candidate_layers));
+            break;
+        };
+        trace!(
+            "inspecting {} for candidate {}, current best {}",
+            l.short_id(),
+            candidate_start_lsn,
+            current_best_start_lsn
+        );
+
+        let r = l.lsn_range();
+
+        // Image layers don't restrict our choice of cutoff LSN
+        if l.is_delta() {
+            // Is this candidate workable? In other words, are there any
+            // delta layers that span across this LSN
+            //
+            // Valid:                 Not valid:
+            //  +                     +
+            //  |                     | +
+            //  +  <- candidate       + |   <- candidate
+            //     +                    +
+            //     |
+            //     +
+            if r.end <= candidate_start_lsn {
+                // Hooray, there are no crossing LSNs. And we have visited
+                // through all the layers within candidate..end_lsn. The
+                // current candidate can be accepted.
+                current_best_start_lsn = r.end;
+                current_best_layers.extend_from_slice(&std::mem::take(&mut candidate_layers));
+                candidate_start_lsn = r.start;
+            }
+
+            // Is it small enough to be considered part of this level?
+            if r.end.0 - r.start.0 > lsn_max_size {
+                // Too large, this layer belongs to next level. Stop.
+                trace!(
+                    "too large {}, size {} vs {}",
+                    l.short_id(),
+                    r.end.0 - r.start.0,
+                    lsn_max_size
+                );
+                break;
+            }
+
+            // If this crosses the candidate lsn, push it down.
+            if r.start < candidate_start_lsn {
+                trace!(
+                    "layer {} prevents from stopping at {}",
+                    l.short_id(),
+                    candidate_start_lsn
+                );
+                candidate_start_lsn = r.start;
+            }
+        }
+
+        // Include this layer in our candidate
+        candidate_layers.push(l);
+    }
+
+    Ok(if current_best_start_lsn == end_lsn {
+        // empty level
+        None
+    } else {
+        Some(Level {
+            lsn_range: current_best_start_lsn..end_lsn,
+            layers: current_best_layers,
+        })
+    })
+}
+
+// helper struct used in depth()
+struct Event<K> {
+    key: K,
+    layer_idx: usize,
+    start: bool,
+}
+
+impl<L> Level<L> {
+    /// Count the number of deltas stacked on each other.
+    pub fn depth<K>(&self) -> u64
+    where
+        K: CompactionKey,
+        L: CompactionLayer<K>,
+    {
+        let mut events: Vec<Event<K>> = Vec::new();
+        for (idx, l) in self.layers.iter().enumerate() {
+            events.push(Event {
+                key: l.key_range().start,
+                layer_idx: idx,
+                start: true,
+            });
+            events.push(Event {
+                key: l.key_range().end,
+                layer_idx: idx,
+                start: false,
+            });
+        }
+        events.sort_by_key(|e| (e.key, e.start));
+
+        // Sweep the key space left to right. Stop at each distinct key, and
+        // count the number of deltas on top of the highest image at that key.
+        //
+        // This is a little enefficient, as we walk through the active_set on
+        // every key. We could increment/decrement a counter on each step
+        // instead, but that'd require a bit more complex bookkeeping.
+        let mut active_set: BTreeSet<(Lsn, bool, usize)> = BTreeSet::new();
+        let mut max_depth = 0;
+        let mut events_iter = events.iter().peekable();
+        while let Some(e) = events_iter.next() {
+            let l = &self.layers[e.layer_idx];
+            let is_image = !l.is_delta();
+
+            // update the active set
+            if e.start {
+                active_set.insert((l.lsn_range().end, is_image, e.layer_idx));
+            } else {
+                active_set.remove(&(l.lsn_range().end, is_image, e.layer_idx));
+            }
+
+            // recalculate depth if this was the last event at this point
+            let more_events_at_this_key = events_iter
+                .peek()
+                .map_or(false, |next_e| next_e.key == e.key);
+            if !more_events_at_this_key {
+                let mut active_depth = 0;
+                for (_end_lsn, is_image, _idx) in active_set.iter().rev() {
+                    if *is_image {
+                        break;
+                    }
+                    active_depth += 1;
+                }
+                if active_depth > max_depth {
+                    max_depth = active_depth;
+                }
+            }
+        }
+        max_depth
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::simulator::{Key, MockDeltaLayer, MockImageLayer, MockLayer};
+    use std::sync::{Arc, Mutex};
+
+    fn delta(key_range: Range<Key>, lsn_range: Range<Lsn>) -> MockLayer {
+        MockLayer::Delta(Arc::new(MockDeltaLayer {
+            key_range,
+            lsn_range,
+            // identify_level() doesn't pay attention to the rest of the fields
+            file_size: 0,
+            deleted: Mutex::new(false),
+            records: vec![],
+        }))
+    }
+
+    fn image(key_range: Range<Key>, lsn: Lsn) -> MockLayer {
+        MockLayer::Image(Arc::new(MockImageLayer {
+            key_range,
+            lsn_range: lsn..(lsn + 1),
+            // identify_level() doesn't pay attention to the rest of the fields
+            file_size: 0,
+            deleted: Mutex::new(false),
+        }))
+    }
+
+    #[tokio::test]
+    async fn test_identify_level() -> anyhow::Result<()> {
+        let layers = vec![
+            delta(Key::MIN..Key::MAX, Lsn(0x8000)..Lsn(0x9000)),
+            delta(Key::MIN..Key::MAX, Lsn(0x5000)..Lsn(0x7000)),
+            delta(Key::MIN..Key::MAX, Lsn(0x4000)..Lsn(0x5000)),
+            delta(Key::MIN..Key::MAX, Lsn(0x3000)..Lsn(0x4000)),
+            delta(Key::MIN..Key::MAX, Lsn(0x2000)..Lsn(0x3000)),
+            delta(Key::MIN..Key::MAX, Lsn(0x1000)..Lsn(0x2000)),
+        ];
+
+        // All layers fit in the max file size
+        let level = identify_level(layers.clone(), Lsn(0x10000), 0x2000)
+            .await?
+            .unwrap();
+        assert_eq!(level.depth(), 6);
+
+        // Same LSN with smaller max file size. The second layer from the top is larger
+        // and belongs to next level.
+        let level = identify_level(layers.clone(), Lsn(0x10000), 0x1000)
+            .await?
+            .unwrap();
+        assert_eq!(level.depth(), 1);
+
+        // Call with a smaller LSN
+        let level = identify_level(layers.clone(), Lsn(0x3000), 0x1000)
+            .await?
+            .unwrap();
+        assert_eq!(level.depth(), 2);
+
+        // Call with an LSN that doesn't partition the space
+        let result = identify_level(layers, Lsn(0x6000), 0x1000).await;
+        assert!(result.is_err());
+        Ok(())
+    }
+
+    #[tokio::test]
+    async fn test_overlapping_lsn_ranges() -> anyhow::Result<()> {
+        // The files LSN ranges overlap, so even though there are more files that
+        // fit under the file size, they are not included in the level because they
+        // overlap so that we'd need to include the oldest file, too, which is
+        // larger
+        let layers = vec![
+            delta(Key::MIN..Key::MAX, Lsn(0x4000)..Lsn(0x5000)),
+            delta(Key::MIN..Key::MAX, Lsn(0x3000)..Lsn(0x4000)), // overlap
+            delta(Key::MIN..Key::MAX, Lsn(0x2500)..Lsn(0x3500)), // overlap
+            delta(Key::MIN..Key::MAX, Lsn(0x2000)..Lsn(0x3000)), // overlap
+            delta(Key::MIN..Key::MAX, Lsn(0x1000)..Lsn(0x2500)), // larger
+        ];
+
+        let level = identify_level(layers.clone(), Lsn(0x10000), 0x1000)
+            .await?
+            .unwrap();
+        assert_eq!(level.depth(), 1);
+
+        Ok(())
+    }
+
+    #[tokio::test]
+    async fn test_depth_nonoverlapping() -> anyhow::Result<()> {
+        // The key ranges don't overlap, so depth is only 1.
+        let layers = vec![
+            delta(4000..5000, Lsn(0x6000)..Lsn(0x7000)),
+            delta(3000..4000, Lsn(0x7000)..Lsn(0x8000)),
+            delta(1000..2000, Lsn(0x8000)..Lsn(0x9000)),
+        ];
+
+        let level = identify_level(layers.clone(), Lsn(0x10000), 0x2000)
+            .await?
+            .unwrap();
+        assert_eq!(level.layers.len(), 3);
+        assert_eq!(level.depth(), 1);
+
+        // Staggered. The 1st and 3rd layer don't overlap with each other.
+        let layers = vec![
+            delta(1000..2000, Lsn(0x8000)..Lsn(0x9000)),
+            delta(1500..2500, Lsn(0x7000)..Lsn(0x8000)),
+            delta(2000..3000, Lsn(0x6000)..Lsn(0x7000)),
+        ];
+
+        let level = identify_level(layers.clone(), Lsn(0x10000), 0x2000)
+            .await?
+            .unwrap();
+        assert_eq!(level.layers.len(), 3);
+        assert_eq!(level.depth(), 2);
+        Ok(())
+    }
+
+    #[tokio::test]
+    async fn test_depth_images() -> anyhow::Result<()> {
+        let layers: Vec<MockLayer> = vec![
+            delta(1000..2000, Lsn(0x8000)..Lsn(0x9000)),
+            delta(1500..2500, Lsn(0x7000)..Lsn(0x8000)),
+            delta(2000..3000, Lsn(0x6000)..Lsn(0x7000)),
+            // This covers the same key range as the 2nd delta layer. The depth
+            // in that key range is therefore 0.
+            image(1500..2500, Lsn(0x9000)),
+        ];
+
+        let level = identify_level(layers.clone(), Lsn(0x10000), 0x2000)
+            .await?
+            .unwrap();
+        assert_eq!(level.layers.len(), 4);
+        assert_eq!(level.depth(), 1);
+        Ok(())
+    }
+}

pageserver/compaction/src/interface.rs

@@ -0,0 +1,152 @@
+//! This is what the compaction implementation needs to know about
+//! layers, keyspace etc.
+//!
+//! All the heavy lifting is done by the create_image and create_delta
+//! functions that the implementor provides.
+//!
+use async_trait::async_trait;
+use std::ops::Range;
+use utils::lsn::Lsn;
+
+/// Public interface. This is the main thing that the implementor needs to provide
+#[async_trait]
+pub trait CompactionJobExecutor {
+    // Type system.
+    //
+    // We assume that there are two kinds of layers, deltas and images. The
+    // compaction doesn't distinguish whether they are stored locally or
+    // remotely.
+    //
+    // The keyspace is defined by CompactionKey trait.
+    //
+    type Key: CompactionKey;
+
+    type Layer: CompactionLayer<Self::Key> + Clone;
+    type DeltaLayer: CompactionDeltaLayer<Self> + Clone;
+    type ImageLayer: CompactionImageLayer<Self> + Clone;
+
+    // This is passed through to all the interface functions. The compaction
+    // implementation doesn't do anything with it, but it might be useful for
+    // the interface implementation.
+    type RequestContext: CompactionRequestContext;
+
+    // ----
+    // Functions that the planner uses to support its decisions
+    // ----
+
+    /// Return all layers that overlap the given bounding box.
+    async fn get_layers(
+        &mut self,
+        key_range: &Range<Self::Key>,
+        lsn_range: &Range<Lsn>,
+        ctx: &Self::RequestContext,
+    ) -> anyhow::Result<Vec<Self::Layer>>;
+
+    async fn get_keyspace(
+        &mut self,
+        key_range: &Range<Self::Key>,
+        lsn: Lsn,
+        ctx: &Self::RequestContext,
+    ) -> anyhow::Result<CompactionKeySpace<Self::Key>>;
+
+    /// NB: This is a pretty expensive operation. In the real pageserver
+    /// implementation, it downloads the layer, and keeps it resident
+    /// until the DeltaLayer is dropped.
+    async fn downcast_delta_layer(
+        &self,
+        layer: &Self::Layer,
+    ) -> anyhow::Result<Option<Self::DeltaLayer>>;
+
+    // ----
+    // Functions to execute the plan
+    // ----
+
+    /// Create a new image layer, materializing all the values in the key range,
+    /// at given 'lsn'.
+    async fn create_image(
+        &mut self,
+        lsn: Lsn,
+        key_range: &Range<Self::Key>,
+        ctx: &Self::RequestContext,
+    ) -> anyhow::Result<()>;
+
+    /// Create a new delta layer, containing all the values from 'input_layers'
+    /// in the given key and LSN range.
+    async fn create_delta(
+        &mut self,
+        lsn_range: &Range<Lsn>,
+        key_range: &Range<Self::Key>,
+        input_layers: &[Self::DeltaLayer],
+        ctx: &Self::RequestContext,
+    ) -> anyhow::Result<()>;
+
+    /// Delete a layer. The compaction implementation will call this only after
+    /// all the create_image() or create_delta() calls that deletion of this
+    /// layer depends on have finished. But if the implementor has extra lazy
+    /// background tasks, like uploading the index json file to remote storage,
+    /// it is the implemenation's responsibility to track those.
+    async fn delete_layer(
+        &mut self,
+        layer: &Self::Layer,
+        ctx: &Self::RequestContext,
+    ) -> anyhow::Result<()>;
+}
+
+pub trait CompactionKey: std::cmp::Ord + Clone + Copy + std::fmt::Display {
+    const MIN: Self;
+    const MAX: Self;
+
+    /// Calculate distance between key_range.start and key_range.end.
+    ///
+    /// This returns u32, for compatibility with Repository::key. If the
+    /// distance is larger, return u32::MAX.
+    fn key_range_size(key_range: &Range<Self>) -> u32;
+
+    // return "self + 1"
+    fn next(&self) -> Self;
+
+    // return "self + <some decent amount to skip>". The amount to skip
+    // is left to the implementation.
+    // FIXME: why not just "add(u32)" ?  This is hard to use
+    fn skip_some(&self) -> Self;
+}
+
+/// Contiguous ranges of keys that belong to the key space. In key order, and
+/// with no overlap.
+pub type CompactionKeySpace<K> = Vec<Range<K>>;
+
+/// Functions needed from all layers.
+pub trait CompactionLayer<K: CompactionKey + ?Sized> {
+    fn key_range(&self) -> &Range<K>;
+    fn lsn_range(&self) -> &Range<Lsn>;
+
+    fn file_size(&self) -> u64;
+
+    /// For debugging, short human-readable representation of the layer. E.g. filename.
+    fn short_id(&self) -> String;
+
+    fn is_delta(&self) -> bool;
+}
+
+#[async_trait]
+pub trait CompactionDeltaLayer<E: CompactionJobExecutor + ?Sized>: CompactionLayer<E::Key> {
+    type DeltaEntry<'a>: CompactionDeltaEntry<'a, E::Key>
+    where
+        Self: 'a;
+
+    /// Return all keys in this delta layer.
+    async fn load_keys<'a>(
+        &self,
+        ctx: &E::RequestContext,
+    ) -> anyhow::Result<Vec<Self::DeltaEntry<'_>>>;
+}
+
+pub trait CompactionImageLayer<E: CompactionJobExecutor + ?Sized>: CompactionLayer<E::Key> {}
+
+pub trait CompactionDeltaEntry<'a, K> {
+    fn key(&self) -> K;
+    fn lsn(&self) -> Lsn;
+    fn size(&self) -> u64;
+}
+
+pub trait CompactionRequestContext {}

pageserver/compaction/src/lib.rs

@@ -0,0 +1,12 @@
+// The main module implementing the compaction algorithm
+pub mod compact_tiered;
+pub(crate) mod identify_levels;
+
+// Traits that the caller of the compaction needs to implement
+pub mod interface;
+
+// Utility functions, useful for the implementation
+pub mod helpers;
+
+// A simulator with mock implementations of 'interface'
+pub mod simulator;

pageserver/compaction/src/simulator.rs

@@ -0,0 +1,613 @@
+mod draw;
+
+use draw::{LayerTraceEvent, LayerTraceFile, LayerTraceOp};
+
+use async_trait::async_trait;
+use futures::StreamExt;
+use rand::Rng;
+use tracing::info;
+
+use utils::lsn::Lsn;
+
+use std::fmt::Write;
+use std::ops::Range;
+use std::sync::Arc;
+use std::sync::Mutex;
+
+use crate::helpers::{merge_delta_keys, overlaps_with};
+
+use crate::interface;
+use crate::interface::CompactionLayer;
+
+//
+// Implementation for the CompactionExecutor interface
+//
+pub struct MockTimeline {
+    // Parameters for the compaction algorithm
+    pub target_file_size: u64,
+    tiers_per_level: u64,
+
+    num_l0_flushes: u64,
+    last_compact_at_flush: u64,
+    last_flush_lsn: Lsn,
+
+    // In-memory layer
+    records: Vec<MockRecord>,
+    total_len: u64,
+    start_lsn: Lsn,
+    end_lsn: Lsn,
+
+    // Current keyspace at `end_lsn`. This is updated on every ingested record.
+    keyspace: KeySpace,
+
+    // historic keyspaces
+    old_keyspaces: Vec<(Lsn, KeySpace)>,
+
+    // "on-disk" layers
+    pub live_layers: Vec<MockLayer>,
+
+    num_deleted_layers: u64,
+
+    // Statistics
+    wal_ingested: u64,
+    bytes_written: u64,
+    bytes_deleted: u64,
+    layers_created: u64,
+    layers_deleted: u64,
+
+    // All the events - creation and deletion of files - are collected
+    // in 'history'. It is used to draw the SVG animation at the end.
+    time: u64,
+    history: Vec<draw::LayerTraceEvent>,
+}
+
+type KeySpace = interface::CompactionKeySpace<Key>;
+
+pub struct MockRequestContext {}
+impl interface::CompactionRequestContext for MockRequestContext {}
+
+pub type Key = u64;
+
+impl interface::CompactionKey for Key {
+    const MIN: Self = u64::MIN;
+    const MAX: Self = u64::MAX;
+
+    fn key_range_size(key_range: &Range<Self>) -> u32 {
+        std::cmp::min(key_range.end - key_range.start, u32::MAX as u64) as u32
+    }
+
+    fn next(&self) -> Self {
+        self + 1
+    }
+    fn skip_some(&self) -> Self {
+        // round up to next xx
+        self + 100
+    }
+}
+
+#[derive(Clone)]
+pub struct MockRecord {
+    lsn: Lsn,
+    key: Key,
+    len: u64,
+}
+
+impl interface::CompactionDeltaEntry<'_, Key> for MockRecord {
+    fn key(&self) -> Key {
+        self.key
+    }
+    fn lsn(&self) -> Lsn {
+        self.lsn
+    }
+    fn size(&self) -> u64 {
+        self.len
+    }
+}
+
+pub struct MockDeltaLayer {
+    pub key_range: Range<Key>,
+    pub lsn_range: Range<Lsn>,
+
+    pub file_size: u64,
+
+    pub deleted: Mutex<bool>,
+
+    pub records: Vec<MockRecord>,
+}
+
+impl interface::CompactionLayer<Key> for Arc<MockDeltaLayer> {
+    fn key_range(&self) -> &Range<Key> {
+        &self.key_range
+    }
+    fn lsn_range(&self) -> &Range<Lsn> {
+        &self.lsn_range
+    }
+
+    fn file_size(&self) -> u64 {
+        self.file_size
+    }
+
+    fn short_id(&self) -> String {
+        format!(
+            "{:016X}-{:016X}__{:08X}-{:08X}",
+            self.key_range.start, self.key_range.end, self.lsn_range.start.0, self.lsn_range.end.0
+        )
+    }
+
+    fn is_delta(&self) -> bool {
+        true
+    }
+}
+
+#[async_trait]
+impl interface::CompactionDeltaLayer<MockTimeline> for Arc<MockDeltaLayer> {
+    type DeltaEntry<'a> = MockRecord;
+
+    async fn load_keys<'a>(&self, _ctx: &MockRequestContext) -> anyhow::Result<Vec<MockRecord>> {
+        Ok(self.records.clone())
+    }
+}
+
+pub struct MockImageLayer {
+    pub key_range: Range<Key>,
+    pub lsn_range: Range<Lsn>,
+
+    pub file_size: u64,
+
+    pub deleted: Mutex<bool>,
+}
+
+impl interface::CompactionImageLayer<MockTimeline> for Arc<MockImageLayer> {}
+
+impl interface::CompactionLayer<Key> for Arc<MockImageLayer> {
+    fn key_range(&self) -> &Range<Key> {
+        &self.key_range
+    }
+    fn lsn_range(&self) -> &Range<Lsn> {
+        &self.lsn_range
+    }
+
+    fn file_size(&self) -> u64 {
+        self.file_size
+    }
+
+    fn short_id(&self) -> String {
+        format!(
+            "{:016X}-{:016X}__{:08X}",
+            self.key_range.start, self.key_range.end, self.lsn_range.start.0,
+        )
+    }
+
+    fn is_delta(&self) -> bool {
+        false
+    }
+}
+
+impl MockTimeline {
+    pub fn new() -> Self {
+        MockTimeline {
+            target_file_size: 256 * 1024 * 1024,
+            tiers_per_level: 4,
+
+            num_l0_flushes: 0,
+            last_compact_at_flush: 0,
+            last_flush_lsn: Lsn(0),
+
+            records: Vec::new(),
+            total_len: 0,
+            start_lsn: Lsn(1000),
+            end_lsn: Lsn(1000),
+            keyspace: KeySpace::new(),
+
+            old_keyspaces: vec![],
+
+            live_layers: vec![],
+
+            num_deleted_layers: 0,
+
+            wal_ingested: 0,
+            bytes_written: 0,
+            bytes_deleted: 0,
+            layers_created: 0,
+            layers_deleted: 0,
+
+            time: 0,
+            history: Vec::new(),
+        }
+    }
+
+    pub async fn compact(&mut self) -> anyhow::Result<()> {
+        let ctx = MockRequestContext {};
+
+        crate::compact_tiered::compact_tiered(
+            self,
+            self.last_flush_lsn,
+            self.target_file_size,
+            self.tiers_per_level,
+            &ctx,
+        )
+        .await?;
+
+        Ok(())
+    }
+
+    // Ingest one record to the timeline
+    pub fn ingest_record(&mut self, key: Key, len: u64) {
+        self.records.push(MockRecord {
+            lsn: self.end_lsn,
+            key,
+            len,
+        });
+        self.total_len += len;
+        self.end_lsn += len;
+
+        if self.total_len > self.target_file_size {
+            self.flush_l0();
+        }
+    }
+
+    pub async fn compact_if_needed(&mut self) -> anyhow::Result<()> {
+        if self.num_l0_flushes - self.last_compact_at_flush >= self.tiers_per_level {
+            self.compact().await?;
+            self.last_compact_at_flush = self.num_l0_flushes;
+        }
+        Ok(())
+    }
+
+    pub fn flush_l0(&mut self) {
+        if self.records.is_empty() {
+            return;
+        }
+
+        let mut records = std::mem::take(&mut self.records);
+        records.sort_by_key(|rec| rec.key);
+
+        let lsn_range = self.start_lsn..self.end_lsn;
+        let new_layer = Arc::new(MockDeltaLayer {
+            key_range: Key::MIN..Key::MAX,
+            lsn_range: lsn_range.clone(),
+            file_size: self.total_len,
+            records,
+            deleted: Mutex::new(false),
+        });
+        info!("flushed L0 layer {}", new_layer.short_id());
+        self.live_layers.push(MockLayer::from(&new_layer));
+
+        // reset L0
+        self.start_lsn = self.end_lsn;
+        self.total_len = 0;
+        self.records = Vec::new();
+
+        self.layers_created += 1;
+        self.bytes_written += new_layer.file_size;
+
+        self.time += 1;
+        self.history.push(LayerTraceEvent {
+            time_rel: self.time,
+            op: LayerTraceOp::Flush,
+            file: LayerTraceFile {
+                filename: new_layer.short_id(),
+                key_range: new_layer.key_range.clone(),
+                lsn_range: new_layer.lsn_range.clone(),
+            },
+        });
+
+        self.num_l0_flushes += 1;
+        self.last_flush_lsn = self.end_lsn;
+    }
+
+    // Ingest `num_records' records to the timeline, with random keys
+    // uniformly distributed in `key_range`
+    pub fn ingest_uniform(
+        &mut self,
+        num_records: u64,
+        len: u64,
+        key_range: &Range<Key>,
+    ) -> anyhow::Result<()> {
+        crate::helpers::union_to_keyspace(&mut self.keyspace, vec![key_range.clone()]);
+        let mut rng = rand::thread_rng();
+        for _ in 0..num_records {
+            self.ingest_record(rng.gen_range(key_range.clone()), len);
+            self.wal_ingested += len;
+        }
+        Ok(())
+    }
+
+    pub fn print_stats(&self) -> anyhow::Result<String> {
+        let mut s = String::new();
+
+        writeln!(s, "STATISTICS:")?;
+        writeln!(
+            s,
+            "WAL ingested:   {:>10} MB",
+            self.wal_ingested / (1024 * 1024)
+        )?;
+        writeln!(
+            s,
+            "size created:   {:>10} MB",
+            self.bytes_written / (1024 * 1024)
+        )?;
+        writeln!(
+            s,
+            "size deleted:   {:>10} MB",
+            self.bytes_deleted / (1024 * 1024)
+        )?;
+        writeln!(s, "files created:     {:>10}", self.layers_created)?;
+        writeln!(s, "files deleted:     {:>10}", self.layers_deleted)?;
+        writeln!(
+            s,
+            "write amp:         {:>10.2}",
+            self.bytes_written as f64 / self.wal_ingested as f64
+        )?;
+        writeln!(
+            s,
+            "storage amp:       {:>10.2}",
+            (self.bytes_written - self.bytes_deleted) as f64 / self.wal_ingested as f64
+        )?;
+
+        Ok(s)
+    }
+
+    pub fn draw_history<W: std::io::Write>(&self, output: W) -> anyhow::Result<()> {
+        draw::draw_history(&self.history, output)
+    }
+}
+
+impl Default for MockTimeline {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+#[derive(Clone)]
+pub enum MockLayer {
+    Delta(Arc<MockDeltaLayer>),
+    Image(Arc<MockImageLayer>),
+}
+
+impl interface::CompactionLayer<Key> for MockLayer {
+    fn key_range(&self) -> &Range<Key> {
+        match self {
+            MockLayer::Delta(this) => this.key_range(),
+            MockLayer::Image(this) => this.key_range(),
+        }
+    }
+    fn lsn_range(&self) -> &Range<Lsn> {
+        match self {
+            MockLayer::Delta(this) => this.lsn_range(),
+            MockLayer::Image(this) => this.lsn_range(),
+        }
+    }
+    fn file_size(&self) -> u64 {
+        match self {
+            MockLayer::Delta(this) => this.file_size(),
+            MockLayer::Image(this) => this.file_size(),
+        }
+    }
+    fn short_id(&self) -> String {
+        match self {
+            MockLayer::Delta(this) => this.short_id(),
+            MockLayer::Image(this) => this.short_id(),
+        }
+    }
+
+    fn is_delta(&self) -> bool {
+        match self {
+            MockLayer::Delta(_) => true,
+            MockLayer::Image(_) => false,
+        }
+    }
+}
+
+impl MockLayer {
+    fn is_deleted(&self) -> bool {
+        let guard = match self {
+            MockLayer::Delta(this) => this.deleted.lock().unwrap(),
+            MockLayer::Image(this) => this.deleted.lock().unwrap(),
+        };
+        *guard
+    }
+    fn mark_deleted(&self) {
+        let mut deleted_guard = match self {
+            MockLayer::Delta(this) => this.deleted.lock().unwrap(),
+            MockLayer::Image(this) => this.deleted.lock().unwrap(),
+        };
+        assert!(!*deleted_guard, "layer already deleted");
+        *deleted_guard = true;
+    }
+}
+
+impl From<&Arc<MockDeltaLayer>> for MockLayer {
+    fn from(l: &Arc<MockDeltaLayer>) -> Self {
+        MockLayer::Delta(l.clone())
+    }
+}
+
+impl From<&Arc<MockImageLayer>> for MockLayer {
+    fn from(l: &Arc<MockImageLayer>) -> Self {
+        MockLayer::Image(l.clone())
+    }
+}
+
+#[async_trait]
+impl interface::CompactionJobExecutor for MockTimeline {
+    type Key = Key;
+    type Layer = MockLayer;
+    type DeltaLayer = Arc<MockDeltaLayer>;
+    type ImageLayer = Arc<MockImageLayer>;
+    type RequestContext = MockRequestContext;
+
+    async fn get_layers(
+        &mut self,
+        key_range: &Range<Self::Key>,
+        lsn_range: &Range<Lsn>,
+        _ctx: &Self::RequestContext,
+    ) -> anyhow::Result<Vec<Self::Layer>> {
+        // Clear any deleted layers from our vec
+        self.live_layers.retain(|l| !l.is_deleted());
+
+        let layers: Vec<MockLayer> = self
+            .live_layers
+            .iter()
+            .filter(|l| {
+                overlaps_with(l.lsn_range(), lsn_range) && overlaps_with(l.key_range(), key_range)
+            })
+            .cloned()
+            .collect();
+
+        Ok(layers)
+    }
+
+    async fn get_keyspace(
+        &mut self,
+        key_range: &Range<Self::Key>,
+        _lsn: Lsn,
+        _ctx: &Self::RequestContext,
+    ) -> anyhow::Result<interface::CompactionKeySpace<Key>> {
+        // find it in the levels
+        if self.old_keyspaces.is_empty() {
+            Ok(crate::helpers::intersect_keyspace(
+                &self.keyspace,
+                key_range,
+            ))
+        } else {
+            // not implemented
+
+            // The mock implementation only allows requesting the
+            // keyspace at the level's end LSN. That's all that the
+            // current implementation needs.
+            panic!("keyspace not available for requested lsn");
+        }
+    }
+
+    async fn downcast_delta_layer(
+        &self,
+        layer: &MockLayer,
+    ) -> anyhow::Result<Option<Arc<MockDeltaLayer>>> {
+        Ok(match layer {
+            MockLayer::Delta(l) => Some(l.clone()),
+            MockLayer::Image(_) => None,
+        })
+    }
+
+    async fn create_image(
+        &mut self,
+        lsn: Lsn,
+        key_range: &Range<Key>,
+        ctx: &MockRequestContext,
+    ) -> anyhow::Result<()> {
+        let keyspace = self.get_keyspace(key_range, lsn, ctx).await?;
+
+        let mut accum_size: u64 = 0;
+        for r in keyspace {
+            accum_size += r.end - r.start;
+        }
+
+        let new_layer = Arc::new(MockImageLayer {
+            key_range: key_range.clone(),
+            lsn_range: lsn..lsn,
+            file_size: accum_size * 8192,
+            deleted: Mutex::new(false),
+        });
+        info!(
+            "created image layer, size {}: {}",
+            new_layer.file_size,
+            new_layer.short_id()
+        );
+        self.live_layers.push(MockLayer::Image(new_layer.clone()));
+
+        // update stats
+        self.bytes_written += new_layer.file_size;
+        self.layers_created += 1;
+
+        self.time += 1;
+        self.history.push(LayerTraceEvent {
+            time_rel: self.time,
+            op: LayerTraceOp::CreateImage,
+            file: LayerTraceFile {
+                filename: new_layer.short_id(),
+                key_range: new_layer.key_range.clone(),
+                lsn_range: new_layer.lsn_range.clone(),
+            },
+        });
+
+        Ok(())
+    }
+
+    async fn create_delta(
+        &mut self,
+        lsn_range: &Range<Lsn>,
+        key_range: &Range<Key>,
+        input_layers: &[Arc<MockDeltaLayer>],
+        ctx: &MockRequestContext,
+    ) -> anyhow::Result<()> {
+        let mut key_value_stream =
+            std::pin::pin!(merge_delta_keys::<MockTimeline>(input_layers, ctx));
+        let mut records: Vec<MockRecord> = Vec::new();
+        let mut total_len = 2;
+        while let Some(delta_entry) = key_value_stream.next().await {
+            let delta_entry: MockRecord = delta_entry?;
+            if key_range.contains(&delta_entry.key) && lsn_range.contains(&delta_entry.lsn) {
+                total_len += delta_entry.len;
+                records.push(delta_entry);
+            }
+        }
+        let total_records = records.len();
+        let new_layer = Arc::new(MockDeltaLayer {
+            key_range: key_range.clone(),
+            lsn_range: lsn_range.clone(),
+            file_size: total_len,
+            records,
+            deleted: Mutex::new(false),
+        });
+        info!(
+            "created delta layer, recs {}, size {}: {}",
+            total_records,
+            total_len,
+            new_layer.short_id()
+        );
+        self.live_layers.push(MockLayer::Delta(new_layer.clone()));
+
+        // update stats
+        self.bytes_written += total_len;
+        self.layers_created += 1;
+
+        self.time += 1;
+        self.history.push(LayerTraceEvent {
+            time_rel: self.time,
+            op: LayerTraceOp::CreateDelta,
+            file: LayerTraceFile {
+                filename: new_layer.short_id(),
+                key_range: new_layer.key_range.clone(),
+                lsn_range: new_layer.lsn_range.clone(),
+            },
+        });
+
+        Ok(())
+    }
+
+    async fn delete_layer(
+        &mut self,
+        layer: &Self::Layer,
+        _ctx: &MockRequestContext,
+    ) -> anyhow::Result<()> {
+        let layer = std::pin::pin!(layer);
+        info!("deleting layer: {}", layer.short_id());
+        self.num_deleted_layers += 1;
+        self.bytes_deleted += layer.file_size();
+        layer.mark_deleted();
+
+        self.time += 1;
+        self.history.push(LayerTraceEvent {
+            time_rel: self.time,
+            op: LayerTraceOp::Delete,
+            file: LayerTraceFile {
+                filename: layer.short_id(),
+                key_range: layer.key_range().clone(),
+                lsn_range: layer.lsn_range().clone(),
+            },
+        });
+
+        Ok(())
+    }
+}

pageserver/compaction/src/simulator/draw.rs

@@ -0,0 +1,411 @@
+use super::Key;
+use anyhow::Result;
+use std::cmp::Ordering;
+use std::{
+    collections::{BTreeMap, BTreeSet, HashSet},
+    fmt::Write,
+    ops::Range,
+};
+use svg_fmt::{rgb, BeginSvg, EndSvg, Fill, Stroke, Style};
+use utils::lsn::Lsn;
+
+// Map values to their compressed coordinate - the index the value
+// would have in a sorted and deduplicated list of all values.
+struct CoordinateMap<T: Ord + Copy> {
+    map: BTreeMap<T, usize>,
+    stretch: f32,
+}
+
+impl<T: Ord + Copy> CoordinateMap<T> {
+    fn new(coords: Vec<T>, stretch: f32) -> Self {
+        let set: BTreeSet<T> = coords.into_iter().collect();
+
+        let mut map: BTreeMap<T, usize> = BTreeMap::new();
+        for (i, e) in set.iter().enumerate() {
+            map.insert(*e, i);
+        }
+
+        Self { map, stretch }
+    }
+
+    // This assumes that the map contains an exact point for this.
+    // Use map_inexact for values inbetween
+    fn map(&self, val: T) -> f32 {
+        *self.map.get(&val).unwrap() as f32 * self.stretch
+    }
+
+    // the value is still assumed to be within the min/max bounds
+    // (this is currently unused)
+    fn _map_inexact(&self, val: T) -> f32 {
+        let prev = *self.map.range(..=val).next().unwrap().1;
+        let next = *self.map.range(val..).next().unwrap().1;
+
+        // interpolate
+        (prev as f32 + (next - prev) as f32) * self.stretch
+    }
+
+    fn max(&self) -> f32 {
+        self.map.len() as f32 * self.stretch
+    }
+}
+
+#[derive(PartialEq, Hash, Eq)]
+pub enum LayerTraceOp {
+    Flush,
+    CreateDelta,
+    CreateImage,
+    Delete,
+}
+
+impl std::fmt::Display for LayerTraceOp {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> Result<(), std::fmt::Error> {
+        let op_str = match self {
+            LayerTraceOp::Flush => "flush",
+            LayerTraceOp::CreateDelta => "create_delta",
+            LayerTraceOp::CreateImage => "create_image",
+            LayerTraceOp::Delete => "delete",
+        };
+        f.write_str(op_str)
+    }
+}
+
+#[derive(PartialEq, Hash, Eq, Clone)]
+pub struct LayerTraceFile {
+    pub filename: String,
+    pub key_range: Range<Key>,
+    pub lsn_range: Range<Lsn>,
+}
+
+impl LayerTraceFile {
+    fn is_image(&self) -> bool {
+        self.lsn_range.end == self.lsn_range.start
+    }
+}
+
+pub struct LayerTraceEvent {
+    pub time_rel: u64,
+    pub op: LayerTraceOp,
+    pub file: LayerTraceFile,
+}
+
+pub fn draw_history<W: std::io::Write>(history: &[LayerTraceEvent], mut output: W) -> Result<()> {
+    let mut files: Vec<LayerTraceFile> = Vec::new();
+
+    for event in history {
+        files.push(event.file.clone());
+    }
+    let last_time_rel = history.last().unwrap().time_rel;
+
+    // Collect all coordinates
+    let mut keys: Vec<Key> = vec![];
+    let mut lsns: Vec<Lsn> = vec![];
+    for f in files.iter() {
+        keys.push(f.key_range.start);
+        keys.push(f.key_range.end);
+        lsns.push(f.lsn_range.start);
+        lsns.push(f.lsn_range.end);
+    }
+
+    // Analyze
+    let key_map = CoordinateMap::new(keys, 2.0);
+    // Stretch out vertically for better visibility
+    let lsn_map = CoordinateMap::new(lsns, 3.0);
+
+    let mut svg = String::new();
+
+    // Draw
+    writeln!(
+        svg,
+        "{}",
+        BeginSvg {
+            w: key_map.max(),
+            h: lsn_map.max(),
+        }
+    )?;
+    let lsn_max = lsn_map.max();
+
+    // Sort the files by LSN, but so that image layers go after all delta layers
+    // The SVG is painted in the order the elements appear, and we want to draw
+    // image layers on top of the delta layers if they overlap
+    //
+    // (This could also be implemented via z coordinates: image layers get one z
+    // coord, delta layers get another z coord.)
+    let mut files_sorted: Vec<LayerTraceFile> = files.into_iter().collect();
+    files_sorted.sort_by(|a, b| {
+        if a.is_image() && !b.is_image() {
+            Ordering::Greater
+        } else if !a.is_image() && b.is_image() {
+            Ordering::Less
+        } else {
+            a.lsn_range.end.cmp(&b.lsn_range.end)
+        }
+    });
+
+    writeln!(svg, "<!-- layers -->")?;
+    let mut files_seen = HashSet::new();
+    for f in files_sorted {
+        if files_seen.contains(&f) {
+            continue;
+        }
+        let key_start = key_map.map(f.key_range.start);
+        let key_end = key_map.map(f.key_range.end);
+        let key_diff = key_end - key_start;
+
+        if key_start >= key_end {
+            panic!("Invalid key range {}-{}", key_start, key_end);
+        }
+
+        let lsn_start = lsn_map.map(f.lsn_range.start);
+        let lsn_end = lsn_map.map(f.lsn_range.end);
+
+        // Fill in and thicken rectangle if it's an
+        // image layer so that we can see it.
+        let mut style = Style::default();
+        style.fill = Fill::Color(rgb(0x80, 0x80, 0x80));
+        style.stroke = Stroke::Color(rgb(0, 0, 0), 0.5);
+
+        let y_start = lsn_max - lsn_start;
+        let y_end = lsn_max - lsn_end;
+
+        let x_margin = 0.25;
+        let y_margin = 0.5;
+
+        match f.lsn_range.start.cmp(&f.lsn_range.end) {
+            Ordering::Less => {
+                write!(
+                    svg,
+                    r#"    <rect id="layer_{}" x="{}" y="{}" width="{}" height="{}" ry="{}" style="{}">"#,
+                    f.filename,
+                    key_start + x_margin,
+                    y_end + y_margin,
+                    key_diff - x_margin * 2.0,
+                    y_start - y_end - y_margin * 2.0,
+                    1.0, // border_radius,
+                    style,
+                )?;
+                write!(svg, "<title>{}</title>", f.filename)?;
+                writeln!(svg, "</rect>")?;
+            }
+            Ordering::Equal => {
+                //lsn_diff = 0.3;
+                //lsn_offset = -lsn_diff / 2.0;
+                //margin = 0.05;
+                style.fill = Fill::Color(rgb(0x80, 0, 0x80));
+                style.stroke = Stroke::Color(rgb(0x80, 0, 0x80), 3.0);
+                write!(
+                    svg,
+                    r#"    <line id="layer_{}" x1="{}" y1="{}" x2="{}" y2="{}" style="{}">"#,
+                    f.filename,
+                    key_start + x_margin,
+                    y_end,
+                    key_end - x_margin,
+                    y_end,
+                    style,
+                )?;
+                write!(
+                    svg,
+                    "<title>{}<br>{} - {}</title>",
+                    f.filename, lsn_end, y_end
+                )?;
+                writeln!(svg, "</line>")?;
+            }
+            Ordering::Greater => panic!("Invalid lsn range {}-{}", lsn_start, lsn_end),
+        }
+        files_seen.insert(f);
+    }
+
+    let mut record_style = Style::default();
+    record_style.fill = Fill::Color(rgb(0x80, 0x80, 0x80));
+    record_style.stroke = Stroke::None;
+
+    writeln!(svg, "{}", EndSvg)?;
+
+    let mut layer_events_str = String::new();
+    let mut first = true;
+    for e in history {
+        if !first {
+            writeln!(layer_events_str, ",")?;
+        }
+        write!(
+            layer_events_str,
+            r#"  {{"time_rel": {}, "filename": "{}", "op": "{}"}}"#,
+            e.time_rel, e.file.filename, e.op
+        )?;
+        first = false;
+    }
+    writeln!(layer_events_str)?;
+
+    writeln!(
+        output,
+        r#"<!DOCTYPE html>
+<html>
+<head>
+<style>
+/* Keep the slider pinned at top */
+.topbar {{
+  display: block;
+  overflow: hidden;
+  background-color: lightgrey;
+  position: fixed;
+  top: 0;
+  width: 100%;
+/*  width: 500px; */
+}}
+.slidercontainer {{
+  float: left;
+  width: 50%;
+  margin-right: 200px;
+}}
+.slider {{
+  float: left;
+  width: 100%;
+}}
+.legend {{
+  width: 200px;
+  float: right;
+}}
+
+/* Main content */
+.main {{
+  margin-top: 50px; /* Add a top margin to avoid content overlay */
+}}
+</style>
+</head>
+
+  <body onload="init()">
+    <script type="text/javascript">
+
+      var layer_events = [{layer_events_str}]
+
+      let ticker;
+
+      function init() {{
+          for (let i = 0; i < layer_events.length; i++) {{
+              var layer = document.getElementById("layer_" + layer_events[i].filename);
+              layer.style.visibility = "hidden";
+          }}
+          last_layer_event = -1;
+          moveSlider(last_slider_pos)
+      }}
+
+      function startAnimation() {{
+          ticker = setInterval(animateStep, 100);
+      }}
+      function stopAnimation() {{
+          clearInterval(ticker);
+      }}
+
+      function animateStep() {{
+          if (last_layer_event < layer_events.length - 1) {{
+              var slider = document.getElementById("time-slider");
+              let prevPos = slider.value
+              let nextEvent = last_layer_event + 1
+              while (nextEvent <= layer_events.length - 1) {{
+                  if (layer_events[nextEvent].time_rel > prevPos) {{
+                      break;
+                  }}
+                  nextEvent += 1;
+              }}
+              let nextPos = layer_events[nextEvent].time_rel
+              slider.value = nextPos
+              moveSlider(nextPos)
+          }}
+      }}
+
+      function redoLayerEvent(n, dir) {{
+          var layer = document.getElementById("layer_" + layer_events[n].filename);
+          switch (layer_events[n].op) {{
+              case "flush":
+                  layer.style.visibility = "visible";
+                  break;
+              case "create_delta":
+                  layer.style.visibility = "visible";
+                  break;
+              case "create_image":
+                  layer.style.visibility = "visible";
+                  break;
+              case "delete":
+                  layer.style.visibility = "hidden";
+                  break;
+          }}
+      }}
+      function undoLayerEvent(n) {{
+          var layer = document.getElementById("layer_" + layer_events[n].filename);
+          switch (layer_events[n].op) {{
+              case "flush":
+                  layer.style.visibility = "hidden";
+                  break;
+              case "create_delta":
+                  layer.style.visibility = "hidden";
+                  break;
+              case "create_image":
+                  layer.style.visibility = "hidden";
+                  break;
+              case "delete":
+                  layer.style.visibility = "visible";
+                  break;
+          }}
+      }}
+
+      var last_slider_pos = 0
+      var last_layer_event = 0
+
+      var moveSlider = function(new_pos) {{
+          if (new_pos > last_slider_pos) {{
+              while (last_layer_event < layer_events.length - 1) {{
+                  if (layer_events[last_layer_event + 1].time_rel > new_pos) {{
+                      break;
+                  }}
+                  last_layer_event += 1;
+                  redoLayerEvent(last_layer_event)
+              }}
+          }}
+          if (new_pos < last_slider_pos) {{
+              while (last_layer_event >= 0) {{
+                  if (layer_events[last_layer_event].time_rel <= new_pos) {{
+                      break;
+                  }}
+                  undoLayerEvent(last_layer_event)
+                  last_layer_event -= 1;
+              }}
+          }}
+          last_slider_pos = new_pos;
+          document.getElementById("debug_pos").textContent=new_pos;
+          if (last_layer_event >= 0) {{
+              document.getElementById("debug_layer_event").textContent=last_layer_event + " " + layer_events[last_layer_event].time_rel + " " + layer_events[last_layer_event].op;
+          }} else {{
+              document.getElementById("debug_layer_event").textContent="begin";
+          }}
+      }}
+    </script>
+
+    <div class="topbar">
+      <div class="slidercontainer">
+        <label for="time-slider">TIME</label>:
+        <input id="time-slider" class="slider" type="range" min="0" max="{last_time_rel}" value="0" oninput="moveSlider(this.value)"><br>
+
+        pos: <span id="debug_pos"></span><br>
+        event: <span id="debug_layer_event"></span><br>
+        gc: <span id="debug_gc_event"></span><br>
+      </div>
+
+      <button onclick="startAnimation()">Play</button>
+      <button onclick="stopAnimation()">Stop</button>
+
+      <svg class="legend">
+        <rect x=5 y=0 width=20 height=20 style="fill:rgb(128,128,128);stroke:rgb(0,0,0);stroke-width:0.5;fill-opacity:1;stroke-opacity:1;"/>
+        <line x1=5 y1=30 x2=25 y2=30 style="fill:rgb(128,0,128);stroke:rgb(128,0,128);stroke-width:3;fill-opacity:1;stroke-opacity:1;"/>
+        <line x1=0 y1=40 x2=30 y2=40 style="fill:none;stroke:rgb(255,0,0);stroke-width:0.5;fill-opacity:1;stroke-opacity:1;"/>
+      </svg>
+    </div>
+
+    <div class="main">
+{svg}
+    </div>
+  </body>
+</html>
+"#
+    )?;
+
+    Ok(())
+}

pageserver/compaction/tests/tests.rs

@@ -0,0 +1,37 @@
+use pageserver_compaction::interface::CompactionLayer;
+use pageserver_compaction::simulator::MockTimeline;
+
+/// Test the extreme case that there are so many updates for a single key that
+/// even if we produce an extremely narrow delta layer, spanning just that one
+/// key, we still too many records to fit in the target file size. We need to
+/// split in the LSN dimension too in that case.
+///
+/// TODO: The code to avoid this problem has not been implemented yet! So the
+/// assertion currently fails, but we need to make it not fail.
+#[ignore]
+#[tokio::test]
+async fn test_many_updates_for_single_key() -> anyhow::Result<()> {
+    let mut executor = MockTimeline::new();
+    executor.target_file_size = 10_000_000; // 10 MB
+
+    // Ingest 100 MB of updates to a single key.
+    for _ in 1..1000 {
+        executor.ingest_uniform(100, 10, &(0..100_000))?;
+        executor.ingest_uniform(10_000, 10, &(0..1))?;
+        executor.compact().await?;
+    }
+
+    // Check that all the layers are smaller than the target size (with some slop)
+    for l in executor.live_layers.iter() {
+        println!("layer {}: {}", l.short_id(), l.file_size());
+    }
+    for l in executor.live_layers.iter() {
+        assert!(l.file_size() < executor.target_file_size * 2);
+        // sanity check that none of the delta layers are stupidly small either
+        if l.is_delta() {
+            assert!(l.file_size() > executor.target_file_size / 2);
+        }
+    }
+
+    Ok(())
+}

pageserver/src/basebackup.rs

@@ -13,6 +13,7 @@
 use anyhow::{anyhow, bail, ensure, Context};
 use bytes::{BufMut, BytesMut};
 use fail::fail_point;
+use postgres_ffi::pg_constants;
 use std::fmt::Write as FmtWrite;
 use std::time::SystemTime;
 use tokio::io;
@@ -180,6 +181,7 @@ where
             }
         }
 
+        let mut min_restart_lsn: Lsn = Lsn::MAX;
         // Create tablespace directories
         for ((spcnode, dbnode), has_relmap_file) in
             self.timeline.list_dbdirs(self.lsn, self.ctx).await?
@@ -213,6 +215,34 @@ where
                     self.add_rel(rel, rel).await?;
                 }
             }
+
+            for (path, content) in self.timeline.list_aux_files(self.lsn, self.ctx).await? {
+                if path.starts_with("pg_replslot") {
+                    let offs = pg_constants::REPL_SLOT_ON_DISK_OFFSETOF_RESTART_LSN;
+                    let restart_lsn = Lsn(u64::from_le_bytes(
+                        content[offs..offs + 8].try_into().unwrap(),
+                    ));
+                    info!("Replication slot {} restart LSN={}", path, restart_lsn);
+                    min_restart_lsn = Lsn::min(min_restart_lsn, restart_lsn);
+                }
+                let header = new_tar_header(&path, content.len() as u64)?;
+                self.ar
+                    .append(&header, &*content)
+                    .await
+                    .context("could not add aux file to basebackup tarball")?;
+            }
+        }
+        if min_restart_lsn != Lsn::MAX {
+            info!(
+                "Min restart LSN for logical replication is {}",
+                min_restart_lsn
+            );
+            let data = min_restart_lsn.0.to_le_bytes();
+            let header = new_tar_header("restart.lsn", data.len() as u64)?;
+            self.ar
+                .append(&header, &data[..])
+                .await
+                .context("could not add restart.lsn file to basebackup tarball")?;
         }
         for xid in self
             .timeline

pageserver/src/bin/pageserver.rs

@@ -2,6 +2,7 @@
 
 use std::env::{var, VarError};
 use std::sync::Arc;
+use std::time::Duration;
 use std::{env, ops::ControlFlow, str::FromStr};
 
 use anyhow::{anyhow, Context};
@@ -33,11 +34,12 @@ use postgres_backend::AuthType;
 use utils::logging::TracingErrorLayerEnablement;
 use utils::signals::ShutdownSignals;
 use utils::{
-    auth::JwtAuth, logging, project_git_version, sentry_init::init_sentry, signals::Signal,
-    tcp_listener,
+    auth::JwtAuth, logging, project_build_tag, project_git_version, sentry_init::init_sentry,
+    signals::Signal, tcp_listener,
 };
 
 project_git_version!(GIT_VERSION);
+project_build_tag!(BUILD_TAG);
 
 const PID_FILE_NAME: &str = "pageserver.pid";
 
@@ -200,6 +202,51 @@ fn initialize_config(
     })
 }
 
+struct WaitForPhaseResult<F: std::future::Future + Unpin> {
+    timeout_remaining: Duration,
+    skipped: Option<F>,
+}
+
+/// During startup, we apply a timeout to our waits for readiness, to avoid
+/// stalling the whole service if one Tenant experiences some problem.  Each
+/// phase may consume some of the timeout: this function returns the updated
+/// timeout for use in the next call.
+async fn wait_for_phase<F>(phase: &str, mut fut: F, timeout: Duration) -> WaitForPhaseResult<F>
+where
+    F: std::future::Future + Unpin,
+{
+    let initial_t = Instant::now();
+    let skipped = match tokio::time::timeout(timeout, &mut fut).await {
+        Ok(_) => None,
+        Err(_) => {
+            tracing::info!(
+                timeout_millis = timeout.as_millis(),
+                %phase,
+                "Startup phase timed out, proceeding anyway"
+            );
+            Some(fut)
+        }
+    };
+
+    WaitForPhaseResult {
+        timeout_remaining: timeout
+            .checked_sub(Instant::now().duration_since(initial_t))
+            .unwrap_or(Duration::ZERO),
+        skipped,
+    }
+}
+
+fn startup_checkpoint(started_at: Instant, phase: &str, human_phase: &str) {
+    let elapsed = started_at.elapsed();
+    let secs = elapsed.as_secs_f64();
+    STARTUP_DURATION.with_label_values(&[phase]).set(secs);
+
+    info!(
+        elapsed_ms = elapsed.as_millis(),
+        "{human_phase} ({secs:.3}s since start)"
+    )
+}
+
 fn start_pageserver(
     launch_ts: &'static LaunchTimestamp,
     conf: &'static PageServerConf,
@@ -207,26 +254,17 @@ fn start_pageserver(
     // Monotonic time for later calculating startup duration
     let started_startup_at = Instant::now();
 
-    let startup_checkpoint = move |phase: &str, human_phase: &str| {
-        let elapsed = started_startup_at.elapsed();
-        let secs = elapsed.as_secs_f64();
-        STARTUP_DURATION.with_label_values(&[phase]).set(secs);
-        info!(
-            elapsed_ms = elapsed.as_millis(),
-            "{human_phase} ({secs:.3}s since start)"
-        )
-    };
-
     // Print version and launch timestamp to the log,
     // and expose them as prometheus metrics.
     // A changed version string indicates changed software.
     // A changed launch timestamp indicates a pageserver restart.
     info!(
-        "version: {} launch_timestamp: {}",
+        "version: {} launch_timestamp: {} build_tag: {}",
         version(),
-        launch_ts.to_string()
+        launch_ts.to_string(),
+        BUILD_TAG,
     );
-    set_build_info_metric(GIT_VERSION);
+    set_build_info_metric(GIT_VERSION, BUILD_TAG);
     set_launch_timestamp_metric(launch_ts);
     pageserver::preinitialize_metrics();
 
@@ -341,7 +379,7 @@ fn start_pageserver(
 
     // Up to this point no significant I/O has been done: this should have been fast.  Record
     // duration prior to starting I/O intensive phase of startup.
-    startup_checkpoint("initial", "Starting loading tenants");
+    startup_checkpoint(started_startup_at, "initial", "Starting loading tenants");
     STARTUP_IS_LOADING.set(1);
 
     // Startup staging or optimizing:
@@ -355,6 +393,7 @@ fn start_pageserver(
     // consumer side) will be dropped once we can start the background jobs. Currently it is behind
     // completing all initial logical size calculations (init_logical_size_done_rx) and a timeout
     // (background_task_maximum_delay).
+    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();
@@ -362,7 +401,8 @@ fn start_pageserver(
     let (background_jobs_can_start, background_jobs_barrier) = utils::completion::channel();
 
     let order = pageserver::InitializationOrder {
-        initial_tenant_load: Some(init_done_tx),
+        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(),
@@ -386,55 +426,93 @@ fn start_pageserver(
         let shutdown_pageserver = shutdown_pageserver.clone();
         let drive_init = async move {
             // NOTE: unlike many futures in pageserver, this one is cancellation-safe
-            let guard = scopeguard::guard_on_success((), |_| tracing::info!("Cancelled before initial load completed"));
+            let guard = scopeguard::guard_on_success((), |_| {
+                tracing::info!("Cancelled before initial load completed")
+            });
 
-            init_done_rx.wait().await;
-            startup_checkpoint("initial_tenant_load", "Initial load completed");
-            STARTUP_IS_LOADING.set(0);
+            let timeout = conf.background_task_maximum_delay;
+
+            let init_remote_done = std::pin::pin!(async {
+                init_remote_done_rx.wait().await;
+                startup_checkpoint(
+                    started_startup_at,
+                    "initial_tenant_load_remote",
+                    "Remote part of initial load completed",
+                );
+            });
+
+            let WaitForPhaseResult {
+                timeout_remaining: timeout,
+                skipped: init_remote_skipped,
+            } = wait_for_phase("initial_tenant_load_remote", init_remote_done, timeout).await;
+
+            let init_load_done = std::pin::pin!(async {
+                init_done_rx.wait().await;
+                startup_checkpoint(
+                    started_startup_at,
+                    "initial_tenant_load",
+                    "Initial load completed",
+                );
+                STARTUP_IS_LOADING.set(0);
+            });
+
+            let WaitForPhaseResult {
+                timeout_remaining: timeout,
+                skipped: init_load_skipped,
+            } = wait_for_phase("initial_tenant_load", init_load_done, timeout).await;
 
             // initial logical sizes can now start, as they were waiting on init_done_rx.
 
             scopeguard::ScopeGuard::into_inner(guard);
 
-            let mut init_sizes_done = std::pin::pin!(init_logical_size_done_rx.wait());
+            let guard = scopeguard::guard_on_success((), |_| {
+                tracing::info!("Cancelled before initial logical sizes completed")
+            });
 
-            let timeout = conf.background_task_maximum_delay;
+            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 guard = scopeguard::guard_on_success((), |_| tracing::info!("Cancelled before initial logical sizes completed"));
-
-            let init_sizes_done = match tokio::time::timeout(timeout, &mut init_sizes_done).await {
-                Ok(_) => {
-                    startup_checkpoint("initial_logical_sizes", "Initial logical sizes completed");
-                    None
-                }
-                Err(_) => {
-                    tracing::info!(
-                        timeout_millis = timeout.as_millis(),
-                        "Initial logical size timeout elapsed; starting background jobs"
-                    );
-                    Some(init_sizes_done)
-                }
-            };
+            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
+            // 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.
             drop(background_jobs_can_start);
-            startup_checkpoint("background_jobs_can_start", "Starting background jobs");
-
-            if let Some(init_sizes_done) = init_sizes_done {
-                // ending up here is not a bug; at the latest logical sizes will be queried by
-                // consumption metrics.
-                let guard = scopeguard::guard_on_success((), |_| tracing::info!("Cancelled before initial logical sizes completed"));
-                init_sizes_done.await;
-
-                scopeguard::ScopeGuard::into_inner(guard);
-
-                startup_checkpoint("initial_logical_sizes", "Initial logical sizes completed after timeout (background jobs already started)");
+            startup_checkpoint(
+                started_startup_at,
+                "background_jobs_can_start",
+                "Starting background jobs",
+            );
 
+            // We are done. If we skipped any phases due to timeout, run them to completion here so that
+            // they will eventually update their startup_checkpoint, and so that we do not declare the
+            // 'complete' stage until all the other stages are really done.
+            let guard = scopeguard::guard_on_success((), |_| {
+                tracing::info!("Cancelled before waiting for skipped phases done")
+            });
+            if let Some(f) = init_remote_skipped {
+                f.await;
             }
+            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("complete", "Startup complete");
+            startup_checkpoint(started_startup_at, "complete", "Startup complete");
         };
 
         async move {
@@ -574,6 +652,7 @@ fn start_pageserver(
                     pageserver_listener,
                     conf.pg_auth_type,
                     libpq_ctx,
+                    task_mgr::shutdown_token(),
                 )
                 .await
             },

pageserver/src/config.rs

@@ -33,8 +33,7 @@ use crate::disk_usage_eviction_task::DiskUsageEvictionTaskConfig;
 use crate::tenant::config::TenantConf;
 use crate::tenant::config::TenantConfOpt;
 use crate::tenant::{
-    TENANTS_SEGMENT_NAME, TENANT_ATTACHING_MARKER_FILENAME, TENANT_DELETED_MARKER_FILE_NAME,
-    TIMELINES_SEGMENT_NAME,
+    TENANTS_SEGMENT_NAME, TENANT_DELETED_MARKER_FILE_NAME, TIMELINES_SEGMENT_NAME,
 };
 use crate::{
     IGNORED_TENANT_FILE_NAME, METADATA_FILE_NAME, TENANT_CONFIG_NAME, TENANT_LOCATION_CONFIG_NAME,
@@ -633,11 +632,6 @@ impl PageServerConf {
         self.tenants_path().join(tenant_id.to_string())
     }
 
-    pub fn tenant_attaching_mark_file_path(&self, tenant_id: &TenantId) -> Utf8PathBuf {
-        self.tenant_path(tenant_id)
-            .join(TENANT_ATTACHING_MARKER_FILENAME)
-    }
-
     pub fn tenant_ignore_mark_file_path(&self, tenant_id: &TenantId) -> Utf8PathBuf {
         self.tenant_path(tenant_id).join(IGNORED_TENANT_FILE_NAME)
     }
@@ -886,6 +880,13 @@ impl PageServerConf {
             );
         }
 
+        if let Some(compaction_algorithm) = item.get("compaction_algorithm") {
+            t_conf.compaction_algorithm = Some(
+                deserialize_from_item("compaction_algorithm", compaction_algorithm)
+                    .context("parse compaction_algorithm")?,
+            );
+        }
+
         if let Some(gc_horizon) = item.get("gc_horizon") {
             t_conf.gc_horizon = Some(parse_toml_u64("gc_horizon", gc_horizon)?);
         }

pageserver/src/consumption_metrics.rs

@@ -2,6 +2,7 @@
 //! and push them to a HTTP endpoint.
 use crate::context::{DownloadBehavior, RequestContext};
 use crate::task_mgr::{self, TaskKind, BACKGROUND_RUNTIME};
+use crate::tenant::tasks::BackgroundLoopKind;
 use crate::tenant::{mgr, LogicalSizeCalculationCause};
 use camino::Utf8PathBuf;
 use consumption_metrics::EventType;
@@ -10,11 +11,12 @@ use reqwest::Url;
 use std::collections::HashMap;
 use std::sync::Arc;
 use std::time::{Duration, SystemTime};
+use tokio::time::Instant;
 use tracing::*;
 use utils::id::NodeId;
 
 mod metrics;
-use metrics::MetricsKey;
+use crate::consumption_metrics::metrics::MetricsKey;
 mod disk_cache;
 mod upload;
 
@@ -87,22 +89,12 @@ pub async fn collect_metrics(
 
     let node_id = node_id.to_string();
 
-    // reminder: ticker is ready immediatedly
-    let mut ticker = tokio::time::interval(metric_collection_interval);
-
     loop {
-        let tick_at = tokio::select! {
-            _ = cancel.cancelled() => return Ok(()),
-            tick_at = ticker.tick() => tick_at,
-        };
+        let started_at = Instant::now();
 
         // these are point in time, with variable "now"
         let metrics = metrics::collect_all_metrics(&cached_metrics, &ctx).await;
 
-        if metrics.is_empty() {
-            continue;
-        }
-
         let metrics = Arc::new(metrics);
 
         // why not race cancellation here? because we are one of the last tasks, and if we are
@@ -141,10 +133,19 @@ pub async fn collect_metrics(
         let (_, _) = tokio::join!(flush, upload);
 
         crate::tenant::tasks::warn_when_period_overrun(
-            tick_at.elapsed(),
+            started_at.elapsed(),
             metric_collection_interval,
-            "consumption_metrics_collect_metrics",
+            BackgroundLoopKind::ConsumptionMetricsCollectMetrics,
         );
+
+        let res = tokio::time::timeout_at(
+            started_at + metric_collection_interval,
+            task_mgr::shutdown_token().cancelled(),
+        )
+        .await;
+        if res.is_ok() {
+            return Ok(());
+        }
     }
 }
 
@@ -243,16 +244,14 @@ async fn calculate_synthetic_size_worker(
     ctx: &RequestContext,
 ) -> anyhow::Result<()> {
     info!("starting calculate_synthetic_size_worker");
+    scopeguard::defer! {
+        info!("calculate_synthetic_size_worker stopped");
+    };
 
-    // reminder: ticker is ready immediatedly
-    let mut ticker = tokio::time::interval(synthetic_size_calculation_interval);
     let cause = LogicalSizeCalculationCause::ConsumptionMetricsSyntheticSize;
 
     loop {
-        let tick_at = tokio::select! {
-            _ = task_mgr::shutdown_watcher() => return Ok(()),
-            tick_at = ticker.tick() => tick_at,
-        };
+        let started_at = Instant::now();
 
         let tenants = match mgr::list_tenants().await {
             Ok(tenants) => tenants,
@@ -268,6 +267,11 @@ async fn calculate_synthetic_size_worker(
             }
 
             if let Ok(tenant) = mgr::get_tenant(tenant_id, true).await {
+                // TODO should we use concurrent_background_tasks_rate_limit() here, like the other background tasks?
+                // We can put in some prioritization for consumption metrics.
+                // Same for the loop that fetches computed metrics.
+                // 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, ctx).await {
                     error!("failed to calculate synthetic size for tenant {tenant_id}: {e:#}");
                 }
@@ -275,9 +279,18 @@ async fn calculate_synthetic_size_worker(
         }
 
         crate::tenant::tasks::warn_when_period_overrun(
-            tick_at.elapsed(),
+            started_at.elapsed(),
             synthetic_size_calculation_interval,
-            "consumption_metrics_synthetic_size_worker",
+            BackgroundLoopKind::ConsumptionMetricsSyntheticSizeWorker,
         );
+
+        let res = tokio::time::timeout_at(
+            started_at + synthetic_size_calculation_interval,
+            task_mgr::shutdown_token().cancelled(),
+        )
+        .await;
+        if res.is_ok() {
+            return Ok(());
+        }
     }
 }

pageserver/src/control_plane_client.rs

@@ -57,7 +57,10 @@ impl ControlPlaneClient {
 
         if let Some(jwt) = &conf.control_plane_api_token {
             let mut headers = hyper::HeaderMap::new();
-            headers.insert("Authorization", jwt.get_contents().parse().unwrap());
+            headers.insert(
+                "Authorization",
+                format!("Bearer {}", jwt.get_contents()).parse().unwrap(),
+            );
             client = client.default_headers(headers);
         }
 
@@ -144,7 +147,7 @@ impl ControlPlaneGenerationsApi for ControlPlaneClient {
         Ok(response
             .tenants
             .into_iter()
-            .map(|t| (t.id, Generation::new(t.generation)))
+            .map(|t| (t.id, Generation::new(t.gen)))
             .collect::<HashMap<_, _>>())
     }
 

pageserver/src/deletion_queue.rs

@@ -10,6 +10,7 @@ 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::virtual_file::MaybeFatalIo;
 use crate::virtual_file::VirtualFile;
 use anyhow::Context;
 use camino::Utf8PathBuf;
@@ -271,7 +272,9 @@ impl DeletionHeader {
         let temp_path = path_with_suffix_extension(&header_path, TEMP_SUFFIX);
         VirtualFile::crashsafe_overwrite(&header_path, &temp_path, &header_bytes)
             .await
-            .map_err(Into::into)
+            .maybe_fatal_err("save deletion header")?;
+
+        Ok(())
     }
 }
 
@@ -360,6 +363,7 @@ impl DeletionList {
         let bytes = serde_json::to_vec(self).expect("Failed to serialize deletion list");
         VirtualFile::crashsafe_overwrite(&path, &temp_path, &bytes)
             .await
+            .maybe_fatal_err("save deletion list")
             .map_err(Into::into)
     }
 }
@@ -1298,10 +1302,6 @@ pub(crate) mod mock {
             }
         }
 
-        pub fn get_executed(&self) -> usize {
-            self.executed.load(Ordering::Relaxed)
-        }
-
         #[allow(clippy::await_holding_lock)]
         pub async fn pump(&self) {
             if let Some(remote_storage) = &self.remote_storage {

pageserver/src/deletion_queue/list_writer.rs

@@ -34,6 +34,8 @@ use crate::deletion_queue::TEMP_SUFFIX;
 use crate::metrics;
 use crate::tenant::remote_timeline_client::remote_layer_path;
 use crate::tenant::storage_layer::LayerFileName;
+use crate::virtual_file::on_fatal_io_error;
+use crate::virtual_file::MaybeFatalIo;
 
 // The number of keys in a DeletionList before we will proactively persist it
 // (without reaching a flush deadline).  This aims to deliver objects of the order
@@ -195,7 +197,7 @@ impl ListWriter {
                     debug!("Deletion header {header_path} not found, first start?");
                     Ok(None)
                 } else {
-                    Err(anyhow::anyhow!(e))
+                    on_fatal_io_error(&e, "reading deletion header");
                 }
             }
         }
@@ -216,16 +218,9 @@ impl ListWriter {
         self.pending.sequence = validated_sequence + 1;
 
         let deletion_directory = self.conf.deletion_prefix();
-        let mut dir = match tokio::fs::read_dir(&deletion_directory).await {
-            Ok(d) => d,
-            Err(e) => {
-                warn!("Failed to open deletion list directory {deletion_directory}: {e:#}");
-
-                // Give up: if we can't read the deletion list directory, we probably can't
-                // write lists into it later, so the queue won't work.
-                return Err(e.into());
-            }
-        };
+        let mut dir = tokio::fs::read_dir(&deletion_directory)
+            .await
+            .fatal_err("read deletion directory");
 
         let list_name_pattern =
             Regex::new("(?<sequence>[a-zA-Z0-9]{16})-(?<version>[a-zA-Z0-9]{2}).list").unwrap();
@@ -233,7 +228,7 @@ impl ListWriter {
         let temp_extension = format!(".{TEMP_SUFFIX}");
         let header_path = self.conf.deletion_header_path();
         let mut seqs: Vec<u64> = Vec::new();
-        while let Some(dentry) = dir.next_entry().await? {
+        while let Some(dentry) = dir.next_entry().await.fatal_err("read deletion dentry") {
             let file_name = dentry.file_name();
             let dentry_str = file_name.to_string_lossy();
 
@@ -246,11 +241,9 @@ impl ListWriter {
                 info!("Cleaning up temporary file {dentry_str}");
                 let absolute_path =
                     deletion_directory.join(dentry.file_name().to_str().expect("non-Unicode path"));
-                if let Err(e) = tokio::fs::remove_file(&absolute_path).await {
-                    // Non-fatal error: we will just leave the file behind but not
-                    // try and load it.
-                    warn!("Failed to clean up temporary file {absolute_path}: {e:#}");
-                }
+                tokio::fs::remove_file(&absolute_path)
+                    .await
+                    .fatal_err("delete temp file");
 
                 continue;
             }
@@ -290,7 +283,9 @@ impl ListWriter {
         for s in seqs {
             let list_path = self.conf.deletion_list_path(s);
 
-            let list_bytes = tokio::fs::read(&list_path).await?;
+            let list_bytes = tokio::fs::read(&list_path)
+                .await
+                .fatal_err("read deletion list");
 
             let mut deletion_list = match serde_json::from_slice::<DeletionList>(&list_bytes) {
                 Ok(l) => l,
@@ -349,7 +344,7 @@ impl ListWriter {
         info!("Started deletion frontend worker");
 
         // Synchronous, but we only do it once per process lifetime so it's tolerable
-        if let Err(e) = create_dir_all(&self.conf.deletion_prefix()) {
+        if let Err(e) = create_dir_all(self.conf.deletion_prefix()) {
             tracing::error!(
                 "Failed to create deletion list directory {}, deletions will not be executed ({e})",
                 self.conf.deletion_prefix(),

pageserver/src/deletion_queue/validator.rs

@@ -28,6 +28,7 @@ use crate::config::PageServerConf;
 use crate::control_plane_client::ControlPlaneGenerationsApi;
 use crate::control_plane_client::RetryForeverError;
 use crate::metrics;
+use crate::virtual_file::MaybeFatalIo;
 
 use super::deleter::DeleterMessage;
 use super::DeletionHeader;
@@ -287,16 +288,9 @@ where
     async fn cleanup_lists(&mut self, list_paths: Vec<Utf8PathBuf>) {
         for list_path in list_paths {
             debug!("Removing deletion list {list_path}");
-
-            if let Err(e) = tokio::fs::remove_file(&list_path).await {
-                // Unexpected: we should have permissions and nothing else should
-                // be touching these files.  We will leave the file behind.  Subsequent
-                // pageservers will try and load it again: hopefully whatever storage
-                // issue (probably permissions) has been fixed by then.
-                tracing::error!("Failed to delete {list_path}: {e:#}");
-                metrics::DELETION_QUEUE.unexpected_errors.inc();
-                break;
-            }
+            tokio::fs::remove_file(&list_path)
+                .await
+                .fatal_err("remove deletion list");
         }
     }
 

pageserver/src/disk_usage_eviction_task.rs

@@ -60,7 +60,11 @@ use utils::serde_percent::Percent;
 use crate::{
     config::PageServerConf,
     task_mgr::{self, TaskKind, BACKGROUND_RUNTIME},
-    tenant::{self, storage_layer::PersistentLayer, timeline::EvictionError, Timeline},
+    tenant::{
+        self,
+        storage_layer::{AsLayerDesc, EvictionError, Layer},
+        Timeline,
+    },
 };
 
 #[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
@@ -108,7 +112,7 @@ pub fn launch_disk_usage_global_eviction_task(
                 _ = background_jobs_barrier.wait() => { }
             };
 
-            disk_usage_eviction_task(&state, task_config, storage, &conf.tenants_path(), cancel)
+            disk_usage_eviction_task(&state, task_config, &storage, &conf.tenants_path(), cancel)
                 .await;
             Ok(())
         },
@@ -121,7 +125,7 @@ pub fn launch_disk_usage_global_eviction_task(
 async fn disk_usage_eviction_task(
     state: &State,
     task_config: &DiskUsageEvictionTaskConfig,
-    storage: GenericRemoteStorage,
+    _storage: &GenericRemoteStorage,
     tenants_dir: &Utf8Path,
     cancel: CancellationToken,
 ) {
@@ -145,14 +149,8 @@ async fn disk_usage_eviction_task(
         let start = Instant::now();
 
         async {
-            let res = disk_usage_eviction_task_iteration(
-                state,
-                task_config,
-                &storage,
-                tenants_dir,
-                &cancel,
-            )
-            .await;
+            let res =
+                disk_usage_eviction_task_iteration(state, task_config, tenants_dir, &cancel).await;
 
             match res {
                 Ok(()) => {}
@@ -183,13 +181,12 @@ pub trait Usage: Clone + Copy + std::fmt::Debug {
 async fn disk_usage_eviction_task_iteration(
     state: &State,
     task_config: &DiskUsageEvictionTaskConfig,
-    storage: &GenericRemoteStorage,
     tenants_dir: &Utf8Path,
     cancel: &CancellationToken,
 ) -> anyhow::Result<()> {
     let usage_pre = filesystem_level_usage::get(tenants_dir, task_config)
         .context("get filesystem-level disk usage before evictions")?;
-    let res = disk_usage_eviction_task_iteration_impl(state, storage, usage_pre, cancel).await;
+    let res = disk_usage_eviction_task_iteration_impl(state, usage_pre, cancel).await;
     match res {
         Ok(outcome) => {
             debug!(?outcome, "disk_usage_eviction_iteration finished");
@@ -273,7 +270,6 @@ struct LayerCount {
 
 pub async fn disk_usage_eviction_task_iteration_impl<U: Usage>(
     state: &State,
-    storage: &GenericRemoteStorage,
     usage_pre: U,
     cancel: &CancellationToken,
 ) -> anyhow::Result<IterationOutcome<U>> {
@@ -330,9 +326,10 @@ pub async fn disk_usage_eviction_task_iteration_impl<U: Usage>(
     // If we get far enough in the list that we start to evict layers that are below
     // the tenant's min-resident-size threshold, print a warning, and memorize the disk
     // usage at that point, in 'usage_planned_min_resident_size_respecting'.
-    let mut batched: HashMap<_, Vec<Arc<dyn PersistentLayer>>> = HashMap::new();
+    let mut batched: HashMap<_, Vec<_>> = HashMap::new();
     let mut warned = None;
     let mut usage_planned = usage_pre;
+    let mut max_batch_size = 0;
     for (i, (partition, candidate)) in candidates.into_iter().enumerate() {
         if !usage_planned.has_pressure() {
             debug!(
@@ -349,10 +346,18 @@ pub async fn disk_usage_eviction_task_iteration_impl<U: Usage>(
 
         usage_planned.add_available_bytes(candidate.layer.layer_desc().file_size);
 
-        batched
-            .entry(TimelineKey(candidate.timeline))
-            .or_default()
-            .push(candidate.layer);
+        // FIXME: batching makes no sense anymore because of no layermap locking, should just spawn
+        // tasks to evict all seen layers until we have evicted enough
+
+        let batch = batched.entry(TimelineKey(candidate.timeline)).or_default();
+
+        // semaphore will later be used to limit eviction concurrency, and we can express at
+        // most u32 number of permits. unlikely we would have u32::MAX layers to be evicted,
+        // but fail gracefully by not making batches larger.
+        if batch.len() < u32::MAX as usize {
+            batch.push(candidate.layer);
+            max_batch_size = max_batch_size.max(batch.len());
+        }
     }
 
     let usage_planned = match warned {
@@ -369,64 +374,101 @@ pub async fn disk_usage_eviction_task_iteration_impl<U: Usage>(
 
     // phase2: evict victims batched by timeline
 
-    // After the loop, `usage_assumed` is the post-eviction usage,
-    // according to internal accounting.
-    let mut usage_assumed = usage_pre;
-    let mut evictions_failed = LayerCount::default();
+    let mut js = tokio::task::JoinSet::new();
+
+    // ratelimit to 1k files or any higher max batch size
+    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 timeline_id = timeline.timeline_id;
-        let batch_size = batch.len();
+        let batch_size =
+            u32::try_from(batch.len()).expect("batch size limited to u32::MAX during partitioning");
+
+        // I dislike naming of `available_permits` but it means current total amount of permits
+        // because permits can be added
+        assert!(batch_size as usize <= limit.available_permits());
 
         debug!(%timeline_id, "evicting batch for timeline");
 
-        async {
-            let results = timeline.evict_layers(storage, &batch, cancel.clone()).await;
+        let evict = {
+            let limit = limit.clone();
+            let cancel = cancel.clone();
+            async move {
+                let mut evicted_bytes = 0;
+                let mut evictions_failed = LayerCount::default();
 
-            match results {
-                Err(e) => {
-                    warn!("failed to evict batch: {:#}", e);
-                }
-                Ok(results) => {
-                    assert_eq!(results.len(), batch.len());
-                    for (result, layer) in results.into_iter().zip(batch.iter()) {
-                        let file_size = layer.layer_desc().file_size;
-                        match result {
-                            Some(Ok(())) => {
-                                usage_assumed.add_available_bytes(file_size);
-                            }
-                            Some(Err(EvictionError::CannotEvictRemoteLayer)) => {
-                                unreachable!("get_local_layers_for_disk_usage_eviction finds only local layers")
-                            }
-                            Some(Err(EvictionError::FileNotFound)) => {
-                                evictions_failed.file_sizes += file_size;
-                                evictions_failed.count += 1;
-                            }
-                            Some(Err(
-                                e @ EvictionError::LayerNotFound(_)
-                                | e @ EvictionError::StatFailed(_),
-                            )) => {
-                                let e = utils::error::report_compact_sources(&e);
-                                warn!(%layer, "failed to evict layer: {e}");
-                                evictions_failed.file_sizes += file_size;
-                                evictions_failed.count += 1;
-                            }
-                            None => {
-                                assert!(cancel.is_cancelled());
-                                return;
+                let Ok(_permit) = limit.acquire_many_owned(batch_size).await else {
+                    // semaphore closing means cancelled
+                    return (evicted_bytes, evictions_failed);
+                };
+
+                let results = timeline.evict_layers(&batch, &cancel).await;
+
+                match results {
+                    Ok(results) => {
+                        assert_eq!(results.len(), batch.len());
+                        for (result, layer) in results.into_iter().zip(batch.iter()) {
+                            let file_size = layer.layer_desc().file_size;
+                            match result {
+                                Some(Ok(())) => {
+                                    evicted_bytes += file_size;
+                                }
+                                Some(Err(EvictionError::NotFound | EvictionError::Downloaded)) => {
+                                    evictions_failed.file_sizes += file_size;
+                                    evictions_failed.count += 1;
+                                }
+                                None => {
+                                    assert!(cancel.is_cancelled());
+                                }
                             }
                         }
                     }
+                    Err(e) => {
+                        warn!("failed to evict batch: {:#}", e);
+                    }
                 }
+                (evicted_bytes, evictions_failed)
             }
         }
-        .instrument(tracing::info_span!("evict_batch", %tenant_id, %timeline_id, batch_size))
-        .await;
+        .instrument(tracing::info_span!("evict_batch", %tenant_id, %timeline_id, batch_size));
 
-        if cancel.is_cancelled() {
+        js.spawn(evict);
+
+        // spwaning multiple thousands of these is essentially blocking, so give already spawned a
+        // chance of making progress
+        tokio::task::yield_now().await;
+    }
+
+    let join_all = async move {
+        // After the evictions, `usage_assumed` is the post-eviction usage,
+        // according to internal accounting.
+        let mut usage_assumed = usage_pre;
+        let mut evictions_failed = LayerCount::default();
+
+        while let Some(res) = js.join_next().await {
+            match res {
+                Ok((evicted_bytes, failed)) => {
+                    usage_assumed.add_available_bytes(evicted_bytes);
+                    evictions_failed.file_sizes += failed.file_sizes;
+                    evictions_failed.count += failed.count;
+                }
+                Err(je) if je.is_cancelled() => unreachable!("not used"),
+                Err(je) if je.is_panic() => { /* already logged */ }
+                Err(je) => tracing::error!("unknown JoinError: {je:?}"),
+            }
+        }
+        (usage_assumed, evictions_failed)
+    };
+
+    let (usage_assumed, evictions_failed) = tokio::select! {
+        tuple = join_all => { tuple },
+        _ = cancel.cancelled() => {
+            // close the semaphore to stop any pending acquires
+            limit.close();
             return Ok(IterationOutcome::Cancelled);
         }
-    }
+    };
 
     Ok(IterationOutcome::Finished(IterationOutcomeFinished {
         before: usage_pre,
@@ -441,7 +483,7 @@ pub async fn disk_usage_eviction_task_iteration_impl<U: Usage>(
 #[derive(Clone)]
 struct EvictionCandidate {
     timeline: Arc<Timeline>,
-    layer: Arc<dyn PersistentLayer>,
+    layer: Layer,
     last_activity_ts: SystemTime,
 }
 

pageserver/src/http/openapi_spec.yml

@@ -306,6 +306,67 @@ paths:
               schema:
                 $ref: "#/components/schemas/ServiceUnavailableError"
 
+  /v1/tenant/{tenant_id}/timeline/{timeline_id}/get_timestamp_of_lsn:
+    parameters:
+      - name: tenant_id
+        in: path
+        required: true
+        schema:
+          type: string
+          format: hex
+      - name: timeline_id
+        in: path
+        required: true
+        schema:
+          type: string
+          format: hex
+    get:
+      description: Get timestamp for a given LSN
+      parameters:
+        - name: lsn
+          in: query
+          required: true
+          schema:
+            type: integer
+          description: A LSN to get the timestamp
+      responses:
+        "200":
+          description: OK
+          content:
+            application/json:
+              schema:
+                type: string
+                format: date-time
+        "400":
+          description: Error when no tenant id found in path, no timeline id or invalid timestamp
+          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"
+        "404":
+          description: Timeline not found, or there is no timestamp information for the given lsn
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/NotFoundError"
+        "500":
+          description: Generic operation error
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/Error"
 
   /v1/tenant/{tenant_id}/timeline/{timeline_id}/get_lsn_by_timestamp:
     parameters:
@@ -331,13 +392,19 @@ paths:
             type: string
             format: date-time
           description: A timestamp to get the LSN
+        - name: version
+          in: query
+          required: false
+          schema:
+            type: integer
+          description: The version of the endpoint to use
       responses:
         "200":
           description: OK
           content:
             application/json:
               schema:
-                type: string
+                $ref: "#/components/schemas/LsnByTimestampResponse"
         "400":
           description: Error when no tenant id found in path, no timeline id or invalid timestamp
           content:
@@ -502,7 +569,17 @@ paths:
               schema:
                 $ref: "#/components/schemas/NotFoundError"
         "409":
-          description: Tenant download is already in progress
+          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:
@@ -646,6 +723,12 @@ paths:
 
         Errors if the tenant is absent on disk, already present in memory or fails to schedule its load.
         Scheduling a load does not mean that the tenant would load successfully, check tenant status to ensure load correctness.
+      requestBody:
+        required: false
+        content:
+          application/json:
+            schema:
+              $ref: "#/components/schemas/TenantLoadRequest"
       responses:
         "202":
           description: Tenant scheduled to load successfully
@@ -1136,6 +1219,15 @@ components:
             new_tenant_id:
               type: string
               format: hex
+            generation:
+              type: integer
+              description: Attachment generation number.
+    TenantLoadRequest:
+      type: object
+      properties:
+        generation:
+          type: integer
+          description: Attachment generation number.
     TenantAttachRequest:
       type: object
       required:
@@ -1323,6 +1415,19 @@ components:
           type: string
           format: hex
 
+    LsnByTimestampResponse:
+      type: object
+      required:
+        - lsn
+        - kind
+      properties:
+        lsn:
+          type: string
+          format: hex
+        kind:
+          type: string
+          enum: [past, present, future, nodata]
+
     Error:
       type: object
       required:

pageserver/src/http/routes.rs

@@ -2,11 +2,13 @@
 //! Management HTTP API
 //!
 use std::collections::HashMap;
+use std::str::FromStr;
 use std::sync::Arc;
 
 use anyhow::{anyhow, Context, Result};
 use futures::TryFutureExt;
-use hyper::header::CONTENT_TYPE;
+use humantime::format_rfc3339;
+use hyper::header;
 use hyper::StatusCode;
 use hyper::{Body, Request, Response, Uri};
 use metrics::launch_timestamp::LaunchTimestamp;
@@ -15,6 +17,7 @@ use pageserver_api::models::{
     TenantLoadRequest, TenantLocationConfigRequest,
 };
 use remote_storage::GenericRemoteStorage;
+use serde_with::{serde_as, DisplayFromStr};
 use tenant_size_model::{SizeResult, StorageModel};
 use tokio_util::sync::CancellationToken;
 use tracing::*;
@@ -136,9 +139,7 @@ impl From<PageReconstructError> for ApiError {
             PageReconstructError::AncestorStopping(_) => {
                 ApiError::ResourceUnavailable(format!("{pre}").into())
             }
-            PageReconstructError::WalRedo(pre) => {
-                ApiError::InternalServerError(anyhow::Error::new(pre))
-            }
+            PageReconstructError::WalRedo(pre) => ApiError::InternalServerError(pre),
         }
     }
 }
@@ -484,6 +485,8 @@ async fn get_lsn_by_timestamp_handler(
     let tenant_id: TenantId = parse_request_param(&request, "tenant_id")?;
     check_permission(&request, Some(tenant_id))?;
 
+    let version: Option<u8> = parse_query_param(&request, "version")?;
+
     let timeline_id: TimelineId = parse_request_param(&request, "timeline_id")?;
     let timestamp_raw = must_get_query_param(&request, "timestamp")?;
     let timestamp = humantime::parse_rfc3339(&timestamp_raw)
@@ -495,13 +498,59 @@ async fn get_lsn_by_timestamp_handler(
     let timeline = active_timeline_of_active_tenant(tenant_id, timeline_id).await?;
     let result = timeline.find_lsn_for_timestamp(timestamp_pg, &ctx).await?;
 
-    let result = match result {
-        LsnForTimestamp::Present(lsn) => format!("{lsn}"),
-        LsnForTimestamp::Future(_lsn) => "future".into(),
-        LsnForTimestamp::Past(_lsn) => "past".into(),
-        LsnForTimestamp::NoData(_lsn) => "nodata".into(),
-    };
-    json_response(StatusCode::OK, result)
+    if version.unwrap_or(0) > 1 {
+        #[serde_as]
+        #[derive(serde::Serialize)]
+        struct Result {
+            #[serde_as(as = "DisplayFromStr")]
+            lsn: Lsn,
+            kind: &'static str,
+        }
+        let (lsn, kind) = match result {
+            LsnForTimestamp::Present(lsn) => (lsn, "present"),
+            LsnForTimestamp::Future(lsn) => (lsn, "future"),
+            LsnForTimestamp::Past(lsn) => (lsn, "past"),
+            LsnForTimestamp::NoData(lsn) => (lsn, "nodata"),
+        };
+        json_response(StatusCode::OK, Result { lsn, kind })
+    } else {
+        // FIXME: this is a temporary crutch not to break backwards compatibility
+        // See https://github.com/neondatabase/neon/pull/5608
+        let result = match result {
+            LsnForTimestamp::Present(lsn) => format!("{lsn}"),
+            LsnForTimestamp::Future(_lsn) => "future".into(),
+            LsnForTimestamp::Past(_lsn) => "past".into(),
+            LsnForTimestamp::NoData(_lsn) => "nodata".into(),
+        };
+        json_response(StatusCode::OK, result)
+    }
+}
+
+async fn get_timestamp_of_lsn_handler(
+    request: Request<Body>,
+    _cancel: CancellationToken,
+) -> Result<Response<Body>, ApiError> {
+    let tenant_id: TenantId = parse_request_param(&request, "tenant_id")?;
+    check_permission(&request, Some(tenant_id))?;
+
+    let timeline_id: TimelineId = parse_request_param(&request, "timeline_id")?;
+
+    let lsn_str = must_get_query_param(&request, "lsn")?;
+    let lsn = Lsn::from_str(&lsn_str)
+        .with_context(|| format!("Invalid LSN: {lsn_str:?}"))
+        .map_err(ApiError::BadRequest)?;
+
+    let ctx = RequestContext::new(TaskKind::MgmtRequest, DownloadBehavior::Download);
+    let timeline = active_timeline_of_active_tenant(tenant_id, timeline_id).await?;
+    let result = timeline.get_timestamp_for_lsn(lsn, &ctx).await?;
+
+    match result {
+        Some(time) => {
+            let time = format_rfc3339(postgres_ffi::from_pg_timestamp(time)).to_string();
+            json_response(StatusCode::OK, time)
+        }
+        None => json_response(StatusCode::NOT_FOUND, ()),
+    }
 }
 
 async fn tenant_attach_handler(
@@ -740,6 +789,10 @@ async fn tenant_size_handler(
         .map_err(ApiError::InternalServerError)?;
 
     let mut sizes = None;
+    let accepts_html = headers
+        .get(header::ACCEPT)
+        .map(|v| v == "text/html")
+        .unwrap_or_default();
     if !inputs_only.unwrap_or(false) {
         let storage_model = inputs
             .calculate_model()
@@ -747,11 +800,11 @@ async fn tenant_size_handler(
         let size = storage_model.calculate();
 
         // If request header expects html, return html
-        if headers["Accept"] == "text/html" {
+        if accepts_html {
             return synthetic_size_html_response(inputs, storage_model, size);
         }
         sizes = Some(size);
-    } else if headers["Accept"] == "text/html" {
+    } else if accepts_html {
         return Err(ApiError::BadRequest(anyhow!(
             "inputs_only parameter is incompatible with html output request"
         )));
@@ -902,7 +955,7 @@ fn synthetic_size_html_response(
 pub fn html_response(status: StatusCode, data: String) -> Result<Response<Body>, ApiError> {
     let response = Response::builder()
         .status(status)
-        .header(hyper::header::CONTENT_TYPE, "text/html")
+        .header(header::CONTENT_TYPE, "text/html")
         .body(Body::from(data.as_bytes().to_vec()))
         .map_err(|e| ApiError::InternalServerError(e.into()))?;
     Ok(response)
@@ -1036,9 +1089,17 @@ async fn put_tenant_location_config_handler(
     // The `Detached` state is special, it doesn't upsert a tenant, it removes
     // its local disk content and drops it from memory.
     if let LocationConfigMode::Detached = request_data.config.mode {
-        mgr::detach_tenant(conf, tenant_id, true, &state.deletion_queue_client)
+        if let Err(e) = mgr::detach_tenant(conf, tenant_id, true, &state.deletion_queue_client)
             .instrument(info_span!("tenant_detach", %tenant_id))
-            .await?;
+            .await
+        {
+            match e {
+                TenantStateError::NotFound(_) => {
+                    // This API is idempotent: a NotFound on a detach is fine.
+                }
+                _ => return Err(e.into()),
+            }
+        }
         return json_response(StatusCode::OK, ());
     }
 
@@ -1144,7 +1205,7 @@ async fn timeline_compact_handler(
         timeline
             .compact(&cancel, &ctx)
             .await
-            .map_err(ApiError::InternalServerError)?;
+            .map_err(|e| ApiError::InternalServerError(e.into()))?;
         json_response(StatusCode::OK, ())
     }
     .instrument(info_span!("manual_compaction", %tenant_id, %timeline_id))
@@ -1169,7 +1230,7 @@ async fn timeline_checkpoint_handler(
         timeline
             .compact(&cancel, &ctx)
             .await
-            .map_err(ApiError::InternalServerError)?;
+            .map_err(|e| ApiError::InternalServerError(e.into()))?;
 
         json_response(StatusCode::OK, ())
     }
@@ -1275,7 +1336,7 @@ async fn getpage_at_lsn_handler(
         Result::<_, ApiError>::Ok(
             Response::builder()
                 .status(StatusCode::OK)
-                .header(CONTENT_TYPE, "application/octet-stream")
+                .header(header::CONTENT_TYPE, "application/octet-stream")
                 .body(hyper::Body::from(page))
                 .unwrap(),
         )
@@ -1439,11 +1500,11 @@ async fn disk_usage_eviction_run(
 
     let state = get_state(&r);
 
-    let Some(storage) = state.remote_storage.clone() else {
+    if state.remote_storage.as_ref().is_none() {
         return Err(ApiError::InternalServerError(anyhow::anyhow!(
             "remote storage not configured, cannot run eviction iteration"
         )));
-    };
+    }
 
     let state = state.disk_usage_eviction_state.clone();
 
@@ -1461,7 +1522,6 @@ async fn disk_usage_eviction_run(
         async move {
             let res = crate::disk_usage_eviction_task::disk_usage_eviction_task_iteration_impl(
                 &state,
-                &storage,
                 usage,
                 &child_cancel,
             )
@@ -1674,6 +1734,10 @@ pub fn make_router(
             "/v1/tenant/:tenant_id/timeline/:timeline_id/get_lsn_by_timestamp",
             |r| api_handler(r, get_lsn_by_timestamp_handler),
         )
+        .get(
+            "/v1/tenant/:tenant_id/timeline/:timeline_id/get_timestamp_of_lsn",
+            |r| api_handler(r, get_timestamp_of_lsn_handler),
+        )
         .put("/v1/tenant/:tenant_id/timeline/:timeline_id/do_gc", |r| {
             api_handler(r, timeline_gc_handler)
         })

pageserver/src/lib.rs

@@ -149,6 +149,10 @@ fn ends_with_suffix(path: &Utf8Path, suffix: &str) -> bool {
     }
 }
 
+// FIXME: DO NOT ADD new query methods like this, which will have a next step of parsing timelineid
+// from the directory name. Instead create type "UninitMark(TimelineId)" and only parse it once
+// from the name.
+
 pub fn is_uninit_mark(path: &Utf8Path) -> bool {
     ends_with_suffix(path, TIMELINE_UNINIT_MARK_SUFFIX)
 }
@@ -173,6 +177,9 @@ fn is_walkdir_io_not_found(e: &walkdir::Error) -> bool {
 /// delaying is needed.
 #[derive(Clone)]
 pub struct InitializationOrder {
+    /// Each initial tenant load task carries this until it is done loading timelines from remote storage
+    pub initial_tenant_load_remote: Option<utils::completion::Completion>,
+
     /// Each initial tenant load task carries this until completion.
     pub initial_tenant_load: Option<utils::completion::Completion>,
 

pageserver/src/metrics.rs

@@ -1067,6 +1067,26 @@ pub(crate) static TENANT_TASK_EVENTS: Lazy<IntCounterVec> = Lazy::new(|| {
     .expect("Failed to register tenant_task_events metric")
 });
 
+pub(crate) static BACKGROUND_LOOP_SEMAPHORE_WAIT_START_COUNT: Lazy<IntCounterVec> =
+    Lazy::new(|| {
+        register_int_counter_vec!(
+            "pageserver_background_loop_semaphore_wait_start_count",
+            "Counter for background loop concurrency-limiting semaphore acquire calls started",
+            &["task"],
+        )
+        .unwrap()
+    });
+
+pub(crate) static BACKGROUND_LOOP_SEMAPHORE_WAIT_FINISH_COUNT: Lazy<IntCounterVec> =
+    Lazy::new(|| {
+        register_int_counter_vec!(
+            "pageserver_background_loop_semaphore_wait_finish_count",
+            "Counter for background loop concurrency-limiting semaphore acquire calls finished",
+            &["task"],
+        )
+        .unwrap()
+    });
+
 pub(crate) static BACKGROUND_LOOP_PERIOD_OVERRUN_COUNT: Lazy<IntCounterVec> = Lazy::new(|| {
     register_int_counter_vec!(
         "pageserver_background_loop_period_overrun_count",
@@ -1368,28 +1388,23 @@ impl TimelineMetrics {
         }
     }
 
-    pub fn record_new_file_metrics(&self, sz: u64) {
+    pub(crate) fn record_new_file_metrics(&self, sz: u64) {
         self.resident_physical_size_add(sz);
         self.num_persistent_files_created.inc_by(1);
         self.persistent_bytes_written.inc_by(sz);
     }
 
-    pub fn resident_physical_size_sub(&self, sz: u64) {
+    pub(crate) fn resident_physical_size_sub(&self, sz: u64) {
         self.resident_physical_size_gauge.sub(sz);
         crate::metrics::RESIDENT_PHYSICAL_SIZE_GLOBAL.sub(sz);
     }
 
-    pub fn resident_physical_size_add(&self, sz: u64) {
+    pub(crate) fn resident_physical_size_add(&self, sz: u64) {
         self.resident_physical_size_gauge.add(sz);
         crate::metrics::RESIDENT_PHYSICAL_SIZE_GLOBAL.add(sz);
     }
 
-    pub fn resident_physical_size_set(&self, sz: u64) {
-        self.resident_physical_size_gauge.set(sz);
-        crate::metrics::RESIDENT_PHYSICAL_SIZE_GLOBAL.set(sz);
-    }
-
-    pub fn resident_physical_size_get(&self) -> u64 {
+    pub(crate) fn resident_physical_size_get(&self) -> u64 {
         self.resident_physical_size_gauge.get()
     }
 }

pageserver/src/page_cache.rs

@@ -318,15 +318,6 @@ impl std::ops::Deref for PageWriteGuard<'_> {
     }
 }
 
-impl AsMut<[u8; PAGE_SZ]> for PageWriteGuard<'_> {
-    fn as_mut(&mut self) -> &mut [u8; PAGE_SZ] {
-        match &mut self.state {
-            PageWriteGuardState::Invalid { inner, _permit } => inner.buf,
-            PageWriteGuardState::Downgraded => unreachable!(),
-        }
-    }
-}
-
 impl<'a> PageWriteGuard<'a> {
     /// Mark that the buffer contents are now valid.
     #[must_use]

pageserver/src/page_service.rs

@@ -122,6 +122,7 @@ pub async fn libpq_listener_main(
     listener: TcpListener,
     auth_type: AuthType,
     listener_ctx: RequestContext,
+    cancel: CancellationToken,
 ) -> anyhow::Result<()> {
     listener.set_nonblocking(true)?;
     let tokio_listener = tokio::net::TcpListener::from_std(listener)?;
@@ -130,7 +131,7 @@ pub async fn libpq_listener_main(
     while let Some(res) = tokio::select! {
         biased;
 
-        _ = task_mgr::shutdown_watcher() => {
+        _ = cancel.cancelled() => {
             // We were requested to shut down.
             None
         }
@@ -299,7 +300,7 @@ impl PageServerHandler {
                 Ok(flush_r?)
             },
             _ = self.cancel.cancelled() => {
-                Err(QueryError::Other(anyhow::anyhow!("Shutting down")))
+                Err(QueryError::Shutdown)
             }
         )
     }
@@ -316,11 +317,11 @@ impl PageServerHandler {
                 let msg = tokio::select! {
                     biased;
 
-                    _ = task_mgr::shutdown_watcher() => {
+                    _ = self.cancel.cancelled() => {
                         // We were requested to shut down.
                         let msg = "pageserver is shutting down";
                         let _ = pgb.write_message_noflush(&BeMessage::ErrorResponse(msg, None));
-                        Err(QueryError::Other(anyhow::anyhow!(msg)))
+                        Err(QueryError::Shutdown)
                     }
 
                     msg = pgb.read_message() => { msg.map_err(QueryError::from)}
@@ -414,10 +415,10 @@ impl PageServerHandler {
             let msg = tokio::select! {
                 biased;
 
-                _ = task_mgr::shutdown_watcher() => {
+                _ = self.cancel.cancelled() => {
                     // We were requested to shut down.
                     info!("shutdown request received in page handler");
-                    break;
+                    return Err(QueryError::Shutdown)
                 }
 
                 msg = pgb.read_message() => { msg }

pageserver/src/pgdatadir_mapping.rs

@@ -19,6 +19,7 @@ use postgres_ffi::BLCKSZ;
 use postgres_ffi::{Oid, TimestampTz, TransactionId};
 use serde::{Deserialize, Serialize};
 use std::collections::{hash_map, HashMap, HashSet};
+use std::ops::ControlFlow;
 use std::ops::Range;
 use tokio_util::sync::CancellationToken;
 use tracing::{debug, trace, warn};
@@ -370,7 +371,6 @@ impl Timeline {
         }
     }
 
-    ///
     /// Subroutine of find_lsn_for_timestamp(). Returns true, if there are any
     /// commits that committed after 'search_timestamp', at LSN 'probe_lsn'.
     ///
@@ -385,6 +385,50 @@ impl Timeline {
         found_larger: &mut bool,
         ctx: &RequestContext,
     ) -> Result<bool, PageReconstructError> {
+        self.map_all_timestamps(probe_lsn, ctx, |timestamp| {
+            if timestamp >= search_timestamp {
+                *found_larger = true;
+                return ControlFlow::Break(true);
+            } else {
+                *found_smaller = true;
+            }
+            ControlFlow::Continue(())
+        })
+        .await
+    }
+
+    /// Obtain the possible timestamp range for the given lsn.
+    ///
+    /// If the lsn has no timestamps, returns None. returns `(min, max, median)` if it has timestamps.
+    pub async fn get_timestamp_for_lsn(
+        &self,
+        probe_lsn: Lsn,
+        ctx: &RequestContext,
+    ) -> Result<Option<TimestampTz>, PageReconstructError> {
+        let mut max: Option<TimestampTz> = None;
+        self.map_all_timestamps(probe_lsn, ctx, |timestamp| {
+            if let Some(max_prev) = max {
+                max = Some(max_prev.max(timestamp));
+            } else {
+                max = Some(timestamp);
+            }
+            ControlFlow::Continue(())
+        })
+        .await?;
+
+        Ok(max)
+    }
+
+    /// Runs the given function on all the timestamps for a given lsn
+    ///
+    /// The return value is either given by the closure, or set to the `Default`
+    /// impl's output.
+    async fn map_all_timestamps<T: Default>(
+        &self,
+        probe_lsn: Lsn,
+        ctx: &RequestContext,
+        mut f: impl FnMut(TimestampTz) -> ControlFlow<T>,
+    ) -> Result<T, PageReconstructError> {
         for segno in self
             .list_slru_segments(SlruKind::Clog, probe_lsn, ctx)
             .await?
@@ -402,16 +446,14 @@ impl Timeline {
                     timestamp_bytes.copy_from_slice(&clog_page[BLCKSZ as usize..]);
                     let timestamp = TimestampTz::from_be_bytes(timestamp_bytes);
 
-                    if timestamp >= search_timestamp {
-                        *found_larger = true;
-                        return Ok(true);
-                    } else {
-                        *found_smaller = true;
+                    match f(timestamp) {
+                        ControlFlow::Break(b) => return Ok(b),
+                        ControlFlow::Continue(()) => (),
                     }
                 }
             }
         }
-        Ok(false)
+        Ok(Default::default())
     }
 
     /// Get a list of SLRU segments
@@ -499,6 +541,24 @@ impl Timeline {
         self.get(CHECKPOINT_KEY, lsn, ctx).await
     }
 
+    pub async fn list_aux_files(
+        &self,
+        lsn: Lsn,
+        ctx: &RequestContext,
+    ) -> Result<HashMap<String, Bytes>, PageReconstructError> {
+        match self.get(AUX_FILES_KEY, lsn, ctx).await {
+            Ok(buf) => match AuxFilesDirectory::des(&buf).context("deserialization failure") {
+                Ok(dir) => Ok(dir.files),
+                Err(e) => Err(PageReconstructError::from(e)),
+            },
+            Err(e) => {
+                // This is expected: historical databases do not have the key.
+                debug!("Failed to get info about AUX files: {}", e);
+                Ok(HashMap::new())
+            }
+        }
+    }
+
     /// Does the same as get_current_logical_size but counted on demand.
     /// Used to initialize the logical size tracking on startup.
     ///
@@ -616,7 +676,9 @@ impl Timeline {
 
         result.add_key(CONTROLFILE_KEY);
         result.add_key(CHECKPOINT_KEY);
-
+        if self.get(AUX_FILES_KEY, lsn, ctx).await.is_ok() {
+            result.add_key(AUX_FILES_KEY);
+        }
         Ok(result.to_keyspace())
     }
 
@@ -692,6 +754,12 @@ impl<'a> DatadirModification<'a> {
         })?;
         self.put(DBDIR_KEY, Value::Image(buf.into()));
 
+        // Create AuxFilesDirectory
+        let buf = AuxFilesDirectory::ser(&AuxFilesDirectory {
+            files: HashMap::new(),
+        })?;
+        self.put(AUX_FILES_KEY, Value::Image(Bytes::from(buf)));
+
         let buf = TwoPhaseDirectory::ser(&TwoPhaseDirectory {
             xids: HashSet::new(),
         })?;
@@ -796,6 +864,12 @@ impl<'a> DatadirModification<'a> {
             // 'true', now write the updated 'dbdirs' map back.
             let buf = DbDirectory::ser(&dbdir)?;
             self.put(DBDIR_KEY, Value::Image(buf.into()));
+
+            // Create AuxFilesDirectory as well
+            let buf = AuxFilesDirectory::ser(&AuxFilesDirectory {
+                files: HashMap::new(),
+            })?;
+            self.put(AUX_FILES_KEY, Value::Image(Bytes::from(buf)));
         }
         if r.is_none() {
             // Create RelDirectory
@@ -1120,6 +1194,37 @@ impl<'a> DatadirModification<'a> {
         Ok(())
     }
 
+    pub async fn put_file(
+        &mut self,
+        path: &str,
+        content: &[u8],
+        ctx: &RequestContext,
+    ) -> anyhow::Result<()> {
+        let mut dir = match self.get(AUX_FILES_KEY, ctx).await {
+            Ok(buf) => AuxFilesDirectory::des(&buf)?,
+            Err(e) => {
+                // This is expected: historical databases do not have the key.
+                debug!("Failed to get info about AUX files: {}", e);
+                AuxFilesDirectory {
+                    files: HashMap::new(),
+                }
+            }
+        };
+        let path = path.to_string();
+        if content.is_empty() {
+            dir.files.remove(&path);
+        } else {
+            dir.files.insert(path, Bytes::copy_from_slice(content));
+        }
+        self.put(
+            AUX_FILES_KEY,
+            Value::Image(Bytes::from(
+                AuxFilesDirectory::ser(&dir).context("serialize")?,
+            )),
+        );
+        Ok(())
+    }
+
     ///
     /// Flush changes accumulated so far to the underlying repository.
     ///
@@ -1255,6 +1360,11 @@ struct RelDirectory {
     rels: HashSet<(Oid, u8)>,
 }
 
+#[derive(Debug, Serialize, Deserialize, Default)]
+struct AuxFilesDirectory {
+    files: HashMap<String, Bytes>,
+}
+
 #[derive(Debug, Serialize, Deserialize)]
 struct RelSizeEntry {
     nblocks: u32,
@@ -1303,10 +1413,12 @@ static ZERO_PAGE: Bytes = Bytes::from_static(&[0u8; BLCKSZ as usize]);
 // 02 pg_twophase
 //
 // 03 misc
-//    controlfile
+//    Controlfile
 //    checkpoint
 //    pg_version
 //
+// 04 aux files
+//
 // Below is a full list of the keyspace allocation:
 //
 // DbDir:
@@ -1344,6 +1456,11 @@ static ZERO_PAGE: Bytes = Bytes::from_static(&[0u8; BLCKSZ as usize]);
 //
 // Checkpoint:
 // 03 00000000 00000000 00000000 00   00000001
+//
+// AuxFiles:
+// 03 00000000 00000000 00000000 00   00000002
+//
+
 //-- Section 01: relation data and metadata
 
 const DBDIR_KEY: Key = Key {
@@ -1567,6 +1684,15 @@ const CHECKPOINT_KEY: Key = Key {
     field6: 1,
 };
 
+const AUX_FILES_KEY: Key = Key {
+    field1: 0x03,
+    field2: 0,
+    field3: 0,
+    field4: 0,
+    field5: 0,
+    field6: 2,
+};
+
 // Reverse mappings for a few Keys.
 // These are needed by WAL redo manager.
 

pageserver/src/repository.rs

@@ -172,6 +172,21 @@ impl Key {
     }
 }
 
+impl pageserver_compaction::interface::CompactionKey for Key {
+    const MIN: Self = Self::MIN;
+    const MAX: Self = Self::MAX;
+
+    fn key_range_size(r: &std::ops::Range<Self>) -> u32 {
+        key_range_size(r)
+    }
+    fn next(&self) -> Key {
+        (self as &Key).next()
+    }
+    fn skip_some(&self) -> Key {
+        self.add(128)
+    }
+}
+
 /// A 'value' stored for a one Key.
 #[derive(Debug, Clone, Serialize, Deserialize)]
 pub enum Value {

pageserver/src/tenant/config.rs

@@ -23,12 +23,17 @@ pub mod defaults {
     pub const DEFAULT_CHECKPOINT_DISTANCE: u64 = 256 * 1024 * 1024;
     pub const DEFAULT_CHECKPOINT_TIMEOUT: &str = "10 m";
 
+    // FIXME the below configs are only used by legacy algorithm. The new algorithm
+    // has different parameters.
+
     // Target file size, when creating image and delta layers.
     // This parameter determines L1 layer file size.
     pub const DEFAULT_COMPACTION_TARGET_SIZE: u64 = 128 * 1024 * 1024;
 
     pub const DEFAULT_COMPACTION_PERIOD: &str = "20 s";
     pub const DEFAULT_COMPACTION_THRESHOLD: usize = 10;
+    pub const DEFAULT_COMPACTION_ALGORITHM: super::CompactionAlgorithm =
+        super::CompactionAlgorithm::Legacy;
 
     pub const DEFAULT_GC_HORIZON: u64 = 64 * 1024 * 1024;
 
@@ -270,6 +275,7 @@ pub struct TenantConf {
     pub compaction_period: Duration,
     // Level0 delta layer threshold for compaction.
     pub compaction_threshold: usize,
+    pub compaction_algorithm: CompactionAlgorithm,
     // Determines how much history is retained, to allow
     // branching and read replicas at an older point in time.
     // The unit is #of bytes of WAL.
@@ -333,6 +339,10 @@ pub struct TenantConfOpt {
     #[serde(default)]
     pub compaction_threshold: Option<usize>,
 
+    #[serde(skip_serializing_if = "Option::is_none")]
+    #[serde(default)]
+    pub compaction_algorithm: Option<CompactionAlgorithm>,
+
     #[serde(skip_serializing_if = "Option::is_none")]
     #[serde(default)]
     pub gc_horizon: Option<u64>,
@@ -387,6 +397,13 @@ pub struct TenantConfOpt {
     pub gc_feedback: Option<bool>,
 }
 
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(tag = "kind")]
+pub enum CompactionAlgorithm {
+    Legacy,
+    Tiered,
+}
+
 #[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
 #[serde(tag = "kind")]
 pub enum EvictionPolicy {
@@ -429,6 +446,9 @@ impl TenantConfOpt {
             compaction_threshold: self
                 .compaction_threshold
                 .unwrap_or(global_conf.compaction_threshold),
+            compaction_algorithm: self
+                .compaction_algorithm
+                .unwrap_or(global_conf.compaction_algorithm),
             gc_horizon: self.gc_horizon.unwrap_or(global_conf.gc_horizon),
             gc_period: self.gc_period.unwrap_or(global_conf.gc_period),
             image_creation_threshold: self
@@ -468,6 +488,7 @@ impl Default for TenantConf {
             compaction_period: humantime::parse_duration(DEFAULT_COMPACTION_PERIOD)
                 .expect("cannot parse default compaction period"),
             compaction_threshold: DEFAULT_COMPACTION_THRESHOLD,
+            compaction_algorithm: DEFAULT_COMPACTION_ALGORITHM,
             gc_horizon: DEFAULT_GC_HORIZON,
             gc_period: humantime::parse_duration(DEFAULT_GC_PERIOD)
                 .expect("cannot parse default gc period"),
@@ -556,6 +577,12 @@ impl TryFrom<&'_ models::TenantConfig> for TenantConfOpt {
 
         tenant_conf.compaction_target_size = request_data.compaction_target_size;
         tenant_conf.compaction_threshold = request_data.compaction_threshold;
+        if let Some(compaction_algorithm) = &request_data.compaction_algorithm {
+            tenant_conf.compaction_algorithm = Some(
+                serde::Deserialize::deserialize(compaction_algorithm)
+                    .context("parse field `compaction_algorithm`")?,
+            );
+        }
 
         if let Some(compaction_period) = &request_data.compaction_period {
             tenant_conf.compaction_period = Some(

pageserver/src/tenant/delete.rs

@@ -3,10 +3,10 @@ use std::sync::Arc;
 use anyhow::Context;
 use camino::{Utf8Path, Utf8PathBuf};
 use pageserver_api::models::TenantState;
-use remote_storage::{DownloadError, GenericRemoteStorage, RemotePath};
+use remote_storage::{GenericRemoteStorage, RemotePath};
 use tokio::sync::OwnedMutexGuard;
 use tokio_util::sync::CancellationToken;
-use tracing::{error, info, instrument, warn, Instrument, Span};
+use tracing::{error, instrument, warn, Instrument, Span};
 
 use utils::{
     backoff, completion, crashsafe, fs_ext,
@@ -25,11 +25,9 @@ use super::{
     remote_timeline_client::{FAILED_REMOTE_OP_RETRIES, FAILED_UPLOAD_WARN_THRESHOLD},
     span,
     timeline::delete::DeleteTimelineFlow,
-    tree_sort_timelines, DeleteTimelineError, Tenant,
+    tree_sort_timelines, DeleteTimelineError, Tenant, TenantPreload,
 };
 
-const SHOULD_RESUME_DELETION_FETCH_MARK_ATTEMPTS: u32 = 3;
-
 #[derive(Debug, thiserror::Error)]
 pub(crate) enum DeleteTenantError {
     #[error("GetTenant {0}")]
@@ -60,7 +58,7 @@ fn remote_tenant_delete_mark_path(
         .context("Failed to strip workdir prefix")
         .and_then(RemotePath::new)
         .context("tenant path")?;
-    Ok(tenant_remote_path.join(Utf8Path::new("deleted")))
+    Ok(tenant_remote_path.join(Utf8Path::new("timelines/deleted")))
 }
 
 async fn create_remote_delete_mark(
@@ -150,7 +148,8 @@ async fn ensure_timelines_dir_empty(timelines_path: &Utf8Path) -> Result<(), Del
     // Assert timelines dir is empty.
     if !fs_ext::is_directory_empty(timelines_path).await? {
         // Display first 10 items in directory
-        let list = &fs_ext::list_dir(timelines_path).await.context("list_dir")?[..10];
+        let list = fs_ext::list_dir(timelines_path).await.context("list_dir")?;
+        let list = &list.into_iter().take(10).collect::<Vec<_>>();
         return Err(DeleteTenantError::Other(anyhow::anyhow!(
             "Timelines directory is not empty after all timelines deletion: {list:?}"
         )));
@@ -239,32 +238,6 @@ async fn cleanup_remaining_fs_traces(
     Ok(())
 }
 
-pub(crate) async fn remote_delete_mark_exists(
-    conf: &PageServerConf,
-    tenant_id: &TenantId,
-    remote_storage: &GenericRemoteStorage,
-) -> anyhow::Result<bool> {
-    // If remote storage is there we rely on it
-    let remote_mark_path = remote_tenant_delete_mark_path(conf, tenant_id).context("path")?;
-
-    let result = backoff::retry(
-        || async { remote_storage.download(&remote_mark_path).await },
-        |e| matches!(e, DownloadError::NotFound),
-        SHOULD_RESUME_DELETION_FETCH_MARK_ATTEMPTS,
-        SHOULD_RESUME_DELETION_FETCH_MARK_ATTEMPTS,
-        "fetch_tenant_deletion_mark",
-        // TODO: use a cancellation token (https://github.com/neondatabase/neon/issues/5066)
-        backoff::Cancel::new(CancellationToken::new(), || unreachable!()),
-    )
-    .await;
-
-    match result {
-        Ok(_) => Ok(true),
-        Err(DownloadError::NotFound) => Ok(false),
-        Err(e) => Err(anyhow::anyhow!(e)).context("remote_delete_mark_exists")?,
-    }
-}
-
 /// Orchestrates tenant shut down of all tasks, removes its in-memory structures,
 /// and deletes its data from both disk and s3.
 /// The sequence of steps:
@@ -276,10 +249,9 @@ pub(crate) async fn remote_delete_mark_exists(
 /// 6. Remove remote mark
 /// 7. Cleanup remaining fs traces, tenant dir, config, timelines dir, local delete mark
 /// It is resumable from any step in case a crash/restart occurs.
-/// There are three entrypoints to the process:
+/// There are two entrypoints to the process:
 /// 1. [`DeleteTenantFlow::run`] this is the main one called by a management api handler.
-/// 2. [`DeleteTenantFlow::resume_from_load`] is called during restarts when local or remote deletion marks are still there.
-/// 3. [`DeleteTenantFlow::resume_from_attach`] is called when deletion is resumed tenant is found to be deleted during attach process.
+/// 2. [`DeleteTenantFlow::resume_from_attach`] is called when deletion is resumed tenant is found to be deleted during attach process.
 ///  Note the only other place that messes around timeline delete mark is the `Tenant::spawn_load` function.
 #[derive(Default)]
 pub enum DeleteTenantFlow {
@@ -378,7 +350,7 @@ impl DeleteTenantFlow {
 
     pub(crate) async fn should_resume_deletion(
         conf: &'static PageServerConf,
-        remote_storage: Option<&GenericRemoteStorage>,
+        remote_mark_exists: bool,
         tenant: &Tenant,
     ) -> Result<Option<DeletionGuard>, DeleteTenantError> {
         let acquire = |t: &Tenant| {
@@ -389,66 +361,25 @@ impl DeleteTenantFlow {
             )
         };
 
-        let tenant_id = tenant.tenant_id;
-        // 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_id).exists() {
+        if remote_mark_exists {
             return Ok(acquire(tenant));
         }
 
-        let remote_storage = match remote_storage {
-            Some(remote_storage) => remote_storage,
-            None => return Ok(None),
-        };
-
-        if remote_delete_mark_exists(conf, &tenant_id, remote_storage).await? {
+        let tenant_id = tenant.tenant_id;
+        // 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_id).exists() {
             Ok(acquire(tenant))
         } else {
             Ok(None)
         }
     }
 
-    pub(crate) async fn resume_from_load(
-        guard: DeletionGuard,
-        tenant: &Arc<Tenant>,
-        init_order: Option<&InitializationOrder>,
-        tenants: &'static tokio::sync::RwLock<TenantsMap>,
-        ctx: &RequestContext,
-    ) -> Result<(), DeleteTenantError> {
-        let (_, progress) = completion::channel();
-
-        tenant
-            .set_stopping(progress, true, false)
-            .await
-            .expect("cant be stopping or broken");
-
-        // Do not consume valuable resources during the load phase, continue deletion once init phase is complete.
-        let background_jobs_can_start = init_order.as_ref().map(|x| &x.background_jobs_can_start);
-        if let Some(background) = background_jobs_can_start {
-            info!("waiting for backgound jobs barrier");
-            background.clone().wait().await;
-            info!("ready for backgound jobs barrier");
-        }
-
-        // Tenant may not be loadable if we fail late in cleanup_remaining_fs_traces (e g remove timelines dir)
-        let timelines_path = tenant.conf.timelines_path(&tenant.tenant_id);
-        if timelines_path.exists() {
-            tenant.load(init_order, ctx).await.context("load")?;
-        }
-
-        Self::background(
-            guard,
-            tenant.conf,
-            tenant.remote_storage.clone(),
-            tenants,
-            tenant,
-        )
-        .await
-    }
-
     pub(crate) async fn resume_from_attach(
         guard: DeletionGuard,
         tenant: &Arc<Tenant>,
+        preload: Option<TenantPreload>,
         tenants: &'static tokio::sync::RwLock<TenantsMap>,
+        init_order: Option<InitializationOrder>,
         ctx: &RequestContext,
     ) -> Result<(), DeleteTenantError> {
         let (_, progress) = completion::channel();
@@ -458,7 +389,10 @@ impl DeleteTenantFlow {
             .await
             .expect("cant be stopping or broken");
 
-        tenant.attach(ctx).await.context("attach")?;
+        tenant
+            .attach(init_order, preload, ctx)
+            .await
+            .context("attach")?;
 
         Self::background(
             guard,

pageserver/src/tenant/ephemeral_file.rs

@@ -354,8 +354,7 @@ mod tests {
         }
 
         // Test a large blob that spans multiple pages
-        let mut large_data = Vec::new();
-        large_data.resize(20000, 0);
+        let mut large_data = vec![0; 20000];
         thread_rng().fill_bytes(&mut large_data);
         let pos_large = file.write_blob(&large_data, &ctx).await?;
         let result = file.block_cursor().read_blob(pos_large, &ctx).await?;

pageserver/src/tenant/layer_map.rs

@@ -639,147 +639,10 @@ impl LayerMap {
         }
 
         println!("historic_layers:");
-        for layer in self.iter_historic_layers() {
-            layer.dump(verbose, ctx)?;
+        for desc in self.iter_historic_layers() {
+            desc.dump();
         }
         println!("End dump LayerMap");
         Ok(())
     }
 }
-
-#[cfg(test)]
-mod tests {
-    use super::LayerMap;
-    use crate::tenant::storage_layer::LayerFileName;
-    use std::str::FromStr;
-    use std::sync::Arc;
-
-    mod l0_delta_layers_updated {
-
-        use crate::tenant::{
-            storage_layer::{AsLayerDesc, PersistentLayerDesc},
-            timeline::layer_manager::LayerFileManager,
-        };
-
-        use super::*;
-
-        struct LayerObject(PersistentLayerDesc);
-
-        impl AsLayerDesc for LayerObject {
-            fn layer_desc(&self) -> &PersistentLayerDesc {
-                &self.0
-            }
-        }
-
-        impl LayerObject {
-            fn new(desc: PersistentLayerDesc) -> Self {
-                LayerObject(desc)
-            }
-        }
-
-        type TestLayerFileManager = LayerFileManager<LayerObject>;
-
-        #[test]
-        fn for_full_range_delta() {
-            // l0_delta_layers are used by compaction, and should observe all buffered updates
-            l0_delta_layers_updated_scenario(
-                 "000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__0000000053423C21-0000000053424D69",
-                 true
-             )
-        }
-
-        #[test]
-        fn for_non_full_range_delta() {
-            // has minimal uncovered areas compared to l0_delta_layers_updated_on_insert_replace_remove_for_full_range_delta
-            l0_delta_layers_updated_scenario(
-                 "000000000000000000000000000000000001-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFE__0000000053423C21-0000000053424D69",
-                 // because not full range
-                 false
-             )
-        }
-
-        #[test]
-        fn for_image() {
-            l0_delta_layers_updated_scenario(
-                 "000000000000000000000000000000000000-000000000000000000000000000000010000__0000000053424D69",
-                 // code only checks if it is a full range layer, doesn't care about images, which must
-                 // mean we should in practice never have full range images
-                 false
-             )
-        }
-
-        #[test]
-        fn replacing_missing_l0_is_notfound() {
-            // original impl had an oversight, and L0 was an anyhow::Error. anyhow::Error should
-            // however only happen for precondition failures.
-
-            let layer = "000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__0000000053423C21-0000000053424D69";
-            let layer = LayerFileName::from_str(layer).unwrap();
-            let layer = PersistentLayerDesc::from(layer);
-
-            // same skeletan construction; see scenario below
-            let not_found = Arc::new(LayerObject::new(layer.clone()));
-            let new_version = Arc::new(LayerObject::new(layer));
-
-            // after the immutable storage state refactor, the replace operation
-            // will not use layer map any more. We keep it here for consistency in test cases
-            // and can remove it in the future.
-            let _map = LayerMap::default();
-
-            let mut mapping = TestLayerFileManager::new();
-
-            mapping
-                .replace_and_verify(not_found, new_version)
-                .unwrap_err();
-        }
-
-        fn l0_delta_layers_updated_scenario(layer_name: &str, expected_l0: bool) {
-            let name = LayerFileName::from_str(layer_name).unwrap();
-            let skeleton = PersistentLayerDesc::from(name);
-
-            let remote = Arc::new(LayerObject::new(skeleton.clone()));
-            let downloaded = Arc::new(LayerObject::new(skeleton));
-
-            let mut map = LayerMap::default();
-            let mut mapping = LayerFileManager::new();
-
-            // two disjoint Arcs in different lifecycle phases. even if it seems they must be the
-            // same layer, we use LayerMap::compare_arced_layers as the identity of layers.
-            assert_eq!(remote.layer_desc(), downloaded.layer_desc());
-
-            let expected_in_counts = (1, usize::from(expected_l0));
-
-            map.batch_update()
-                .insert_historic(remote.layer_desc().clone());
-            mapping.insert(remote.clone());
-            assert_eq!(
-                count_layer_in(&map, remote.layer_desc()),
-                expected_in_counts
-            );
-
-            mapping
-                .replace_and_verify(remote, downloaded.clone())
-                .expect("name derived attributes are the same");
-            assert_eq!(
-                count_layer_in(&map, downloaded.layer_desc()),
-                expected_in_counts
-            );
-
-            map.batch_update().remove_historic(downloaded.layer_desc());
-            assert_eq!(count_layer_in(&map, downloaded.layer_desc()), (0, 0));
-        }
-
-        fn count_layer_in(map: &LayerMap, layer: &PersistentLayerDesc) -> (usize, usize) {
-            let historic = map
-                .iter_historic_layers()
-                .filter(|x| x.key() == layer.key())
-                .count();
-            let l0s = map
-                .get_level0_deltas()
-                .expect("why does this return a result");
-            let l0 = l0s.iter().filter(|x| x.key() == layer.key()).count();
-
-            (historic, l0)
-        }
-    }
-}

pageserver/src/tenant/mgr.rs

@@ -1,7 +1,7 @@
 //! This module acts as a switchboard to access different repositories managed by this
 //! page server.
 
-use camino::{Utf8Path, Utf8PathBuf};
+use camino::{Utf8DirEntry, Utf8Path, Utf8PathBuf};
 use rand::{distributions::Alphanumeric, Rng};
 use std::collections::{hash_map, HashMap};
 use std::sync::Arc;
@@ -26,9 +26,7 @@ use crate::deletion_queue::DeletionQueueClient;
 use crate::task_mgr::{self, TaskKind};
 use crate::tenant::config::{AttachmentMode, LocationConf, LocationMode, TenantConfOpt};
 use crate::tenant::delete::DeleteTenantFlow;
-use crate::tenant::{
-    create_tenant_files, AttachedTenantConf, CreateTenantFilesMode, Tenant, TenantState,
-};
+use crate::tenant::{create_tenant_files, AttachedTenantConf, SpawnMode, Tenant, TenantState};
 use crate::{InitializationOrder, IGNORED_TENANT_FILE_NAME, TEMP_FILE_SUFFIX};
 
 use utils::crashsafe::path_with_suffix_extension;
@@ -151,6 +149,49 @@ async fn safe_rename_tenant_dir(path: impl AsRef<Utf8Path>) -> std::io::Result<U
 
 static TENANTS: Lazy<RwLock<TenantsMap>> = Lazy::new(|| 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(())
+}
+
 fn emergency_generations(
     tenant_confs: &HashMap<TenantId, anyhow::Result<LocationConf>>,
 ) -> HashMap<TenantId, Generation> {
@@ -212,83 +253,99 @@ async fn init_load_generations(
     Ok(Some(generations))
 }
 
+/// Given a directory discovered in the pageserver's tenants/ directory, attempt
+/// to load a tenant config from it.
+///
+/// If file is missing, return Ok(None)
+fn load_tenant_config(
+    conf: &'static PageServerConf,
+    dentry: Utf8DirEntry,
+) -> anyhow::Result<Option<(TenantId, 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}");
+        // No need to use safe_remove_tenant_dir_all because this is already
+        // a temporary path
+        if let Err(e) = std::fs::remove_dir_all(&tenant_dir_path) {
+            error!(
+                "Failed to remove temporary directory '{}': {:?}",
+                tenant_dir_path, e
+            );
+        }
+        return Ok(None);
+    }
+
+    // This case happens if we crash during attachment before writing a config into the dir
+    let is_empty = tenant_dir_path
+        .is_empty_dir()
+        .with_context(|| format!("Failed to check whether {tenant_dir_path:?} is an empty dir"))?;
+    if is_empty {
+        info!("removing empty tenant directory {tenant_dir_path:?}");
+        if let Err(e) = std::fs::remove_dir(&tenant_dir_path) {
+            error!(
+                "Failed to remove empty tenant directory '{}': {e:#}",
+                tenant_dir_path
+            )
+        }
+        return Ok(None);
+    }
+
+    let tenant_ignore_mark_file = tenant_dir_path.join(IGNORED_TENANT_FILE_NAME);
+    if tenant_ignore_mark_file.exists() {
+        info!("Found an ignore mark file {tenant_ignore_mark_file:?}, skipping the tenant");
+        return Ok(None);
+    }
+
+    let tenant_id = match tenant_dir_path
+        .file_name()
+        .unwrap_or_default()
+        .parse::<TenantId>()
+    {
+        Ok(id) => id,
+        Err(_) => {
+            warn!("Invalid tenant path (garbage in our repo directory?): {tenant_dir_path}",);
+            return Ok(None);
+        }
+    };
+
+    Ok(Some((
+        tenant_id,
+        Tenant::load_tenant_config(conf, &tenant_id),
+    )))
+}
+
 /// Initial stage of load: walk the local tenants directory, clean up any temp files,
 /// and load configurations for the tenants we found.
+///
+/// Do this in parallel, because we expect 10k+ tenants, so serial execution can take
+/// seconds even on reasonably fast drives.
 async fn init_load_tenant_configs(
     conf: &'static PageServerConf,
 ) -> anyhow::Result<HashMap<TenantId, anyhow::Result<LocationConf>>> {
     let tenants_dir = conf.tenants_path();
 
-    let mut dir_entries = tenants_dir
-        .read_dir_utf8()
-        .with_context(|| format!("Failed to list tenants dir {tenants_dir:?}"))?;
+    let dentries = tokio::task::spawn_blocking(move || -> anyhow::Result<Vec<Utf8DirEntry>> {
+        let dir_entries = tenants_dir
+            .read_dir_utf8()
+            .with_context(|| format!("Failed to list tenants dir {tenants_dir:?}"))?;
+
+        Ok(dir_entries.collect::<Result<Vec<_>, std::io::Error>>()?)
+    })
+    .await??;
 
     let mut configs = HashMap::new();
 
-    loop {
-        match dir_entries.next() {
-            None => break,
-            Some(Ok(dentry)) => {
-                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}");
-                    // No need to use safe_remove_tenant_dir_all because this is already
-                    // a temporary path
-                    if let Err(e) = fs::remove_dir_all(&tenant_dir_path).await {
-                        error!(
-                            "Failed to remove temporary directory '{}': {:?}",
-                            tenant_dir_path, e
-                        );
-                    }
-                    continue;
-                }
+    let mut join_set = JoinSet::new();
+    for dentry in dentries {
+        join_set.spawn_blocking(move || load_tenant_config(conf, dentry));
+    }
 
-                // This case happens if we:
-                // * crash during attach before creating the attach marker file
-                // * crash during tenant delete before removing tenant directory
-                let is_empty = tenant_dir_path.is_empty_dir().with_context(|| {
-                    format!("Failed to check whether {tenant_dir_path:?} is an empty dir")
-                })?;
-                if is_empty {
-                    info!("removing empty tenant directory {tenant_dir_path:?}");
-                    if let Err(e) = fs::remove_dir(&tenant_dir_path).await {
-                        error!(
-                            "Failed to remove empty tenant directory '{}': {e:#}",
-                            tenant_dir_path
-                        )
-                    }
-                    continue;
-                }
-
-                let tenant_ignore_mark_file = tenant_dir_path.join(IGNORED_TENANT_FILE_NAME);
-                if tenant_ignore_mark_file.exists() {
-                    info!("Found an ignore mark file {tenant_ignore_mark_file:?}, skipping the tenant");
-                    continue;
-                }
-
-                let tenant_id = match tenant_dir_path
-                    .file_name()
-                    .unwrap_or_default()
-                    .parse::<TenantId>()
-                {
-                    Ok(id) => id,
-                    Err(_) => {
-                        warn!(
-                            "Invalid tenant path (garbage in our repo directory?): {tenant_dir_path}",
-                        );
-                        continue;
-                    }
-                };
-
-                configs.insert(tenant_id, Tenant::load_tenant_config(conf, &tenant_id));
-            }
-            Some(Err(e)) => {
-                // An error listing the top level directory indicates serious problem
-                // with local filesystem: we will fail to load, and fail to start.
-                anyhow::bail!(e);
-            }
+    while let Some(r) = join_set.join_next().await {
+        if let Some((tenant_id, tenant_config)) = r?? {
+            configs.insert(tenant_id, tenant_config);
         }
     }
+
     Ok(configs)
 }
 
@@ -377,14 +434,15 @@ pub async fn init_tenant_mgr(
         location_conf.attach_in_generation(generation);
         Tenant::persist_tenant_config(conf, &tenant_id, &location_conf).await?;
 
-        match schedule_local_tenant_processing(
+        match tenant_spawn(
             conf,
             tenant_id,
             &tenant_dir_path,
-            AttachedTenantConf::try_from(location_conf)?,
             resources.clone(),
+            AttachedTenantConf::try_from(location_conf)?,
             Some(init_order.clone()),
             &TENANTS,
+            SpawnMode::Normal,
             &ctx,
         ) {
             Ok(tenant) => {
@@ -404,15 +462,18 @@ pub async fn init_tenant_mgr(
     Ok(())
 }
 
+/// Wrapper for Tenant::spawn that checks invariants before running, and inserts
+/// a broken tenant in the map if Tenant::spawn fails.
 #[allow(clippy::too_many_arguments)]
-pub(crate) fn schedule_local_tenant_processing(
+pub(crate) fn tenant_spawn(
     conf: &'static PageServerConf,
     tenant_id: TenantId,
     tenant_path: &Utf8Path,
-    location_conf: AttachedTenantConf,
     resources: TenantSharedResources,
+    location_conf: AttachedTenantConf,
     init_order: Option<InitializationOrder>,
     tenants: &'static tokio::sync::RwLock<TenantsMap>,
+    mode: SpawnMode,
     ctx: &RequestContext,
 ) -> anyhow::Result<Arc<Tenant>> {
     anyhow::ensure!(
@@ -436,37 +497,24 @@ pub(crate) fn schedule_local_tenant_processing(
         "Cannot load tenant, ignore mark found at {tenant_ignore_mark:?}"
     );
 
-    let tenant = if conf.tenant_attaching_mark_file_path(&tenant_id).exists() {
-        info!("tenant {tenant_id} has attaching mark file, resuming its attach operation");
-        if resources.remote_storage.is_none() {
-            warn!("tenant {tenant_id} has attaching mark file, but pageserver has no remote storage configured");
-            Tenant::create_broken_tenant(
-                conf,
-                tenant_id,
-                "attaching mark file present but no remote storage configured".to_string(),
-            )
-        } else {
-            match Tenant::spawn_attach(conf, tenant_id, resources, location_conf, tenants, ctx) {
-                Ok(tenant) => tenant,
-                Err(e) => {
-                    error!("Failed to spawn_attach tenant {tenant_id}, reason: {e:#}");
-                    Tenant::create_broken_tenant(conf, tenant_id, format!("{e:#}"))
-                }
-            }
+    info!("Attaching tenant {tenant_id}");
+    let tenant = match Tenant::spawn(
+        conf,
+        tenant_id,
+        resources,
+        location_conf,
+        init_order,
+        tenants,
+        mode,
+        ctx,
+    ) {
+        Ok(tenant) => tenant,
+        Err(e) => {
+            error!("Failed to spawn tenant {tenant_id}, reason: {e:#}");
+            Tenant::create_broken_tenant(conf, tenant_id, format!("{e:#}"))
         }
-    } else {
-        info!("tenant {tenant_id} is assumed to be loadable, starting load operation");
-        // Start loading the tenant into memory. It will initially be in Loading state.
-        Tenant::spawn_load(
-            conf,
-            tenant_id,
-            location_conf,
-            resources,
-            init_order,
-            tenants,
-            ctx,
-        )
     };
+
     Ok(tenant)
 }
 
@@ -602,29 +650,41 @@ pub(crate) async fn create_tenant(
     ctx: &RequestContext,
 ) -> Result<Arc<Tenant>, TenantMapInsertError> {
     tenant_map_insert(tenant_id, || async {
-
         let location_conf = LocationConf::attached_single(tenant_conf, generation);
 
         // We're holding the tenants lock in write mode while doing local IO.
         // If this section ever becomes contentious, introduce a new `TenantState::Creating`
         // and do the work in that state.
-        let tenant_directory = super::create_tenant_files(conf, &location_conf, &tenant_id, CreateTenantFilesMode::Create).await?;
+        super::create_tenant_files(conf, &location_conf, &tenant_id).await?;
+
         // TODO: tenant directory remains on disk if we bail out from here on.
         //       See https://github.com/neondatabase/neon/issues/4233
 
-        let created_tenant =
-            schedule_local_tenant_processing(conf, tenant_id, &tenant_directory,
-                AttachedTenantConf::try_from(location_conf)?, resources, None, &TENANTS, ctx)?;
+        let tenant_path = conf.tenant_path(&tenant_id);
+
+        let created_tenant = tenant_spawn(
+            conf,
+            tenant_id,
+            &tenant_path,
+            resources,
+            AttachedTenantConf::try_from(location_conf)?,
+            None,
+            &TENANTS,
+            SpawnMode::Create,
+            ctx,
+        )?;
         // TODO: tenant object & its background loops remain, untracked in tenant map, if we fail here.
         //      See https://github.com/neondatabase/neon/issues/4233
 
         let crated_tenant_id = created_tenant.tenant_id();
         anyhow::ensure!(
-                tenant_id == crated_tenant_id,
-                "loaded created tenant has unexpected tenant id (expect {tenant_id} != actual {crated_tenant_id})",
-            );
+            tenant_id == crated_tenant_id,
+            "loaded created tenant has unexpected tenant id \
+                (expect {tenant_id} != actual {crated_tenant_id})",
+        );
         Ok(created_tenant)
-    }).await
+    })
+    .await
 }
 
 #[derive(Debug, thiserror::Error)]
@@ -655,7 +715,7 @@ pub(crate) async fn set_new_tenant_config(
     Ok(())
 }
 
-#[instrument(skip_all, fields(tenant_id, new_location_config))]
+#[instrument(skip_all, fields(%tenant_id))]
 pub(crate) async fn upsert_location(
     conf: &'static PageServerConf,
     tenant_id: TenantId,
@@ -733,37 +793,56 @@ pub(crate) async fn upsert_location(
                 }
             }
 
+            let tenant_path = conf.tenant_path(&tenant_id);
+
             let new_slot = match &new_location_config.mode {
-                LocationMode::Secondary(_) => TenantSlot::Secondary,
-                LocationMode::Attached(_attach_config) => {
-                    // Do a schedule_local_tenant_processing
-                    // FIXME: should avoid doing this disk I/O inside the TenantsMap lock,
-                    // we have the same problem in load_tenant/attach_tenant.  Probably
-                    // need a lock in TenantSlot to fix this.
+                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)
+                        .await
+                        .with_context(|| format!("Creating {tenant_path}"))?;
+
                     Tenant::persist_tenant_config(conf, &tenant_id, &new_location_config)
                         .await
                         .map_err(SetNewTenantConfigError::Persist)?;
-                    let tenant_path = conf.tenant_path(&tenant_id);
-                    let resources = TenantSharedResources {
-                        broker_client,
-                        remote_storage,
-                        deletion_queue_client,
-                    };
-                    let new_tenant = schedule_local_tenant_processing(
+
+                    TenantSlot::Secondary
+                }
+                LocationMode::Attached(_attach_config) => {
+                    // FIXME: should avoid doing this disk I/O inside the TenantsMap lock,
+                    // we have the same problem in load_tenant/attach_tenant.  Probably
+                    // need a lock in TenantSlot to fix this.
+                    let timelines_path = conf.timelines_path(&tenant_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)
+                        .await
+                        .with_context(|| format!("Creating {timelines_path}"))?;
+
+                    Tenant::persist_tenant_config(conf, &tenant_id, &new_location_config)
+                        .await
+                        .map_err(SetNewTenantConfigError::Persist)?;
+
+                    let tenant = tenant_spawn(
                         conf,
                         tenant_id,
                         &tenant_path,
+                        TenantSharedResources {
+                            broker_client,
+                            remote_storage,
+                            deletion_queue_client,
+                        },
                         AttachedTenantConf::try_from(new_location_config)?,
-                        resources,
                         None,
                         &TENANTS,
+                        SpawnMode::Normal,
                         ctx,
-                    )
-                    .with_context(|| {
-                        format!("Failed to schedule tenant processing in path {tenant_path:?}")
-                    })?;
+                    )?;
 
-                    TenantSlot::Attached(new_tenant)
+                    TenantSlot::Attached(tenant)
                 }
             };
 
@@ -771,7 +850,6 @@ pub(crate) async fn upsert_location(
         })
         .await?;
     }
-
     Ok(())
 }
 
@@ -951,7 +1029,7 @@ pub(crate) async fn load_tenant(
         location_conf.attach_in_generation(generation);
         Tenant::persist_tenant_config(conf, &tenant_id, &location_conf).await?;
 
-        let new_tenant = schedule_local_tenant_processing(conf, tenant_id, &tenant_path, AttachedTenantConf::try_from(location_conf)?, resources, None,  &TENANTS, ctx)
+        let new_tenant = tenant_spawn(conf, tenant_id, &tenant_path, resources, AttachedTenantConf::try_from(location_conf)?, None,  &TENANTS, SpawnMode::Normal, ctx)
             .with_context(|| {
                 format!("Failed to schedule tenant processing in path {tenant_path:?}")
             })?;
@@ -1025,18 +1103,12 @@ pub(crate) async fn attach_tenant(
 ) -> Result<(), TenantMapInsertError> {
     tenant_map_insert(tenant_id, || async {
         let location_conf = LocationConf::attached_single(tenant_conf, generation);
-        let tenant_dir = create_tenant_files(conf, &location_conf, &tenant_id, CreateTenantFilesMode::Attach).await?;
+        let tenant_dir = create_tenant_files(conf, &location_conf, &tenant_id).await?;
         // TODO: tenant directory remains on disk if we bail out from here on.
         //       See https://github.com/neondatabase/neon/issues/4233
 
-        // Without the attach marker, schedule_local_tenant_processing will treat the attached tenant as fully attached
-        let marker_file_exists = conf
-            .tenant_attaching_mark_file_path(&tenant_id)
-            .try_exists()
-            .context("check for attach marker file existence")?;
-        anyhow::ensure!(marker_file_exists, "create_tenant_files should have created the attach marker file");
-
-        let attached_tenant = schedule_local_tenant_processing(conf, tenant_id, &tenant_dir, AttachedTenantConf::try_from(location_conf)?, resources, None, &TENANTS, ctx)?;
+        let attached_tenant = tenant_spawn(conf, tenant_id, &tenant_dir,
+            resources, AttachedTenantConf::try_from(location_conf)?, None, &TENANTS, SpawnMode::Normal, ctx)?;
         // TODO: tenant object & its background loops remain, untracked in tenant map, if we fail here.
         //      See https://github.com/neondatabase/neon/issues/4233
 

pageserver/src/tenant/par_fsync.rs

@@ -57,8 +57,7 @@ pub fn par_fsync(paths: &[Utf8PathBuf]) -> io::Result<()> {
     fsync_in_thread_pool(paths)
 }
 
-/// Parallel fsync asynchronously. If number of files are less than PARALLEL_PATH_THRESHOLD, fsync is done in the current
-/// execution thread. Otherwise, we will spawn_blocking and run it in tokio.
+/// Parallel fsync asynchronously.
 pub async fn par_fsync_async(paths: &[Utf8PathBuf]) -> io::Result<()> {
     const MAX_CONCURRENT_FSYNC: usize = 64;
     let mut next = paths.iter().peekable();