update dashmap with new unsafe raw shards api

The new image iterator and delta iterator uses an iterator-based API.
https://github.com/neondatabase/neon/pull/8006 / part of
https://github.com/neondatabase/neon/issues/8002

This requires the underlying thing (the btree) to have an iterator API,
and the iterator should have a type name so that it can be stored
somewhere.

```rust
pub struct DeltaLayerIterator {
  index_iterator: BTreeIterator
}
```

versus

```rust
pub struct DeltaLayerIterator {
  index_iterator: impl Stream<....>
}
```

(this requires nightly flag and still buggy in the Rust compiler)


There are multiple ways to achieve this:

1. Either write a BTreeIterator from scratch that provides `async next`.
This is the most efficient way to do that.
2. Or wrap the current `get_stream` API, which is the current approach
in the pull request.

In the future, we should do (1), and the `get_stream` API should be
refactored to use the iterator API. With (2), we have to wrap the
`get_stream` API with `Pin<Box<dyn Stream>>`, where we have the overhead
of dynamic dispatch. However, (2) needs a rewrite of the `visit`
function, which would take some time to write and review. I'd like to
define this iterator API first and work on a real iterator API later.

## Summary of changes

Add `DiskBtreeIterator` and related tests.

Signed-off-by: Alex Chi Z <chi@neon.tech>

.github/workflows/build-build-tools-image.yml

@@ -30,7 +30,6 @@ jobs:
   check-image:
     uses: ./.github/workflows/check-build-tools-image.yml
 
-  # This job uses older version of GitHub Actions because it's run on gen2 runners, which don't support node 20 (for newer versions)
   build-image:
     needs: [ check-image ]
     if: needs.check-image.outputs.found == 'false'

.github/workflows/build_and_test.yml

@@ -299,21 +299,21 @@ jobs:
         uses: actions/cache@v4
         with:
           path: pg_install/v14
-          key: v1-${{ runner.os }}-${{ matrix.build_type }}-pg-${{ steps.pg_v14_rev.outputs.pg_rev }}-${{ hashFiles('Makefile') }}
+          key: v1-${{ runner.os }}-${{ matrix.build_type }}-pg-${{ steps.pg_v14_rev.outputs.pg_rev }}-${{ hashFiles('Makefile', 'Dockerfile.build-tools') }}
 
       - name: Cache postgres v15 build
         id: cache_pg_15
         uses: actions/cache@v4
         with:
           path: pg_install/v15
-          key: v1-${{ runner.os }}-${{ matrix.build_type }}-pg-${{ steps.pg_v15_rev.outputs.pg_rev }}-${{ hashFiles('Makefile') }}
+          key: v1-${{ runner.os }}-${{ matrix.build_type }}-pg-${{ steps.pg_v15_rev.outputs.pg_rev }}-${{ hashFiles('Makefile', 'Dockerfile.build-tools') }}
 
       - name: Cache postgres v16 build
         id: cache_pg_16
         uses: actions/cache@v4
         with:
           path: pg_install/v16
-          key: v1-${{ runner.os }}-${{ matrix.build_type }}-pg-${{ steps.pg_v16_rev.outputs.pg_rev }}-${{ hashFiles('Makefile') }}
+          key: v1-${{ runner.os }}-${{ matrix.build_type }}-pg-${{ steps.pg_v16_rev.outputs.pg_rev }}-${{ hashFiles('Makefile', 'Dockerfile.build-tools') }}
 
       - name: Build postgres v14
         if: steps.cache_pg_14.outputs.cache-hit != 'true'

.github/workflows/check-build-tools-image.yml

@@ -25,26 +25,17 @@ jobs:
       found: ${{ steps.check-image.outputs.found }}
 
     steps:
+      - uses: actions/checkout@v4
+
       - name: Get build-tools image tag for the current commit
         id: get-build-tools-tag
         env:
-          # Usually, for COMMIT_SHA, we use `github.event.pull_request.head.sha || github.sha`, but here, even for PRs,
-          # we want to use `github.sha` i.e. point to a phantom merge commit to determine the image tag correctly.
-          COMMIT_SHA: ${{ github.sha }}
-          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          IMAGE_TAG: |
+            ${{ hashFiles('Dockerfile.build-tools',
+                          '.github/workflows/check-build-tools-image.yml',
+                          '.github/workflows/build-build-tools-image.yml') }}
         run: |
-          LAST_BUILD_TOOLS_SHA=$(
-            gh api \
-              -H "Accept: application/vnd.github+json" \
-              -H "X-GitHub-Api-Version: 2022-11-28" \
-              --method GET \
-              --field path=Dockerfile.build-tools \
-              --field sha=${COMMIT_SHA} \
-              --field per_page=1 \
-              --jq ".[0].sha" \
-              "/repos/${GITHUB_REPOSITORY}/commits"
-          )
-          echo "image-tag=${LAST_BUILD_TOOLS_SHA}" | tee -a $GITHUB_OUTPUT
+          echo "image-tag=${IMAGE_TAG}" | tee -a $GITHUB_OUTPUT
 
       - name: Check if such tag found in the registry
         id: check-image

Cargo.lock

@@ -1598,6 +1598,20 @@ dependencies = [
  "parking_lot_core 0.9.8",
 ]
 
+[[package]]
+name = "dashmap"
+version = "6.0.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "23fadfd577acfd4485fb258011b0fd080882ea83359b6fd41304900b94ccf487"
+dependencies = [
+ "cfg-if",
+ "crossbeam-utils",
+ "hashbrown 0.14.5",
+ "lock_api",
+ "once_cell",
+ "parking_lot_core 0.9.8",
+]
+
 [[package]]
 name = "data-encoding"
 version = "2.4.0"
@@ -2848,7 +2862,7 @@ version = "0.7.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "4644821e1c3d7a560fe13d842d13f587c07348a1a05d3a797152d41c90c56df2"
 dependencies = [
- "dashmap",
+ "dashmap 5.5.0",
  "hashbrown 0.13.2",
 ]
 
@@ -4296,7 +4310,7 @@ dependencies = [
  "clap",
  "consumption_metrics",
  "crossbeam-deque",
- "dashmap",
+ "dashmap 6.0.0",
  "env_logger",
  "fallible-iterator",
  "framed-websockets",
@@ -5158,6 +5172,7 @@ dependencies = [
  "tokio-io-timeout",
  "tokio-postgres",
  "tokio-stream",
+ "tokio-tar",
  "tokio-util",
  "toml_edit",
  "tracing",

Cargo.toml

@@ -77,7 +77,7 @@ const_format = "0.2"
 crc32c = "0.6"
 crossbeam-deque = "0.8.5"
 crossbeam-utils = "0.8.5"
-dashmap = { version = "5.5.0", features = ["raw-api"] }
+dashmap = { version = "6.0", features = ["raw-api"] }
 either = "1.8"
 enum-map = "2.4.2"
 enumset = "1.0.12"

Dockerfile

@@ -69,8 +69,6 @@ RUN set -e \
     && apt install -y \
         libreadline-dev \
         libseccomp-dev \
-        libicu67 \
-        openssl \
         ca-certificates \
     && rm -rf /var/lib/apt/lists/* /tmp/* /var/tmp/* \
     && useradd -d /data neon \

Dockerfile.build-tools

@@ -112,6 +112,45 @@ RUN for package in Capture::Tiny DateTime Devel::Cover Digest::MD5 File::Spec JS
     && make install \
     && rm -rf ../lcov.tar.gz
 
+# Compile and install the static OpenSSL library
+ENV OPENSSL_VERSION=3.2.2
+ENV OPENSSL_PREFIX=/usr/local/openssl
+RUN wget -O /tmp/openssl-${OPENSSL_VERSION}.tar.gz https://www.openssl.org/source/openssl-${OPENSSL_VERSION}.tar.gz && \
+    echo "197149c18d9e9f292c43f0400acaba12e5f52cacfe050f3d199277ea738ec2e7 /tmp/openssl-${OPENSSL_VERSION}.tar.gz" | sha256sum --check && \
+    cd /tmp && \
+    tar xzvf /tmp/openssl-${OPENSSL_VERSION}.tar.gz && \
+    rm /tmp/openssl-${OPENSSL_VERSION}.tar.gz && \
+    cd /tmp/openssl-${OPENSSL_VERSION} && \
+    ./config --prefix=${OPENSSL_PREFIX}  -static --static no-shared -fPIC && \
+    make -j "$(nproc)" && \
+    make install && \
+    cd /tmp && \
+    rm -rf /tmp/openssl-${OPENSSL_VERSION}
+
+# Use the same version of libicu as the compute nodes so that
+# clusters created using inidb on pageserver can be used by computes.
+#
+# TODO: at this time, Dockerfile.compute-node uses the debian bullseye libicu
+# package, which is 67.1. We're duplicating that knowledge here, and also, technically,
+# Debian has a few patches on top of 67.1 that we're not adding here.
+ENV ICU_VERSION=67.1
+ENV ICU_PREFIX=/usr/local/icu
+
+# Download and build static ICU
+RUN wget -O /tmp/libicu-${ICU_VERSION}.tgz https://github.com/unicode-org/icu/releases/download/release-${ICU_VERSION//./-}/icu4c-${ICU_VERSION//./_}-src.tgz && \
+    echo "94a80cd6f251a53bd2a997f6f1b5ac6653fe791dfab66e1eb0227740fb86d5dc /tmp/libicu-${ICU_VERSION}.tgz" | sha256sum --check && \
+    mkdir /tmp/icu && \
+    pushd /tmp/icu && \
+    tar -xzf /tmp/libicu-${ICU_VERSION}.tgz && \
+    pushd icu/source && \
+    ./configure --prefix=${ICU_PREFIX}  --enable-static --enable-shared=no CXXFLAGS="-fPIC" CFLAGS="-fPIC" && \
+    make -j "$(nproc)" && \
+    make install && \
+    popd && \
+    rm -rf icu && \
+    rm -f /tmp/libicu-${ICU_VERSION}.tgz && \
+    popd
+
 # Switch to nonroot user
 USER nonroot:nonroot
 WORKDIR /home/nonroot
@@ -170,3 +209,6 @@ RUN whoami \
     && rustup --version --verbose \
     && rustc --version --verbose \
     && clang --version
+
+# Set following flag to check in Makefile if its running in Docker
+RUN touch /home/nonroot/.docker_build

Makefile

@@ -3,6 +3,9 @@ ROOT_PROJECT_DIR := $(dir $(abspath $(lastword $(MAKEFILE_LIST))))
 # Where to install Postgres, default is ./pg_install, maybe useful for package managers
 POSTGRES_INSTALL_DIR ?= $(ROOT_PROJECT_DIR)/pg_install/
 
+OPENSSL_PREFIX_DIR := /usr/local/openssl
+ICU_PREFIX_DIR := /usr/local/icu
+
 #
 # We differentiate between release / debug build types using the BUILD_TYPE
 # environment variable.
@@ -20,6 +23,16 @@ else
 	$(error Bad build type '$(BUILD_TYPE)', see Makefile for options)
 endif
 
+ifeq ($(shell test -e /home/nonroot/.docker_build && echo -n yes),yes)
+	# Exclude static build openssl, icu for local build (MacOS, Linux)
+	# Only keep for build type release and debug
+	PG_CFLAGS += -I$(OPENSSL_PREFIX_DIR)/include
+	PG_CONFIGURE_OPTS += --with-icu
+	PG_CONFIGURE_OPTS += ICU_CFLAGS='-I/$(ICU_PREFIX_DIR)/include -DU_STATIC_IMPLEMENTATION'
+	PG_CONFIGURE_OPTS += ICU_LIBS='-L$(ICU_PREFIX_DIR)/lib -L$(ICU_PREFIX_DIR)/lib64 -licui18n -licuuc -licudata -lstdc++ -Wl,-Bdynamic -lm'
+	PG_CONFIGURE_OPTS += LDFLAGS='-L$(OPENSSL_PREFIX_DIR)/lib -L$(OPENSSL_PREFIX_DIR)/lib64 -L$(ICU_PREFIX_DIR)/lib -L$(ICU_PREFIX_DIR)/lib64 -Wl,-Bstatic -lssl -lcrypto -Wl,-Bdynamic -lrt -lm -ldl -lpthread'
+endif
+
 UNAME_S := $(shell uname -s)
 ifeq ($(UNAME_S),Linux)
 	# Seccomp BPF is only available for Linux
@@ -28,7 +41,7 @@ else ifeq ($(UNAME_S),Darwin)
 	ifndef DISABLE_HOMEBREW
 		# macOS with brew-installed openssl requires explicit paths
 		# It can be configured with OPENSSL_PREFIX variable
-		OPENSSL_PREFIX ?= $(shell brew --prefix openssl@3)
+		OPENSSL_PREFIX := $(shell brew --prefix openssl@3)
 		PG_CONFIGURE_OPTS += --with-includes=$(OPENSSL_PREFIX)/include --with-libraries=$(OPENSSL_PREFIX)/lib
 		PG_CONFIGURE_OPTS += PKG_CONFIG_PATH=$(shell brew --prefix icu4c)/lib/pkgconfig
 		# macOS already has bison and flex in the system, but they are old and result in postgres-v14 target failure

compute_tools/src/compute.rs

@@ -918,38 +918,39 @@ impl ComputeNode {
         // temporarily reset max_cluster_size in config
         // to avoid the possibility of hitting the limit, while we are reconfiguring:
         // creating new extensions, roles, etc...
-        config::compute_ctl_temp_override_create(pgdata_path, "neon.max_cluster_size=-1")?;
-        self.pg_reload_conf()?;
+        config::with_compute_ctl_tmp_override(pgdata_path, "neon.max_cluster_size=-1", || {
+            self.pg_reload_conf()?;
 
-        let mut client = Client::connect(self.connstr.as_str(), NoTls)?;
+            let mut client = Client::connect(self.connstr.as_str(), NoTls)?;
 
-        // Proceed with post-startup configuration. Note, that order of operations is important.
-        // 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,
-                &mut client,
-                self.connstr.as_str(),
-                self.has_feature(ComputeFeature::AnonExtension),
-            )?;
-            handle_extensions(&spec, &mut client)?;
-            handle_extension_neon(&mut client)?;
-            // We can skip handle_migrations here because a new migration can only appear
-            // if we have a new version of the compute_ctl binary, which can only happen
-            // if compute got restarted, in which case we'll end up inside of apply_config
-            // instead of reconfigure.
-        }
+            // Proceed with post-startup configuration. Note, that order of operations is important.
+            // 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,
+                    &mut client,
+                    self.connstr.as_str(),
+                    self.has_feature(ComputeFeature::AnonExtension),
+                )?;
+                handle_extensions(&spec, &mut client)?;
+                handle_extension_neon(&mut client)?;
+                // We can skip handle_migrations here because a new migration can only appear
+                // if we have a new version of the compute_ctl binary, which can only happen
+                // if compute got restarted, in which case we'll end up inside of apply_config
+                // instead of reconfigure.
+            }
 
-        // 'Close' connection
-        drop(client);
+            // 'Close' connection
+            drop(client);
+
+            Ok(())
+        })?;
 
-        // reset max_cluster_size in config back to original value and reload config
-        config::compute_ctl_temp_override_remove(pgdata_path)?;
         self.pg_reload_conf()?;
 
         let unknown_op = "unknown".to_string();
@@ -1040,12 +1041,17 @@ impl ComputeNode {
                 // temporarily reset max_cluster_size in config
                 // to avoid the possibility of hitting the limit, while we are applying config:
                 // creating new extensions, roles, etc...
-                config::compute_ctl_temp_override_create(pgdata_path, "neon.max_cluster_size=-1")?;
-                self.pg_reload_conf()?;
+                config::with_compute_ctl_tmp_override(
+                    pgdata_path,
+                    "neon.max_cluster_size=-1",
+                    || {
+                        self.pg_reload_conf()?;
 
-                self.apply_config(&compute_state)?;
+                        self.apply_config(&compute_state)?;
 
-                config::compute_ctl_temp_override_remove(pgdata_path)?;
+                        Ok(())
+                    },
+                )?;
                 self.pg_reload_conf()?;
             }
             self.post_apply_config()?;

compute_tools/src/config.rs

@@ -131,18 +131,17 @@ pub fn write_postgres_conf(
     Ok(())
 }
 
-/// create file compute_ctl_temp_override.conf in pgdata_dir
-/// add provided options to this file
-pub fn compute_ctl_temp_override_create(pgdata_path: &Path, options: &str) -> Result<()> {
+pub fn with_compute_ctl_tmp_override<F>(pgdata_path: &Path, options: &str, exec: F) -> Result<()>
+where
+    F: FnOnce() -> Result<()>,
+{
     let path = pgdata_path.join("compute_ctl_temp_override.conf");
     let mut file = File::create(path)?;
     write!(file, "{}", options)?;
-    Ok(())
-}
 
-/// remove file compute_ctl_temp_override.conf in pgdata_dir
-pub fn compute_ctl_temp_override_remove(pgdata_path: &Path) -> Result<()> {
-    let path = pgdata_path.join("compute_ctl_temp_override.conf");
-    std::fs::remove_file(path)?;
-    Ok(())
+    let res = exec();
+
+    file.set_len(0)?;
+
+    res
 }

compute_tools/src/http/api.rs

@@ -17,7 +17,7 @@ use hyper::header::CONTENT_TYPE;
 use hyper::service::{make_service_fn, service_fn};
 use hyper::{Body, Method, Request, Response, Server, StatusCode};
 use tokio::task;
-use tracing::{error, info, warn};
+use tracing::{debug, error, info, warn};
 use tracing_utils::http::OtelName;
 use utils::http::request::must_get_query_param;
 
@@ -48,7 +48,7 @@ async fn routes(req: Request<Body>, compute: &Arc<ComputeNode>) -> Response<Body
     match (req.method(), req.uri().path()) {
         // Serialized compute state.
         (&Method::GET, "/status") => {
-            info!("serving /status GET request");
+            debug!("serving /status GET request");
             let state = compute.state.lock().unwrap();
             let status_response = status_response_from_state(&state);
             Response::new(Body::from(serde_json::to_string(&status_response).unwrap()))

control_plane/src/pageserver.rs

@@ -383,6 +383,10 @@ impl PageServerNode {
                 .map(|x| x.parse::<AuxFilePolicy>())
                 .transpose()
                 .context("Failed to parse 'switch_aux_file_policy'")?,
+            lsn_lease_length: settings.remove("lsn_lease_length").map(|x| x.to_string()),
+            lsn_lease_length_for_ts: settings
+                .remove("lsn_lease_length_for_ts")
+                .map(|x| x.to_string()),
         };
         if !settings.is_empty() {
             bail!("Unrecognized tenant settings: {settings:?}")
@@ -506,6 +510,10 @@ impl PageServerNode {
                     .map(|x| x.parse::<AuxFilePolicy>())
                     .transpose()
                     .context("Failed to parse 'switch_aux_file_policy'")?,
+                lsn_lease_length: settings.remove("lsn_lease_length").map(|x| x.to_string()),
+                lsn_lease_length_for_ts: settings
+                    .remove("lsn_lease_length_for_ts")
+                    .map(|x| x.to_string()),
             }
         };
 

control_plane/src/safekeeper.rs

@@ -14,6 +14,7 @@ use camino::Utf8PathBuf;
 use postgres_connection::PgConnectionConfig;
 use reqwest::{IntoUrl, Method};
 use thiserror::Error;
+use utils::auth::{Claims, Scope};
 use utils::{http::error::HttpErrorBody, id::NodeId};
 
 use crate::{
@@ -197,7 +198,7 @@ impl SafekeeperNode {
             &datadir,
             &self.env.safekeeper_bin(),
             &args,
-            [],
+            self.safekeeper_env_variables()?,
             background_process::InitialPidFile::Expect(self.pid_file()),
             || async {
                 match self.check_status().await {
@@ -210,6 +211,18 @@ impl SafekeeperNode {
         .await
     }
 
+    fn safekeeper_env_variables(&self) -> anyhow::Result<Vec<(String, String)>> {
+        // Generate a token to connect from safekeeper to peers
+        if self.conf.auth_enabled {
+            let token = self
+                .env
+                .generate_auth_token(&Claims::new(None, Scope::SafekeeperData))?;
+            Ok(vec![("SAFEKEEPER_AUTH_TOKEN".to_owned(), token)])
+        } else {
+            Ok(Vec::new())
+        }
+    }
+
     ///
     /// Stop the server.
     ///

docs/core_changes.md

@@ -11,15 +11,28 @@ page server. We currently use the same binary for both, with --wal-redo runtime
 the WAL redo mode. Some PostgreSQL changes are needed in the compute node, while others are just for
 the WAL redo process.
 
-In addition to core PostgreSQL changes, there is a Neon extension in contrib/neon, to hook into the
-smgr interface. Once all the core changes have been submitted to upstream or eliminated some other
-way, the extension could live outside the postgres repository and build against vanilla PostgreSQL.
+In addition to core PostgreSQL changes, there is a Neon extension in the pgxn/neon directory that
+hooks into the smgr interface, and rmgr extension in pgxn/neon_rmgr. The extensions are loaded into
+the Postgres processes with shared_preload_libraries. Most of the Neon-specific code is in the
+extensions, and for any new features, that is preferred over modifying core PostgreSQL code.
 
 Below is a list of all the PostgreSQL source code changes, categorized into changes needed for
 compute, and changes needed for the WAL redo process:
 
 # Changes for Compute node
 
+## Prefetching
+
+There are changes in many places to perform prefetching, for example for sequential scans. Neon
+doesn't benefit from OS readahead, and the latency to pageservers is quite high compared to local
+disk, so prefetching is critical for performance, also for sequential scans.
+
+### How to get rid of the patch
+
+Upcoming "streaming read" work in v17 might simplify this. And async I/O work in v18 will hopefully
+do more.
+
+
 ## Add t_cid to heap WAL records
 
 ```
@@ -37,54 +50,11 @@ The problem is that the XLOG_HEAP_INSERT record does not include the command id
 
 Bite the bullet and submit the patch to PostgreSQL, to add the t_cid to the WAL records. It makes the WAL records larger, which could make this unpopular in the PostgreSQL community. However, it might simplify some logical decoding code; Andres Freund briefly mentioned in PGCon 2022 discussion on Heikki's Neon presentation that logical decoding currently needs to jump through some hoops to reconstruct the same information.
 
+Update from Heikki (2024-04-17): I tried to write an upstream patch for that, to use the t_cid field for logical decoding, but it was not as straightforward as it first sounded.
 
 ### Alternatives
 Perhaps we could write an extra WAL record with the t_cid information, when a page is evicted that contains rows that were touched a transaction that's still running. However, that seems very complicated.
 
-## ginfast.c
-
-```
-diff --git a/src/backend/access/gin/ginfast.c b/src/backend/access/gin/ginfast.c
-index e0d9940946..2d964c02e9 100644
---- a/src/backend/access/gin/ginfast.c
-+++ b/src/backend/access/gin/ginfast.c
-@@ -285,6 +285,17 @@ ginHeapTupleFastInsert(GinState *ginstate, GinTupleCollector *collector)
-                memset(&sublist, 0, sizeof(GinMetaPageData));
-                makeSublist(index, collector->tuples, collector->ntuples, &sublist);
- 
-+               if (metadata->head != InvalidBlockNumber)
-+               {
-+                       /*
-+                        * ZENITH: Get buffer before XLogBeginInsert() to avoid recursive call
-+                        * of XLogBeginInsert(). Reading a new buffer might evict a dirty page from
-+                        * the buffer cache, and if that page happens to be an FSM or VM page, zenith_write()
-+                        * will try to WAL-log an image of the page.
-+                        */
-+                       buffer = ReadBuffer(index, metadata->tail);
-+               }
-+
-                if (needWal)
-                        XLogBeginInsert();
- 
-@@ -316,7 +327,6 @@ ginHeapTupleFastInsert(GinState *ginstate, GinTupleCollector *collector)
-                        data.prevTail = metadata->tail;
-                        data.newRightlink = sublist.head;
- 
--                       buffer = ReadBuffer(index, metadata->tail);
-                        LockBuffer(buffer, GIN_EXCLUSIVE);
-                        page = BufferGetPage(buffer);
-```
-
-The problem is explained in the comment above
-
-### How to get rid of the patch
-
-Can we stop WAL-logging FSM or VM pages? Or delay the WAL logging until we're out of the critical
-section or something.
-
-Maybe some bigger rewrite of FSM and VM would help to avoid WAL-logging FSM and VM page images?
-
-
 ## Mark index builds that use buffer manager without logging explicitly
 
 ```
@@ -95,6 +65,8 @@ Maybe some bigger rewrite of FSM and VM would help to avoid WAL-logging FSM and
 also some changes in src/backend/storage/smgr/smgr.c
 ```
 
+pgvector 0.6.0 also needs a similar change, which would be very nice to get rid of too.
+
 When a GIN index is built, for example, it is built by inserting the entries into the index more or
 less normally, but without WAL-logging anything. After the index has been built, we iterate through
 all pages and write them to the WAL. That doesn't work for Neon, because if a page is not WAL-logged
@@ -109,6 +81,10 @@ an operation: `smgr_start_unlogged_build`, `smgr_finish_unlogged_build_phase_1`
 I think it would make sense to be more explicit about that in PostgreSQL too. So extract these
 changes to a patch and post to pgsql-hackers.
 
+Perhaps we could deduce that an unlogged index build has started when we see a page being evicted
+with zero LSN. How to be sure it's an unlogged index build rather than a bug? Currently we have a
+check for that and PANIC if we see page with zero LSN being evicted. And how do we detect when the
+index build has finished? See https://github.com/neondatabase/neon/pull/7440 for an attempt at that.
 
 ## Track last-written page LSN
 
@@ -140,57 +116,6 @@ The old method is still available, though.
 Wait until v15?
 
 
-## Cache relation sizes
-
-The Neon extension contains a little cache for smgrnblocks() and smgrexists() calls, to avoid going
-to the page server every time. It might be useful to cache those in PostgreSQL, maybe in the
-relcache? (I think we do cache nblocks in relcache already, check why that's not good enough for
-Neon)
-
-
-## Use buffer manager when extending VM or FSM
-
-```
- src/backend/storage/freespace/freespace.c                   |   14 +-
- src/backend/access/heap/visibilitymap.c                     |   15 +-
-
-diff --git a/src/backend/access/heap/visibilitymap.c b/src/backend/access/heap/visibilitymap.c
-index e198df65d8..addfe93eac 100644
---- a/src/backend/access/heap/visibilitymap.c
-+++ b/src/backend/access/heap/visibilitymap.c
-@@ -652,10 +652,19 @@ vm_extend(Relation rel, BlockNumber vm_nblocks)
-        /* Now extend the file */
-        while (vm_nblocks_now < vm_nblocks)
-        {
--               PageSetChecksumInplace((Page) pg.data, vm_nblocks_now);
-+               /*
-+                * ZENITH: Initialize VM pages through buffer cache to prevent loading
-+                * them from pageserver.
-+                */
-+               Buffer  buffer = ReadBufferExtended(rel, VISIBILITYMAP_FORKNUM, P_NEW,
-+                                                                                       RBM_ZERO_AND_LOCK, NULL);
-+               Page    page = BufferGetPage(buffer);
-+
-+               PageInit((Page) page, BLCKSZ, 0);
-+               PageSetChecksumInplace(page, vm_nblocks_now);
-+               MarkBufferDirty(buffer);
-+               UnlockReleaseBuffer(buffer);
- 
--               smgrextend(rel->rd_smgr, VISIBILITYMAP_FORKNUM, vm_nblocks_now,
--                                  pg.data, false);
-                vm_nblocks_now++;
-        }
-```
-
-### Problem we're trying to solve
-
-???
-
-### How to get rid of the patch
-
-Maybe this would be a reasonable change in PostgreSQL too?
-
-
 ## Allow startup without reading checkpoint record
 
 In Neon, the compute node is stateless. So when we are launching compute node, we need to provide
@@ -231,7 +156,7 @@ index 0415df9ccb..9f9db3c8bc 100644
   * crash we can lose (skip over) as many values as we pre-logged.
   */
 -#define SEQ_LOG_VALS   32
-+/* Zenith XXX: to ensure sequence order of sequence in Zenith we need to WAL log each sequence update. */
++/* Neon XXX: to ensure sequence order of sequence in Zenith we need to WAL log each sequence update. */
 +/* #define SEQ_LOG_VALS        32 */
 +#define SEQ_LOG_VALS   0
 ```
@@ -250,66 +175,6 @@ would be weird if the sequence moved backwards though, think of PITR.
 Or add a GUC for the amount to prefix to PostgreSQL, and force it to 1 in Neon.
 
 
-## Walproposer
-
-```
- src/Makefile                                                |    1 +
- src/backend/replication/libpqwalproposer/Makefile           |   37 +
- src/backend/replication/libpqwalproposer/libpqwalproposer.c |  416 ++++++++++++
- src/backend/postmaster/bgworker.c                           |    4 +
- src/backend/postmaster/postmaster.c                         |    6 +
- src/backend/replication/Makefile                            |    4 +-
- src/backend/replication/walproposer.c                       | 2350 +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
- src/backend/replication/walproposer_utils.c                 |  402 +++++++++++
- src/backend/replication/walreceiver.c                       |    7 +
- src/backend/replication/walsender.c                         |  320 ++++++---
- src/backend/storage/ipc/ipci.c                              |    6 +
- src/include/replication/walproposer.h                       |  565 ++++++++++++++++
-```
-
-WAL proposer is communicating with safekeeper and ensures WAL durability by quorum writes.  It is
-currently implemented as patch to standard WAL sender.
-
-### How to get rid of the patch
-
-Refactor into an extension. Submit hooks or APIs into upstream if necessary.
-
-@MMeent did some work on this already: https://github.com/neondatabase/postgres/pull/96
-
-## Ignore unexpected data beyond EOF in bufmgr.c
-
-```
-@@ -922,11 +928,14 @@ ReadBuffer_common(SMgrRelation smgr, char relpersistence, ForkNumber forkNum,
-                 */
-                bufBlock = isLocalBuf ? LocalBufHdrGetBlock(bufHdr) : BufHdrGetBlock(bufHdr);
-                if (!PageIsNew((Page) bufBlock))
--                       ereport(ERROR,
-+               {
-+                        // XXX-ZENITH
-+                        MemSet((char *) bufBlock, 0, BLCKSZ);
-+                        ereport(DEBUG1,
-                                        (errmsg("unexpected data beyond EOF in block %u of relation %s",
-                                                        blockNum, relpath(smgr->smgr_rnode, forkNum)),
-                                         errhint("This has been seen to occur with buggy kernels; consider updating your system.")));
--
-+               }
-                /*
-                 * We *must* do smgrextend before succeeding, else the page will not
-                 * be reserved by the kernel, and the next P_NEW call will decide to
-```
-
-PostgreSQL is a bit sloppy with extending relations. Usually, the relation is extended with zeros
-first, then the page is filled, and finally the new page WAL-logged. But if multiple backends extend
-a relation at the same time, the pages can be WAL-logged in different order.
-
-I'm not sure what scenario exactly required this change in Neon, though.
-
-### How to get rid of the patch
-
-Submit patches to pgsql-hackers, to tighten up the WAL-logging around relation extension. It's a bit
-confusing even in PostgreSQL. Maybe WAL log the intention to extend first, then extend the relation,
-and finally WAL-log that the extension succeeded.
-
 ## Make smgr interface available to extensions
 
 ```
@@ -321,6 +186,8 @@ and finally WAL-log that the extension succeeded.
 
 Submit to upstream. This could be useful for the Disk Encryption patches too, or for compression.
 
+We have submitted this to upstream, but it's moving at glacial a speed.
+https://commitfest.postgresql.org/47/4428/
 
 ## Added relpersistence argument to smgropen()
 
@@ -444,6 +311,148 @@ Ignore it. This is only needed for disaster recovery, so once we've eliminated a
 patches, we can just keep it around as a patch or as separate branch in a repo.
 
 
+## pg_waldump flags to ignore errors
+
+After creating a new project or branch in Neon, the first timeline can begin in the middle of a WAL segment. pg_waldump chokes on that, so we added some flags to make it possible to ignore errors.
+
+### How to get rid of the patch
+
+Like previous one, ignore it.
+
+
+
+## Backpressure if pageserver doesn't ingest WAL fast enough
+
+```
+@@ -3200,6 +3202,7 @@ ProcessInterrupts(void)
+                return;
+        InterruptPending = false;
+ 
++retry:
+        if (ProcDiePending)
+        {
+                ProcDiePending = false;
+@@ -3447,6 +3450,13 @@ ProcessInterrupts(void)
+ 
+        if (ParallelApplyMessagePending)
+                HandleParallelApplyMessages();
++
++       /* Call registered callback if any */
++       if (ProcessInterruptsCallback)
++       {
++               if (ProcessInterruptsCallback())
++                       goto retry;
++       }
+ }
+```
+
+
+### How to get rid of the patch
+
+Submit a patch to upstream, for a hook in ProcessInterrupts. Could be useful for other extensions
+too.
+
+
+## SLRU on-demand download
+
+```
+ src/backend/access/transam/slru.c | 105 ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++-------------
+ 1 file changed, 92 insertions(+), 13 deletions(-)
+```
+
+### Problem we're trying to solve
+
+Previously, SLRU files were included in the basebackup, but the total size of them can be large,
+several GB, and downloading them all made the startup time too long.
+
+### Alternatives
+
+FUSE hook or LD_PRELOAD trick to intercept the reads on SLRU files
+
+
+## WAL-log an all-zeros page as one large hole
+
+- In XLogRecordAssemble()
+
+### Problem we're trying to solve
+
+This change was made in v16. Starting with v16, when PostgreSQL extends a relation, it first extends
+it with zeros, and it can extend the relation more than one block at a time. The all-zeros page is WAL-ogged, but it's very wasteful to include 8 kB of zeros in the WAL for that. This hack was made so that we WAL logged a compact record with a whole-page "hole". However, PostgreSQL has assertions that prevent that such WAL records from being replayed, so this breaks compatibility such that unmodified PostreSQL cannot process Neon-generated WAL.
+
+### How to get rid of the patch
+
+Find another compact representation for a full-page image of an all-zeros page. A compressed image perhaps.
+
+
+## Shut down walproposer after checkpointer
+
+```
++                       /* Neon: Also allow walproposer background worker to be treated like a WAL sender, so that it's shut down last */
++                       if ((bp->bkend_type == BACKEND_TYPE_NORMAL || bp->bkend_type == BACKEND_TYPE_BGWORKER) &&
+```
+
+This changes was needed so that postmaster shuts down the walproposer process only after the shutdown checkpoint record is written. Otherwise, the shutdown record will never make it to the safekeepers.
+
+### How to get rid of the patch
+
+Do a bigger refactoring of the postmaster state machine, such that a background worker can specify
+the shutdown ordering by itself. The postmaster state machine has grown pretty complicated, and
+would benefit from a refactoring for the sake of readability anyway.
+
+
+## EXPLAIN changes for prefetch and LFC
+
+### How to get rid of the patch
+
+Konstantin submitted a patch to -hackers already: https://commitfest.postgresql.org/47/4643/. Get that into a committable state.
+
+
+## On-demand download of extensions
+
+### How to get rid of the patch
+
+FUSE or LD_PRELOAD trickery to intercept reads?
+
+
+## Publication superuser checks
+
+We have hacked CreatePublication so that also neon_superuser can create them.
+
+### How to get rid of the patch
+
+Create an upstream patch with more fine-grained privileges for publications CREATE/DROP that can be GRANTed to users.
+
+
+## WAL log replication slots
+
+### How to get rid of the patch
+
+Utilize the upcoming v17 "slot sync worker", or a similar neon-specific background worker process, to periodically WAL-log the slots, or to export them somewhere else.
+
+
+## WAL-log replication snapshots
+
+### How to get rid of the patch
+
+WAL-log them periodically, from a backgound worker.
+
+
+## WAL-log relmapper files
+
+Similarly to replications snapshot files, the CID mapping files generated during VACUUM FULL of a catalog table are WAL-logged
+
+### How to get rid of the patch
+
+WAL-log them periodically, from a backgound worker.
+
+
+## XLogWaitForReplayOf()
+
+??
+
+
+
+
 # Not currently committed but proposed
 
 ## Disable ring buffer buffer manager strategies
@@ -472,23 +481,10 @@ hint bits are set. Wal logging hint bits updates requires FPI which significantl
 
 Add special WAL record for setting page hints.
 
-## Prefetching
-
-### Why?
-
-As far as pages in Neon are loaded on demand, to reduce node startup time
-and also speedup some massive queries we need some mechanism for bulk loading to
-reduce page request round-trip overhead.
-
-Currently Postgres is supporting prefetching only for bitmap scan.
-In Neon we should also use prefetch for sequential and index scans, because the OS is not doing it for us.
-For sequential scan we could prefetch some number of following pages. For index scan we could prefetch pages
-of heap relation addressed by TIDs.
-
 ## Prewarming
 
 ### Why?
 
-Short downtime (or, in other words, fast compute node restart time) is one of the key feature of Zenith.
+Short downtime (or, in other words, fast compute node restart time) is one of the key feature of Neon.
 But overhead of request-response round-trip for loading pages on demand can make started node warm-up quite slow.
 We can capture state of compute node buffer cache and send bulk request for this pages at startup.

docs/pageserver-services.md

@@ -101,11 +101,12 @@ or
 ```toml
 [remote_storage]
 container_name = 'some-container-name'
+storage_account = 'somestorageaccnt'
 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.
+The `AZURE_STORAGE_ACCESS_KEY` env variable can be used to specify the azure credentials if needed.
 
 ## Repository background tasks
 

libs/pageserver_api/src/models.rs

@@ -177,6 +177,20 @@ serde_with::serde_conv!(
     |value: String| -> Result<_, humantime::TimestampError> { humantime::parse_rfc3339(&value) }
 );
 
+impl LsnLease {
+    /// The default length for an explicit LSN lease request (10 minutes).
+    pub const DEFAULT_LENGTH: Duration = Duration::from_secs(10 * 60);
+
+    /// The default length for an implicit LSN lease granted during
+    /// `get_lsn_by_timestamp` request (1 minutes).
+    pub const DEFAULT_LENGTH_FOR_TS: Duration = Duration::from_secs(60);
+
+    /// Checks whether the lease is expired.
+    pub fn is_expired(&self, now: &SystemTime) -> bool {
+        now > &self.valid_until
+    }
+}
+
 /// The only [`TenantState`] variants we could be `TenantState::Activating` from.
 #[derive(Clone, Copy, Debug, PartialEq, Eq, serde::Serialize, serde::Deserialize)]
 pub enum ActivatingFrom {
@@ -322,6 +336,8 @@ pub struct TenantConfig {
     pub timeline_get_throttle: Option<ThrottleConfig>,
     pub image_layer_creation_check_threshold: Option<u8>,
     pub switch_aux_file_policy: Option<AuxFilePolicy>,
+    pub lsn_lease_length: Option<String>,
+    pub lsn_lease_length_for_ts: Option<String>,
 }
 
 /// The policy for the aux file storage. It can be switched through `switch_aux_file_policy`

libs/remote_storage/src/azure_blob.rs

@@ -54,7 +54,10 @@ impl AzureBlobStorage {
             azure_config.container_name
         );
 
-        let account = env::var("AZURE_STORAGE_ACCOUNT").expect("missing AZURE_STORAGE_ACCOUNT");
+        // Use the storage account from the config by default, fall back to env var if not present.
+        let account = azure_config.storage_account.clone().unwrap_or_else(|| {
+            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.

libs/remote_storage/src/lib.rs

@@ -466,7 +466,11 @@ impl GenericRemoteStorage {
                 Self::AwsS3(Arc::new(S3Bucket::new(s3_config, timeout)?))
             }
             RemoteStorageKind::AzureContainer(azure_config) => {
-                info!("Using azure container '{}' in region '{}' as a remote storage, prefix in container: '{:?}'",
+                let storage_account = azure_config
+                    .storage_account
+                    .as_deref()
+                    .unwrap_or("<AZURE_STORAGE_ACCOUNT>");
+                info!("Using azure container '{}' in account '{storage_account}' 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, timeout)?))
             }
@@ -589,6 +593,8 @@ impl Debug for S3Config {
 pub struct AzureConfig {
     /// Name of the container to connect to.
     pub container_name: String,
+    /// Name of the storage account the container is inside of
+    pub storage_account: Option<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.
@@ -603,8 +609,9 @@ 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("storage_account", &self.storage_account)
             .field("bucket_region", &self.container_region)
-            .field("prefix_in_bucket", &self.prefix_in_container)
+            .field("prefix_in_container", &self.prefix_in_container)
             .field("concurrency_limit", &self.concurrency_limit)
             .field(
                 "max_keys_per_list_response",
@@ -718,6 +725,12 @@ impl RemoteStorageConfig {
             (None, None, None, Some(container_name), Some(container_region)) => {
                 RemoteStorageKind::AzureContainer(AzureConfig {
                     container_name: parse_toml_string("container_name", container_name)?,
+                    storage_account: toml
+                        .get("storage_account")
+                        .map(|storage_account| {
+                            parse_toml_string("storage_account", storage_account)
+                        })
+                        .transpose()?,
                     container_region: parse_toml_string("container_region", container_region)?,
                     prefix_in_container: toml
                         .get("prefix_in_container")

libs/remote_storage/tests/test_real_azure.rs

@@ -212,6 +212,7 @@ fn create_azure_client(
     let remote_storage_config = RemoteStorageConfig {
         storage: RemoteStorageKind::AzureContainer(AzureConfig {
             container_name: remote_storage_azure_container,
+            storage_account: None,
             container_region: remote_storage_azure_region,
             prefix_in_container: Some(format!("test_{millis}_{random:08x}/")),
             concurrency_limit: NonZeroUsize::new(100).unwrap(),

libs/vm_monitor/src/cgroup.rs

@@ -25,6 +25,8 @@ pub struct Config {
     ///
     /// For simplicity, this value must be greater than or equal to `memory_history_len`.
     memory_history_log_interval: usize,
+    /// The max number of iterations to skip before logging the next iteration
+    memory_history_log_noskip_interval: Duration,
 }
 
 impl Default for Config {
@@ -33,6 +35,7 @@ impl Default for Config {
             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)
+            memory_history_log_noskip_interval: Duration::from_secs(15), // but only if it's changed, or 60 seconds have passed
         }
     }
 }
@@ -85,7 +88,12 @@ impl CgroupWatcher {
 
         // buffer for samples that will be logged. once full, it remains so.
         let history_log_len = self.config.memory_history_log_interval;
+        let max_skip = self.config.memory_history_log_noskip_interval;
         let mut history_log_buf = vec![MemoryStatus::zeroed(); history_log_len];
+        let mut last_logged_memusage = MemoryStatus::zeroed();
+
+        // Ensure that we're tracking a value that's definitely in the past, as Instant::now is only guaranteed to be non-decreasing on Rust's T1-supported systems.
+        let mut can_skip_logs_until = Instant::now() - max_skip;
 
         for t in 0_u64.. {
             ticker.tick().await;
@@ -115,12 +123,24 @@ impl CgroupWatcher {
             // 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 {
+            //
+            // We skip logging the data if data hasn't meaningfully changed in a while, unless
+            // we've already ignored previous iterations for the last max_skip period.
+            if i == history_log_len - 1
+                && (now > can_skip_logs_until
+                    || !history_log_buf
+                        .iter()
+                        .all(|usage| last_logged_memusage.status_is_close_or_similar(usage)))
+            {
                 info!(
                     history = ?MemoryStatus::debug_slice(&history_log_buf),
                     summary = ?summary,
                     "Recent cgroup memory statistics history"
                 );
+
+                can_skip_logs_until = now + max_skip;
+
+                last_logged_memusage = *history_log_buf.last().unwrap();
             }
 
             updates
@@ -232,6 +252,24 @@ impl MemoryStatus {
 
         DS(slice)
     }
+
+    /// Check if the other memory status is a close or similar result.
+    /// Returns true if the larger value is not larger than the smaller value
+    /// by 1/8 of the smaller value, and within 128MiB.
+    /// See tests::check_similarity_behaviour for examples of behaviour
+    fn status_is_close_or_similar(&self, other: &MemoryStatus) -> bool {
+        let margin;
+        let diff;
+        if self.non_reclaimable >= other.non_reclaimable {
+            margin = other.non_reclaimable / 8;
+            diff = self.non_reclaimable - other.non_reclaimable;
+        } else {
+            margin = self.non_reclaimable / 8;
+            diff = other.non_reclaimable - self.non_reclaimable;
+        }
+
+        diff < margin && diff < 128 * 1024 * 1024
+    }
 }
 
 #[cfg(test)]
@@ -261,4 +299,65 @@ mod tests {
         assert_eq!(values(2, 4), [9, 0, 1, 2]);
         assert_eq!(values(2, 10), [3, 4, 5, 6, 7, 8, 9, 0, 1, 2]);
     }
+
+    #[test]
+    fn check_similarity_behaviour() {
+        // This all accesses private methods, so we can't actually run this
+        // as doctests, because doctests run as an external crate.
+        let mut small = super::MemoryStatus {
+            non_reclaimable: 1024,
+        };
+        let mut large = super::MemoryStatus {
+            non_reclaimable: 1024 * 1024 * 1024 * 1024,
+        };
+
+        // objects are self-similar, no matter the size
+        assert!(small.status_is_close_or_similar(&small));
+        assert!(large.status_is_close_or_similar(&large));
+
+        // inequality is symmetric
+        assert!(!small.status_is_close_or_similar(&large));
+        assert!(!large.status_is_close_or_similar(&small));
+
+        small.non_reclaimable = 64;
+        large.non_reclaimable = (small.non_reclaimable / 8) * 9;
+
+        // objects are self-similar, no matter the size
+        assert!(small.status_is_close_or_similar(&small));
+        assert!(large.status_is_close_or_similar(&large));
+
+        // values are similar if the larger value is larger by less than
+        // 12.5%, i.e. 1/8 of the smaller value.
+        // In the example above, large is exactly 12.5% larger, so this doesn't
+        // match.
+        assert!(!small.status_is_close_or_similar(&large));
+        assert!(!large.status_is_close_or_similar(&small));
+
+        large.non_reclaimable -= 1;
+        assert!(large.status_is_close_or_similar(&large));
+
+        assert!(small.status_is_close_or_similar(&large));
+        assert!(large.status_is_close_or_similar(&small));
+
+        // The 1/8 rule only applies up to 128MiB of difference
+        small.non_reclaimable = 1024 * 1024 * 1024 * 1024;
+        large.non_reclaimable = small.non_reclaimable / 8 * 9;
+        assert!(small.status_is_close_or_similar(&small));
+        assert!(large.status_is_close_or_similar(&large));
+
+        assert!(!small.status_is_close_or_similar(&large));
+        assert!(!large.status_is_close_or_similar(&small));
+        // the large value is put just above the threshold
+        large.non_reclaimable = small.non_reclaimable + 128 * 1024 * 1024;
+        assert!(large.status_is_close_or_similar(&large));
+
+        assert!(!small.status_is_close_or_similar(&large));
+        assert!(!large.status_is_close_or_similar(&small));
+        // now below
+        large.non_reclaimable -= 1;
+        assert!(large.status_is_close_or_similar(&large));
+
+        assert!(small.status_is_close_or_similar(&large));
+        assert!(large.status_is_close_or_similar(&small));
+    }
 }

libs/vm_monitor/src/dispatcher.rs

@@ -12,11 +12,11 @@ use futures::{
     stream::{SplitSink, SplitStream},
     SinkExt, StreamExt,
 };
-use tracing::info;
+use tracing::{debug, info};
 
 use crate::protocol::{
-    OutboundMsg, ProtocolRange, ProtocolResponse, ProtocolVersion, PROTOCOL_MAX_VERSION,
-    PROTOCOL_MIN_VERSION,
+    OutboundMsg, OutboundMsgKind, ProtocolRange, ProtocolResponse, ProtocolVersion,
+    PROTOCOL_MAX_VERSION, PROTOCOL_MIN_VERSION,
 };
 
 /// The central handler for all communications in the monitor.
@@ -118,7 +118,12 @@ impl Dispatcher {
     /// serialize the wrong thing and send it, since `self.sink.send` will take
     /// any string.
     pub async fn send(&mut self, message: OutboundMsg) -> anyhow::Result<()> {
-        info!(?message, "sending message");
+        if matches!(&message.inner, OutboundMsgKind::HealthCheck { .. }) {
+            debug!(?message, "sending message");
+        } else {
+            info!(?message, "sending message");
+        }
+
         let json = serde_json::to_string(&message).context("failed to serialize message")?;
         self.sink
             .send(Message::Text(json))

libs/vm_monitor/src/runner.rs

@@ -12,7 +12,7 @@ use axum::extract::ws::{Message, WebSocket};
 use futures::StreamExt;
 use tokio::sync::{broadcast, watch};
 use tokio_util::sync::CancellationToken;
-use tracing::{error, info, warn};
+use tracing::{debug, error, info, warn};
 
 use crate::cgroup::{self, CgroupWatcher};
 use crate::dispatcher::Dispatcher;
@@ -474,26 +474,29 @@ impl Runner {
                 // there is a message from the agent
                 msg = self.dispatcher.source.next() => {
                     if let Some(msg) = msg {
-                        // Don't use 'message' as a key as the string also uses
-                        // that for its key
-                        info!(?msg, "received message");
-                        match msg {
+                        match &msg {
                             Ok(msg) => {
                                 let message: InboundMsg = match msg {
                                     Message::Text(text) => {
-                                        serde_json::from_str(&text).context("failed to deserialize text message")?
+                                        serde_json::from_str(text).context("failed to deserialize text message")?
                                     }
                                     other => {
                                         warn!(
                                             // Don't use 'message' as a key as the
                                             // string also uses that for its key
                                             msg = ?other,
-                                            "agent should only send text messages but received different type"
+                                            "problem processing incoming message: agent should only send text messages but received different type"
                                         );
                                         continue
                                     },
                                 };
 
+                                if matches!(&message.inner, InboundMsgKind::HealthCheck { .. }) {
+                                    debug!(?msg, "received message");
+                                } else {
+                                    info!(?msg, "received message");
+                                }
+
                                 let out = match self.process_message(message.clone()).await {
                                     Ok(Some(out)) => out,
                                     Ok(None) => continue,
@@ -517,7 +520,11 @@ impl Runner {
                                     .await
                                     .context("failed to send message")?;
                             }
-                            Err(e) => warn!("{e}"),
+                            Err(e) => warn!(
+                                error = format!("{e}"),
+                                msg = ?msg,
+                                "received error message"
+                            ),
                         }
                     } else {
                         anyhow::bail!("dispatcher connection closed")

pageserver/src/http/routes.rs

@@ -1730,7 +1730,7 @@ async fn lsn_lease_handler(
         active_timeline_of_active_tenant(&state.tenant_manager, tenant_shard_id, timeline_id)
             .await?;
     let result = timeline
-        .make_lsn_lease(lsn, &ctx)
+        .make_lsn_lease(lsn, timeline.get_lsn_lease_length(), &ctx)
         .map_err(|e| ApiError::InternalServerError(e.context("lsn lease http handler")))?;
 
     json_response(StatusCode::OK, result)

pageserver/src/page_service.rs

@@ -935,7 +935,7 @@ impl PageServerHandler {
         let timeline = self
             .get_active_tenant_timeline(tenant_shard_id.tenant_id, timeline_id, shard_selector)
             .await?;
-        let lease = timeline.make_lsn_lease(lsn, ctx)?;
+        let lease = timeline.make_lsn_lease(lsn, timeline.get_lsn_lease_length(), ctx)?;
         let valid_until = lease
             .valid_until
             .duration_since(SystemTime::UNIX_EPOCH)

pageserver/src/repository.rs

@@ -240,6 +240,7 @@ pub struct GcResult {
     pub layers_needed_by_cutoff: u64,
     pub layers_needed_by_pitr: u64,
     pub layers_needed_by_branches: u64,
+    pub layers_needed_by_leases: u64,
     pub layers_not_updated: u64,
     pub layers_removed: u64, // # of layer files removed because they have been made obsolete by newer ondisk files.
 
@@ -269,6 +270,7 @@ impl AddAssign for GcResult {
         self.layers_needed_by_pitr += other.layers_needed_by_pitr;
         self.layers_needed_by_cutoff += other.layers_needed_by_cutoff;
         self.layers_needed_by_branches += other.layers_needed_by_branches;
+        self.layers_needed_by_leases += other.layers_needed_by_leases;
         self.layers_not_updated += other.layers_not_updated;
         self.layers_removed += other.layers_removed;
 

pageserver/src/tenant.rs

@@ -31,6 +31,7 @@ use remote_storage::DownloadError;
 use remote_storage::GenericRemoteStorage;
 use remote_storage::TimeoutOrCancel;
 use std::fmt;
+use std::time::SystemTime;
 use storage_broker::BrokerClientChannel;
 use tokio::io::BufReader;
 use tokio::sync::watch;
@@ -65,9 +66,9 @@ use self::timeline::uninit::TimelineCreateGuard;
 use self::timeline::uninit::TimelineExclusionError;
 use self::timeline::uninit::UninitializedTimeline;
 use self::timeline::EvictionTaskTenantState;
+use self::timeline::GcCutoffs;
 use self::timeline::TimelineResources;
 use self::timeline::WaitLsnError;
-use self::timeline::{GcCutoffs, GcInfo};
 use crate::config::PageServerConf;
 use crate::context::{DownloadBehavior, RequestContext};
 use crate::deletion_queue::DeletionQueueClient;
@@ -2428,6 +2429,13 @@ impl Tenant {
         }
     }
 
+    pub fn get_lsn_lease_length(&self) -> Duration {
+        let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
+        tenant_conf
+            .lsn_lease_length
+            .unwrap_or(self.conf.default_tenant_conf.lsn_lease_length)
+    }
+
     pub fn set_new_tenant_config(&self, new_tenant_conf: TenantConfOpt) {
         // Use read-copy-update in order to avoid overwriting the location config
         // state if this races with [`Tenant::set_new_location_config`]. Note that
@@ -3010,12 +3018,13 @@ impl Tenant {
             {
                 let mut target = timeline.gc_info.write().unwrap();
 
+                let now = SystemTime::now();
+                target.leases.retain(|_, lease| !lease.is_expired(&now));
+
                 match gc_cutoffs.remove(&timeline.timeline_id) {
                     Some(cutoffs) => {
-                        *target = GcInfo {
-                            retain_lsns: branchpoints,
-                            cutoffs,
-                        };
+                        target.retain_lsns = branchpoints;
+                        target.cutoffs = cutoffs;
                     }
                     None => {
                         // reasons for this being unavailable:
@@ -3833,6 +3842,8 @@ pub(crate) mod harness {
                     tenant_conf.image_layer_creation_check_threshold,
                 ),
                 switch_aux_file_policy: Some(tenant_conf.switch_aux_file_policy),
+                lsn_lease_length: Some(tenant_conf.lsn_lease_length),
+                lsn_lease_length_for_ts: Some(tenant_conf.lsn_lease_length_for_ts),
             }
         }
     }
@@ -6939,4 +6950,93 @@ mod tests {
 
         Ok(())
     }
+
+    #[tokio::test]
+    async fn test_lsn_lease() -> anyhow::Result<()> {
+        let (tenant, ctx) = TenantHarness::create("test_lsn_lease")?.load().await;
+        let key = Key::from_hex("010000000033333333444444445500000000").unwrap();
+
+        let end_lsn = Lsn(0x100);
+        let image_layers = (0x20..=0x90)
+            .step_by(0x10)
+            .map(|n| {
+                (
+                    Lsn(n),
+                    vec![(key, test_img(&format!("data key at {:x}", n)))],
+                )
+            })
+            .collect();
+
+        let timeline = tenant
+            .create_test_timeline_with_layers(
+                TIMELINE_ID,
+                Lsn(0x10),
+                DEFAULT_PG_VERSION,
+                &ctx,
+                Vec::new(),
+                image_layers,
+                end_lsn,
+            )
+            .await?;
+
+        let leased_lsns = [0x30, 0x50, 0x70];
+        let mut leases = Vec::new();
+        let _: anyhow::Result<_> = leased_lsns.iter().try_for_each(|n| {
+            leases.push(timeline.make_lsn_lease(Lsn(*n), timeline.get_lsn_lease_length(), &ctx)?);
+            Ok(())
+        });
+
+        // Renewing with shorter lease should not change the lease.
+        let updated_lease_0 =
+            timeline.make_lsn_lease(Lsn(leased_lsns[0]), Duration::from_secs(0), &ctx)?;
+        assert_eq!(updated_lease_0.valid_until, leases[0].valid_until);
+
+        // Renewing with a long lease should renew lease with later expiration time.
+        let updated_lease_1 = timeline.make_lsn_lease(
+            Lsn(leased_lsns[1]),
+            timeline.get_lsn_lease_length() * 2,
+            &ctx,
+        )?;
+
+        assert!(updated_lease_1.valid_until > leases[1].valid_until);
+
+        // Force set disk consistent lsn so we can get the cutoff at `end_lsn`.
+        info!(
+            "latest_gc_cutoff_lsn: {}",
+            *timeline.get_latest_gc_cutoff_lsn()
+        );
+        timeline.force_set_disk_consistent_lsn(end_lsn);
+
+        let res = tenant
+            .gc_iteration(
+                Some(TIMELINE_ID),
+                0,
+                Duration::ZERO,
+                &CancellationToken::new(),
+                &ctx,
+            )
+            .await?;
+
+        // Keeping everything <= Lsn(0x80) b/c leases:
+        // 0/10: initdb layer
+        // (0/20..=0/70).step_by(0x10): image layers added when creating the timeline.
+        assert_eq!(res.layers_needed_by_leases, 7);
+        // Keeping 0/90 b/c it is the latest layer.
+        assert_eq!(res.layers_not_updated, 1);
+        // Removed 0/80.
+        assert_eq!(res.layers_removed, 1);
+
+        // Make lease on a already GC-ed LSN.
+        // 0/80 does not have a valid lease + is below latest_gc_cutoff
+        assert!(Lsn(0x80) < *timeline.get_latest_gc_cutoff_lsn());
+        let res = timeline.make_lsn_lease(Lsn(0x80), timeline.get_lsn_lease_length(), &ctx);
+        assert!(res.is_err());
+
+        // Should still be able to renew a currently valid lease
+        // Assumption: original lease to is still valid for 0/50.
+        let _ =
+            timeline.make_lsn_lease(Lsn(leased_lsns[1]), timeline.get_lsn_lease_length(), &ctx)?;
+
+        Ok(())
+    }
 }

pageserver/src/tenant/config.rs

@@ -13,6 +13,7 @@ use pageserver_api::models::AuxFilePolicy;
 use pageserver_api::models::CompactionAlgorithm;
 use pageserver_api::models::CompactionAlgorithmSettings;
 use pageserver_api::models::EvictionPolicy;
+use pageserver_api::models::LsnLease;
 use pageserver_api::models::{self, ThrottleConfig};
 use pageserver_api::shard::{ShardCount, ShardIdentity, ShardNumber, ShardStripeSize};
 use serde::de::IntoDeserializer;
@@ -377,6 +378,16 @@ pub struct TenantConf {
     /// There is a `last_aux_file_policy` flag which gets persisted in `index_part.json` once the first aux
     /// file is written.
     pub switch_aux_file_policy: AuxFilePolicy,
+
+    /// The length for an explicit LSN lease request.
+    /// Layers needed to reconstruct pages at LSN will not be GC-ed during this interval.
+    #[serde(with = "humantime_serde")]
+    pub lsn_lease_length: Duration,
+
+    /// The length for an implicit LSN lease granted as part of `get_lsn_by_timestamp` request.
+    /// Layers needed to reconstruct pages at LSN will not be GC-ed during this interval.
+    #[serde(with = "humantime_serde")]
+    pub lsn_lease_length_for_ts: Duration,
 }
 
 /// Same as TenantConf, but this struct preserves the information about
@@ -476,6 +487,16 @@ pub struct TenantConfOpt {
     #[serde(skip_serializing_if = "Option::is_none")]
     #[serde(default)]
     pub switch_aux_file_policy: Option<AuxFilePolicy>,
+
+    #[serde(skip_serializing_if = "Option::is_none")]
+    #[serde(with = "humantime_serde")]
+    #[serde(default)]
+    pub lsn_lease_length: Option<Duration>,
+
+    #[serde(skip_serializing_if = "Option::is_none")]
+    #[serde(with = "humantime_serde")]
+    #[serde(default)]
+    pub lsn_lease_length_for_ts: Option<Duration>,
 }
 
 impl TenantConfOpt {
@@ -538,6 +559,12 @@ impl TenantConfOpt {
             switch_aux_file_policy: self
                 .switch_aux_file_policy
                 .unwrap_or(global_conf.switch_aux_file_policy),
+            lsn_lease_length: self
+                .lsn_lease_length
+                .unwrap_or(global_conf.lsn_lease_length),
+            lsn_lease_length_for_ts: self
+                .lsn_lease_length_for_ts
+                .unwrap_or(global_conf.lsn_lease_length_for_ts),
         }
     }
 }
@@ -582,6 +609,8 @@ impl Default for TenantConf {
             timeline_get_throttle: crate::tenant::throttle::Config::disabled(),
             image_layer_creation_check_threshold: DEFAULT_IMAGE_LAYER_CREATION_CHECK_THRESHOLD,
             switch_aux_file_policy: AuxFilePolicy::default_tenant_config(),
+            lsn_lease_length: LsnLease::DEFAULT_LENGTH,
+            lsn_lease_length_for_ts: LsnLease::DEFAULT_LENGTH_FOR_TS,
         }
     }
 }
@@ -657,6 +686,8 @@ impl From<TenantConfOpt> for models::TenantConfig {
             timeline_get_throttle: value.timeline_get_throttle.map(ThrottleConfig::from),
             image_layer_creation_check_threshold: value.image_layer_creation_check_threshold,
             switch_aux_file_policy: value.switch_aux_file_policy,
+            lsn_lease_length: value.lsn_lease_length.map(humantime),
+            lsn_lease_length_for_ts: value.lsn_lease_length_for_ts.map(humantime),
         }
     }
 }

pageserver/src/tenant/disk_btree.rs

@@ -22,7 +22,7 @@ use async_stream::try_stream;
 use byteorder::{ReadBytesExt, BE};
 use bytes::{BufMut, Bytes, BytesMut};
 use either::Either;
-use futures::Stream;
+use futures::{Stream, StreamExt};
 use hex;
 use std::{
     cmp::Ordering,
@@ -259,6 +259,16 @@ where
         Ok(result)
     }
 
+    pub fn iter<'a>(
+        &'a self,
+        start_key: &'a [u8; L],
+        ctx: &'a RequestContext,
+    ) -> DiskBtreeIterator<'a> {
+        DiskBtreeIterator {
+            stream: Box::pin(self.get_stream_from(start_key, ctx)),
+        }
+    }
+
     /// Return a stream which yields all key, value pairs from the index
     /// starting from the first key greater or equal to `start_key`.
     ///
@@ -496,6 +506,19 @@ where
     }
 }
 
+pub struct DiskBtreeIterator<'a> {
+    #[allow(clippy::type_complexity)]
+    stream: std::pin::Pin<
+        Box<dyn Stream<Item = std::result::Result<(Vec<u8>, u64), DiskBtreeError>> + 'a>,
+    >,
+}
+
+impl<'a> DiskBtreeIterator<'a> {
+    pub async fn next(&mut self) -> Option<std::result::Result<(Vec<u8>, u64), DiskBtreeError>> {
+        self.stream.next().await
+    }
+}
+
 ///
 /// Public builder object, for creating a new tree.
 ///
@@ -1088,6 +1111,17 @@ pub(crate) mod tests {
                 == all_data.get(&u128::MAX).cloned()
         );
 
+        // Test iterator and get_stream API
+        let mut iter = reader.iter(&[0; 16], &ctx);
+        let mut cnt = 0;
+        while let Some(res) = iter.next().await {
+            let (key, val) = res?;
+            let key = u128::from_be_bytes(key.as_slice().try_into().unwrap());
+            assert_eq!(val, *all_data.get(&key).unwrap());
+            cnt += 1;
+        }
+        assert_eq!(cnt, all_data.len());
+
         Ok(())
     }
 

pageserver/src/tenant/size.rs

@@ -60,10 +60,6 @@ pub(crate) enum CalculateSyntheticSizeError {
     #[error(transparent)]
     Fatal(anyhow::Error),
 
-    /// The LSN we are trying to calculate a size at no longer exists at the point we query it
-    #[error("Could not find size at {lsn} in timeline {timeline_id}")]
-    LsnNotFound { timeline_id: TimelineId, lsn: Lsn },
-
     /// Tenant shut down while calculating size
     #[error("Cancelled")]
     Cancelled,
@@ -375,9 +371,8 @@ pub(super) async fn gather_inputs(
 
 /// Augment 'segments' with logical sizes
 ///
-/// this will probably conflict with on-demand downloaded layers, or at least force them all
-/// to be downloaded
-///
+/// This will leave segments' sizes as None if the Timeline associated with the segment is deleted concurrently
+/// (i.e. we cannot read its logical size at a particular LSN).
 async fn fill_logical_sizes(
     timelines: &[Arc<Timeline>],
     segments: &mut [SegmentMeta],
@@ -498,8 +493,6 @@ async fn fill_logical_sizes(
 
         if let Some(Some(size)) = sizes_needed.get(&(timeline_id, lsn)) {
             seg.segment.size = Some(*size);
-        } else {
-            return Err(CalculateSyntheticSizeError::LsnNotFound { timeline_id, lsn });
         }
     }
     Ok(())

pageserver/src/tenant/tasks.rs

@@ -346,6 +346,7 @@ async fn gc_loop(tenant: Arc<Tenant>, cancel: CancellationToken) {
         // cutoff specified as time.
         let ctx =
             RequestContext::todo_child(TaskKind::GarbageCollector, DownloadBehavior::Download);
+
         let mut first = true;
         loop {
             tokio::select! {
@@ -362,6 +363,14 @@ async fn gc_loop(tenant: Arc<Tenant>, cancel: CancellationToken) {
 
             if first {
                 first = false;
+
+                if delay_by_lease_length(tenant.get_lsn_lease_length(), &cancel)
+                    .await
+                    .is_err()
+                {
+                    break;
+                }
+
                 if random_init_delay(period, &cancel).await.is_err() {
                     break;
                 }
@@ -531,6 +540,21 @@ pub(crate) async fn random_init_delay(
     }
 }
 
+/// Delays GC by defaul lease length at restart.
+///
+/// We do this as the leases mapping are not persisted to disk. By delaying GC by default
+/// length, we gurantees that all the leases we granted before the restart will expire
+/// when we run GC for the first time after the restart.
+pub(crate) async fn delay_by_lease_length(
+    length: Duration,
+    cancel: &CancellationToken,
+) -> Result<(), Cancelled> {
+    match tokio::time::timeout(length, cancel.cancelled()).await {
+        Ok(_) => Err(Cancelled),
+        Err(_) => Ok(()),
+    }
+}
+
 /// Attention: the `task` and `period` beocme labels of a pageserver-wide prometheus metric.
 pub(crate) fn warn_when_period_overrun(
     elapsed: Duration,

pageserver/src/tenant/timeline.rs

@@ -47,7 +47,6 @@ use utils::{
     vec_map::VecMap,
 };
 
-use std::ops::{Deref, Range};
 use std::pin::pin;
 use std::sync::atomic::Ordering as AtomicOrdering;
 use std::sync::{Arc, Mutex, RwLock, Weak};
@@ -61,6 +60,10 @@ use std::{
     cmp::{max, min, Ordering},
     ops::ControlFlow,
 };
+use std::{
+    collections::btree_map::Entry,
+    ops::{Deref, Range},
+};
 
 use crate::metrics::GetKind;
 use crate::pgdatadir_mapping::MAX_AUX_FILE_V2_DELTAS;
@@ -454,6 +457,9 @@ pub(crate) struct GcInfo {
 
     /// The cutoff coordinates, which are combined by selecting the minimum.
     pub(crate) cutoffs: GcCutoffs,
+
+    /// Leases granted to particular LSNs.
+    pub(crate) leases: BTreeMap<Lsn, LsnLease>,
 }
 
 impl GcInfo {
@@ -1555,17 +1561,46 @@ impl Timeline {
         Ok(())
     }
 
-    /// Obtains a temporary lease blocking garbage collection for the given LSN
+    /// Obtains a temporary lease blocking garbage collection for the given LSN.
+    ///
+    /// This function will error if the requesting LSN is less than the `latest_gc_cutoff_lsn` and there is also
+    /// no existing lease to renew. If there is an existing lease in the map, the lease will be renewed only if
+    /// the request extends the lease. The returned lease is therefore the maximum between the existing lease and
+    /// the requesting lease.
     pub(crate) fn make_lsn_lease(
         &self,
-        _lsn: Lsn,
+        lsn: Lsn,
+        length: Duration,
         _ctx: &RequestContext,
     ) -> anyhow::Result<LsnLease> {
-        const LEASE_LENGTH: Duration = Duration::from_secs(5 * 60);
-        let lease = LsnLease {
-            valid_until: SystemTime::now() + LEASE_LENGTH,
+        let lease = {
+            let mut gc_info = self.gc_info.write().unwrap();
+
+            let valid_until = SystemTime::now() + length;
+
+            let entry = gc_info.leases.entry(lsn);
+
+            let lease = {
+                if let Entry::Occupied(mut occupied) = entry {
+                    let existing_lease = occupied.get_mut();
+                    if valid_until > existing_lease.valid_until {
+                        existing_lease.valid_until = valid_until;
+                    }
+                    existing_lease.clone()
+                } else {
+                    // Reject already GC-ed LSN (lsn < latest_gc_cutoff)
+                    let latest_gc_cutoff_lsn = self.get_latest_gc_cutoff_lsn();
+                    if lsn < *latest_gc_cutoff_lsn {
+                        bail!("tried to request a page version that was garbage collected. requested at {} gc cutoff {}", lsn, *latest_gc_cutoff_lsn);
+                    }
+
+                    entry.or_insert(LsnLease { valid_until }).clone()
+                }
+            };
+
+            lease
         };
-        // TODO: dummy implementation
+
         Ok(lease)
     }
 
@@ -2082,6 +2117,24 @@ const REPARTITION_FREQ_IN_CHECKPOINT_DISTANCE: u64 = 10;
 
 // Private functions
 impl Timeline {
+    pub(crate) fn get_lsn_lease_length(&self) -> Duration {
+        let tenant_conf = self.tenant_conf.load();
+        tenant_conf
+            .tenant_conf
+            .lsn_lease_length
+            .unwrap_or(self.conf.default_tenant_conf.lsn_lease_length)
+    }
+
+    // TODO(yuchen): remove unused flag after implementing https://github.com/neondatabase/neon/issues/8072
+    #[allow(unused)]
+    pub(crate) fn get_lsn_lease_length_for_ts(&self) -> Duration {
+        let tenant_conf = self.tenant_conf.load();
+        tenant_conf
+            .tenant_conf
+            .lsn_lease_length_for_ts
+            .unwrap_or(self.conf.default_tenant_conf.lsn_lease_length_for_ts)
+    }
+
     pub(crate) fn get_switch_aux_file_policy(&self) -> AuxFilePolicy {
         let tenant_conf = self.tenant_conf.load();
         tenant_conf
@@ -4907,13 +4960,25 @@ impl Timeline {
             return Err(GcError::TimelineCancelled);
         }
 
-        let (horizon_cutoff, pitr_cutoff, retain_lsns) = {
+        let (horizon_cutoff, pitr_cutoff, retain_lsns, max_lsn_with_valid_lease) = {
             let gc_info = self.gc_info.read().unwrap();
 
             let horizon_cutoff = min(gc_info.cutoffs.horizon, self.get_disk_consistent_lsn());
             let pitr_cutoff = gc_info.cutoffs.pitr;
             let retain_lsns = gc_info.retain_lsns.clone();
-            (horizon_cutoff, pitr_cutoff, retain_lsns)
+
+            // Gets the maximum LSN that holds the valid lease.
+            //
+            // Caveat: `refresh_gc_info` is in charged of updating the lease map.
+            // Here, we do not check for stale leases again.
+            let max_lsn_with_valid_lease = gc_info.leases.last_key_value().map(|(lsn, _)| *lsn);
+
+            (
+                horizon_cutoff,
+                pitr_cutoff,
+                retain_lsns,
+                max_lsn_with_valid_lease,
+            )
         };
 
         let mut new_gc_cutoff = Lsn::min(horizon_cutoff, pitr_cutoff);
@@ -4944,7 +5009,13 @@ impl Timeline {
             .set(Lsn::INVALID.0 as i64);
 
         let res = self
-            .gc_timeline(horizon_cutoff, pitr_cutoff, retain_lsns, new_gc_cutoff)
+            .gc_timeline(
+                horizon_cutoff,
+                pitr_cutoff,
+                retain_lsns,
+                max_lsn_with_valid_lease,
+                new_gc_cutoff,
+            )
             .instrument(
                 info_span!("gc_timeline", timeline_id = %self.timeline_id, cutoff = %new_gc_cutoff),
             )
@@ -4961,6 +5032,7 @@ impl Timeline {
         horizon_cutoff: Lsn,
         pitr_cutoff: Lsn,
         retain_lsns: Vec<Lsn>,
+        max_lsn_with_valid_lease: Option<Lsn>,
         new_gc_cutoff: Lsn,
     ) -> Result<GcResult, GcError> {
         // FIXME: if there is an ongoing detach_from_ancestor, we should just skip gc
@@ -5009,7 +5081,8 @@ impl Timeline {
         // 1. it is older than cutoff LSN;
         // 2. it is older than PITR interval;
         // 3. it doesn't need to be retained for 'retain_lsns';
-        // 4. newer on-disk image layers cover the layer's whole key range
+        // 4. it does not need to be kept for LSNs holding valid leases.
+        // 5. newer on-disk image layers cover the layer's whole key range
         //
         // TODO holding a write lock is too agressive and avoidable
         let mut guard = self.layers.write().await;
@@ -5060,7 +5133,21 @@ impl Timeline {
                 }
             }
 
-            // 4. Is there a later on-disk layer for this relation?
+            // 4. Is there a valid lease that requires us to keep this layer?
+            if let Some(lsn) = &max_lsn_with_valid_lease {
+                // keep if layer start <= any of the lease
+                if &l.get_lsn_range().start <= lsn {
+                    debug!(
+                        "keeping {} because there is a valid lease preventing GC at {}",
+                        l.layer_name(),
+                        lsn,
+                    );
+                    result.layers_needed_by_leases += 1;
+                    continue 'outer;
+                }
+            }
+
+            // 5. Is there a later on-disk layer for this relation?
             //
             // The end-LSN is exclusive, while disk_consistent_lsn is
             // inclusive. For example, if disk_consistent_lsn is 100, it is
@@ -5438,6 +5525,11 @@ impl Timeline {
         self.last_record_lsn.advance(new_lsn);
     }
 
+    #[cfg(test)]
+    pub(super) fn force_set_disk_consistent_lsn(&self, new_value: Lsn) {
+        self.disk_consistent_lsn.store(new_value);
+    }
+
     /// Force create an image layer and place it into the layer map.
     ///
     /// DO NOT use this function directly. Use [`Tenant::branch_timeline_test_with_layers`]

pgxn/neon/walproposer_pg.c

@@ -100,17 +100,12 @@ static void StartProposerReplication(WalProposer *wp, StartReplicationCmd *cmd);
 static void WalSndLoop(WalProposer *wp);
 static void XLogBroadcastWalProposer(WalProposer *wp);
 
-static void XLogWalPropWrite(WalProposer *wp, char *buf, Size nbytes, XLogRecPtr recptr);
-static void XLogWalPropClose(XLogRecPtr recptr);
-
 static void add_nwr_event_set(Safekeeper *sk, uint32 events);
 static void update_nwr_event_set(Safekeeper *sk, uint32 events);
 static void rm_safekeeper_event_set(Safekeeper *to_remove, bool is_sk);
 
 static void CheckGracefulShutdown(WalProposer *wp);
 
-static XLogRecPtr GetLogRepRestartLSN(WalProposer *wp);
-
 static void
 init_walprop_config(bool syncSafekeepers)
 {
@@ -1236,8 +1231,6 @@ StartProposerReplication(WalProposer *wp, StartReplicationCmd *cmd)
 static void
 WalSndLoop(WalProposer *wp)
 {
-	XLogRecPtr	flushPtr;
-
 	/* Clear any already-pending wakeups */
 	ResetLatch(MyLatch);
 
@@ -1333,8 +1326,9 @@ XLogBroadcastWalProposer(WalProposer *wp)
 }
 
 /*
-  Used to download WAL before basebackup for logical walsenders from sk, no longer
-  needed because walsender always uses neon_walreader.
+  Used to download WAL before basebackup for walproposer/logical walsenders. No
+  longer used, replaced by neon_walreader; but callback still exists because
+  simulation tests use it.
  */
 static bool
 WalProposerRecovery(WalProposer *wp, Safekeeper *sk)
@@ -1342,136 +1336,6 @@ WalProposerRecovery(WalProposer *wp, Safekeeper *sk)
 	return true;
 }
 
-/*
- * These variables are used similarly to openLogFile/SegNo,
- * but for walproposer to write the XLOG during recovery. walpropFileTLI is the TimeLineID
- * corresponding the filename of walpropFile.
- */
-static int	walpropFile = -1;
-static TimeLineID walpropFileTLI = 0;
-static XLogSegNo walpropSegNo = 0;
-
-/*
- * Write XLOG data to disk.
- */
-static void
-XLogWalPropWrite(WalProposer *wp, char *buf, Size nbytes, XLogRecPtr recptr)
-{
-	int			startoff;
-	int			byteswritten;
-
-	/*
-	 * Apart from walproposer, basebackup LSN page is also written out by
-	 * postgres itself which writes WAL only in pages, and in basebackup it is
-	 * inherently dummy (only safekeepers have historic WAL). Update WAL
-	 * buffers here to avoid dummy page overwriting correct one we download
-	 * here. Ugly, but alternatives are about the same ugly. We won't need
-	 * that if we switch to on-demand WAL download from safekeepers, without
-	 * writing to disk.
-	 *
-	 * https://github.com/neondatabase/neon/issues/5749
-	 */
-	if (!wp->config->syncSafekeepers)
-		XLogUpdateWalBuffers(buf, recptr, nbytes);
-
-	while (nbytes > 0)
-	{
-		int			segbytes;
-
-		/* Close the current segment if it's completed */
-		if (walpropFile >= 0 && !XLByteInSeg(recptr, walpropSegNo, wal_segment_size))
-			XLogWalPropClose(recptr);
-
-		if (walpropFile < 0)
-		{
-#if PG_VERSION_NUM >= 150000
-			/* FIXME Is it ok to use hardcoded value here? */
-			TimeLineID	tli = 1;
-#else
-			bool		use_existent = true;
-#endif
-			/* Create/use new log file */
-			XLByteToSeg(recptr, walpropSegNo, wal_segment_size);
-#if PG_VERSION_NUM >= 150000
-			walpropFile = XLogFileInit(walpropSegNo, tli);
-			walpropFileTLI = tli;
-#else
-			walpropFile = XLogFileInit(walpropSegNo, &use_existent, false);
-			walpropFileTLI = ThisTimeLineID;
-#endif
-		}
-
-		/* Calculate the start offset of the received logs */
-		startoff = XLogSegmentOffset(recptr, wal_segment_size);
-
-		if (startoff + nbytes > wal_segment_size)
-			segbytes = wal_segment_size - startoff;
-		else
-			segbytes = nbytes;
-
-		/* OK to write the logs */
-		errno = 0;
-
-		byteswritten = pg_pwrite(walpropFile, buf, segbytes, (off_t) startoff);
-		if (byteswritten <= 0)
-		{
-			char		xlogfname[MAXFNAMELEN];
-			int			save_errno;
-
-			/* if write didn't set errno, assume no disk space */
-			if (errno == 0)
-				errno = ENOSPC;
-
-			save_errno = errno;
-			XLogFileName(xlogfname, walpropFileTLI, walpropSegNo, wal_segment_size);
-			errno = save_errno;
-			ereport(PANIC,
-					(errcode_for_file_access(),
-					 errmsg("could not write to log segment %s "
-							"at offset %u, length %lu: %m",
-							xlogfname, startoff, (unsigned long) segbytes)));
-		}
-
-		/* Update state for write */
-		recptr += byteswritten;
-
-		nbytes -= byteswritten;
-		buf += byteswritten;
-	}
-
-	/*
-	 * Close the current segment if it's fully written up in the last cycle of
-	 * the loop.
-	 */
-	if (walpropFile >= 0 && !XLByteInSeg(recptr, walpropSegNo, wal_segment_size))
-	{
-		XLogWalPropClose(recptr);
-	}
-}
-
-/*
- * Close the current segment.
- */
-static void
-XLogWalPropClose(XLogRecPtr recptr)
-{
-	Assert(walpropFile >= 0 && !XLByteInSeg(recptr, walpropSegNo, wal_segment_size));
-
-	if (close(walpropFile) != 0)
-	{
-		char		xlogfname[MAXFNAMELEN];
-
-		XLogFileName(xlogfname, walpropFileTLI, walpropSegNo, wal_segment_size);
-
-		ereport(PANIC,
-				(errcode_for_file_access(),
-				 errmsg("could not close log segment %s: %m",
-						xlogfname)));
-	}
-
-	walpropFile = -1;
-}
-
 static void
 walprop_pg_wal_reader_allocate(Safekeeper *sk)
 {
@@ -1987,58 +1851,6 @@ walprop_pg_log_internal(WalProposer *wp, int level, const char *line)
 	elog(FATAL, "unexpected log_internal message at level %d: %s", level, line);
 }
 
-static XLogRecPtr
-GetLogRepRestartLSN(WalProposer *wp)
-{
-	FILE	   *f;
-	XLogRecPtr	lrRestartLsn = InvalidXLogRecPtr;
-
-	/* We don't need to do anything in syncSafekeepers mode. */
-	if (wp->config->syncSafekeepers)
-		return InvalidXLogRecPtr;
-
-	/*
-	 * If there are active logical replication subscription we need to provide
-	 * enough WAL for their WAL senders based on th position of their
-	 * replication slots.
-	 */
-	f = fopen("restart.lsn", "rb");
-	if (f != NULL)
-	{
-		size_t		rc = fread(&lrRestartLsn, sizeof(lrRestartLsn), 1, f);
-
-		fclose(f);
-		if (rc == 1 && lrRestartLsn != InvalidXLogRecPtr)
-		{
-			uint64		download_range_mb;
-
-			wpg_log(LOG, "logical replication restart LSN %X/%X", LSN_FORMAT_ARGS(lrRestartLsn));
-
-			/*
-			 * If we need to download more than a max_slot_wal_keep_size,
-			 * don't do it to avoid risk of exploding pg_wal. Logical
-			 * replication won't work until recreated, but at least compute
-			 * would start; this also follows max_slot_wal_keep_size
-			 * semantics.
-			 */
-			download_range_mb = (wp->propEpochStartLsn - lrRestartLsn) / MB;
-			if (max_slot_wal_keep_size_mb > 0 && download_range_mb >= max_slot_wal_keep_size_mb)
-			{
-				wpg_log(WARNING, "not downloading WAL for logical replication since %X/%X as max_slot_wal_keep_size=%dMB",
-						LSN_FORMAT_ARGS(lrRestartLsn), max_slot_wal_keep_size_mb);
-				return InvalidXLogRecPtr;
-			}
-
-			/*
-			 * start from the beginning of the segment to fetch page headers
-			 * verifed by XLogReader
-			 */
-			lrRestartLsn = lrRestartLsn - XLogSegmentOffset(lrRestartLsn, wal_segment_size);
-		}
-	}
-	return lrRestartLsn;
-}
-
 void
 SetNeonCurrentClusterSize(uint64 size)
 {

poetry.lock

@@ -2806,13 +2806,13 @@ files = [
 
 [[package]]
 name = "urllib3"
-version = "1.26.18"
+version = "1.26.19"
 description = "HTTP library with thread-safe connection pooling, file post, and more."
 optional = false
-python-versions = ">=2.7, !=3.0.*, !=3.1.*, !=3.2.*, !=3.3.*, !=3.4.*, !=3.5.*"
+python-versions = "!=3.0.*,!=3.1.*,!=3.2.*,!=3.3.*,!=3.4.*,!=3.5.*,>=2.7"
 files = [
-    {file = "urllib3-1.26.18-py2.py3-none-any.whl", hash = "sha256:34b97092d7e0a3a8cf7cd10e386f401b3737364026c45e622aa02903dffe0f07"},
-    {file = "urllib3-1.26.18.tar.gz", hash = "sha256:f8ecc1bba5667413457c529ab955bf8c67b45db799d159066261719e328580a0"},
+    {file = "urllib3-1.26.19-py2.py3-none-any.whl", hash = "sha256:37a0344459b199fce0e80b0d3569837ec6b6937435c5244e7fd73fa6006830f3"},
+    {file = "urllib3-1.26.19.tar.gz", hash = "sha256:3e3d753a8618b86d7de333b4223005f68720bcd6a7d2bcb9fbd2229ec7c1e429"},
 ]
 
 [package.extras]

proxy/src/cache/project_info.rs

@@ -305,7 +305,7 @@ impl ProjectInfoCacheImpl {
         // acquire a random shard lock
         let mut removed = 0;
         let shard = self.project2ep.shards()[shard].write();
-        for (_, endpoints) in shard.iter() {
+        for (_, endpoints) in crate::rawtable::iter(&*shard) {
             for endpoint in endpoints.get().iter() {
                 self.cache.remove(endpoint);
                 removed += 1;

proxy/src/console/provider.rs

@@ -517,11 +517,18 @@ impl<K: Hash + Eq + Clone> ApiLocks<K> {
                 );
                 let mut lock = shard.write();
                 let timer = self.metrics.reclamation_lag_seconds.start_timer();
-                let count = lock
-                    .extract_if(|_, semaphore| Arc::strong_count(semaphore.get_mut()) == 1)
-                    .count();
+
+                let mut removed = 0;
+                crate::rawtable::retain(&mut *lock, |_, semaphore| {
+                    let remove = Arc::strong_count(semaphore.get_mut()) == 1;
+                    if remove {
+                        removed += 1;
+                    }
+                    !remove
+                });
+
                 drop(lock);
-                self.metrics.semaphores_unregistered.inc_by(count as u64);
+                self.metrics.semaphores_unregistered.inc_by(removed as u64);
                 timer.observe();
             }
         }

proxy/src/lib.rs

@@ -25,6 +25,7 @@ pub mod parse;
 pub mod protocol2;
 pub mod proxy;
 pub mod rate_limiter;
+mod rawtable;
 pub mod redis;
 pub mod sasl;
 pub mod scram;

proxy/src/rawtable.rs

@@ -0,0 +1,61 @@
+//! Dashmap moved to using RawTable for the shards.
+//! Some of the APIs we used before are unsafe to access, but we can copy the implementations from the safe
+//! HashMap wrappers for our needs.
+
+// Safety info: All implementations here are taken directly from hashbrown HashMap impl.
+
+use std::marker::PhantomData;
+
+use hashbrown::raw;
+
+// taken from https://docs.rs/hashbrown/0.14.5/src/hashbrown/map.rs.html#919-932
+pub fn retain<K, V, F>(table: &mut raw::RawTable<(K, V)>, mut f: F)
+where
+    F: FnMut(&K, &mut V) -> bool,
+{
+    // SAFETY: Here we only use `iter` as a temporary, preventing use-after-free
+    unsafe {
+        for item in table.iter() {
+            let &mut (ref key, ref mut value) = item.as_mut();
+            if !f(key, value) {
+                table.erase(item);
+            }
+        }
+    }
+}
+
+// taken from https://docs.rs/hashbrown/0.14.5/src/hashbrown/map.rs.html#756-764
+pub fn iter<K, V>(table: &raw::RawTable<(K, V)>) -> impl Iterator<Item = (&K, &V)> + '_ {
+    pub struct Iter<'a, K, V> {
+        inner: raw::RawIter<(K, V)>,
+        marker: PhantomData<(&'a K, &'a V)>,
+    }
+
+    impl<'a, K, V> Iterator for Iter<'a, K, V> {
+        type Item = (&'a K, &'a V);
+
+        #[cfg_attr(feature = "inline-more", inline)]
+        fn next(&mut self) -> Option<(&'a K, &'a V)> {
+            let x = self.inner.next()?;
+            // SAFETY: the borrows do not outlive the rawtable
+            unsafe {
+                let r = x.as_ref();
+                Some((&r.0, &r.1))
+            }
+        }
+        #[cfg_attr(feature = "inline-more", inline)]
+        fn size_hint(&self) -> (usize, Option<usize>) {
+            self.inner.size_hint()
+        }
+    }
+
+    // SAFETY:
+    // > It is up to the caller to ensure that the RawTable outlives the RawIter
+    // Here we tie the lifetime of self to the iter.
+    unsafe {
+        Iter {
+            inner: table.iter(),
+            marker: PhantomData,
+        }
+    }
+}

proxy/src/serverless/conn_pool.rs

@@ -324,7 +324,8 @@ impl<C: ClientInnerExt> GlobalConnPool<C> {
             .start_timer();
         let current_len = shard.len();
         let mut clients_removed = 0;
-        shard.retain(|endpoint, x| {
+
+        crate::rawtable::retain(&mut *shard, |endpoint, x| {
             // if the current endpoint pool is unique (no other strong or weak references)
             // then it is currently not in use by any connections.
             if let Some(pool) = Arc::get_mut(x.get_mut()) {

safekeeper/Cargo.toml

@@ -46,6 +46,7 @@ tokio = { workspace = true, features = ["fs"] }
 tokio-util = { workspace = true }
 tokio-io-timeout.workspace = true
 tokio-postgres.workspace = true
+tokio-tar.workspace = true
 toml_edit.workspace = true
 tracing.workspace = true
 url.workspace = true

safekeeper/src/bin/safekeeper.rs

@@ -13,7 +13,9 @@ use tokio::runtime::Handle;
 use tokio::signal::unix::{signal, SignalKind};
 use tokio::task::JoinError;
 use toml_edit::Document;
+use utils::logging::SecretString;
 
+use std::env::{var, VarError};
 use std::fs::{self, File};
 use std::io::{ErrorKind, Write};
 use std::str::FromStr;
@@ -287,6 +289,22 @@ async fn main() -> anyhow::Result<()> {
         }
     };
 
+    // Load JWT auth token to connect to other safekeepers for pull_timeline.
+    let sk_auth_token = match var("SAFEKEEPER_AUTH_TOKEN") {
+        Ok(v) => {
+            info!("loaded JWT token for authentication with safekeepers");
+            Some(SecretString::from(v))
+        }
+        Err(VarError::NotPresent) => {
+            info!("no JWT token for authentication with safekeepers detected");
+            None
+        }
+        Err(_) => {
+            warn!("JWT token for authentication with safekeepers is not unicode");
+            None
+        }
+    };
+
     let conf = SafeKeeperConf {
         workdir,
         my_id: id,
@@ -307,6 +325,7 @@ async fn main() -> anyhow::Result<()> {
         pg_auth,
         pg_tenant_only_auth,
         http_auth,
+        sk_auth_token,
         current_thread_runtime: args.current_thread_runtime,
         walsenders_keep_horizon: args.walsenders_keep_horizon,
         partial_backup_enabled: args.partial_backup_enabled,

safekeeper/src/control_file.rs

@@ -23,7 +23,7 @@ pub const SK_MAGIC: u32 = 0xcafeceefu32;
 pub const SK_FORMAT_VERSION: u32 = 8;
 
 // contains persistent metadata for safekeeper
-const CONTROL_FILE_NAME: &str = "safekeeper.control";
+pub const CONTROL_FILE_NAME: &str = "safekeeper.control";
 // needed to atomically update the state using `rename`
 const CONTROL_FILE_NAME_PARTIAL: &str = "safekeeper.control.partial";
 pub const CHECKSUM_SIZE: usize = std::mem::size_of::<u32>();

safekeeper/src/http/client.rs

@@ -0,0 +1,139 @@
+//! Safekeeper http client.
+//!
+//! Partially copied from pageserver client; some parts might be better to be
+//! united.
+//!
+//! It would be also good to move it out to separate crate, but this needs
+//! duplication of internal-but-reported structs like WalSenderState, ServerInfo
+//! etc.
+
+use reqwest::{IntoUrl, Method, StatusCode};
+use utils::{
+    http::error::HttpErrorBody,
+    id::{TenantId, TimelineId},
+    logging::SecretString,
+};
+
+use super::routes::TimelineStatus;
+
+#[derive(Debug, Clone)]
+pub struct Client {
+    mgmt_api_endpoint: String,
+    authorization_header: Option<SecretString>,
+    client: reqwest::Client,
+}
+
+#[derive(thiserror::Error, Debug)]
+pub enum Error {
+    /// Failed to receive body (reqwest error).
+    #[error("receive body: {0}")]
+    ReceiveBody(reqwest::Error),
+
+    /// Status is not ok, but failed to parse body as `HttpErrorBody`.
+    #[error("receive error body: {0}")]
+    ReceiveErrorBody(String),
+
+    /// Status is not ok; parsed error in body as `HttpErrorBody`.
+    #[error("safekeeper API: {1}")]
+    ApiError(StatusCode, String),
+}
+
+pub type Result<T> = std::result::Result<T, Error>;
+
+pub trait ResponseErrorMessageExt: Sized {
+    fn error_from_body(self) -> impl std::future::Future<Output = Result<Self>> + Send;
+}
+
+/// If status is not ok, try to extract error message from the body.
+impl ResponseErrorMessageExt for reqwest::Response {
+    async fn error_from_body(self) -> Result<Self> {
+        let status = self.status();
+        if !(status.is_client_error() || status.is_server_error()) {
+            return Ok(self);
+        }
+
+        let url = self.url().to_owned();
+        Err(match self.json::<HttpErrorBody>().await {
+            Ok(HttpErrorBody { msg }) => Error::ApiError(status, msg),
+            Err(_) => {
+                Error::ReceiveErrorBody(format!("http error ({}) at {}.", status.as_u16(), url))
+            }
+        })
+    }
+}
+
+impl Client {
+    pub fn new(mgmt_api_endpoint: String, jwt: Option<SecretString>) -> Self {
+        Self::from_client(reqwest::Client::new(), mgmt_api_endpoint, jwt)
+    }
+
+    pub fn from_client(
+        client: reqwest::Client,
+        mgmt_api_endpoint: String,
+        jwt: Option<SecretString>,
+    ) -> Self {
+        Self {
+            mgmt_api_endpoint,
+            authorization_header: jwt
+                .map(|jwt| SecretString::from(format!("Bearer {}", jwt.get_contents()))),
+            client,
+        }
+    }
+
+    pub async fn timeline_status(
+        &self,
+        tenant_id: TenantId,
+        timeline_id: TimelineId,
+    ) -> Result<TimelineStatus> {
+        let uri = format!(
+            "{}/v1/tenant/{}/timeline/{}",
+            self.mgmt_api_endpoint, tenant_id, timeline_id
+        );
+        let resp = self.get(&uri).await?;
+        resp.json().await.map_err(Error::ReceiveBody)
+    }
+
+    pub async fn snapshot(
+        &self,
+        tenant_id: TenantId,
+        timeline_id: TimelineId,
+    ) -> Result<reqwest::Response> {
+        let uri = format!(
+            "{}/v1/tenant/{}/timeline/{}/snapshot",
+            self.mgmt_api_endpoint, tenant_id, timeline_id
+        );
+        self.get(&uri).await
+    }
+
+    async fn get<U: IntoUrl>(&self, uri: U) -> Result<reqwest::Response> {
+        self.request(Method::GET, uri, ()).await
+    }
+
+    /// Send the request and check that the status code is good.
+    async fn request<B: serde::Serialize, U: reqwest::IntoUrl>(
+        &self,
+        method: Method,
+        uri: U,
+        body: B,
+    ) -> Result<reqwest::Response> {
+        let res = self.request_noerror(method, uri, body).await?;
+        let response = res.error_from_body().await?;
+        Ok(response)
+    }
+
+    /// Just send the request.
+    async fn request_noerror<B: serde::Serialize, U: reqwest::IntoUrl>(
+        &self,
+        method: Method,
+        uri: U,
+        body: B,
+    ) -> Result<reqwest::Response> {
+        let req = self.client.request(method, uri);
+        let req = if let Some(value) = &self.authorization_header {
+            req.header(reqwest::header::AUTHORIZATION, value.get_contents())
+        } else {
+            req
+        };
+        req.json(&body).send().await.map_err(Error::ReceiveBody)
+    }
+}

safekeeper/src/http/mod.rs

@@ -1,3 +1,4 @@
+pub mod client;
 pub mod routes;
 pub use routes::make_router;
 

safekeeper/src/http/routes.rs

@@ -1,38 +1,25 @@
 use hyper::{Body, Request, Response, StatusCode, Uri};
-
 use once_cell::sync::Lazy;
-use postgres_ffi::WAL_SEGMENT_SIZE;
-use safekeeper_api::models::{SkTimelineInfo, TimelineCopyRequest};
 use serde::{Deserialize, Serialize};
 use std::collections::{HashMap, HashSet};
 use std::fmt;
+use std::io::Write as _;
 use std::str::FromStr;
 use std::sync::Arc;
 use storage_broker::proto::SafekeeperTimelineInfo;
 use storage_broker::proto::TenantTimelineId as ProtoTenantTimelineId;
-use tokio::fs::File;
-use tokio::io::AsyncReadExt;
+use tokio::sync::mpsc;
+use tokio::task;
+use tokio_stream::wrappers::ReceiverStream;
 use tokio_util::sync::CancellationToken;
+use tracing::{info_span, Instrument};
 use utils::failpoint_support::failpoints_handler;
+use utils::http::endpoint::{prometheus_metrics_handler, request_span, ChannelWriter};
 use utils::http::request::parse_query_param;
 
-use std::io::Write as _;
-use tokio::sync::mpsc;
-use tokio_stream::wrappers::ReceiverStream;
-use tracing::{info_span, Instrument};
-use utils::http::endpoint::{prometheus_metrics_handler, request_span, ChannelWriter};
-
-use crate::debug_dump::TimelineDigestRequest;
-use crate::receive_wal::WalReceiverState;
-use crate::safekeeper::Term;
-use crate::safekeeper::{ServerInfo, TermLsn};
-use crate::send_wal::WalSenderState;
-use crate::timeline::PeerInfo;
-use crate::{copy_timeline, debug_dump, patch_control_file, pull_timeline};
-
-use crate::timelines_global_map::TimelineDeleteForceResult;
-use crate::GlobalTimelines;
-use crate::SafeKeeperConf;
+use postgres_ffi::WAL_SEGMENT_SIZE;
+use safekeeper_api::models::TimelineCreateRequest;
+use safekeeper_api::models::{SkTimelineInfo, TimelineCopyRequest};
 use utils::{
     auth::SwappableJwtAuth,
     http::{
@@ -46,7 +33,16 @@ use utils::{
     lsn::Lsn,
 };
 
-use super::models::TimelineCreateRequest;
+use crate::debug_dump::TimelineDigestRequest;
+use crate::receive_wal::WalReceiverState;
+use crate::safekeeper::Term;
+use crate::safekeeper::{ServerInfo, TermLsn};
+use crate::send_wal::WalSenderState;
+use crate::timeline::PeerInfo;
+use crate::timelines_global_map::TimelineDeleteForceResult;
+use crate::GlobalTimelines;
+use crate::SafeKeeperConf;
+use crate::{copy_timeline, debug_dump, patch_control_file, pull_timeline};
 
 #[derive(Debug, Serialize)]
 struct SafekeeperStatus {
@@ -199,13 +195,50 @@ async fn timeline_pull_handler(mut request: Request<Body>) -> Result<Response<Bo
     check_permission(&request, None)?;
 
     let data: pull_timeline::Request = json_request(&mut request).await?;
+    let conf = get_conf(&request);
 
-    let resp = pull_timeline::handle_request(data)
+    let resp = pull_timeline::handle_request(data, conf.sk_auth_token.clone())
         .await
         .map_err(ApiError::InternalServerError)?;
     json_response(StatusCode::OK, resp)
 }
 
+/// Stream tar archive with all timeline data.
+async fn timeline_snapshot_handler(request: Request<Body>) -> Result<Response<Body>, ApiError> {
+    let ttid = TenantTimelineId::new(
+        parse_request_param(&request, "tenant_id")?,
+        parse_request_param(&request, "timeline_id")?,
+    );
+    check_permission(&request, Some(ttid.tenant_id))?;
+
+    let tli = GlobalTimelines::get(ttid).map_err(ApiError::from)?;
+    // Note: with evicted timelines it should work better then de-evict them and
+    // stream; probably start_snapshot would copy partial s3 file to dest path
+    // and stream control file, or return FullAccessTimeline if timeline is not
+    // evicted.
+    let tli = tli
+        .full_access_guard()
+        .await
+        .map_err(ApiError::InternalServerError)?;
+
+    // To stream the body use wrap_stream which wants Stream of Result<Bytes>,
+    // so create the chan and write to it in another task.
+    let (tx, rx) = mpsc::channel(1);
+
+    task::spawn(pull_timeline::stream_snapshot(tli, tx));
+
+    let rx_stream = ReceiverStream::new(rx);
+    let body = Body::wrap_stream(rx_stream);
+
+    let response = Response::builder()
+        .status(200)
+        .header(hyper::header::CONTENT_TYPE, "application/octet-stream")
+        .body(body)
+        .unwrap();
+
+    Ok(response)
+}
+
 async fn timeline_copy_handler(mut request: Request<Body>) -> Result<Response<Body>, ApiError> {
     check_permission(&request, None)?;
 
@@ -260,41 +293,6 @@ async fn timeline_digest_handler(request: Request<Body>) -> Result<Response<Body
     json_response(StatusCode::OK, response)
 }
 
-/// Download a file from the timeline directory.
-// TODO: figure out a better way to copy files between safekeepers
-async fn timeline_files_handler(request: Request<Body>) -> Result<Response<Body>, ApiError> {
-    let ttid = TenantTimelineId::new(
-        parse_request_param(&request, "tenant_id")?,
-        parse_request_param(&request, "timeline_id")?,
-    );
-    check_permission(&request, Some(ttid.tenant_id))?;
-
-    let filename: String = parse_request_param(&request, "filename")?;
-
-    let tli = GlobalTimelines::get(ttid).map_err(ApiError::from)?;
-    let tli = tli
-        .full_access_guard()
-        .await
-        .map_err(ApiError::InternalServerError)?;
-
-    let filepath = tli.get_timeline_dir().join(filename);
-    let mut file = File::open(&filepath)
-        .await
-        .map_err(|e| ApiError::InternalServerError(e.into()))?;
-
-    let mut content = Vec::new();
-    // TODO: don't store files in memory
-    file.read_to_end(&mut content)
-        .await
-        .map_err(|e| ApiError::InternalServerError(e.into()))?;
-
-    Response::builder()
-        .status(StatusCode::OK)
-        .header("Content-Type", "application/octet-stream")
-        .body(Body::from(content))
-        .map_err(|e| ApiError::InternalServerError(e.into()))
-}
-
 /// Force persist control file.
 async fn timeline_checkpoint_handler(request: Request<Body>) -> Result<Response<Body>, ApiError> {
     check_permission(&request, None)?;
@@ -566,13 +564,13 @@ pub fn make_router(conf: SafeKeeperConf) -> RouterBuilder<hyper::Body, ApiError>
         .delete("/v1/tenant/:tenant_id", |r| {
             request_span(r, tenant_delete_handler)
         })
+        .get(
+            "/v1/tenant/:tenant_id/timeline/:timeline_id/snapshot",
+            |r| request_span(r, timeline_snapshot_handler),
+        )
         .post("/v1/pull_timeline", |r| {
             request_span(r, timeline_pull_handler)
         })
-        .get(
-            "/v1/tenant/:tenant_id/timeline/:timeline_id/file/:filename",
-            |r| request_span(r, timeline_files_handler),
-        )
         .post(
             "/v1/tenant/:tenant_id/timeline/:source_timeline_id/copy",
             |r| request_span(r, timeline_copy_handler),

safekeeper/src/lib.rs

@@ -7,7 +7,7 @@ use tokio::runtime::Runtime;
 use std::time::Duration;
 use storage_broker::Uri;
 
-use utils::{auth::SwappableJwtAuth, id::NodeId};
+use utils::{auth::SwappableJwtAuth, id::NodeId, logging::SecretString};
 
 mod auth;
 pub mod broker;
@@ -78,6 +78,8 @@ pub struct SafeKeeperConf {
     pub pg_auth: Option<Arc<JwtAuth>>,
     pub pg_tenant_only_auth: Option<Arc<JwtAuth>>,
     pub http_auth: Option<Arc<SwappableJwtAuth>>,
+    /// JWT token to connect to other safekeepers with.
+    pub sk_auth_token: Option<SecretString>,
     pub current_thread_runtime: bool,
     pub walsenders_keep_horizon: bool,
     pub partial_backup_enabled: bool,
@@ -114,6 +116,7 @@ impl SafeKeeperConf {
             pg_auth: None,
             pg_tenant_only_auth: None,
             http_auth: None,
+            sk_auth_token: None,
             heartbeat_timeout: Duration::new(5, 0),
             max_offloader_lag_bytes: defaults::DEFAULT_MAX_OFFLOADER_LAG_BYTES,
             current_thread_runtime: false,

safekeeper/src/pull_timeline.rs

@@ -1,28 +1,244 @@
-use std::sync::Arc;
-
+use anyhow::{anyhow, bail, Context, Result};
+use bytes::Bytes;
 use camino::Utf8PathBuf;
 use camino_tempfile::Utf8TempDir;
 use chrono::{DateTime, Utc};
+use futures::{SinkExt, StreamExt, TryStreamExt};
+use postgres_ffi::{XLogFileName, XLogSegNo, PG_TLI};
 use serde::{Deserialize, Serialize};
+use std::{
+    cmp::min,
+    io::{self, ErrorKind},
+    sync::Arc,
+};
+use tokio::{
+    fs::{File, OpenOptions},
+    io::AsyncWrite,
+    sync::mpsc,
+    task,
+};
+use tokio_tar::{Archive, Builder};
+use tokio_util::{
+    io::{CopyToBytes, SinkWriter},
+    sync::PollSender,
+};
+use tracing::{error, info, instrument};
 
-use anyhow::{bail, Context, Result};
-use tokio::io::AsyncWriteExt;
-use tracing::info;
+use crate::{
+    control_file::{self, CONTROL_FILE_NAME},
+    debug_dump,
+    http::{
+        client::{self, Client},
+        routes::TimelineStatus,
+    },
+    safekeeper::Term,
+    timeline::{get_tenant_dir, get_timeline_dir, FullAccessTimeline, Timeline, TimelineError},
+    wal_storage::{self, open_wal_file, Storage},
+    GlobalTimelines, SafeKeeperConf,
+};
 use utils::{
+    crashsafe::{durable_rename, fsync_async_opt},
     id::{TenantId, TenantTimelineId, TimelineId},
+    logging::SecretString,
     lsn::Lsn,
     pausable_failpoint,
 };
 
-use crate::{
-    control_file, debug_dump,
-    http::routes::TimelineStatus,
-    timeline::{get_tenant_dir, get_timeline_dir, Timeline, TimelineError},
-    wal_storage::{self, Storage},
-    GlobalTimelines, SafeKeeperConf,
-};
+/// Stream tar archive of timeline to tx.
+#[instrument(name = "snapshot", skip_all, fields(ttid = %tli.ttid))]
+pub async fn stream_snapshot(tli: FullAccessTimeline, tx: mpsc::Sender<Result<Bytes>>) {
+    if let Err(e) = stream_snapshot_guts(tli, tx.clone()).await {
+        // Error type/contents don't matter as they won't can't reach the client
+        // (hyper likely doesn't do anything with it), but http stream will be
+        // prematurely terminated. It would be nice to try to send the error in
+        // trailers though.
+        tx.send(Err(anyhow!("snapshot failed"))).await.ok();
+        error!("snapshot failed: {:#}", e);
+    }
+}
 
-/// Info about timeline on safekeeper ready for reporting.
+/// State needed while streaming the snapshot.
+pub struct SnapshotContext {
+    pub from_segno: XLogSegNo, // including
+    pub upto_segno: XLogSegNo, // including
+    pub term: Term,
+    pub last_log_term: Term,
+    pub flush_lsn: Lsn,
+    pub wal_seg_size: usize,
+    // used to remove WAL hold off in Drop.
+    pub tli: FullAccessTimeline,
+}
+
+impl Drop for SnapshotContext {
+    fn drop(&mut self) {
+        let tli = self.tli.clone();
+        task::spawn(async move {
+            let mut shared_state = tli.write_shared_state().await;
+            shared_state.wal_removal_on_hold = false;
+        });
+    }
+}
+
+pub async fn stream_snapshot_guts(
+    tli: FullAccessTimeline,
+    tx: mpsc::Sender<Result<Bytes>>,
+) -> Result<()> {
+    // tokio-tar wants Write implementor, but we have mpsc tx <Result<Bytes>>;
+    // use SinkWriter as a Write impl. That is,
+    // - create Sink from the tx. It returns PollSendError if chan is closed.
+    let sink = PollSender::new(tx);
+    // - SinkWriter needs sink error to be io one, map it.
+    let sink_io_err = sink.sink_map_err(|_| io::Error::from(ErrorKind::BrokenPipe));
+    // - SinkWriter wants sink type to be just Bytes, not Result<Bytes>, so map
+    //   it with with(). Note that with() accepts async function which we don't
+    //   need and allows the map to fail, which we don't need either, but hence
+    //   two Oks.
+    let oksink = sink_io_err.with(|b: Bytes| async { io::Result::Ok(Result::Ok(b)) });
+    // - SinkWriter (not surprisingly) wants sink of &[u8], not bytes, so wrap
+    // into CopyToBytes. This is a data copy.
+    let copy_to_bytes = CopyToBytes::new(oksink);
+    let mut writer = SinkWriter::new(copy_to_bytes);
+    let pinned_writer = std::pin::pin!(writer);
+
+    // Note that tokio_tar append_* funcs use tokio::io::copy with 8KB buffer
+    // which is also likely suboptimal.
+    let mut ar = Builder::new_non_terminated(pinned_writer);
+
+    let bctx = tli.start_snapshot(&mut ar).await?;
+    pausable_failpoint!("sk-snapshot-after-list-pausable");
+
+    let tli_dir = tli.get_timeline_dir();
+    info!(
+        "sending {} segments [{:#X}-{:#X}], term={}, last_log_term={}, flush_lsn={}",
+        bctx.upto_segno - bctx.from_segno + 1,
+        bctx.from_segno,
+        bctx.upto_segno,
+        bctx.term,
+        bctx.last_log_term,
+        bctx.flush_lsn,
+    );
+    for segno in bctx.from_segno..=bctx.upto_segno {
+        let (mut sf, is_partial) = open_wal_file(&tli_dir, segno, bctx.wal_seg_size).await?;
+        let mut wal_file_name = XLogFileName(PG_TLI, segno, bctx.wal_seg_size);
+        if is_partial {
+            wal_file_name.push_str(".partial");
+        }
+        ar.append_file(&wal_file_name, &mut sf).await?;
+    }
+
+    // Do the term check before ar.finish to make archive corrupted in case of
+    // term change. Client shouldn't ignore abrupt stream end, but to be sure.
+    tli.finish_snapshot(&bctx).await?;
+
+    ar.finish().await?;
+
+    Ok(())
+}
+
+impl FullAccessTimeline {
+    /// Start streaming tar archive with timeline:
+    /// 1) stream control file under lock;
+    /// 2) hold off WAL removal;
+    /// 3) collect SnapshotContext to understand which WAL segments should be
+    ///    streamed.
+    ///
+    /// Snapshot streams data up to flush_lsn. To make this safe, we must check
+    /// that term doesn't change during the procedure, or we risk sending mix of
+    /// WAL from different histories. Term is remembered in the SnapshotContext
+    /// and checked in finish_snapshot. Note that in the last segment some WAL
+    /// higher than flush_lsn set here might be streamed; that's fine as long as
+    /// terms doesn't change.
+    ///
+    /// Alternatively we could send only up to commit_lsn to get some valid
+    /// state which later will be recovered by compute, in this case term check
+    /// is not needed, but we likely don't want that as there might be no
+    /// compute which could perform the recovery.
+    ///
+    /// When returned SnapshotContext is dropped WAL hold is removed.
+    async fn start_snapshot<W: AsyncWrite + Unpin + Send>(
+        &self,
+        ar: &mut tokio_tar::Builder<W>,
+    ) -> Result<SnapshotContext> {
+        let mut shared_state = self.write_shared_state().await;
+
+        let cf_path = self.get_timeline_dir().join(CONTROL_FILE_NAME);
+        let mut cf = File::open(cf_path).await?;
+        ar.append_file(CONTROL_FILE_NAME, &mut cf).await?;
+
+        // We need to stream since the oldest segment someone (s3 or pageserver)
+        // still needs. This duplicates calc_horizon_lsn logic.
+        //
+        // We know that WAL wasn't removed up to this point because it cannot be
+        // removed further than `backup_lsn`. Since we're holding shared_state
+        // lock and setting `wal_removal_on_hold` later, it guarantees that WAL
+        // won't be removed until we're done.
+        let from_lsn = min(
+            shared_state.sk.state.remote_consistent_lsn,
+            shared_state.sk.state.backup_lsn,
+        );
+        if from_lsn == Lsn::INVALID {
+            // this is possible if snapshot is called before handling first
+            // elected message
+            bail!("snapshot is called on uninitialized timeline");
+        }
+        let from_segno = from_lsn.segment_number(shared_state.get_wal_seg_size());
+        let term = shared_state.sk.get_term();
+        let last_log_term = shared_state.sk.get_last_log_term();
+        let flush_lsn = shared_state.sk.flush_lsn();
+        let upto_segno = flush_lsn.segment_number(shared_state.get_wal_seg_size());
+        // have some limit on max number of segments as a sanity check
+        const MAX_ALLOWED_SEGS: u64 = 1000;
+        let num_segs = upto_segno - from_segno + 1;
+        if num_segs > MAX_ALLOWED_SEGS {
+            bail!(
+                "snapshot is called on timeline with {} segments, but the limit is {}",
+                num_segs,
+                MAX_ALLOWED_SEGS
+            );
+        }
+
+        // Prevent WAL removal while we're streaming data.
+        //
+        // Since this a flag, not a counter just bail out if already set; we
+        // shouldn't need concurrent snapshotting.
+        if shared_state.wal_removal_on_hold {
+            bail!("wal_removal_on_hold is already true");
+        }
+        shared_state.wal_removal_on_hold = true;
+
+        let bctx = SnapshotContext {
+            from_segno,
+            upto_segno,
+            term,
+            last_log_term,
+            flush_lsn,
+            wal_seg_size: shared_state.get_wal_seg_size(),
+            tli: self.clone(),
+        };
+
+        Ok(bctx)
+    }
+
+    /// Finish snapshotting: check that term(s) hasn't changed.
+    ///
+    /// Note that WAL gc hold off is removed in Drop of SnapshotContext to not
+    /// forget this if snapshotting fails mid the way.
+    pub async fn finish_snapshot(&self, bctx: &SnapshotContext) -> Result<()> {
+        let shared_state = self.read_shared_state().await;
+        let term = shared_state.sk.get_term();
+        let last_log_term = shared_state.sk.get_last_log_term();
+        // There are some cases to relax this check (e.g. last_log_term might
+        // change, but as long as older history is strictly part of new that's
+        // fine), but there is no need to do it.
+        if bctx.term != term || bctx.last_log_term != last_log_term {
+            bail!("term(s) changed during snapshot: were term={}, last_log_term={}, now term={}, last_log_term={}",
+              bctx.term, bctx.last_log_term, term, last_log_term);
+        }
+        Ok(())
+    }
+}
+
+/// pull_timeline request body.
 #[derive(Debug, Serialize, Deserialize)]
 pub struct Request {
     pub tenant_id: TenantId,
@@ -48,7 +264,10 @@ pub struct DebugDumpResponse {
 }
 
 /// Find the most advanced safekeeper and pull timeline from it.
-pub async fn handle_request(request: Request) -> Result<Response> {
+pub async fn handle_request(
+    request: Request,
+    sk_auth_token: Option<SecretString>,
+) -> Result<Response> {
     let existing_tli = GlobalTimelines::get(TenantTimelineId::new(
         request.tenant_id,
         request.timeline_id,
@@ -57,28 +276,26 @@ pub async fn handle_request(request: Request) -> Result<Response> {
         bail!("Timeline {} already exists", request.timeline_id);
     }
 
-    let client = reqwest::Client::new();
     let http_hosts = request.http_hosts.clone();
 
-    // Send request to /v1/tenant/:tenant_id/timeline/:timeline_id
-    let responses = futures::future::join_all(http_hosts.iter().map(|url| {
-        let url = format!(
-            "{}/v1/tenant/{}/timeline/{}",
-            url, request.tenant_id, request.timeline_id
-        );
-        client.get(url).send()
-    }))
-    .await;
+    // Figure out statuses of potential donors.
+    let responses: Vec<Result<TimelineStatus, client::Error>> =
+        futures::future::join_all(http_hosts.iter().map(|url| async {
+            let cclient = Client::new(url.clone(), sk_auth_token.clone());
+            let info = cclient
+                .timeline_status(request.tenant_id, request.timeline_id)
+                .await?;
+            Ok(info)
+        }))
+        .await;
 
     let mut statuses = Vec::new();
     for (i, response) in responses.into_iter().enumerate() {
-        let response = response.context(format!("Failed to get status from {}", http_hosts[i]))?;
-        let status: crate::http::routes::TimelineStatus = response.json().await?;
+        let status = response.context(format!("fetching status from {}", http_hosts[i]))?;
         statuses.push((status, i));
     }
 
     // Find the most advanced safekeeper
-    // TODO: current logic may be wrong, fix it later
     let (status, i) = statuses
         .into_iter()
         .max_by_key(|(status, _)| {
@@ -94,10 +311,14 @@ pub async fn handle_request(request: Request) -> Result<Response> {
     assert!(status.tenant_id == request.tenant_id);
     assert!(status.timeline_id == request.timeline_id);
 
-    pull_timeline(status, safekeeper_host).await
+    pull_timeline(status, safekeeper_host, sk_auth_token).await
 }
 
-async fn pull_timeline(status: TimelineStatus, host: String) -> Result<Response> {
+async fn pull_timeline(
+    status: TimelineStatus,
+    host: String,
+    sk_auth_token: Option<SecretString>,
+) -> Result<Response> {
     let ttid = TenantTimelineId::new(status.tenant_id, status.timeline_id);
     info!(
         "pulling timeline {} from safekeeper {}, commit_lsn={}, flush_lsn={}, term={}, epoch={}",
@@ -111,95 +332,53 @@ async fn pull_timeline(status: TimelineStatus, host: String) -> Result<Response>
 
     let conf = &GlobalTimelines::get_global_config();
 
-    let client = reqwest::Client::new();
-    // TODO: don't use debug dump, it should be used only in tests.
-    //      This is a proof of concept, we should figure out a way
-    //      to use scp without implementing it manually.
-
-    // Implementing our own scp over HTTP.
-    // At first, we need to fetch list of files from safekeeper.
-    let dump: DebugDumpResponse = client
-        .get(format!(
-            "{}/v1/debug_dump?dump_all=true&tenant_id={}&timeline_id={}",
-            host, status.tenant_id, status.timeline_id
-        ))
-        .send()
-        .await?
-        .json()
-        .await?;
-
-    if dump.timelines.len() != 1 {
-        bail!(
-            "expected to fetch single timeline, got {} timelines",
-            dump.timelines.len()
-        );
-    }
-
-    let timeline = dump.timelines.into_iter().next().unwrap();
-    let disk_content = timeline.disk_content.ok_or(anyhow::anyhow!(
-        "timeline {} doesn't have disk content",
-        ttid
-    ))?;
-
-    let mut filenames = disk_content
-        .files
-        .iter()
-        .map(|file| file.name.clone())
-        .collect::<Vec<_>>();
-
-    // Sort filenames to make sure we pull files in correct order
-    // After sorting, we should have:
-    // - 000000010000000000000001
-    // - ...
-    // - 000000010000000000000002.partial
-    // - safekeeper.control
-    filenames.sort();
-
-    // safekeeper.control should be the first file, so we need to move it to the beginning
-    let control_file_index = filenames
-        .iter()
-        .position(|name| name == "safekeeper.control")
-        .ok_or(anyhow::anyhow!("safekeeper.control not found"))?;
-    filenames.remove(control_file_index);
-    filenames.insert(0, "safekeeper.control".to_string());
-
-    pausable_failpoint!("sk-pull-timeline-after-list-pausable");
-
-    info!(
-        "downloading {} files from safekeeper {}",
-        filenames.len(),
-        host
-    );
-
     let (_tmp_dir, tli_dir_path) = create_temp_timeline_dir(conf, ttid).await?;
 
-    // Note: some time happens between fetching list of files and fetching files themselves.
-    //       It's possible that some files will be removed from safekeeper and we will fail to fetch them.
-    //       This function will fail in this case, should be retried by the caller.
-    for filename in filenames {
-        let file_path = tli_dir_path.join(&filename);
-        // /v1/tenant/:tenant_id/timeline/:timeline_id/file/:filename
-        let http_url = format!(
-            "{}/v1/tenant/{}/timeline/{}/file/{}",
-            host, status.tenant_id, status.timeline_id, filename
-        );
+    let client = Client::new(host.clone(), sk_auth_token.clone());
+    // Request stream with basebackup archive.
+    let bb_resp = client
+        .snapshot(status.tenant_id, status.timeline_id)
+        .await?;
 
-        let mut file = tokio::fs::File::create(&file_path).await?;
-        let mut response = client.get(&http_url).send().await?;
-        if response.status() != reqwest::StatusCode::OK {
-            bail!(
-                "pulling file {} failed: status is {}",
-                filename,
-                response.status()
-            );
-        }
-        while let Some(chunk) = response.chunk().await? {
-            file.write_all(&chunk).await?;
-            file.flush().await?;
+    // Make Stream of Bytes from it...
+    let bb_stream = bb_resp.bytes_stream().map_err(std::io::Error::other);
+    // and turn it into StreamReader implementing AsyncRead.
+    let bb_reader = tokio_util::io::StreamReader::new(bb_stream);
+
+    // Extract it on the fly to the disk. We don't use simple unpack() to fsync
+    // files.
+    let mut entries = Archive::new(bb_reader).entries()?;
+    while let Some(base_tar_entry) = entries.next().await {
+        let mut entry = base_tar_entry?;
+        let header = entry.header();
+        let file_path = header.path()?.into_owned();
+        match header.entry_type() {
+            tokio_tar::EntryType::Regular => {
+                let utf8_file_path =
+                    Utf8PathBuf::from_path_buf(file_path).expect("non-Unicode path");
+                let dst_path = tli_dir_path.join(utf8_file_path);
+                let mut f = OpenOptions::new()
+                    .create(true)
+                    .truncate(true)
+                    .write(true)
+                    .open(&dst_path)
+                    .await?;
+                tokio::io::copy(&mut entry, &mut f).await?;
+                // fsync the file
+                f.sync_all().await?;
+            }
+            _ => {
+                bail!(
+                    "entry {} in backup tar archive is of unexpected type: {:?}",
+                    file_path.display(),
+                    header.entry_type()
+                );
+            }
         }
     }
 
-    // TODO: fsync?
+    // fsync temp timeline directory to remember its contents.
+    fsync_async_opt(&tli_dir_path, !conf.no_sync).await?;
 
     // Let's create timeline from temp directory and verify that it's correct
     let (commit_lsn, flush_lsn) = validate_temp_timeline(conf, ttid, &tli_dir_path).await?;
@@ -290,7 +469,9 @@ pub async fn load_temp_timeline(
         ttid, tmp_path, timeline_path
     );
     tokio::fs::create_dir_all(get_tenant_dir(conf, &ttid.tenant_id)).await?;
-    tokio::fs::rename(tmp_path, &timeline_path).await?;
+    // fsync tenant dir creation
+    fsync_async_opt(&conf.workdir, !conf.no_sync).await?;
+    durable_rename(tmp_path, &timeline_path, !conf.no_sync).await?;
 
     let tli = GlobalTimelines::load_timeline(&guard, ttid)
         .await

safekeeper/src/safekeeper.rs

@@ -780,6 +780,9 @@ where
 
             // Initializing backup_lsn is useful to avoid making backup think it should upload 0 segment.
             state.backup_lsn = max(state.backup_lsn, state.timeline_start_lsn);
+            // similar for remote_consistent_lsn
+            state.remote_consistent_lsn =
+                max(state.remote_consistent_lsn, state.timeline_start_lsn);
 
             state.acceptor_state.term_history = msg.term_history.clone();
             self.state.finish_change(&state).await?;

safekeeper/src/timeline.rs

@@ -4,7 +4,7 @@
 use anyhow::{anyhow, bail, Result};
 use camino::Utf8PathBuf;
 use serde::{Deserialize, Serialize};
-use tokio::fs;
+use tokio::fs::{self};
 use tokio_util::sync::CancellationToken;
 use utils::id::TenantId;
 
@@ -168,6 +168,9 @@ pub struct SharedState {
     pub(crate) sk: SafeKeeper<control_file::FileStorage, wal_storage::PhysicalStorage>,
     /// In memory list containing state of peers sent in latest messages from them.
     pub(crate) peers_info: PeersInfo,
+    // True value hinders old WAL removal; this is used by snapshotting. We
+    // could make it a counter, but there is no need to.
+    pub(crate) wal_removal_on_hold: bool,
 }
 
 impl SharedState {
@@ -205,6 +208,7 @@ impl SharedState {
         Ok(Self {
             sk,
             peers_info: PeersInfo(vec![]),
+            wal_removal_on_hold: false,
         })
     }
 
@@ -222,10 +226,11 @@ impl SharedState {
         Ok(Self {
             sk: SafeKeeper::new(control_store, wal_store, conf.my_id)?,
             peers_info: PeersInfo(vec![]),
+            wal_removal_on_hold: false,
         })
     }
 
-    fn get_wal_seg_size(&self) -> usize {
+    pub(crate) fn get_wal_seg_size(&self) -> usize {
         self.sk.state.server.wal_seg_size as usize
     }
 

safekeeper/src/timeline_manager.rs

@@ -39,6 +39,7 @@ pub struct StateSnapshot {
     // misc
     pub cfile_last_persist_at: Instant,
     pub inmem_flush_pending: bool,
+    pub wal_removal_on_hold: bool,
     pub peers: Vec<PeerInfo>,
 }
 
@@ -54,6 +55,7 @@ impl StateSnapshot {
             cfile_backup_lsn: read_guard.sk.state.backup_lsn,
             cfile_last_persist_at: read_guard.sk.state.pers.last_persist_at(),
             inmem_flush_pending: Self::has_unflushed_inmem_state(&read_guard),
+            wal_removal_on_hold: read_guard.wal_removal_on_hold,
             peers: read_guard.get_peers(heartbeat_timeout),
         }
     }
@@ -324,8 +326,8 @@ async fn update_wal_removal(
     last_removed_segno: u64,
     wal_removal_task: &mut Option<JoinHandle<anyhow::Result<u64>>>,
 ) {
-    if wal_removal_task.is_some() {
-        // WAL removal is already in progress
+    if wal_removal_task.is_some() || state.wal_removal_on_hold {
+        // WAL removal is already in progress or hold off
         return;
     }
 

safekeeper/src/wal_storage.rs

@@ -684,13 +684,12 @@ impl WalReader {
         let xlogoff = self.pos.segment_offset(self.wal_seg_size);
         let segno = self.pos.segment_number(self.wal_seg_size);
         let wal_file_name = XLogFileName(PG_TLI, segno, self.wal_seg_size);
-        let wal_file_path = self.timeline_dir.join(&wal_file_name);
 
         // Try to open local file, if we may have WAL locally
         if self.pos >= self.local_start_lsn {
-            let res = Self::open_wal_file(&wal_file_path).await;
+            let res = open_wal_file(&self.timeline_dir, segno, self.wal_seg_size).await;
             match res {
-                Ok(mut file) => {
+                Ok((mut file, _)) => {
                     file.seek(SeekFrom::Start(xlogoff as u64)).await?;
                     return Ok(Box::pin(file));
                 }
@@ -718,25 +717,6 @@ impl WalReader {
 
         bail!("WAL segment is not found")
     }
-
-    /// Helper function for opening a wal file.
-    async fn open_wal_file(wal_file_path: &Utf8Path) -> Result<tokio::fs::File> {
-        // First try to open the .partial file.
-        let mut partial_path = wal_file_path.to_owned();
-        partial_path.set_extension("partial");
-        if let Ok(opened_file) = tokio::fs::File::open(&partial_path).await {
-            return Ok(opened_file);
-        }
-
-        // If that failed, try it without the .partial extension.
-        tokio::fs::File::open(&wal_file_path)
-            .await
-            .with_context(|| format!("Failed to open WAL file {:?}", wal_file_path))
-            .map_err(|e| {
-                warn!("{}", e);
-                e
-            })
-    }
 }
 
 /// Zero block for filling created WAL segments.
@@ -758,6 +738,34 @@ async fn write_zeroes(file: &mut File, mut count: usize) -> Result<()> {
     Ok(())
 }
 
+/// Helper function for opening WAL segment `segno` in `dir`. Returns file and
+/// whether it is .partial.
+pub(crate) async fn open_wal_file(
+    timeline_dir: &Utf8Path,
+    segno: XLogSegNo,
+    wal_seg_size: usize,
+) -> Result<(tokio::fs::File, bool)> {
+    let (wal_file_path, wal_file_partial_path) = wal_file_paths(timeline_dir, segno, wal_seg_size)?;
+
+    // First try to open the .partial file.
+    let mut partial_path = wal_file_path.to_owned();
+    partial_path.set_extension("partial");
+    if let Ok(opened_file) = tokio::fs::File::open(&wal_file_partial_path).await {
+        return Ok((opened_file, true));
+    }
+
+    // If that failed, try it without the .partial extension.
+    let pf = tokio::fs::File::open(&wal_file_path)
+        .await
+        .with_context(|| format!("failed to open WAL file {:#}", wal_file_path))
+        .map_err(|e| {
+            warn!("{}", e);
+            e
+        })?;
+
+    Ok((pf, false))
+}
+
 /// Helper returning full path to WAL segment file and its .partial brother.
 pub fn wal_file_paths(
     timeline_dir: &Utf8Path,

safekeeper/tests/walproposer_sim/safekeeper.rs

@@ -174,6 +174,7 @@ pub fn run_server(os: NodeOs, disk: Arc<SafekeeperDisk>) -> Result<()> {
         pg_auth: None,
         pg_tenant_only_auth: None,
         http_auth: None,
+        sk_auth_token: None,
         current_thread_runtime: false,
         walsenders_keep_horizon: false,
         partial_backup_enabled: false,

storage_controller/src/heartbeater.rs

@@ -31,6 +31,7 @@ pub(crate) enum PageserverState {
     Available {
         last_seen_at: Instant,
         utilization: PageserverUtilization,
+        new: bool,
     },
     Offline,
 }
@@ -127,6 +128,7 @@ impl HeartbeaterTask {
             heartbeat_futs.push({
                 let jwt_token = self.jwt_token.clone();
                 let cancel = self.cancel.clone();
+                let new_node = !self.state.contains_key(node_id);
 
                 // Clone the node and mark it as available such that the request
                 // goes through to the pageserver even when the node is marked offline.
@@ -159,6 +161,7 @@ impl HeartbeaterTask {
                         PageserverState::Available {
                             last_seen_at: Instant::now(),
                             utilization,
+                            new: new_node,
                         }
                     } else {
                         PageserverState::Offline
@@ -220,6 +223,7 @@ impl HeartbeaterTask {
                     }
                 },
                 Vacant(_) => {
+                    // This is a new node. Don't generate a delta for it.
                     deltas.push((node_id, ps_state.clone()));
                 }
             }

storage_controller/src/node.rs

@@ -3,7 +3,7 @@ use std::{str::FromStr, time::Duration};
 use pageserver_api::{
     controller_api::{
         NodeAvailability, NodeDescribeResponse, NodeRegisterRequest, NodeSchedulingPolicy,
-        TenantLocateResponseShard,
+        TenantLocateResponseShard, UtilizationScore,
     },
     shard::TenantShardId,
 };
@@ -116,6 +116,16 @@ impl Node {
         match (self.availability, availability) {
             (Offline, Active(_)) => ToActive,
             (Active(_), Offline) => ToOffline,
+            // Consider the case when the storage controller handles the re-attach of a node
+            // before the heartbeats detect that the node is back online. We still need
+            // [`Service::node_configure`] to attempt reconciliations for shards with an
+            // unknown observed location.
+            // The unsavoury match arm below handles this situation.
+            (Active(lhs), Active(rhs))
+                if lhs == UtilizationScore::worst() && rhs < UtilizationScore::worst() =>
+            {
+                ToActive
+            }
             _ => Unchanged,
         }
     }

storage_controller/src/service.rs

@@ -12,7 +12,7 @@ use crate::{
     id_lock_map::{trace_exclusive_lock, trace_shared_lock, IdLockMap, WrappedWriteGuard},
     persistence::{AbortShardSplitStatus, TenantFilter},
     reconciler::{ReconcileError, ReconcileUnits},
-    scheduler::{ScheduleContext, ScheduleMode},
+    scheduler::{MaySchedule, ScheduleContext, ScheduleMode},
     tenant_shard::{
         MigrateAttachment, ReconcileNeeded, ScheduleOptimization, ScheduleOptimizationAction,
     },
@@ -747,29 +747,61 @@ impl Service {
             let res = self.heartbeater.heartbeat(nodes).await;
             if let Ok(deltas) = res {
                 for (node_id, state) in deltas.0 {
-                    let new_availability = match state {
-                        PageserverState::Available { utilization, .. } => NodeAvailability::Active(
-                            UtilizationScore(utilization.utilization_score),
+                    let (new_node, new_availability) = match state {
+                        PageserverState::Available {
+                            utilization, new, ..
+                        } => (
+                            new,
+                            NodeAvailability::Active(UtilizationScore(
+                                utilization.utilization_score,
+                            )),
                         ),
-                        PageserverState::Offline => NodeAvailability::Offline,
+                        PageserverState::Offline => (false, NodeAvailability::Offline),
                     };
-                    let res = self
-                        .node_configure(node_id, Some(new_availability), None)
-                        .await;
 
-                    match res {
-                        Ok(()) => {}
-                        Err(ApiError::NotFound(_)) => {
-                            // This should be rare, but legitimate since the heartbeats are done
-                            // on a snapshot of the nodes.
-                            tracing::info!("Node {} was not found after heartbeat round", node_id);
+                    if new_node {
+                        // When the heartbeats detect a newly added node, we don't wish
+                        // to attempt to reconcile the shards assigned to it. The node
+                        // is likely handling it's re-attach response, so reconciling now
+                        // would be counterproductive.
+                        //
+                        // Instead, update the in-memory state with the details learned about the
+                        // node.
+                        let mut locked = self.inner.write().unwrap();
+                        let (nodes, _tenants, scheduler) = locked.parts_mut();
+
+                        let mut new_nodes = (**nodes).clone();
+
+                        if let Some(node) = new_nodes.get_mut(&node_id) {
+                            node.set_availability(new_availability);
+                            scheduler.node_upsert(node);
                         }
-                        Err(err) => {
-                            tracing::error!(
-                                "Failed to update node {} after heartbeat round: {}",
-                                node_id,
-                                err
-                            );
+
+                        locked.nodes = Arc::new(new_nodes);
+                    } else {
+                        // This is the code path for geniune availability transitions (i.e node
+                        // goes unavailable and/or comes back online).
+                        let res = self
+                            .node_configure(node_id, Some(new_availability), None)
+                            .await;
+
+                        match res {
+                            Ok(()) => {}
+                            Err(ApiError::NotFound(_)) => {
+                                // This should be rare, but legitimate since the heartbeats are done
+                                // on a snapshot of the nodes.
+                                tracing::info!(
+                                    "Node {} was not found after heartbeat round",
+                                    node_id
+                                );
+                            }
+                            Err(err) => {
+                                tracing::error!(
+                                    "Failed to update node {} after heartbeat round: {}",
+                                    node_id,
+                                    err
+                                );
+                            }
                         }
                     }
                 }
@@ -4316,6 +4348,16 @@ impl Service {
                         continue;
                     }
 
+                    if !new_nodes
+                        .values()
+                        .any(|n| matches!(n.may_schedule(), MaySchedule::Yes(_)))
+                    {
+                        // Special case for when all nodes are unavailable and/or unschedulable: there is no point
+                        // trying to reschedule since there's nowhere else to go. Without this
+                        // branch we incorrectly detach tenants in response to node unavailability.
+                        continue;
+                    }
+
                     if tenant_shard.intent.demote_attached(scheduler, node_id) {
                         tenant_shard.sequence = tenant_shard.sequence.next();
 
@@ -4353,6 +4395,12 @@ impl Service {
                 // When a node comes back online, we must reconcile any tenant that has a None observed
                 // location on the node.
                 for tenant_shard in locked.tenants.values_mut() {
+                    // If a reconciliation is already in progress, rely on the previous scheduling
+                    // decision and skip triggering a new reconciliation.
+                    if tenant_shard.reconciler.is_some() {
+                        continue;
+                    }
+
                     if let Some(observed_loc) = tenant_shard.observed.locations.get_mut(&node_id) {
                         if observed_loc.conf.is_none() {
                             self.maybe_reconcile_shard(tenant_shard, &new_nodes);

test_runner/fixtures/neon_fixtures.py

@@ -3847,7 +3847,15 @@ class Safekeeper(LogUtils):
                 assert isinstance(res, dict)
                 return res
 
-    def http_client(self, auth_token: Optional[str] = None) -> SafekeeperHttpClient:
+    def http_client(
+        self, auth_token: Optional[str] = None, gen_sk_wide_token: bool = True
+    ) -> SafekeeperHttpClient:
+        """
+        When auth_token is None but gen_sk_wide is True creates safekeeper wide
+        token, which is a reasonable default.
+        """
+        if auth_token is None and gen_sk_wide_token:
+            auth_token = self.env.auth_keys.generate_safekeeper_token()
         is_testing_enabled = '"testing"' in self.env.get_binary_version("safekeeper")
         return SafekeeperHttpClient(
             port=self.port.http, auth_token=auth_token, is_testing_enabled=is_testing_enabled
@@ -3897,11 +3905,13 @@ class Safekeeper(LogUtils):
         segments.sort()
         return segments
 
-    def checkpoint_up_to(self, tenant_id: TenantId, timeline_id: TimelineId, lsn: Lsn):
+    def checkpoint_up_to(
+        self, tenant_id: TenantId, timeline_id: TimelineId, lsn: Lsn, wait_wal_removal=True
+    ):
         """
         Assuming pageserver(s) uploaded to s3 up to `lsn`,
         1) wait for remote_consistent_lsn and wal_backup_lsn on safekeeper to reach it.
-        2) checkpoint timeline on safekeeper, which should remove WAL before this LSN.
+        2) checkpoint timeline on safekeeper, which should remove WAL before this LSN; optionally wait for that.
         """
         cli = self.http_client()
 
@@ -3925,7 +3935,8 @@ class Safekeeper(LogUtils):
         # pageserver to this safekeeper
         wait_until(30, 1, are_lsns_advanced)
         cli.checkpoint(tenant_id, timeline_id)
-        wait_until(30, 1, are_segments_removed)
+        if wait_wal_removal:
+            wait_until(30, 1, are_segments_removed)
 
     def wait_until_paused(self, failpoint: str):
         msg = f"at failpoint {failpoint}"
@@ -4447,6 +4458,7 @@ def wait_for_last_flush_lsn(
     tenant: TenantId,
     timeline: TimelineId,
     pageserver_id: Optional[int] = None,
+    auth_token: Optional[str] = None,
 ) -> Lsn:
     """Wait for pageserver to catch up the latest flush LSN, returns the last observed lsn."""
 
@@ -4460,7 +4472,7 @@ def wait_for_last_flush_lsn(
             f"wait_for_last_flush_lsn: waiting for {last_flush_lsn} on shard {tenant_shard_id} on pageserver {pageserver.id})"
         )
         waited = wait_for_last_record_lsn(
-            pageserver.http_client(), tenant_shard_id, timeline, last_flush_lsn
+            pageserver.http_client(auth_token=auth_token), tenant_shard_id, timeline, last_flush_lsn
         )
 
         assert waited >= last_flush_lsn
@@ -4556,6 +4568,7 @@ def last_flush_lsn_upload(
     tenant_id: TenantId,
     timeline_id: TimelineId,
     pageserver_id: Optional[int] = None,
+    auth_token: Optional[str] = None,
 ) -> Lsn:
     """
     Wait for pageserver to catch to the latest flush LSN of given endpoint,
@@ -4563,11 +4576,11 @@ def last_flush_lsn_upload(
     reaching flush LSN).
     """
     last_flush_lsn = wait_for_last_flush_lsn(
-        env, endpoint, tenant_id, timeline_id, pageserver_id=pageserver_id
+        env, endpoint, tenant_id, timeline_id, pageserver_id=pageserver_id, auth_token=auth_token
     )
     shards = tenant_get_shards(env, tenant_id, pageserver_id)
     for tenant_shard_id, pageserver in shards:
-        ps_http = pageserver.http_client()
+        ps_http = pageserver.http_client(auth_token=auth_token)
         wait_for_last_record_lsn(ps_http, tenant_shard_id, timeline_id, last_flush_lsn)
         # force a checkpoint to trigger upload
         ps_http.timeline_checkpoint(tenant_shard_id, timeline_id)

test_runner/pg_clients/typescript/serverless-driver/package-lock.json

@@ -6,7 +6,7 @@
     "": {
       "dependencies": {
         "@neondatabase/serverless": "0.9.0",
-        "ws": "8.16.0"
+        "ws": "8.17.1"
       }
     },
     "node_modules/@neondatabase/serverless": {
@@ -96,9 +96,9 @@
       }
     },
     "node_modules/ws": {
-      "version": "8.16.0",
-      "resolved": "https://registry.npmjs.org/ws/-/ws-8.16.0.tgz",
-      "integrity": "sha512-HS0c//TP7Ina87TfiPUz1rQzMhHrl/SG2guqRcTOIUYD2q8uhUdNHZYJUaQ8aTGPzCh+c6oawMKW35nFl1dxyQ==",
+      "version": "8.17.1",
+      "resolved": "https://registry.npmjs.org/ws/-/ws-8.17.1.tgz",
+      "integrity": "sha512-6XQFvXTkbfUOZOKKILFG1PDK2NDQs4azKQl26T0YS5CxqWLgXajbPZ+h4gZekJyRqFU8pvnbAbbs/3TgRPy+GQ==",
       "engines": {
         "node": ">=10.0.0"
       },

test_runner/pg_clients/typescript/serverless-driver/package.json

@@ -2,6 +2,6 @@
   "type": "module",
   "dependencies": {
     "@neondatabase/serverless": "0.9.0",
-    "ws": "8.16.0"
+    "ws": "8.17.1"
   }
 }

test_runner/regress/test_attach_tenant_config.py

@@ -195,6 +195,8 @@ def test_fully_custom_config(positive_env: NeonEnv):
         "walreceiver_connect_timeout": "13m",
         "image_layer_creation_check_threshold": 1,
         "switch_aux_file_policy": "cross-validation",
+        "lsn_lease_length": "1m",
+        "lsn_lease_length_for_ts": "5s",
     }
 
     ps_http = env.pageserver.http_client()

test_runner/regress/test_pageserver_metric_collection.py

@@ -75,9 +75,6 @@ def test_metric_collection(
     env.pageserver.allowed_errors.extend(
         [
             ".*metrics endpoint refused the sent metrics*",
-            # we have a fast rate of calculation, these can happen at shutdown
-            ".*synthetic_size_worker:calculate_synthetic_size.*:gather_size_inputs.*: failed to calculate logical size at .*: cancelled.*",
-            ".*synthetic_size_worker: failed to calculate synthetic size for tenant .*: failed to calculate some logical_sizes",
             ".*metrics_collection: failed to upload to S3: Failed to upload data of length .* to storage path.*",
         ]
     )
@@ -238,9 +235,6 @@ def test_metric_collection_cleans_up_tempfile(
     env.pageserver.allowed_errors.extend(
         [
             ".*metrics endpoint refused the sent metrics*",
-            # we have a fast rate of calculation, these can happen at shutdown
-            ".*synthetic_size_worker:calculate_synthetic_size.*:gather_size_inputs.*: failed to calculate logical size at .*: cancelled.*",
-            ".*synthetic_size_worker: failed to calculate synthetic size for tenant .*: failed to calculate some logical_sizes",
         ]
     )
 

test_runner/regress/test_storage_controller.py

@@ -934,19 +934,27 @@ class Failure:
     def clear(self, env: NeonEnv):
         raise NotImplementedError()
 
+    def nodes(self):
+        raise NotImplementedError()
+
 
 class NodeStop(Failure):
-    def __init__(self, pageserver_id, immediate):
-        self.pageserver_id = pageserver_id
+    def __init__(self, pageserver_ids, immediate):
+        self.pageserver_ids = pageserver_ids
         self.immediate = immediate
 
     def apply(self, env: NeonEnv):
-        pageserver = env.get_pageserver(self.pageserver_id)
-        pageserver.stop(immediate=self.immediate)
+        for ps_id in self.pageserver_ids:
+            pageserver = env.get_pageserver(ps_id)
+            pageserver.stop(immediate=self.immediate)
 
     def clear(self, env: NeonEnv):
-        pageserver = env.get_pageserver(self.pageserver_id)
-        pageserver.start()
+        for ps_id in self.pageserver_ids:
+            pageserver = env.get_pageserver(ps_id)
+            pageserver.start()
+
+    def nodes(self):
+        return self.pageserver_ids
 
 
 class PageserverFailpoint(Failure):
@@ -962,6 +970,9 @@ class PageserverFailpoint(Failure):
         pageserver = env.get_pageserver(self.pageserver_id)
         pageserver.http_client().configure_failpoints((self.failpoint, "off"))
 
+    def nodes(self):
+        return [self.pageserver_id]
+
 
 def build_node_to_tenants_map(env: NeonEnv) -> dict[int, list[TenantId]]:
     tenants = env.storage_controller.tenant_list()
@@ -985,8 +996,9 @@ def build_node_to_tenants_map(env: NeonEnv) -> dict[int, list[TenantId]]:
 @pytest.mark.parametrize(
     "failure",
     [
-        NodeStop(pageserver_id=1, immediate=False),
-        NodeStop(pageserver_id=1, immediate=True),
+        NodeStop(pageserver_ids=[1], immediate=False),
+        NodeStop(pageserver_ids=[1], immediate=True),
+        NodeStop(pageserver_ids=[1, 2], immediate=True),
         PageserverFailpoint(pageserver_id=1, failpoint="get-utilization-http-handler"),
     ],
 )
@@ -1039,33 +1051,50 @@ def test_storage_controller_heartbeats(
     wait_until(10, 1, tenants_placed)
 
     # ... then we apply the failure
-    offline_node_id = failure.pageserver_id
-    online_node_id = (set(range(1, len(env.pageservers) + 1)) - {offline_node_id}).pop()
-    env.get_pageserver(offline_node_id).allowed_errors.append(
-        # In the case of the failpoint failure, the impacted pageserver
-        # still believes it has the tenant attached since location
-        # config calls into it will fail due to being marked offline.
-        ".*Dropped remote consistent LSN updates.*",
-    )
+    offline_node_ids = set(failure.nodes())
+    online_node_ids = set(range(1, len(env.pageservers) + 1)) - offline_node_ids
+
+    for node_id in offline_node_ids:
+        env.get_pageserver(node_id).allowed_errors.append(
+            # In the case of the failpoint failure, the impacted pageserver
+            # still believes it has the tenant attached since location
+            # config calls into it will fail due to being marked offline.
+            ".*Dropped remote consistent LSN updates.*",
+        )
+
+        if len(offline_node_ids) > 1:
+            env.get_pageserver(node_id).allowed_errors.append(
+                ".*Scheduling error when marking pageserver.*offline.*",
+            )
 
     failure.apply(env)
 
     # ... expecting the heartbeats to mark it offline
-    def node_offline():
+    def nodes_offline():
         nodes = env.storage_controller.node_list()
         log.info(f"{nodes=}")
-        target = next(n for n in nodes if n["id"] == offline_node_id)
-        assert target["availability"] == "Offline"
+        for node in nodes:
+            if node["id"] in offline_node_ids:
+                assert node["availability"] == "Offline"
 
     # A node is considered offline if the last successful heartbeat
     # was more than 10 seconds ago (hardcoded in the storage controller).
-    wait_until(20, 1, node_offline)
+    wait_until(20, 1, nodes_offline)
 
     # .. expecting the tenant on the offline node to be migrated
     def tenant_migrated():
+        if len(online_node_ids) == 0:
+            time.sleep(5)
+            return
+
         node_to_tenants = build_node_to_tenants_map(env)
         log.info(f"{node_to_tenants=}")
-        assert set(node_to_tenants[online_node_id]) == set(tenant_ids)
+
+        observed_tenants = set()
+        for node_id in online_node_ids:
+            observed_tenants |= set(node_to_tenants[node_id])
+
+        assert observed_tenants == set(tenant_ids)
 
     wait_until(10, 1, tenant_migrated)
 
@@ -1073,31 +1102,24 @@ def test_storage_controller_heartbeats(
     failure.clear(env)
 
     # ... expecting the offline node to become active again
-    def node_online():
+    def nodes_online():
         nodes = env.storage_controller.node_list()
-        target = next(n for n in nodes if n["id"] == offline_node_id)
-        assert target["availability"] == "Active"
+        for node in nodes:
+            if node["id"] in online_node_ids:
+                assert node["availability"] == "Active"
 
-    wait_until(10, 1, node_online)
+    wait_until(10, 1, nodes_online)
 
     time.sleep(5)
 
-    # ... then we create a new tenant
-    tid = TenantId.generate()
-    env.storage_controller.tenant_create(tid)
-
-    # ... expecting it to be placed on the node that just came back online
-    tenants = env.storage_controller.tenant_list()
-    newest_tenant = next(t for t in tenants if t["tenant_shard_id"] == str(tid))
-    locations = list(newest_tenant["observed"]["locations"].keys())
-    locations = [int(node_id) for node_id in locations]
-    assert locations == [offline_node_id]
+    node_to_tenants = build_node_to_tenants_map(env)
+    log.info(f"Back online: {node_to_tenants=}")
 
     # ... expecting the storage controller to reach a consistent state
     def storage_controller_consistent():
         env.storage_controller.consistency_check()
 
-    wait_until(10, 1, storage_controller_consistent)
+    wait_until(30, 1, storage_controller_consistent)
 
 
 def test_storage_controller_re_attach(neon_env_builder: NeonEnvBuilder):

test_runner/regress/test_wal_acceptor.py

@@ -317,9 +317,9 @@ def test_broker(neon_env_builder: NeonEnvBuilder):
         time.sleep(1)
 
     # Ensure that safekeepers don't lose remote_consistent_lsn on restart.
-    # Control file is persisted each 5s. TODO: do that on shutdown and remove sleep.
-    time.sleep(6)
     for sk in env.safekeepers:
+        # force persist cfile
+        sk.http_client().checkpoint(tenant_id, timeline_id)
         sk.stop()
         sk.start()
     stat_after_restart = [cli.timeline_status(tenant_id, timeline_id) for cli in clients]
@@ -374,7 +374,7 @@ def test_wal_removal(neon_env_builder: NeonEnvBuilder, auth_enabled: bool):
         http_cli_other = env.safekeepers[0].http_client(
             auth_token=env.auth_keys.generate_tenant_token(TenantId.generate())
         )
-        http_cli_noauth = env.safekeepers[0].http_client()
+        http_cli_noauth = env.safekeepers[0].http_client(gen_sk_wide_token=False)
 
     # Pretend WAL is offloaded to s3.
     if auth_enabled:
@@ -830,7 +830,7 @@ def test_timeline_status(neon_env_builder: NeonEnvBuilder, auth_enabled: bool):
             auth_token=env.auth_keys.generate_tenant_token(TenantId.generate())
         )
         wa_http_cli_bad.check_status()
-        wa_http_cli_noauth = wa.http_client()
+        wa_http_cli_noauth = wa.http_client(gen_sk_wide_token=False)
         wa_http_cli_noauth.check_status()
 
         # debug endpoint requires safekeeper scope
@@ -964,7 +964,7 @@ def test_sk_auth(neon_env_builder: NeonEnvBuilder):
 
     # By default, neon_local enables auth on all services if auth is configured,
     # so http must require the token.
-    sk_http_cli_noauth = sk.http_client()
+    sk_http_cli_noauth = sk.http_client(gen_sk_wide_token=False)
     sk_http_cli_auth = sk.http_client(auth_token=env.auth_keys.generate_tenant_token(tenant_id))
     with pytest.raises(sk_http_cli_noauth.HTTPError, match="Forbidden|Unauthorized"):
         sk_http_cli_noauth.timeline_status(tenant_id, timeline_id)
@@ -1640,7 +1640,7 @@ def test_delete_force(neon_env_builder: NeonEnvBuilder, auth_enabled: bool):
         sk_http_other = sk.http_client(
             auth_token=env.auth_keys.generate_tenant_token(tenant_id_other)
         )
-        sk_http_noauth = sk.http_client()
+        sk_http_noauth = sk.http_client(gen_sk_wide_token=False)
     assert (sk_data_dir / str(tenant_id) / str(timeline_id_1)).is_dir()
     assert (sk_data_dir / str(tenant_id) / str(timeline_id_2)).is_dir()
     assert (sk_data_dir / str(tenant_id) / str(timeline_id_3)).is_dir()
@@ -1723,7 +1723,10 @@ def test_delete_force(neon_env_builder: NeonEnvBuilder, auth_enabled: bool):
             cur.execute("INSERT INTO t (key) VALUES (123)")
 
 
+# Basic pull_timeline test.
 def test_pull_timeline(neon_env_builder: NeonEnvBuilder):
+    neon_env_builder.auth_enabled = True
+
     def execute_payload(endpoint: Endpoint):
         with closing(endpoint.connect()) as conn:
             with conn.cursor() as cur:
@@ -1739,7 +1742,7 @@ def test_pull_timeline(neon_env_builder: NeonEnvBuilder):
 
     def show_statuses(safekeepers: List[Safekeeper], tenant_id: TenantId, timeline_id: TimelineId):
         for sk in safekeepers:
-            http_cli = sk.http_client()
+            http_cli = sk.http_client(auth_token=env.auth_keys.generate_tenant_token(tenant_id))
             try:
                 status = http_cli.timeline_status(tenant_id, timeline_id)
                 log.info(f"Safekeeper {sk.id} status: {status}")
@@ -1749,11 +1752,11 @@ def test_pull_timeline(neon_env_builder: NeonEnvBuilder):
     neon_env_builder.num_safekeepers = 4
     env = neon_env_builder.init_start()
     tenant_id = env.initial_tenant
-    timeline_id = env.neon_cli.create_branch("test_pull_timeline")
+    timeline_id = env.initial_timeline
 
     log.info("Use only first 3 safekeepers")
     env.safekeepers[3].stop()
-    endpoint = env.endpoints.create("test_pull_timeline")
+    endpoint = env.endpoints.create("main")
     endpoint.active_safekeepers = [1, 2, 3]
     endpoint.start()
 
@@ -1769,7 +1772,7 @@ def test_pull_timeline(neon_env_builder: NeonEnvBuilder):
 
     res = (
         env.safekeepers[3]
-        .http_client()
+        .http_client(auth_token=env.auth_keys.generate_safekeeper_token())
         .pull_timeline(
             {
                 "tenant_id": str(tenant_id),
@@ -1787,7 +1790,7 @@ def test_pull_timeline(neon_env_builder: NeonEnvBuilder):
     show_statuses(env.safekeepers, tenant_id, timeline_id)
 
     log.info("Restarting compute with new config to verify that it works")
-    endpoint.stop_and_destroy().create("test_pull_timeline")
+    endpoint.stop_and_destroy().create("main")
     endpoint.active_safekeepers = [1, 3, 4]
     endpoint.start()
 
@@ -1816,8 +1819,8 @@ def test_pull_timeline(neon_env_builder: NeonEnvBuilder):
 # 4) Do some write, verify integrity with timeline_digest.
 # Expected to fail while holding off WAL gc plus fetching commit_lsn WAL
 # segment is not implemented.
-@pytest.mark.xfail
 def test_pull_timeline_gc(neon_env_builder: NeonEnvBuilder):
+    neon_env_builder.auth_enabled = True
     neon_env_builder.num_safekeepers = 3
     neon_env_builder.enable_safekeeper_remote_storage(default_remote_storage())
     env = neon_env_builder.init_start()
@@ -1836,27 +1839,36 @@ def test_pull_timeline_gc(neon_env_builder: NeonEnvBuilder):
     src_flush_lsn = src_sk.get_flush_lsn(tenant_id, timeline_id)
     log.info(f"flush_lsn on src before pull_timeline: {src_flush_lsn}")
 
-    dst_http = dst_sk.http_client()
+    src_http = src_sk.http_client()
     # run pull_timeline which will halt before downloading files
-    dst_http.configure_failpoints(("sk-pull-timeline-after-list-pausable", "pause"))
+    src_http.configure_failpoints(("sk-snapshot-after-list-pausable", "pause"))
     pt_handle = PropagatingThread(
         target=dst_sk.pull_timeline, args=([src_sk], tenant_id, timeline_id)
     )
     pt_handle.start()
-    dst_sk.wait_until_paused("sk-pull-timeline-after-list-pausable")
+    src_sk.wait_until_paused("sk-snapshot-after-list-pausable")
 
     # ensure segment exists
     endpoint.safe_psql("insert into t select generate_series(1, 180000), 'papaya'")
-    lsn = last_flush_lsn_upload(env, endpoint, tenant_id, timeline_id)
+    lsn = last_flush_lsn_upload(
+        env,
+        endpoint,
+        tenant_id,
+        timeline_id,
+        auth_token=env.auth_keys.generate_tenant_token(tenant_id),
+    )
     assert lsn > Lsn("0/2000000")
     # Checkpoint timeline beyond lsn.
-    src_sk.checkpoint_up_to(tenant_id, timeline_id, lsn)
+    src_sk.checkpoint_up_to(tenant_id, timeline_id, lsn, wait_wal_removal=False)
     first_segment_p = src_sk.timeline_dir(tenant_id, timeline_id) / "000000010000000000000001"
     log.info(f"first segment exist={os.path.exists(first_segment_p)}")
 
-    dst_http.configure_failpoints(("sk-pull-timeline-after-list-pausable", "off"))
+    src_http.configure_failpoints(("sk-snapshot-after-list-pausable", "off"))
     pt_handle.join()
 
+    # after pull_timeline is finished WAL should be removed on donor
+    src_sk.checkpoint_up_to(tenant_id, timeline_id, lsn, wait_wal_removal=True)
+
     timeline_start_lsn = src_sk.get_timeline_start_lsn(tenant_id, timeline_id)
     dst_flush_lsn = dst_sk.get_flush_lsn(tenant_id, timeline_id)
     log.info(f"flush_lsn on dst after pull_timeline: {dst_flush_lsn}")
@@ -1883,8 +1895,8 @@ def test_pull_timeline_gc(neon_env_builder: NeonEnvBuilder):
 # enough, so it won't be affected by term change anymore.
 #
 # Expected to fail while term check is not implemented.
-@pytest.mark.xfail
 def test_pull_timeline_term_change(neon_env_builder: NeonEnvBuilder):
+    neon_env_builder.auth_enabled = True
     neon_env_builder.num_safekeepers = 3
     neon_env_builder.enable_safekeeper_remote_storage(default_remote_storage())
     env = neon_env_builder.init_start()
@@ -1900,14 +1912,14 @@ def test_pull_timeline_term_change(neon_env_builder: NeonEnvBuilder):
     ep.safe_psql("create table t(key int, value text)")
     ep.safe_psql("insert into t select generate_series(1, 1000), 'pear'")
 
-    dst_http = dst_sk.http_client()
+    src_http = src_sk.http_client()
     # run pull_timeline which will halt before downloading files
-    dst_http.configure_failpoints(("sk-pull-timeline-after-list-pausable", "pause"))
+    src_http.configure_failpoints(("sk-snapshot-after-list-pausable", "pause"))
     pt_handle = PropagatingThread(
         target=dst_sk.pull_timeline, args=([src_sk], tenant_id, timeline_id)
     )
     pt_handle.start()
-    dst_sk.wait_until_paused("sk-pull-timeline-after-list-pausable")
+    src_sk.wait_until_paused("sk-snapshot-after-list-pausable")
 
     src_http = src_sk.http_client()
     term_before = src_http.timeline_status(tenant_id, timeline_id).term
@@ -1922,7 +1934,7 @@ def test_pull_timeline_term_change(neon_env_builder: NeonEnvBuilder):
     term_after = src_http.timeline_status(tenant_id, timeline_id).term
     assert term_after > term_before, f"term_after={term_after}, term_before={term_before}"
 
-    dst_http.configure_failpoints(("sk-pull-timeline-after-list-pausable", "off"))
+    src_http.configure_failpoints(("sk-snapshot-after-list-pausable", "off"))
     with pytest.raises(requests.exceptions.HTTPError):
         pt_handle.join()
 

vm-image-spec.yaml

@@ -324,14 +324,15 @@ files:
         help: 'Whether or not the replication slot wal_status is lost'
         key_labels:
           - slot_name
-        values: [wal_status_is_lost]
+        values: [wal_is_lost]
         query: |
           SELECT slot_name,
           CASE
             WHEN wal_status = 'lost' THEN 1
             ELSE 0
-          END AS wal_status_is_lost
+          END AS wal_is_lost
           FROM pg_replication_slots;
+
   - filename: neon_collector_autoscaling.yml
     content: |
       collector_name: neon_collector_autoscaling