Increase neon_local http client to compute timeout in reconfigure.

Seems like 30s sometimes not enough when CI runners are overloaded,
causing pull_timeline flakiness.

ref https://github.com/neondatabase/neon/issues/9731#issuecomment-2532143008

.github/actionlint.yml

@@ -21,3 +21,5 @@ config-variables:
   - SLACK_UPCOMING_RELEASE_CHANNEL_ID
   - DEV_AWS_OIDC_ROLE_ARN
   - BENCHMARK_INGEST_TARGET_PROJECTID
+  - PGREGRESS_PG16_PROJECT_ID
+  - PGREGRESS_PG17_PROJECT_ID

.github/workflows/cloud-regress.yml

@@ -23,11 +23,14 @@ jobs:
   regress:
     env:
       POSTGRES_DISTRIB_DIR: /tmp/neon/pg_install
-      DEFAULT_PG_VERSION: 16
       TEST_OUTPUT: /tmp/test_output
       BUILD_TYPE: remote
       AWS_ACCESS_KEY_ID: ${{ secrets.AWS_ACCESS_KEY_DEV }}
       AWS_SECRET_ACCESS_KEY: ${{ secrets.AWS_SECRET_KEY_DEV }}
+    strategy:
+      fail-fast: false
+      matrix:
+        pg-version: [16, 17]
 
     runs-on: us-east-2
     container:
@@ -40,9 +43,11 @@ jobs:
           submodules: true
 
       - name: Patch the test
+        env:
+          PG_VERSION: ${{matrix.pg-version}}
         run: |
-          cd "vendor/postgres-v${DEFAULT_PG_VERSION}"
-          patch -p1 < "../../compute/patches/cloud_regress_pg${DEFAULT_PG_VERSION}.patch"
+          cd "vendor/postgres-v${PG_VERSION}"
+          patch -p1 < "../../compute/patches/cloud_regress_pg${PG_VERSION}.patch"
 
       - name: Generate a random password
         id: pwgen
@@ -55,8 +60,9 @@ jobs:
       - name: Change tests according to the generated password
         env:
           DBPASS: ${{ steps.pwgen.outputs.DBPASS }}
+          PG_VERSION: ${{matrix.pg-version}}
         run: |
-          cd vendor/postgres-v"${DEFAULT_PG_VERSION}"/src/test/regress
+          cd vendor/postgres-v"${PG_VERSION}"/src/test/regress
           for fname in sql/*.sql expected/*.out; do
             sed -i.bak s/NEON_PASSWORD_PLACEHOLDER/"'${DBPASS}'"/ "${fname}"
           done
@@ -73,15 +79,29 @@ jobs:
           path: /tmp/neon/
           prefix: latest
 
+      - name: Create a new branch
+        id: create-branch
+        uses: ./.github/actions/neon-branch-create
+        with:
+          api_key: ${{ secrets.NEON_STAGING_API_KEY }}
+          project_id: ${{ vars[format('PGREGRESS_PG{0}_PROJECT_ID', matrix.pg-version)] }}
+
       - name: Run the regression tests
         uses: ./.github/actions/run-python-test-set
         with:
           build_type: ${{ env.BUILD_TYPE }}
           test_selection: cloud_regress
-          pg_version: ${{ env.DEFAULT_PG_VERSION }}
+          pg_version: ${{matrix.pg-version}}
           extra_params: -m remote_cluster
         env:
-          BENCHMARK_CONNSTR: ${{ secrets.PG_REGRESS_CONNSTR }}
+          BENCHMARK_CONNSTR: ${{steps.create-branch.outputs.dsn}}
+
+      - name: Delete branch
+        uses: ./.github/actions/neon-branch-delete
+        with:
+          api_key: ${{ secrets.NEON_STAGING_API_KEY }}
+          project_id: ${{ vars[format('PGREGRESS_PG{0}_PROJECT_ID', matrix.pg-version)] }}
+          branch_id: ${{steps.create-branch.outputs.branch_id}}
 
       - name: Create Allure report
         id: create-allure-report

CODEOWNERS

@@ -1,16 +1,29 @@
-/.github/ @neondatabase/developer-productivity
-/compute_tools/ @neondatabase/control-plane @neondatabase/compute
-/libs/pageserver_api/ @neondatabase/storage
-/libs/postgres_ffi/ @neondatabase/compute @neondatabase/storage
-/libs/proxy/ @neondatabase/proxy
-/libs/remote_storage/ @neondatabase/storage
-/libs/safekeeper_api/ @neondatabase/storage
+# Autoscaling
 /libs/vm_monitor/ @neondatabase/autoscaling
-/pageserver/ @neondatabase/storage
+
+# DevProd
+/.github/ @neondatabase/developer-productivity
+
+# Compute
 /pgxn/ @neondatabase/compute
-/pgxn/neon/ @neondatabase/compute @neondatabase/storage
+/vendor/ @neondatabase/compute
+/compute/ @neondatabase/compute
+/compute_tools/ @neondatabase/compute
+
+# Proxy
+/libs/proxy/ @neondatabase/proxy
 /proxy/ @neondatabase/proxy
+
+# Storage
+/pageserver/ @neondatabase/storage
 /safekeeper/ @neondatabase/storage
 /storage_controller @neondatabase/storage
 /storage_scrubber @neondatabase/storage
-/vendor/ @neondatabase/compute
+/libs/pageserver_api/ @neondatabase/storage
+/libs/remote_storage/ @neondatabase/storage
+/libs/safekeeper_api/ @neondatabase/storage
+
+# Shared
+/pgxn/neon/ @neondatabase/compute @neondatabase/storage
+/libs/compute_api/ @neondatabase/compute @neondatabase/control-plane
+/libs/postgres_ffi/ @neondatabase/compute @neondatabase/storage

Cargo.toml

@@ -51,10 +51,6 @@ anyhow = { version = "1.0", features = ["backtrace"] }
 arc-swap = "1.6"
 async-compression = { version = "0.4.0", features = ["tokio", "gzip", "zstd"] }
 atomic-take = "1.1.0"
-azure_core = { version = "0.19", default-features = false, features = ["enable_reqwest_rustls", "hmac_rust"] }
-azure_identity = { version = "0.19", default-features = false, features = ["enable_reqwest_rustls"] }
-azure_storage = { version = "0.19", default-features = false, features = ["enable_reqwest_rustls"] }
-azure_storage_blobs = { version = "0.19", default-features = false, features = ["enable_reqwest_rustls"] }
 flate2 = "1.0.26"
 async-stream = "0.3"
 async-trait = "0.1"
@@ -216,6 +212,12 @@ postgres-protocol = { git = "https://github.com/neondatabase/rust-postgres.git",
 postgres-types = { git = "https://github.com/neondatabase/rust-postgres.git", branch = "neon" }
 tokio-postgres = { git = "https://github.com/neondatabase/rust-postgres.git", branch = "neon" }
 
+## Azure SDK crates
+azure_core = { git = "https://github.com/neondatabase/azure-sdk-for-rust.git", branch = "neon", default-features = false, features = ["enable_reqwest_rustls", "hmac_rust"] }
+azure_identity = { git = "https://github.com/neondatabase/azure-sdk-for-rust.git", branch = "neon", default-features = false, features = ["enable_reqwest_rustls"] }
+azure_storage = { git = "https://github.com/neondatabase/azure-sdk-for-rust.git", branch = "neon", default-features = false, features = ["enable_reqwest_rustls"] }
+azure_storage_blobs = { git = "https://github.com/neondatabase/azure-sdk-for-rust.git", branch = "neon", default-features = false, features = ["enable_reqwest_rustls"] }
+
 ## Local libraries
 compute_api = { version = "0.1", path = "./libs/compute_api/" }
 consumption_metrics = { version = "0.1", path = "./libs/consumption_metrics/" }

build-tools.Dockerfile

@@ -115,7 +115,7 @@ RUN set -e \
 
 # Keep the version the same as in compute/compute-node.Dockerfile and
 # test_runner/regress/test_compute_metrics.py.
-ENV SQL_EXPORTER_VERSION=0.13.1
+ENV SQL_EXPORTER_VERSION=0.16.0
 RUN curl -fsSL \
     "https://github.com/burningalchemist/sql_exporter/releases/download/${SQL_EXPORTER_VERSION}/sql_exporter-${SQL_EXPORTER_VERSION}.linux-$(case "$(uname -m)" in x86_64) echo amd64;; aarch64) echo arm64;; esac).tar.gz" \
     --output sql_exporter.tar.gz \

compute/compute-node.Dockerfile

@@ -1324,7 +1324,7 @@ FROM quay.io/prometheuscommunity/postgres-exporter:v0.12.1 AS postgres-exporter
 
 # Keep the version the same as in build-tools.Dockerfile and
 # test_runner/regress/test_compute_metrics.py.
-FROM burningalchemist/sql_exporter:0.13.1 AS sql-exporter
+FROM burningalchemist/sql_exporter:0.16.0 AS sql-exporter
 
 #########################################################################################
 #

compute/etc/neon_collector.jsonnet

@@ -6,6 +6,7 @@
     import 'sql_exporter/compute_backpressure_throttling_seconds.libsonnet',
     import 'sql_exporter/compute_current_lsn.libsonnet',
     import 'sql_exporter/compute_logical_snapshot_files.libsonnet',
+    import 'sql_exporter/compute_logical_snapshots_bytes.libsonnet',
     import 'sql_exporter/compute_max_connections.libsonnet',
     import 'sql_exporter/compute_receive_lsn.libsonnet',
     import 'sql_exporter/compute_subscriptions_count.libsonnet',

compute/etc/sql_exporter/compute_logical_snapshots_bytes.15.sql

@@ -0,0 +1,7 @@
+SELECT
+  (SELECT current_setting('neon.timeline_id')) AS timeline_id,
+  -- Postgres creates temporary snapshot files of the form %X-%X.snap.%d.tmp.
+  -- These temporary snapshot files are renamed to the actual snapshot files
+  -- after they are completely built. We only WAL-log the completely built
+  -- snapshot files
+  (SELECT COALESCE(sum(size), 0) FROM pg_ls_logicalsnapdir() WHERE name LIKE '%.snap') AS logical_snapshots_bytes;

compute/etc/sql_exporter/compute_logical_snapshots_bytes.libsonnet

@@ -0,0 +1,17 @@
+local neon = import 'neon.libsonnet';
+
+local pg_ls_logicalsnapdir = importstr 'sql_exporter/compute_logical_snapshots_bytes.15.sql';
+local pg_ls_dir = importstr 'sql_exporter/compute_logical_snapshots_bytes.sql';
+
+{
+  metric_name: 'compute_logical_snapshots_bytes',
+  type: 'gauge',
+  help: 'Size of the pg_logical/snapshots directory, not including temporary files',
+  key_labels: [
+    'timeline_id',
+  ],
+  values: [
+    'logical_snapshots_bytes',
+  ],
+  query: if neon.PG_MAJORVERSION_NUM < 15 then pg_ls_dir else pg_ls_logicalsnapdir,
+}

compute/etc/sql_exporter/compute_logical_snapshots_bytes.sql

@@ -0,0 +1,9 @@
+SELECT
+  (SELECT setting FROM pg_settings WHERE name = 'neon.timeline_id') AS timeline_id,
+  -- Postgres creates temporary snapshot files of the form %X-%X.snap.%d.tmp.
+  -- These temporary snapshot files are renamed to the actual snapshot files
+  -- after they are completely built. We only WAL-log the completely built
+  -- snapshot files
+  (SELECT COALESCE(sum((pg_stat_file('pg_logical/snapshots/' || name, missing_ok => true)).size), 0)
+    FROM (SELECT * FROM pg_ls_dir('pg_logical/snapshots') WHERE pg_ls_dir LIKE '%.snap') AS name
+  ) AS logical_snapshots_bytes;

compute_tools/src/compute.rs

@@ -1243,12 +1243,7 @@ impl ComputeNode {
         let postgresql_conf_path = pgdata_path.join("postgresql.conf");
         config::write_postgres_conf(&postgresql_conf_path, &spec, self.http_port)?;
 
-        // TODO(ololobus): We need a concurrency during reconfiguration as well,
-        // but DB is already running and used by user. We can easily get out of
-        // `max_connections` limit, and the current code won't handle that.
-        // let compute_state = self.state.lock().unwrap().clone();
-        // let max_concurrent_connections = self.max_service_connections(&compute_state, &spec);
-        let max_concurrent_connections = 1;
+        let max_concurrent_connections = spec.reconfigure_concurrency;
 
         // Temporarily reset max_cluster_size in config
         // to avoid the possibility of hitting the limit, while we are reconfiguring:

control_plane/src/endpoint.rs

@@ -53,6 +53,7 @@ use compute_api::spec::Role;
 use nix::sys::signal::kill;
 use nix::sys::signal::Signal;
 use pageserver_api::shard::ShardStripeSize;
+use reqwest::header::CONTENT_TYPE;
 use serde::{Deserialize, Serialize};
 use url::Host;
 use utils::id::{NodeId, TenantId, TimelineId};
@@ -618,6 +619,7 @@ impl Endpoint {
             pgbouncer_settings: None,
             shard_stripe_size: Some(shard_stripe_size),
             local_proxy_config: None,
+            reconfigure_concurrency: 1,
         };
         let spec_path = self.endpoint_path().join("spec.json");
         std::fs::write(spec_path, serde_json::to_string_pretty(&spec)?)?;
@@ -808,7 +810,7 @@ impl Endpoint {
         }
 
         let client = reqwest::Client::builder()
-            .timeout(Duration::from_secs(30))
+            .timeout(Duration::from_secs(120))
             .build()
             .unwrap();
         let response = client
@@ -817,6 +819,7 @@ impl Endpoint {
                 self.http_address.ip(),
                 self.http_address.port()
             ))
+            .header(CONTENT_TYPE.as_str(), "application/json")
             .body(format!(
                 "{{\"spec\":{}}}",
                 serde_json::to_string_pretty(&spec)?

deny.toml

@@ -42,6 +42,7 @@ allow = [
     "MPL-2.0",
     "OpenSSL",
     "Unicode-DFS-2016",
+    "Unicode-3.0",
 ]
 confidence-threshold = 0.8
 exceptions = [

libs/compute_api/src/spec.rs

@@ -19,6 +19,10 @@ pub type PgIdent = String;
 /// String type alias representing Postgres extension version
 pub type ExtVersion = String;
 
+fn default_reconfigure_concurrency() -> usize {
+    1
+}
+
 /// Cluster spec or configuration represented as an optional number of
 /// delta operations + final cluster state description.
 #[derive(Clone, Debug, Default, Deserialize, Serialize)]
@@ -67,7 +71,7 @@ pub struct ComputeSpec {
     pub cluster: Cluster,
     pub delta_operations: Option<Vec<DeltaOp>>,
 
-    /// An optinal hint that can be passed to speed up startup time if we know
+    /// An optional hint that can be passed to speed up startup time if we know
     /// that no pg catalog mutations (like role creation, database creation,
     /// extension creation) need to be done on the actual database to start.
     #[serde(default)] // Default false
@@ -86,9 +90,7 @@ pub struct ComputeSpec {
     // etc. GUCs in cluster.settings. TODO: Once the control plane has been
     // updated to fill these fields, we can make these non optional.
     pub tenant_id: Option<TenantId>,
-
     pub timeline_id: Option<TimelineId>,
-
     pub pageserver_connstring: Option<String>,
 
     #[serde(default)]
@@ -113,6 +115,20 @@ pub struct ComputeSpec {
     /// Local Proxy configuration used for JWT authentication
     #[serde(default)]
     pub local_proxy_config: Option<LocalProxySpec>,
+
+    /// Number of concurrent connections during the parallel RunInEachDatabase
+    /// phase of the apply config process.
+    ///
+    /// We need a higher concurrency during reconfiguration in case of many DBs,
+    /// but instance is already running and used by client. We can easily get out of
+    /// `max_connections` limit, and the current code won't handle that.
+    ///
+    /// Default is 1, but also allow control plane to override this value for specific
+    /// projects. It's also recommended to bump `superuser_reserved_connections` +=
+    /// `reconfigure_concurrency` for such projects to ensure that we always have
+    /// enough spare connections for reconfiguration process to succeed.
+    #[serde(default = "default_reconfigure_concurrency")]
+    pub reconfigure_concurrency: usize,
 }
 
 /// Feature flag to signal `compute_ctl` to enable certain experimental functionality.
@@ -315,6 +331,9 @@ mod tests {
 
         // Features list defaults to empty vector.
         assert!(spec.features.is_empty());
+
+        // Reconfigure concurrency defaults to 1.
+        assert_eq!(spec.reconfigure_concurrency, 1);
     }
 
     #[test]

libs/pageserver_api/src/controller_api.rs

@@ -245,6 +245,17 @@ impl From<NodeAvailability> for NodeAvailabilityWrapper {
     }
 }
 
+/// Scheduling policy enables us to selectively disable some automatic actions that the
+/// controller performs on a tenant shard. This is only set to a non-default value by
+/// human intervention, and it is reset to the default value (Active) when the tenant's
+/// placement policy is modified away from Attached.
+///
+/// The typical use of a non-Active scheduling policy is one of:
+/// - Pinnning a shard to a node (i.e. migrating it there & setting a non-Active scheduling policy)
+/// - Working around a bug (e.g. if something is flapping and we need to stop it until the bug is fixed)
+///
+/// If you're not sure which policy to use to pin a shard to its current location, you probably
+/// want Pause.
 #[derive(Serialize, Deserialize, Clone, Copy, Eq, PartialEq, Debug)]
 pub enum ShardSchedulingPolicy {
     // Normal mode: the tenant's scheduled locations may be updated at will, including

libs/pageserver_api/src/shard.rs

@@ -158,7 +158,8 @@ impl ShardIdentity {
         key_to_shard_number(self.count, self.stripe_size, key)
     }
 
-    /// Return true if the key should be ingested by this shard
+    /// Return true if the key is stored only on this shard. This does not include
+    /// global keys, see is_key_global().
     ///
     /// Shards must ingest _at least_ keys which return true from this check.
     pub fn is_key_local(&self, key: &Key) -> bool {
@@ -171,7 +172,7 @@ impl ShardIdentity {
     }
 
     /// Return true if the key should be stored on all shards, not just one.
-    fn is_key_global(&self, key: &Key) -> bool {
+    pub fn is_key_global(&self, key: &Key) -> bool {
         if key.is_slru_block_key() || key.is_slru_segment_size_key() || key.is_aux_file_key() {
             // Special keys that are only stored on shard 0
             false

libs/remote_storage/src/azure_blob.rs

@@ -8,15 +8,14 @@ use std::io;
 use std::num::NonZeroU32;
 use std::pin::Pin;
 use std::str::FromStr;
-use std::sync::Arc;
 use std::time::Duration;
 use std::time::SystemTime;
 
 use super::REMOTE_STORAGE_PREFIX_SEPARATOR;
+use anyhow::Context;
 use anyhow::Result;
 use azure_core::request_options::{IfMatchCondition, MaxResults, Metadata, Range};
 use azure_core::{Continuable, RetryOptions};
-use azure_identity::DefaultAzureCredential;
 use azure_storage::StorageCredentials;
 use azure_storage_blobs::blob::CopyStatus;
 use azure_storage_blobs::prelude::ClientBuilder;
@@ -76,8 +75,9 @@ impl AzureBlobStorage {
         let credentials = if let Ok(access_key) = env::var("AZURE_STORAGE_ACCESS_KEY") {
             StorageCredentials::access_key(account.clone(), access_key)
         } else {
-            let token_credential = DefaultAzureCredential::default();
-            StorageCredentials::token_credential(Arc::new(token_credential))
+            let token_credential = azure_identity::create_default_credential()
+                .context("trying to obtain Azure default credentials")?;
+            StorageCredentials::token_credential(token_credential)
         };
 
         // we have an outer retry
@@ -624,6 +624,10 @@ impl RemoteStorage for AzureBlobStorage {
         res
     }
 
+    fn max_keys_per_delete(&self) -> usize {
+        super::MAX_KEYS_PER_DELETE_AZURE
+    }
+
     async fn copy(
         &self,
         from: &RemotePath,

libs/remote_storage/src/lib.rs

@@ -70,7 +70,14 @@ pub const DEFAULT_REMOTE_STORAGE_AZURE_CONCURRENCY_LIMIT: usize = 100;
 pub const DEFAULT_MAX_KEYS_PER_LIST_RESPONSE: Option<i32> = None;
 
 /// As defined in S3 docs
-pub const MAX_KEYS_PER_DELETE: usize = 1000;
+///
+/// <https://docs.aws.amazon.com/AmazonS3/latest/API/API_DeleteObjects.html>
+pub const MAX_KEYS_PER_DELETE_S3: usize = 1000;
+
+/// As defined in Azure docs
+///
+/// <https://learn.microsoft.com/en-us/rest/api/storageservices/blob-batch>
+pub const MAX_KEYS_PER_DELETE_AZURE: usize = 256;
 
 const REMOTE_STORAGE_PREFIX_SEPARATOR: char = '/';
 
@@ -340,6 +347,14 @@ pub trait RemoteStorage: Send + Sync + 'static {
         cancel: &CancellationToken,
     ) -> anyhow::Result<()>;
 
+    /// Returns the maximum number of keys that a call to [`Self::delete_objects`] can delete without chunking
+    ///
+    /// The value returned is only an optimization hint, One can pass larger number of objects to
+    /// `delete_objects` as well.
+    ///
+    /// The value is guaranteed to be >= 1.
+    fn max_keys_per_delete(&self) -> usize;
+
     /// Deletes all objects matching the given prefix.
     ///
     /// NB: this uses NoDelimiter and will match partial prefixes. For example, the prefix /a/b will
@@ -533,6 +548,16 @@ impl<Other: RemoteStorage> GenericRemoteStorage<Arc<Other>> {
         }
     }
 
+    /// [`RemoteStorage::max_keys_per_delete`]
+    pub fn max_keys_per_delete(&self) -> usize {
+        match self {
+            Self::LocalFs(s) => s.max_keys_per_delete(),
+            Self::AwsS3(s) => s.max_keys_per_delete(),
+            Self::AzureBlob(s) => s.max_keys_per_delete(),
+            Self::Unreliable(s) => s.max_keys_per_delete(),
+        }
+    }
+
     /// See [`RemoteStorage::delete_prefix`]
     pub async fn delete_prefix(
         &self,

libs/remote_storage/src/local_fs.rs

@@ -573,6 +573,10 @@ impl RemoteStorage for LocalFs {
         Ok(())
     }
 
+    fn max_keys_per_delete(&self) -> usize {
+        super::MAX_KEYS_PER_DELETE_S3
+    }
+
     async fn copy(
         &self,
         from: &RemotePath,

libs/remote_storage/src/s3_bucket.rs

@@ -48,7 +48,7 @@ use crate::{
     metrics::{start_counting_cancelled_wait, start_measuring_requests},
     support::PermitCarrying,
     ConcurrencyLimiter, Download, DownloadError, DownloadOpts, Listing, ListingMode, ListingObject,
-    RemotePath, RemoteStorage, TimeTravelError, TimeoutOrCancel, MAX_KEYS_PER_DELETE,
+    RemotePath, RemoteStorage, TimeTravelError, TimeoutOrCancel, MAX_KEYS_PER_DELETE_S3,
     REMOTE_STORAGE_PREFIX_SEPARATOR,
 };
 
@@ -355,7 +355,7 @@ impl S3Bucket {
         let kind = RequestKind::Delete;
         let mut cancel = std::pin::pin!(cancel.cancelled());
 
-        for chunk in delete_objects.chunks(MAX_KEYS_PER_DELETE) {
+        for chunk in delete_objects.chunks(MAX_KEYS_PER_DELETE_S3) {
             let started_at = start_measuring_requests(kind);
 
             let req = self
@@ -832,6 +832,10 @@ impl RemoteStorage for S3Bucket {
         self.delete_oids(&permit, &delete_objects, cancel).await
     }
 
+    fn max_keys_per_delete(&self) -> usize {
+        MAX_KEYS_PER_DELETE_S3
+    }
+
     async fn delete(&self, path: &RemotePath, cancel: &CancellationToken) -> anyhow::Result<()> {
         let paths = std::array::from_ref(path);
         self.delete_objects(paths, cancel).await

libs/remote_storage/src/simulate_failures.rs

@@ -203,6 +203,10 @@ impl RemoteStorage for UnreliableWrapper {
         Ok(())
     }
 
+    fn max_keys_per_delete(&self) -> usize {
+        self.inner.max_keys_per_delete()
+    }
+
     async fn copy(
         &self,
         from: &RemotePath,

libs/utils/src/shard.rs

@@ -164,6 +164,12 @@ impl TenantShardId {
     }
 }
 
+impl std::fmt::Display for ShardNumber {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        self.0.fmt(f)
+    }
+}
+
 impl std::fmt::Display for ShardSlug<'_> {
     fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
         write!(

pageserver/src/deletion_queue/deleter.rs

@@ -9,7 +9,6 @@
 use remote_storage::GenericRemoteStorage;
 use remote_storage::RemotePath;
 use remote_storage::TimeoutOrCancel;
-use remote_storage::MAX_KEYS_PER_DELETE;
 use std::time::Duration;
 use tokio_util::sync::CancellationToken;
 use tracing::info;
@@ -131,7 +130,8 @@ impl Deleter {
     }
 
     pub(super) async fn background(&mut self) -> Result<(), DeletionQueueError> {
-        self.accumulator.reserve(MAX_KEYS_PER_DELETE);
+        let max_keys_per_delete = self.remote_storage.max_keys_per_delete();
+        self.accumulator.reserve(max_keys_per_delete);
 
         loop {
             if self.cancel.is_cancelled() {
@@ -156,14 +156,14 @@ impl Deleter {
 
             match msg {
                 DeleterMessage::Delete(mut list) => {
-                    while !list.is_empty() || self.accumulator.len() == MAX_KEYS_PER_DELETE {
-                        if self.accumulator.len() == MAX_KEYS_PER_DELETE {
+                    while !list.is_empty() || self.accumulator.len() == max_keys_per_delete {
+                        if self.accumulator.len() == max_keys_per_delete {
                             self.flush().await?;
                             // If we have received this number of keys, proceed with attempting to execute
                             assert_eq!(self.accumulator.len(), 0);
                         }
 
-                        let available_slots = MAX_KEYS_PER_DELETE - self.accumulator.len();
+                        let available_slots = max_keys_per_delete - self.accumulator.len();
                         let take_count = std::cmp::min(available_slots, list.len());
                         for path in list.drain(list.len() - take_count..) {
                             self.accumulator.push(path);

pageserver/src/http/routes.rs

@@ -87,7 +87,7 @@ use crate::tenant::timeline::offload::offload_timeline;
 use crate::tenant::timeline::offload::OffloadError;
 use crate::tenant::timeline::CompactFlags;
 use crate::tenant::timeline::CompactOptions;
-use crate::tenant::timeline::CompactRange;
+use crate::tenant::timeline::CompactRequest;
 use crate::tenant::timeline::CompactionError;
 use crate::tenant::timeline::Timeline;
 use crate::tenant::GetTimelineError;
@@ -1978,6 +1978,26 @@ async fn timeline_gc_handler(
     json_response(StatusCode::OK, gc_result)
 }
 
+// Cancel scheduled compaction tasks
+async fn timeline_cancel_compact_handler(
+    request: Request<Body>,
+    _cancel: CancellationToken,
+) -> Result<Response<Body>, ApiError> {
+    let tenant_shard_id: TenantShardId = parse_request_param(&request, "tenant_shard_id")?;
+    let timeline_id: TimelineId = parse_request_param(&request, "timeline_id")?;
+    check_permission(&request, Some(tenant_shard_id.tenant_id))?;
+    let state = get_state(&request);
+    async {
+        let tenant = state
+            .tenant_manager
+            .get_attached_tenant_shard(tenant_shard_id)?;
+        tenant.cancel_scheduled_compaction(timeline_id);
+        json_response(StatusCode::OK, ())
+    }
+    .instrument(info_span!("timeline_cancel_compact", tenant_id = %tenant_shard_id.tenant_id, shard_id = %tenant_shard_id.shard_slug(), %timeline_id))
+    .await
+}
+
 // Run compaction immediately on given timeline.
 async fn timeline_compact_handler(
     mut request: Request<Body>,
@@ -1987,7 +2007,7 @@ async fn timeline_compact_handler(
     let timeline_id: TimelineId = parse_request_param(&request, "timeline_id")?;
     check_permission(&request, Some(tenant_shard_id.tenant_id))?;
 
-    let compact_range = json_request_maybe::<Option<CompactRange>>(&mut request).await?;
+    let compact_request = json_request_maybe::<Option<CompactRequest>>(&mut request).await?;
 
     let state = get_state(&request);
 
@@ -2012,22 +2032,50 @@ async fn timeline_compact_handler(
     let wait_until_uploaded =
         parse_query_param::<_, bool>(&request, "wait_until_uploaded")?.unwrap_or(false);
 
+    let wait_until_scheduled_compaction_done =
+        parse_query_param::<_, bool>(&request, "wait_until_scheduled_compaction_done")?
+            .unwrap_or(false);
+
+    let sub_compaction = compact_request
+        .as_ref()
+        .map(|r| r.sub_compaction)
+        .unwrap_or(false);
     let options = CompactOptions {
-        compact_range,
+        compact_range: compact_request
+            .as_ref()
+            .and_then(|r| r.compact_range.clone()),
+        compact_below_lsn: compact_request.as_ref().and_then(|r| r.compact_below_lsn),
         flags,
+        sub_compaction,
     };
 
+    let scheduled = compact_request
+        .as_ref()
+        .map(|r| r.scheduled)
+        .unwrap_or(false);
+
     async {
         let ctx = RequestContext::new(TaskKind::MgmtRequest, DownloadBehavior::Download);
         let timeline = active_timeline_of_active_tenant(&state.tenant_manager, tenant_shard_id, timeline_id).await?;
-        timeline
-            .compact_with_options(&cancel, options, &ctx)
-            .await
-            .map_err(|e| ApiError::InternalServerError(e.into()))?;
-        if wait_until_uploaded {
-            timeline.remote_client.wait_completion().await
-            // XXX map to correct ApiError for the cases where it's due to shutdown
-            .context("wait completion").map_err(ApiError::InternalServerError)?;
+        if scheduled {
+            let tenant = state
+                .tenant_manager
+                .get_attached_tenant_shard(tenant_shard_id)?;
+            let rx = tenant.schedule_compaction(timeline_id, options).await.map_err(ApiError::InternalServerError)?;
+            if wait_until_scheduled_compaction_done {
+                // It is possible that this will take a long time, dropping the HTTP request will not cancel the compaction.
+                rx.await.ok();
+            }
+        } else {
+            timeline
+                .compact_with_options(&cancel, options, &ctx)
+                .await
+                .map_err(|e| ApiError::InternalServerError(e.into()))?;
+            if wait_until_uploaded {
+                timeline.remote_client.wait_completion().await
+                // XXX map to correct ApiError for the cases where it's due to shutdown
+                .context("wait completion").map_err(ApiError::InternalServerError)?;
+            }
         }
         json_response(StatusCode::OK, ())
     }
@@ -2108,16 +2156,20 @@ async fn timeline_checkpoint_handler(
     // By default, checkpoints come with a compaction, but this may be optionally disabled by tests that just want to flush + upload.
     let compact = parse_query_param::<_, bool>(&request, "compact")?.unwrap_or(true);
 
+    let wait_until_flushed: bool =
+        parse_query_param(&request, "wait_until_flushed")?.unwrap_or(true);
+
     let wait_until_uploaded =
         parse_query_param::<_, bool>(&request, "wait_until_uploaded")?.unwrap_or(false);
 
     async {
         let ctx = RequestContext::new(TaskKind::MgmtRequest, DownloadBehavior::Download);
         let timeline = active_timeline_of_active_tenant(&state.tenant_manager, tenant_shard_id, timeline_id).await?;
-        timeline
-            .freeze_and_flush()
-            .await
-            .map_err(|e| {
+        if wait_until_flushed {
+            timeline.freeze_and_flush().await
+        } else {
+            timeline.freeze().await.and(Ok(()))
+        }.map_err(|e| {
                 match e {
                     tenant::timeline::FlushLayerError::Cancelled => ApiError::ShuttingDown,
                     other => ApiError::InternalServerError(other.into()),
@@ -3301,6 +3353,10 @@ pub fn make_router(
             "/v1/tenant/:tenant_shard_id/timeline/:timeline_id/compact",
             |r| api_handler(r, timeline_compact_handler),
         )
+        .delete(
+            "/v1/tenant/:tenant_shard_id/timeline/:timeline_id/compact",
+            |r| api_handler(r, timeline_cancel_compact_handler),
+        )
         .put(
             "/v1/tenant/:tenant_shard_id/timeline/:timeline_id/offload",
             |r| testing_api_handler("attempt timeline offload", r, timeline_offload_handler),

pageserver/src/metrics.rs

@@ -464,6 +464,24 @@ static LAST_RECORD_LSN: Lazy<IntGaugeVec> = Lazy::new(|| {
     .expect("failed to define a metric")
 });
 
+static DISK_CONSISTENT_LSN: Lazy<IntGaugeVec> = Lazy::new(|| {
+    register_int_gauge_vec!(
+        "pageserver_disk_consistent_lsn",
+        "Disk consistent LSN grouped by timeline",
+        &["tenant_id", "shard_id", "timeline_id"]
+    )
+    .expect("failed to define a metric")
+});
+
+pub(crate) static PROJECTED_REMOTE_CONSISTENT_LSN: Lazy<UIntGaugeVec> = Lazy::new(|| {
+    register_uint_gauge_vec!(
+        "pageserver_projected_remote_consistent_lsn",
+        "Projected remote consistent LSN grouped by timeline",
+        &["tenant_id", "shard_id", "timeline_id"]
+    )
+    .expect("failed to define a metric")
+});
+
 static PITR_HISTORY_SIZE: Lazy<UIntGaugeVec> = Lazy::new(|| {
     register_uint_gauge_vec!(
         "pageserver_pitr_history_size",
@@ -1205,31 +1223,60 @@ pub(crate) mod virtual_file_io_engine {
     });
 }
 
-pub(crate) struct SmgrOpTimer {
+pub(crate) struct SmgrOpTimer(Option<SmgrOpTimerInner>);
+pub(crate) struct SmgrOpTimerInner {
     global_latency_histo: Histogram,
 
     // Optional because not all op types are tracked per-timeline
     per_timeline_latency_histo: Option<Histogram>,
 
+    global_flush_in_progress_micros: IntCounter,
+    per_timeline_flush_in_progress_micros: IntCounter,
+
     start: Instant,
     throttled: Duration,
     op: SmgrQueryType,
 }
 
+pub(crate) struct SmgrOpFlushInProgress {
+    base: Instant,
+    global_micros: IntCounter,
+    per_timeline_micros: IntCounter,
+}
+
 impl SmgrOpTimer {
     pub(crate) fn deduct_throttle(&mut self, throttle: &Option<Duration>) {
         let Some(throttle) = throttle else {
             return;
         };
-        self.throttled += *throttle;
+        let inner = self.0.as_mut().expect("other public methods consume self");
+        inner.throttled += *throttle;
     }
-}
 
-impl Drop for SmgrOpTimer {
-    fn drop(&mut self) {
-        let elapsed = self.start.elapsed();
+    pub(crate) fn observe_smgr_op_completion_and_start_flushing(mut self) -> SmgrOpFlushInProgress {
+        let (flush_start, inner) = self
+            .smgr_op_end()
+            .expect("this method consume self, and the only other caller is drop handler");
+        let SmgrOpTimerInner {
+            global_flush_in_progress_micros,
+            per_timeline_flush_in_progress_micros,
+            ..
+        } = inner;
+        SmgrOpFlushInProgress {
+            base: flush_start,
+            global_micros: global_flush_in_progress_micros,
+            per_timeline_micros: per_timeline_flush_in_progress_micros,
+        }
+    }
 
-        let elapsed = match elapsed.checked_sub(self.throttled) {
+    /// Returns `None`` if this method has already been called, `Some` otherwise.
+    fn smgr_op_end(&mut self) -> Option<(Instant, SmgrOpTimerInner)> {
+        let inner = self.0.take()?;
+
+        let now = Instant::now();
+        let elapsed = now - inner.start;
+
+        let elapsed = match elapsed.checked_sub(inner.throttled) {
             Some(elapsed) => elapsed,
             None => {
                 use utils::rate_limit::RateLimit;
@@ -1240,9 +1287,9 @@ impl Drop for SmgrOpTimer {
                         })))
                     });
                 let mut guard = LOGGED.lock().unwrap();
-                let rate_limit = &mut guard[self.op];
+                let rate_limit = &mut guard[inner.op];
                 rate_limit.call(|| {
-                    warn!(op=?self.op, ?elapsed, ?self.throttled, "implementation error: time spent throttled exceeds total request wall clock time");
+                    warn!(op=?inner.op, ?elapsed, ?inner.throttled, "implementation error: time spent throttled exceeds total request wall clock time");
                 });
                 elapsed // un-throttled time, more info than just saturating to 0
             }
@@ -1250,10 +1297,54 @@ impl Drop for SmgrOpTimer {
 
         let elapsed = elapsed.as_secs_f64();
 
-        self.global_latency_histo.observe(elapsed);
-        if let Some(per_timeline_getpage_histo) = &self.per_timeline_latency_histo {
+        inner.global_latency_histo.observe(elapsed);
+        if let Some(per_timeline_getpage_histo) = &inner.per_timeline_latency_histo {
             per_timeline_getpage_histo.observe(elapsed);
         }
+
+        Some((now, inner))
+    }
+}
+
+impl Drop for SmgrOpTimer {
+    fn drop(&mut self) {
+        self.smgr_op_end();
+    }
+}
+
+impl SmgrOpFlushInProgress {
+    pub(crate) async fn measure<Fut, O>(mut self, mut fut: Fut) -> O
+    where
+        Fut: std::future::Future<Output = O>,
+    {
+        let mut fut = std::pin::pin!(fut);
+
+        let now = Instant::now();
+        // Whenever observe_guard gets called, or dropped,
+        // it adds the time elapsed since its last call to metrics.
+        // Last call is tracked in `now`.
+        let mut observe_guard = scopeguard::guard(
+            || {
+                let elapsed = now - self.base;
+                self.global_micros
+                    .inc_by(u64::try_from(elapsed.as_micros()).unwrap());
+                self.per_timeline_micros
+                    .inc_by(u64::try_from(elapsed.as_micros()).unwrap());
+                self.base = now;
+            },
+            |mut observe| {
+                observe();
+            },
+        );
+
+        loop {
+            match tokio::time::timeout(Duration::from_secs(10), &mut fut).await {
+                Ok(v) => return v,
+                Err(_timeout) => {
+                    (*observe_guard)();
+                }
+            }
+        }
     }
 }
 
@@ -1284,6 +1375,8 @@ pub(crate) struct SmgrQueryTimePerTimeline {
     per_timeline_getpage_latency: Histogram,
     global_batch_size: Histogram,
     per_timeline_batch_size: Histogram,
+    global_flush_in_progress_micros: IntCounter,
+    per_timeline_flush_in_progress_micros: IntCounter,
 }
 
 static SMGR_QUERY_STARTED_GLOBAL: Lazy<IntCounterVec> = Lazy::new(|| {
@@ -1446,6 +1539,26 @@ fn set_page_service_config_max_batch_size(conf: &PageServicePipeliningConfig) {
         .set(value.try_into().unwrap());
 }
 
+static PAGE_SERVICE_SMGR_FLUSH_INPROGRESS_MICROS: Lazy<IntCounterVec> = Lazy::new(|| {
+    register_int_counter_vec!(
+        "pageserver_page_service_pagestream_flush_in_progress_micros",
+        "Counter that sums up the microseconds that a pagestream response was being flushed into the TCP connection. \
+         If the flush is particularly slow, this counter will be updated periodically to make slow flushes \
+         easily discoverable in monitoring. \
+         Hence, this is NOT a completion latency historgram.",
+        &["tenant_id", "shard_id", "timeline_id"],
+    )
+    .expect("failed to define a metric")
+});
+
+static PAGE_SERVICE_SMGR_FLUSH_INPROGRESS_MICROS_GLOBAL: Lazy<IntCounter> = Lazy::new(|| {
+    register_int_counter!(
+        "pageserver_page_service_pagestream_flush_in_progress_micros_global",
+        "Like pageserver_page_service_pagestream_flush_in_progress_seconds, but instance-wide.",
+    )
+    .expect("failed to define a metric")
+});
+
 impl SmgrQueryTimePerTimeline {
     pub(crate) fn new(tenant_shard_id: &TenantShardId, timeline_id: &TimelineId) -> Self {
         let tenant_id = tenant_shard_id.tenant_id.to_string();
@@ -1486,6 +1599,12 @@ impl SmgrQueryTimePerTimeline {
             .get_metric_with_label_values(&[&tenant_id, &shard_slug, &timeline_id])
             .unwrap();
 
+        let global_flush_in_progress_micros =
+            PAGE_SERVICE_SMGR_FLUSH_INPROGRESS_MICROS_GLOBAL.clone();
+        let per_timeline_flush_in_progress_micros = PAGE_SERVICE_SMGR_FLUSH_INPROGRESS_MICROS
+            .get_metric_with_label_values(&[&tenant_id, &shard_slug, &timeline_id])
+            .unwrap();
+
         Self {
             global_started,
             global_latency,
@@ -1493,6 +1612,8 @@ impl SmgrQueryTimePerTimeline {
             per_timeline_getpage_started,
             global_batch_size,
             per_timeline_batch_size,
+            global_flush_in_progress_micros,
+            per_timeline_flush_in_progress_micros,
         }
     }
     pub(crate) fn start_smgr_op(&self, op: SmgrQueryType, started_at: Instant) -> SmgrOpTimer {
@@ -1505,13 +1626,17 @@ impl SmgrQueryTimePerTimeline {
             None
         };
 
-        SmgrOpTimer {
+        SmgrOpTimer(Some(SmgrOpTimerInner {
             global_latency_histo: self.global_latency[op as usize].clone(),
             per_timeline_latency_histo,
             start: started_at,
             op,
             throttled: Duration::ZERO,
-        }
+            global_flush_in_progress_micros: self.global_flush_in_progress_micros.clone(),
+            per_timeline_flush_in_progress_micros: self
+                .per_timeline_flush_in_progress_micros
+                .clone(),
+        }))
     }
 
     pub(crate) fn observe_getpage_batch_start(&self, batch_size: usize) {
@@ -2186,6 +2311,15 @@ pub(crate) static WAL_INGEST: Lazy<WalIngestMetrics> = Lazy::new(|| WalIngestMet
     .expect("failed to define a metric"),
 });
 
+pub(crate) static PAGESERVER_TIMELINE_WAL_RECORDS_RECEIVED: Lazy<IntCounterVec> = Lazy::new(|| {
+    register_int_counter_vec!(
+        "pageserver_timeline_wal_records_received",
+        "Number of WAL records received per shard",
+        &["tenant_id", "shard_id", "timeline_id"]
+    )
+    .expect("failed to define a metric")
+});
+
 pub(crate) static WAL_REDO_TIME: Lazy<Histogram> = Lazy::new(|| {
     register_histogram!(
         "pageserver_wal_redo_seconds",
@@ -2394,7 +2528,8 @@ pub(crate) struct TimelineMetrics {
     pub load_layer_map_histo: StorageTimeMetrics,
     pub garbage_collect_histo: StorageTimeMetrics,
     pub find_gc_cutoffs_histo: StorageTimeMetrics,
-    pub last_record_gauge: IntGauge,
+    pub last_record_lsn_gauge: IntGauge,
+    pub disk_consistent_lsn_gauge: IntGauge,
     pub pitr_history_size: UIntGauge,
     pub archival_size: UIntGauge,
     pub(crate) layer_size_image: UIntGauge,
@@ -2412,6 +2547,7 @@ pub(crate) struct TimelineMetrics {
     pub evictions_with_low_residence_duration: std::sync::RwLock<EvictionsWithLowResidenceDuration>,
     /// Number of valid LSN leases.
     pub valid_lsn_lease_count_gauge: UIntGauge,
+    pub wal_records_received: IntCounter,
     shutdown: std::sync::atomic::AtomicBool,
 }
 
@@ -2475,7 +2611,11 @@ impl TimelineMetrics {
             &shard_id,
             &timeline_id,
         );
-        let last_record_gauge = LAST_RECORD_LSN
+        let last_record_lsn_gauge = LAST_RECORD_LSN
+            .get_metric_with_label_values(&[&tenant_id, &shard_id, &timeline_id])
+            .unwrap();
+
+        let disk_consistent_lsn_gauge = DISK_CONSISTENT_LSN
             .get_metric_with_label_values(&[&tenant_id, &shard_id, &timeline_id])
             .unwrap();
 
@@ -2565,6 +2705,10 @@ impl TimelineMetrics {
             .get_metric_with_label_values(&[&tenant_id, &shard_id, &timeline_id])
             .unwrap();
 
+        let wal_records_received = PAGESERVER_TIMELINE_WAL_RECORDS_RECEIVED
+            .get_metric_with_label_values(&[&tenant_id, &shard_id, &timeline_id])
+            .unwrap();
+
         TimelineMetrics {
             tenant_id,
             shard_id,
@@ -2578,7 +2722,8 @@ impl TimelineMetrics {
             garbage_collect_histo,
             find_gc_cutoffs_histo,
             load_layer_map_histo,
-            last_record_gauge,
+            last_record_lsn_gauge,
+            disk_consistent_lsn_gauge,
             pitr_history_size,
             archival_size,
             layer_size_image,
@@ -2596,6 +2741,7 @@ impl TimelineMetrics {
                 evictions_with_low_residence_duration,
             ),
             valid_lsn_lease_count_gauge,
+            wal_records_received,
             shutdown: std::sync::atomic::AtomicBool::default(),
         }
     }
@@ -2642,6 +2788,7 @@ impl TimelineMetrics {
         let timeline_id = &self.timeline_id;
         let shard_id = &self.shard_id;
         let _ = LAST_RECORD_LSN.remove_label_values(&[tenant_id, shard_id, timeline_id]);
+        let _ = DISK_CONSISTENT_LSN.remove_label_values(&[tenant_id, shard_id, timeline_id]);
         let _ = FLUSH_WAIT_UPLOAD_TIME.remove_label_values(&[tenant_id, shard_id, timeline_id]);
         let _ = STANDBY_HORIZON.remove_label_values(&[tenant_id, shard_id, timeline_id]);
         {
@@ -2732,6 +2879,16 @@ impl TimelineMetrics {
             shard_id,
             timeline_id,
         ]);
+        let _ = PAGESERVER_TIMELINE_WAL_RECORDS_RECEIVED.remove_label_values(&[
+            tenant_id,
+            shard_id,
+            timeline_id,
+        ]);
+        let _ = PAGE_SERVICE_SMGR_FLUSH_INPROGRESS_MICROS.remove_label_values(&[
+            tenant_id,
+            shard_id,
+            timeline_id,
+        ]);
     }
 }
 
@@ -2805,6 +2962,7 @@ pub(crate) struct RemoteTimelineClientMetrics {
     calls: Mutex<HashMap<(&'static str, &'static str), IntCounterPair>>,
     bytes_started_counter: Mutex<HashMap<(&'static str, &'static str), IntCounter>>,
     bytes_finished_counter: Mutex<HashMap<(&'static str, &'static str), IntCounter>>,
+    pub(crate) projected_remote_consistent_lsn_gauge: UIntGauge,
 }
 
 impl RemoteTimelineClientMetrics {
@@ -2819,6 +2977,10 @@ impl RemoteTimelineClientMetrics {
                 .unwrap(),
         );
 
+        let projected_remote_consistent_lsn_gauge = PROJECTED_REMOTE_CONSISTENT_LSN
+            .get_metric_with_label_values(&[&tenant_id_str, &shard_id_str, &timeline_id_str])
+            .unwrap();
+
         RemoteTimelineClientMetrics {
             tenant_id: tenant_id_str,
             shard_id: shard_id_str,
@@ -2827,6 +2989,7 @@ impl RemoteTimelineClientMetrics {
             bytes_started_counter: Mutex::new(HashMap::default()),
             bytes_finished_counter: Mutex::new(HashMap::default()),
             remote_physical_size_gauge,
+            projected_remote_consistent_lsn_gauge,
         }
     }
 
@@ -3040,6 +3203,7 @@ impl Drop for RemoteTimelineClientMetrics {
             calls,
             bytes_started_counter,
             bytes_finished_counter,
+            projected_remote_consistent_lsn_gauge,
         } = self;
         for ((a, b), _) in calls.get_mut().unwrap().drain() {
             let mut res = [Ok(()), Ok(())];
@@ -3069,6 +3233,14 @@ impl Drop for RemoteTimelineClientMetrics {
             let _ = remote_physical_size_gauge; // use to avoid 'unused' warning in desctructuring above
             let _ = REMOTE_PHYSICAL_SIZE.remove_label_values(&[tenant_id, shard_id, timeline_id]);
         }
+        {
+            let _ = projected_remote_consistent_lsn_gauge;
+            let _ = PROJECTED_REMOTE_CONSISTENT_LSN.remove_label_values(&[
+                tenant_id,
+                shard_id,
+                timeline_id,
+            ]);
+        }
     }
 }
 

pageserver/src/page_service.rs

@@ -1017,10 +1017,8 @@ impl PageServerHandler {
         // Map handler result to protocol behavior.
         // Some handler errors cause exit from pagestream protocol.
         // Other handler errors are sent back as an error message and we stay in pagestream protocol.
-        let mut timers: smallvec::SmallVec<[_; 1]> =
-            smallvec::SmallVec::with_capacity(handler_results.len());
         for handler_result in handler_results {
-            let response_msg = match handler_result {
+            let (response_msg, timer) = match handler_result {
                 Err(e) => match &e {
                     PageStreamError::Shutdown => {
                         // If we fail to fulfil a request during shutdown, which may be _because_ of
@@ -1044,34 +1042,66 @@ impl PageServerHandler {
                         span.in_scope(|| {
                             error!("error reading relation or page version: {full:#}")
                         });
-                        PagestreamBeMessage::Error(PagestreamErrorResponse {
-                            message: e.to_string(),
-                        })
+                        (
+                            PagestreamBeMessage::Error(PagestreamErrorResponse {
+                                message: e.to_string(),
+                            }),
+                            None, // TODO: measure errors
+                        )
                     }
                 },
-                Ok((response_msg, timer)) => {
-                    // Extending the lifetime of the timers so observations on drop
-                    // include the flush time.
-                    timers.push(timer);
-                    response_msg
-                }
+                Ok((response_msg, timer)) => (response_msg, Some(timer)),
             };
 
+            //
             // marshal & transmit response message
+            //
+
             pgb_writer.write_message_noflush(&BeMessage::CopyData(&response_msg.serialize()))?;
-        }
-        tokio::select! {
-            biased;
-            _ = cancel.cancelled() => {
-                // We were requested to shut down.
-                info!("shutdown request received in page handler");
-                return Err(QueryError::Shutdown)
-            }
-            res = pgb_writer.flush() => {
-                res?;
+
+            // We purposefully don't count flush time into the timer.
+            //
+            // The reason is that current compute client will not perform protocol processing
+            // if the postgres backend process is doing things other than `->smgr_read()`.
+            // This is especially the case for prefetch.
+            //
+            // If the compute doesn't read from the connection, eventually TCP will backpressure
+            // all the way into our flush call below.
+            //
+            // The timer's underlying metric is used for a storage-internal latency SLO and
+            // we don't want to include latency in it that we can't control.
+            // And as pointed out above, in this case, we don't control the time that flush will take.
+            let flushing_timer =
+                timer.map(|timer| timer.observe_smgr_op_completion_and_start_flushing());
+
+            // what we want to do
+            let flush_fut = pgb_writer.flush();
+            // metric for how long flushing takes
+            let flush_fut = match flushing_timer {
+                Some(flushing_timer) => {
+                    futures::future::Either::Left(flushing_timer.measure(flush_fut))
+                }
+                None => futures::future::Either::Right(flush_fut),
+            };
+            // do it while respecting cancellation
+            let _: () = async move {
+                tokio::select! {
+                    biased;
+                    _ = cancel.cancelled() => {
+                        // We were requested to shut down.
+                        info!("shutdown request received in page handler");
+                        return Err(QueryError::Shutdown)
+                    }
+                    res = flush_fut => {
+                        res?;
+                    }
+                }
+                Ok(())
             }
+            // and log the info! line inside the request span
+            .instrument(span.clone())
+            .await?;
         }
-        drop(timers);
         Ok(())
     }
 

pageserver/src/tenant.rs

@@ -37,14 +37,19 @@ use remote_timeline_client::manifest::{
 };
 use remote_timeline_client::UploadQueueNotReadyError;
 use std::collections::BTreeMap;
+use std::collections::VecDeque;
 use std::fmt;
 use std::future::Future;
 use std::sync::atomic::AtomicBool;
 use std::sync::Weak;
 use std::time::SystemTime;
 use storage_broker::BrokerClientChannel;
+use timeline::compaction::ScheduledCompactionTask;
 use timeline::import_pgdata;
 use timeline::offload::offload_timeline;
+use timeline::CompactFlags;
+use timeline::CompactOptions;
+use timeline::CompactionError;
 use timeline::ShutdownMode;
 use tokio::io::BufReader;
 use tokio::sync::watch;
@@ -339,6 +344,11 @@ pub struct Tenant {
     /// Overhead of mutex is acceptable because compaction is done with a multi-second period.
     compaction_circuit_breaker: std::sync::Mutex<CircuitBreaker>,
 
+    /// Scheduled compaction tasks. Currently, this can only be populated by triggering
+    /// a manual gc-compaction from the manual compaction API.
+    scheduled_compaction_tasks:
+        std::sync::Mutex<HashMap<TimelineId, VecDeque<ScheduledCompactionTask>>>,
+
     /// If the tenant is in Activating state, notify this to encourage it
     /// to proceed to Active as soon as possible, rather than waiting for lazy
     /// background warmup.
@@ -2953,27 +2963,109 @@ impl Tenant {
 
         for (timeline_id, timeline, (can_compact, can_offload)) in &timelines_to_compact_or_offload
         {
+            // pending_task_left == None: cannot compact, maybe still pending tasks
+            // pending_task_left == Some(true): compaction task left
+            // pending_task_left == Some(false): no compaction task left
             let pending_task_left = if *can_compact {
-                Some(
-                    timeline
-                        .compact(cancel, EnumSet::empty(), ctx)
-                        .instrument(info_span!("compact_timeline", %timeline_id))
-                        .await
-                        .inspect_err(|e| match e {
-                            timeline::CompactionError::ShuttingDown => (),
-                            timeline::CompactionError::Offload(_) => {
-                                // Failures to offload timelines do not trip the circuit breaker, because
-                                // they do not do lots of writes the way compaction itself does: it is cheap
-                                // to retry, and it would be bad to stop all compaction because of an issue with offloading.
+                let has_pending_l0_compaction_task = timeline
+                    .compact(cancel, EnumSet::empty(), ctx)
+                    .instrument(info_span!("compact_timeline", %timeline_id))
+                    .await
+                    .inspect_err(|e| match e {
+                        timeline::CompactionError::ShuttingDown => (),
+                        timeline::CompactionError::Offload(_) => {
+                            // Failures to offload timelines do not trip the circuit breaker, because
+                            // they do not do lots of writes the way compaction itself does: it is cheap
+                            // to retry, and it would be bad to stop all compaction because of an issue with offloading.
+                        }
+                        timeline::CompactionError::Other(e) => {
+                            self.compaction_circuit_breaker
+                                .lock()
+                                .unwrap()
+                                .fail(&CIRCUIT_BREAKERS_BROKEN, e);
+                        }
+                    })?;
+                if has_pending_l0_compaction_task {
+                    Some(true)
+                } else {
+                    let mut has_pending_scheduled_compaction_task;
+                    let next_scheduled_compaction_task = {
+                        let mut guard = self.scheduled_compaction_tasks.lock().unwrap();
+                        if let Some(tline_pending_tasks) = guard.get_mut(timeline_id) {
+                            if !tline_pending_tasks.is_empty() {
+                                info!(
+                                    "{} tasks left in the compaction schedule queue",
+                                    tline_pending_tasks.len()
+                                );
                             }
-                            timeline::CompactionError::Other(e) => {
-                                self.compaction_circuit_breaker
-                                    .lock()
-                                    .unwrap()
-                                    .fail(&CIRCUIT_BREAKERS_BROKEN, e);
+                            let next_task = tline_pending_tasks.pop_front();
+                            has_pending_scheduled_compaction_task = !tline_pending_tasks.is_empty();
+                            next_task
+                        } else {
+                            has_pending_scheduled_compaction_task = false;
+                            None
+                        }
+                    };
+                    if let Some(mut next_scheduled_compaction_task) = next_scheduled_compaction_task
+                    {
+                        if !next_scheduled_compaction_task
+                            .options
+                            .flags
+                            .contains(CompactFlags::EnhancedGcBottomMostCompaction)
+                        {
+                            warn!("ignoring scheduled compaction task: scheduled task must be gc compaction: {:?}", next_scheduled_compaction_task.options);
+                        } else if next_scheduled_compaction_task.options.sub_compaction {
+                            info!("running scheduled enhanced gc bottom-most compaction with sub-compaction, splitting compaction jobs");
+                            let jobs = timeline
+                                .gc_compaction_split_jobs(next_scheduled_compaction_task.options)
+                                .await
+                                .map_err(CompactionError::Other)?;
+                            if jobs.is_empty() {
+                                info!("no jobs to run, skipping scheduled compaction task");
+                            } else {
+                                has_pending_scheduled_compaction_task = true;
+                                let jobs_len = jobs.len();
+                                let mut guard = self.scheduled_compaction_tasks.lock().unwrap();
+                                let tline_pending_tasks = guard.entry(*timeline_id).or_default();
+                                for (idx, job) in jobs.into_iter().enumerate() {
+                                    tline_pending_tasks.push_back(if idx == jobs_len - 1 {
+                                        ScheduledCompactionTask {
+                                            options: job,
+                                            // The last job in the queue sends the signal and releases the gc guard
+                                            result_tx: next_scheduled_compaction_task
+                                                .result_tx
+                                                .take(),
+                                            gc_block: next_scheduled_compaction_task
+                                                .gc_block
+                                                .take(),
+                                        }
+                                    } else {
+                                        ScheduledCompactionTask {
+                                            options: job,
+                                            result_tx: None,
+                                            gc_block: None,
+                                        }
+                                    });
+                                }
+                                info!("scheduled enhanced gc bottom-most compaction with sub-compaction, split into {} jobs", jobs_len);
                             }
-                        })?,
-                )
+                        } else {
+                            let _ = timeline
+                                .compact_with_options(
+                                    cancel,
+                                    next_scheduled_compaction_task.options,
+                                    ctx,
+                                )
+                                .instrument(info_span!("scheduled_compact_timeline", %timeline_id))
+                                .await?;
+                            if let Some(tx) = next_scheduled_compaction_task.result_tx.take() {
+                                // TODO: we can send compaction statistics in the future
+                                tx.send(()).ok();
+                            }
+                        }
+                    }
+                    Some(has_pending_scheduled_compaction_task)
+                }
             } else {
                 None
             };
@@ -2993,6 +3085,43 @@ impl Tenant {
         Ok(has_pending_task)
     }
 
+    /// Cancel scheduled compaction tasks
+    pub(crate) fn cancel_scheduled_compaction(
+        &self,
+        timeline_id: TimelineId,
+    ) -> Vec<ScheduledCompactionTask> {
+        let mut guard = self.scheduled_compaction_tasks.lock().unwrap();
+        if let Some(tline_pending_tasks) = guard.get_mut(&timeline_id) {
+            let current_tline_pending_tasks = std::mem::take(tline_pending_tasks);
+            current_tline_pending_tasks.into_iter().collect()
+        } else {
+            Vec::new()
+        }
+    }
+
+    /// Schedule a compaction task for a timeline.
+    pub(crate) async fn schedule_compaction(
+        &self,
+        timeline_id: TimelineId,
+        options: CompactOptions,
+    ) -> anyhow::Result<tokio::sync::oneshot::Receiver<()>> {
+        let gc_guard = match self.gc_block.start().await {
+            Ok(guard) => guard,
+            Err(e) => {
+                bail!("cannot run gc-compaction because gc is blocked: {}", e);
+            }
+        };
+        let (tx, rx) = tokio::sync::oneshot::channel();
+        let mut guard = self.scheduled_compaction_tasks.lock().unwrap();
+        let tline_pending_tasks = guard.entry(timeline_id).or_default();
+        tline_pending_tasks.push_back(ScheduledCompactionTask {
+            options,
+            result_tx: Some(tx),
+            gc_block: Some(gc_guard),
+        });
+        Ok(rx)
+    }
+
     // Call through to all timelines to freeze ephemeral layers if needed.  Usually
     // this happens during ingest: this background housekeeping is for freezing layers
     // that are open but haven't been written to for some time.
@@ -4005,6 +4134,7 @@ impl Tenant {
                 // use an extremely long backoff.
                 Some(Duration::from_secs(3600 * 24)),
             )),
+            scheduled_compaction_tasks: Mutex::new(Default::default()),
             activate_now_sem: tokio::sync::Semaphore::new(0),
             attach_wal_lag_cooldown: Arc::new(std::sync::OnceLock::new()),
             cancel: CancellationToken::default(),
@@ -4376,7 +4506,12 @@ impl Tenant {
                 // - this timeline was created while we were finding cutoffs
                 // - lsn for timestamp search fails for this timeline repeatedly
                 if let Some(cutoffs) = gc_cutoffs.get(&timeline.timeline_id) {
-                    target.cutoffs = cutoffs.clone();
+                    let original_cutoffs = target.cutoffs.clone();
+                    // GC cutoffs should never go back
+                    target.cutoffs = GcCutoffs {
+                        space: Lsn(cutoffs.space.0.max(original_cutoffs.space.0)),
+                        time: Lsn(cutoffs.time.0.max(original_cutoffs.time.0)),
+                    }
                 }
             }
 
@@ -8036,6 +8171,12 @@ mod tests {
             )
             .await?;
         {
+            tline
+                .latest_gc_cutoff_lsn
+                .lock_for_write()
+                .store_and_unlock(Lsn(0x30))
+                .wait()
+                .await;
             // Update GC info
             let mut guard = tline.gc_info.write().unwrap();
             guard.cutoffs.time = Lsn(0x30);
@@ -8138,6 +8279,12 @@ mod tests {
 
         // increase GC horizon and compact again
         {
+            tline
+                .latest_gc_cutoff_lsn
+                .lock_for_write()
+                .store_and_unlock(Lsn(0x40))
+                .wait()
+                .await;
             // Update GC info
             let mut guard = tline.gc_info.write().unwrap();
             guard.cutoffs.time = Lsn(0x40);
@@ -8518,6 +8665,12 @@ mod tests {
                 .await?
         };
         {
+            tline
+                .latest_gc_cutoff_lsn
+                .lock_for_write()
+                .store_and_unlock(Lsn(0x30))
+                .wait()
+                .await;
             // Update GC info
             let mut guard = tline.gc_info.write().unwrap();
             *guard = GcInfo {
@@ -8599,6 +8752,12 @@ mod tests {
 
         // increase GC horizon and compact again
         {
+            tline
+                .latest_gc_cutoff_lsn
+                .lock_for_write()
+                .store_and_unlock(Lsn(0x40))
+                .wait()
+                .await;
             // Update GC info
             let mut guard = tline.gc_info.write().unwrap();
             guard.cutoffs.time = Lsn(0x40);
@@ -9046,6 +9205,12 @@ mod tests {
             )
             .await?;
         {
+            tline
+                .latest_gc_cutoff_lsn
+                .lock_for_write()
+                .store_and_unlock(Lsn(0x30))
+                .wait()
+                .await;
             // Update GC info
             let mut guard = tline.gc_info.write().unwrap();
             *guard = GcInfo {
@@ -9163,6 +9328,7 @@ mod tests {
                 CompactOptions {
                     flags: dryrun_flags,
                     compact_range: None,
+                    ..Default::default()
                 },
                 &ctx,
             )
@@ -9187,6 +9353,12 @@ mod tests {
 
         // increase GC horizon and compact again
         {
+            tline
+                .latest_gc_cutoff_lsn
+                .lock_for_write()
+                .store_and_unlock(Lsn(0x38))
+                .wait()
+                .await;
             // Update GC info
             let mut guard = tline.gc_info.write().unwrap();
             guard.cutoffs.time = Lsn(0x38);
@@ -9282,6 +9454,12 @@ mod tests {
             )
             .await?;
         {
+            tline
+                .latest_gc_cutoff_lsn
+                .lock_for_write()
+                .store_and_unlock(Lsn(0x30))
+                .wait()
+                .await;
             // Update GC info
             let mut guard = tline.gc_info.write().unwrap();
             *guard = GcInfo {
@@ -9399,6 +9577,7 @@ mod tests {
                 CompactOptions {
                     flags: dryrun_flags,
                     compact_range: None,
+                    ..Default::default()
                 },
                 &ctx,
             )
@@ -9525,6 +9704,12 @@ mod tests {
         branch_tline.add_extra_test_dense_keyspace(KeySpace::single(get_key(0)..get_key(10)));
 
         {
+            parent_tline
+                .latest_gc_cutoff_lsn
+                .lock_for_write()
+                .store_and_unlock(Lsn(0x10))
+                .wait()
+                .await;
             // Update GC info
             let mut guard = parent_tline.gc_info.write().unwrap();
             *guard = GcInfo {
@@ -9539,6 +9724,12 @@ mod tests {
         }
 
         {
+            branch_tline
+                .latest_gc_cutoff_lsn
+                .lock_for_write()
+                .store_and_unlock(Lsn(0x50))
+                .wait()
+                .await;
             // Update GC info
             let mut guard = branch_tline.gc_info.write().unwrap();
             *guard = GcInfo {
@@ -9868,6 +10059,12 @@ mod tests {
             .await?;
 
         {
+            tline
+                .latest_gc_cutoff_lsn
+                .lock_for_write()
+                .store_and_unlock(Lsn(0x30))
+                .wait()
+                .await;
             // Update GC info
             let mut guard = tline.gc_info.write().unwrap();
             *guard = GcInfo {
@@ -9885,7 +10082,15 @@ mod tests {
 
         // Do a partial compaction on key range 0..2
         tline
-            .partial_compact_with_gc(get_key(0)..get_key(2), &cancel, EnumSet::new(), &ctx)
+            .compact_with_gc(
+                &cancel,
+                CompactOptions {
+                    flags: EnumSet::new(),
+                    compact_range: Some((get_key(0)..get_key(2)).into()),
+                    ..Default::default()
+                },
+                &ctx,
+            )
             .await
             .unwrap();
         let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
@@ -9924,7 +10129,15 @@ mod tests {
 
         // Do a partial compaction on key range 2..4
         tline
-            .partial_compact_with_gc(get_key(2)..get_key(4), &cancel, EnumSet::new(), &ctx)
+            .compact_with_gc(
+                &cancel,
+                CompactOptions {
+                    flags: EnumSet::new(),
+                    compact_range: Some((get_key(2)..get_key(4)).into()),
+                    ..Default::default()
+                },
+                &ctx,
+            )
             .await
             .unwrap();
         let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
@@ -9968,7 +10181,15 @@ mod tests {
 
         // Do a partial compaction on key range 4..9
         tline
-            .partial_compact_with_gc(get_key(4)..get_key(9), &cancel, EnumSet::new(), &ctx)
+            .compact_with_gc(
+                &cancel,
+                CompactOptions {
+                    flags: EnumSet::new(),
+                    compact_range: Some((get_key(4)..get_key(9)).into()),
+                    ..Default::default()
+                },
+                &ctx,
+            )
             .await
             .unwrap();
         let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
@@ -10011,7 +10232,15 @@ mod tests {
 
         // Do a partial compaction on key range 9..10
         tline
-            .partial_compact_with_gc(get_key(9)..get_key(10), &cancel, EnumSet::new(), &ctx)
+            .compact_with_gc(
+                &cancel,
+                CompactOptions {
+                    flags: EnumSet::new(),
+                    compact_range: Some((get_key(9)..get_key(10)).into()),
+                    ..Default::default()
+                },
+                &ctx,
+            )
             .await
             .unwrap();
         let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;
@@ -10059,7 +10288,15 @@ mod tests {
 
         // Do a partial compaction on key range 0..10, all image layers below LSN 20 can be replaced with new ones.
         tline
-            .partial_compact_with_gc(get_key(0)..get_key(10), &cancel, EnumSet::new(), &ctx)
+            .compact_with_gc(
+                &cancel,
+                CompactOptions {
+                    flags: EnumSet::new(),
+                    compact_range: Some((get_key(0)..get_key(10)).into()),
+                    ..Default::default()
+                },
+                &ctx,
+            )
             .await
             .unwrap();
         let all_layers = inspect_and_sort(&tline, Some(get_key(0)..get_key(10))).await;

pageserver/src/tenant/gc_block.rs

@@ -1,4 +1,4 @@
-use std::collections::HashMap;
+use std::{collections::HashMap, sync::Arc};
 
 use utils::id::TimelineId;
 
@@ -20,7 +20,7 @@ pub(crate) struct GcBlock {
     /// Do not add any more features taking and forbidding taking this lock. It should be
     /// `tokio::sync::Notify`, but that is rarely used. On the other side, [`GcBlock::insert`]
     /// synchronizes with gc attempts by locking and unlocking this mutex.
-    blocking: tokio::sync::Mutex<()>,
+    blocking: Arc<tokio::sync::Mutex<()>>,
 }
 
 impl GcBlock {
@@ -30,7 +30,7 @@ impl GcBlock {
     /// it's ending, or if not currently possible, a value describing the reasons why not.
     ///
     /// Cancellation safe.
-    pub(super) async fn start(&self) -> Result<Guard<'_>, BlockingReasons> {
+    pub(super) async fn start(&self) -> Result<Guard, BlockingReasons> {
         let reasons = {
             let g = self.reasons.lock().unwrap();
 
@@ -44,7 +44,7 @@ impl GcBlock {
             Err(reasons)
         } else {
             Ok(Guard {
-                _inner: self.blocking.lock().await,
+                _inner: self.blocking.clone().lock_owned().await,
             })
         }
     }
@@ -170,8 +170,8 @@ impl GcBlock {
     }
 }
 
-pub(super) struct Guard<'a> {
-    _inner: tokio::sync::MutexGuard<'a, ()>,
+pub(crate) struct Guard {
+    _inner: tokio::sync::OwnedMutexGuard<()>,
 }
 
 #[derive(Debug)]

pageserver/src/tenant/remote_timeline_client.rs

@@ -2192,6 +2192,9 @@ impl RemoteTimelineClient {
                     upload_queue.clean.1 = Some(task.task_id);
 
                     let lsn = upload_queue.clean.0.metadata.disk_consistent_lsn();
+                    self.metrics
+                        .projected_remote_consistent_lsn_gauge
+                        .set(lsn.0);
 
                     if self.generation.is_none() {
                         // Legacy mode: skip validating generation

pageserver/src/tenant/timeline.rs

@@ -53,7 +53,7 @@ use utils::{
     postgres_client::PostgresClientProtocol,
     sync::gate::{Gate, GateGuard},
 };
-use wal_decoder::serialized_batch::SerializedValueBatch;
+use wal_decoder::serialized_batch::{SerializedValueBatch, ValueMeta};
 
 use std::sync::atomic::Ordering as AtomicOrdering;
 use std::sync::{Arc, Mutex, RwLock, Weak};
@@ -768,7 +768,7 @@ pub enum GetLogicalSizePriority {
     Background,
 }
 
-#[derive(enumset::EnumSetType)]
+#[derive(Debug, enumset::EnumSetType)]
 pub(crate) enum CompactFlags {
     ForceRepartition,
     ForceImageLayerCreation,
@@ -777,6 +777,19 @@ pub(crate) enum CompactFlags {
     DryRun,
 }
 
+#[serde_with::serde_as]
+#[derive(Debug, Clone, serde::Deserialize)]
+pub(crate) struct CompactRequest {
+    pub compact_range: Option<CompactRange>,
+    pub compact_below_lsn: Option<Lsn>,
+    /// Whether the compaction job should be scheduled.
+    #[serde(default)]
+    pub scheduled: bool,
+    /// Whether the compaction job should be split across key ranges.
+    #[serde(default)]
+    pub sub_compaction: bool,
+}
+
 #[serde_with::serde_as]
 #[derive(Debug, Clone, serde::Deserialize)]
 pub(crate) struct CompactRange {
@@ -786,10 +799,27 @@ pub(crate) struct CompactRange {
     pub end: Key,
 }
 
-#[derive(Clone, Default)]
+impl From<Range<Key>> for CompactRange {
+    fn from(range: Range<Key>) -> Self {
+        CompactRange {
+            start: range.start,
+            end: range.end,
+        }
+    }
+}
+
+#[derive(Debug, Clone, Default)]
 pub(crate) struct CompactOptions {
     pub flags: EnumSet<CompactFlags>,
+    /// If set, the compaction will only compact the key range specified by this option.
+    /// This option is only used by GC compaction.
     pub compact_range: Option<CompactRange>,
+    /// If set, the compaction will only compact the LSN below this value.
+    /// This option is only used by GC compaction.
+    pub compact_below_lsn: Option<Lsn>,
+    /// Enable sub-compaction (split compaction job across key ranges).
+    /// This option is only used by GC compaction.
+    pub sub_compaction: bool,
 }
 
 impl std::fmt::Debug for Timeline {
@@ -1433,23 +1463,31 @@ impl Timeline {
         Ok(lease)
     }
 
-    /// Flush to disk all data that was written with the put_* functions
+    /// Freeze the current open in-memory layer. It will be written to disk on next iteration.
+    /// Returns the flush request ID which can be awaited with wait_flush_completion().
+    #[instrument(skip(self), fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug(), timeline_id=%self.timeline_id))]
+    pub(crate) async fn freeze(&self) -> Result<u64, FlushLayerError> {
+        self.freeze0().await
+    }
+
+    /// Freeze and flush the open in-memory layer, waiting for it to be written to disk.
     #[instrument(skip(self), fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug(), timeline_id=%self.timeline_id))]
     pub(crate) async fn freeze_and_flush(&self) -> Result<(), FlushLayerError> {
         self.freeze_and_flush0().await
     }
 
+    /// Freeze the current open in-memory layer. It will be written to disk on next iteration.
+    /// Returns the flush request ID which can be awaited with wait_flush_completion().
+    pub(crate) async fn freeze0(&self) -> Result<u64, FlushLayerError> {
+        let mut g = self.write_lock.lock().await;
+        let to_lsn = self.get_last_record_lsn();
+        self.freeze_inmem_layer_at(to_lsn, &mut g).await
+    }
+
     // This exists to provide a non-span creating version of `freeze_and_flush` we can call without
     // polluting the span hierarchy.
     pub(crate) async fn freeze_and_flush0(&self) -> Result<(), FlushLayerError> {
-        let token = {
-            // Freeze the current open in-memory layer. It will be written to disk on next
-            // iteration.
-            let mut g = self.write_lock.lock().await;
-
-            let to_lsn = self.get_last_record_lsn();
-            self.freeze_inmem_layer_at(to_lsn, &mut g).await?
-        };
+        let token = self.freeze0().await?;
         self.wait_flush_completion(token).await
     }
 
@@ -1604,6 +1642,8 @@ impl Timeline {
             CompactOptions {
                 flags,
                 compact_range: None,
+                compact_below_lsn: None,
+                sub_compaction: false,
             },
             ctx,
         )
@@ -2359,7 +2399,7 @@ impl Timeline {
 
             result
                 .metrics
-                .last_record_gauge
+                .last_record_lsn_gauge
                 .set(disk_consistent_lsn.0 as i64);
             result
         })
@@ -3481,7 +3521,7 @@ impl Timeline {
     pub(crate) fn finish_write(&self, new_lsn: Lsn) {
         assert!(new_lsn.is_aligned());
 
-        self.metrics.last_record_gauge.set(new_lsn.0 as i64);
+        self.metrics.last_record_lsn_gauge.set(new_lsn.0 as i64);
         self.last_record_lsn.advance(new_lsn);
     }
 
@@ -3849,6 +3889,10 @@ impl Timeline {
     fn set_disk_consistent_lsn(&self, new_value: Lsn) -> bool {
         let old_value = self.disk_consistent_lsn.fetch_max(new_value);
         assert!(new_value >= old_value, "disk_consistent_lsn must be growing monotonously at runtime; current {old_value}, offered {new_value}");
+
+        self.metrics
+            .disk_consistent_lsn_gauge
+            .set(new_value.0 as i64);
         new_value != old_value
     }
 
@@ -5887,6 +5931,23 @@ impl<'a> TimelineWriter<'a> {
             return Ok(());
         }
 
+        // In debug builds, assert that we don't write any keys that don't belong to this shard.
+        // We don't assert this in release builds, since key ownership policies may change over
+        // time. Stray keys will be removed during compaction.
+        if cfg!(debug_assertions) {
+            for metadata in &batch.metadata {
+                if let ValueMeta::Serialized(metadata) = metadata {
+                    let key = Key::from_compact(metadata.key);
+                    assert!(
+                        self.shard_identity.is_key_local(&key)
+                            || self.shard_identity.is_key_global(&key),
+                        "key {key} does not belong on shard {}",
+                        self.shard_identity.shard_index()
+                    );
+                }
+            }
+        }
+
         let batch_max_lsn = batch.max_lsn;
         let buf_size: u64 = batch.buffer_size() as u64;
 

pageserver/src/tenant/timeline/compaction.rs

@@ -10,13 +10,12 @@ use std::sync::Arc;
 
 use super::layer_manager::LayerManager;
 use super::{
-    CompactFlags, CompactOptions, CreateImageLayersError, DurationRecorder, ImageLayerCreationMode,
-    RecordedDuration, Timeline,
+    CompactFlags, CompactOptions, CompactRange, CreateImageLayersError, DurationRecorder,
+    ImageLayerCreationMode, RecordedDuration, Timeline,
 };
 
 use anyhow::{anyhow, bail, Context};
 use bytes::Bytes;
-use enumset::EnumSet;
 use fail::fail_point;
 use itertools::Itertools;
 use pageserver_api::key::KEY_SIZE;
@@ -30,7 +29,6 @@ use utils::id::TimelineId;
 use crate::context::{AccessStatsBehavior, RequestContext, RequestContextBuilder};
 use crate::page_cache;
 use crate::statvfs::Statvfs;
-use crate::tenant::checks::check_valid_layermap;
 use crate::tenant::remote_timeline_client::WaitCompletionError;
 use crate::tenant::storage_layer::batch_split_writer::{
     BatchWriterResult, SplitDeltaLayerWriter, SplitImageLayerWriter,
@@ -43,7 +41,7 @@ use crate::tenant::storage_layer::{
 use crate::tenant::timeline::ImageLayerCreationOutcome;
 use crate::tenant::timeline::{drop_rlock, DeltaLayerWriter, ImageLayerWriter};
 use crate::tenant::timeline::{Layer, ResidentLayer};
-use crate::tenant::{DeltaLayer, MaybeOffloaded};
+use crate::tenant::{gc_block, DeltaLayer, MaybeOffloaded};
 use crate::virtual_file::{MaybeFatalIo, VirtualFile};
 use pageserver_api::config::tenant_conf_defaults::{
     DEFAULT_CHECKPOINT_DISTANCE, DEFAULT_COMPACTION_THRESHOLD,
@@ -64,6 +62,15 @@ use super::CompactionError;
 /// Maximum number of deltas before generating an image layer in bottom-most compaction.
 const COMPACTION_DELTA_THRESHOLD: usize = 5;
 
+/// A scheduled compaction task.
+pub(crate) struct ScheduledCompactionTask {
+    pub options: CompactOptions,
+    /// The channel to send the compaction result. If this is a subcompaction, the last compaction job holds the sender.
+    pub result_tx: Option<tokio::sync::oneshot::Sender<()>>,
+    /// Hold the GC block. If this is a subcompaction, the last compaction job holds the gc block guard.
+    pub gc_block: Option<gc_block::Guard>,
+}
+
 pub struct GcCompactionJobDescription {
     /// All layers to read in the compaction job
     selected_layers: Vec<Layer>,
@@ -1174,11 +1181,12 @@ impl Timeline {
                     .await
                     .map_err(CompactionError::Other)?;
             } else {
-                debug!(
-                    "Dropping key {} during compaction (it belongs on shard {:?})",
-                    key,
-                    self.shard_identity.get_shard_number(&key)
-                );
+                let shard = self.shard_identity.shard_index();
+                let owner = self.shard_identity.get_shard_number(&key);
+                if cfg!(debug_assertions) {
+                    panic!("key {key} does not belong on shard {shard}, owned by {owner}");
+                }
+                debug!("dropping key {key} during compaction (it belongs on shard {owner})");
             }
 
             if !new_layers.is_empty() {
@@ -1746,22 +1754,113 @@ impl Timeline {
         Ok(())
     }
 
-    pub(crate) async fn compact_with_gc(
+    /// Split a gc-compaction job into multiple compaction jobs. Optimally, this function should return a vector of
+    /// `GcCompactionJobDesc`. But we want to keep it simple on the tenant scheduling side without exposing too much
+    /// ad-hoc information about gc compaction itself.
+    pub(crate) async fn gc_compaction_split_jobs(
         self: &Arc<Self>,
-        cancel: &CancellationToken,
         options: CompactOptions,
-        ctx: &RequestContext,
-    ) -> anyhow::Result<()> {
-        self.partial_compact_with_gc(
-            options
-                .compact_range
-                .map(|range| range.start..range.end)
-                .unwrap_or_else(|| Key::MIN..Key::MAX),
-            cancel,
-            options.flags,
-            ctx,
-        )
-        .await
+    ) -> anyhow::Result<Vec<CompactOptions>> {
+        if !options.sub_compaction {
+            return Ok(vec![options]);
+        }
+        let compact_range = options.compact_range.clone().unwrap_or(CompactRange {
+            start: Key::MIN,
+            end: Key::MAX,
+        });
+        let compact_below_lsn = if let Some(compact_below_lsn) = options.compact_below_lsn {
+            compact_below_lsn
+        } else {
+            *self.get_latest_gc_cutoff_lsn() // use the real gc cutoff
+        };
+        let mut compact_jobs = Vec::new();
+        // For now, we simply use the key partitioning information; we should do a more fine-grained partitioning
+        // by estimating the amount of files read for a compaction job. We should also partition on LSN.
+        let Ok(partition) = self.partitioning.try_lock() else {
+            bail!("failed to acquire partition lock");
+        };
+        let ((dense_ks, sparse_ks), _) = &*partition;
+        // Truncate the key range to be within user specified compaction range.
+        fn truncate_to(
+            source_start: &Key,
+            source_end: &Key,
+            target_start: &Key,
+            target_end: &Key,
+        ) -> Option<(Key, Key)> {
+            let start = source_start.max(target_start);
+            let end = source_end.min(target_end);
+            if start < end {
+                Some((*start, *end))
+            } else {
+                None
+            }
+        }
+        let mut split_key_ranges = Vec::new();
+        let ranges = dense_ks
+            .parts
+            .iter()
+            .map(|partition| partition.ranges.iter())
+            .chain(sparse_ks.parts.iter().map(|x| x.0.ranges.iter()))
+            .flatten()
+            .cloned()
+            .collect_vec();
+        for range in ranges.iter() {
+            let Some((start, end)) = truncate_to(
+                &range.start,
+                &range.end,
+                &compact_range.start,
+                &compact_range.end,
+            ) else {
+                continue;
+            };
+            split_key_ranges.push((start, end));
+        }
+        split_key_ranges.sort();
+        let guard = self.layers.read().await;
+        let layer_map = guard.layer_map()?;
+        let mut current_start = None;
+        // Split compaction job to about 2GB each
+        const GC_COMPACT_MAX_SIZE_MB: u64 = 4 * 1024; // 4GB, TODO: should be configuration in the future
+        let ranges_num = split_key_ranges.len();
+        for (idx, (start, end)) in split_key_ranges.into_iter().enumerate() {
+            if current_start.is_none() {
+                current_start = Some(start);
+            }
+            let start = current_start.unwrap();
+            if start >= end {
+                // We have already processed this partition.
+                continue;
+            }
+            let res = layer_map.range_search(start..end, compact_below_lsn);
+            let total_size = res.found.keys().map(|x| x.layer.file_size()).sum::<u64>();
+            if total_size > GC_COMPACT_MAX_SIZE_MB * 1024 * 1024 || ranges_num == idx + 1 {
+                let mut compact_options = options.clone();
+                // Try to extend the compaction range so that we include at least one full layer file.
+                let extended_end = res
+                    .found
+                    .keys()
+                    .map(|layer| layer.layer.key_range.end)
+                    .min();
+                // It is possible that the search range does not contain any layer files when we reach the end of the loop.
+                // In this case, we simply use the specified key range end.
+                let end = if let Some(extended_end) = extended_end {
+                    extended_end.max(end)
+                } else {
+                    end
+                };
+                info!(
+                    "splitting compaction job: {}..{}, estimated_size={}",
+                    start, end, total_size
+                );
+                compact_options.compact_range = Some(CompactRange { start, end });
+                compact_options.compact_below_lsn = Some(compact_below_lsn);
+                compact_options.sub_compaction = false;
+                compact_jobs.push(compact_options);
+                current_start = Some(end);
+            }
+        }
+        drop(guard);
+        Ok(compact_jobs)
     }
 
     /// An experimental compaction building block that combines compaction with garbage collection.
@@ -1771,19 +1870,51 @@ impl Timeline {
     /// layers and image layers, which generates image layers on the gc horizon, drop deltas below gc horizon,
     /// and create delta layers with all deltas >= gc horizon.
     ///
-    /// If `key_range` is provided, it will only compact the keys within the range, aka partial compaction.
+    /// If `options.compact_range` is provided, it will only compact the keys within the range, aka partial compaction.
     /// Partial compaction will read and process all layers overlapping with the key range, even if it might
     /// contain extra keys. After the gc-compaction phase completes, delta layers that are not fully contained
     /// within the key range will be rewritten to ensure they do not overlap with the delta layers. Providing
     /// Key::MIN..Key..MAX to the function indicates a full compaction, though technically, `Key::MAX` is not
     /// part of the range.
-    pub(crate) async fn partial_compact_with_gc(
+    ///
+    /// If `options.compact_below_lsn` is provided, the compaction will only compact layers below or intersect with
+    /// the LSN. Otherwise, it will use the gc cutoff by default.
+    pub(crate) async fn compact_with_gc(
         self: &Arc<Self>,
-        compaction_key_range: Range<Key>,
         cancel: &CancellationToken,
-        flags: EnumSet<CompactFlags>,
+        options: CompactOptions,
         ctx: &RequestContext,
     ) -> anyhow::Result<()> {
+        if options.sub_compaction {
+            info!("running enhanced gc bottom-most compaction with sub-compaction, splitting compaction jobs");
+            let jobs = self.gc_compaction_split_jobs(options).await?;
+            let jobs_len = jobs.len();
+            for (idx, job) in jobs.into_iter().enumerate() {
+                info!(
+                    "running enhanced gc bottom-most compaction, sub-compaction {}/{}",
+                    idx + 1,
+                    jobs_len
+                );
+                self.compact_with_gc_inner(cancel, job, ctx).await?;
+            }
+            if jobs_len == 0 {
+                info!("no jobs to run, skipping gc bottom-most compaction");
+            }
+            return Ok(());
+        }
+        self.compact_with_gc_inner(cancel, options, ctx).await
+    }
+
+    async fn compact_with_gc_inner(
+        self: &Arc<Self>,
+        cancel: &CancellationToken,
+        options: CompactOptions,
+        ctx: &RequestContext,
+    ) -> anyhow::Result<()> {
+        assert!(
+            !options.sub_compaction,
+            "sub-compaction should be handled by the outer function"
+        );
         // Block other compaction/GC tasks from running for now. GC-compaction could run along
         // with legacy compaction tasks in the future. Always ensure the lock order is compaction -> gc.
         // Note that we already acquired the compaction lock when the outer `compact` function gets called.
@@ -1803,6 +1934,12 @@ impl Timeline {
         )
         .await?;
 
+        let flags = options.flags;
+        let compaction_key_range = options
+            .compact_range
+            .map(|range| range.start..range.end)
+            .unwrap_or_else(|| Key::MIN..Key::MAX);
+
         let dry_run = flags.contains(CompactFlags::DryRun);
 
         if compaction_key_range == (Key::MIN..Key::MAX) {
@@ -1826,7 +1963,22 @@ impl Timeline {
             let layers = guard.layer_map()?;
             let gc_info = self.gc_info.read().unwrap();
             let mut retain_lsns_below_horizon = Vec::new();
-            let gc_cutoff = gc_info.cutoffs.select_min();
+            let gc_cutoff = {
+                // Currently, gc-compaction only kicks in after the legacy gc has updated the gc_cutoff.
+                // Therefore, it can only clean up data that cannot be cleaned up with legacy gc, instead of
+                // cleaning everything that theoritically it could. In the future, it should use `self.gc_info`
+                // to get the truth data.
+                let real_gc_cutoff = *self.get_latest_gc_cutoff_lsn();
+                // The compaction algorithm will keep all keys above the gc_cutoff while keeping only necessary keys below the gc_cutoff for
+                // each of the retain_lsn. Therefore, if the user-provided `compact_below_lsn` is larger than the real gc cutoff, we will use
+                // the real cutoff.
+                let mut gc_cutoff = options.compact_below_lsn.unwrap_or(real_gc_cutoff);
+                if gc_cutoff > real_gc_cutoff {
+                    warn!("provided compact_below_lsn={} is larger than the real_gc_cutoff={}, using the real gc cutoff", gc_cutoff, real_gc_cutoff);
+                    gc_cutoff = real_gc_cutoff;
+                }
+                gc_cutoff
+            };
             for (lsn, _timeline_id, _is_offloaded) in &gc_info.retain_lsns {
                 if lsn < &gc_cutoff {
                     retain_lsns_below_horizon.push(*lsn);
@@ -1846,7 +1998,7 @@ impl Timeline {
                 .map(|desc| desc.get_lsn_range().end)
                 .max()
             else {
-                info!("no layers to compact with gc");
+                info!("no layers to compact with gc: no historic layers below gc_cutoff, gc_cutoff={}", gc_cutoff);
                 return Ok(());
             };
             // Then, pick all the layers that are below the max_layer_lsn. This is to ensure we can pick all single-key
@@ -1869,7 +2021,7 @@ impl Timeline {
                 }
             }
             if selected_layers.is_empty() {
-                info!("no layers to compact with gc");
+                info!("no layers to compact with gc: no layers within the key range, gc_cutoff={}, key_range={}..{}", gc_cutoff, compaction_key_range.start, compaction_key_range.end);
                 return Ok(());
             }
             retain_lsns_below_horizon.sort();
@@ -1936,14 +2088,15 @@ impl Timeline {
 
         // Step 1: construct a k-merge iterator over all layers.
         // Also, verify if the layer map can be split by drawing a horizontal line at every LSN start/end split point.
-        let layer_names = job_desc
-            .selected_layers
-            .iter()
-            .map(|layer| layer.layer_desc().layer_name())
-            .collect_vec();
-        if let Some(err) = check_valid_layermap(&layer_names) {
-            warn!("gc-compaction layer map check failed because {}, this is normal if partial compaction is not finished yet", err);
-        }
+        // disable the check for now because we need to adjust the check for partial compactions, will enable later.
+        // let layer_names = job_desc
+        //     .selected_layers
+        //     .iter()
+        //     .map(|layer| layer.layer_desc().layer_name())
+        //     .collect_vec();
+        // if let Some(err) = check_valid_layermap(&layer_names) {
+        //     warn!("gc-compaction layer map check failed because {}, this is normal if partial compaction is not finished yet", err);
+        // }
         // The maximum LSN we are processing in this compaction loop
         let end_lsn = job_desc
             .selected_layers
@@ -2048,6 +2201,11 @@ impl Timeline {
                 // This is not handled in the filter iterator because shard is determined by hash.
                 // Therefore, it does not give us any performance benefit to do things like skip
                 // a whole layer file as handling key spaces (ranges).
+                if cfg!(debug_assertions) {
+                    let shard = self.shard_identity.shard_index();
+                    let owner = self.shard_identity.get_shard_number(&key);
+                    panic!("key {key} does not belong on shard {shard}, owned by {owner}");
+                }
                 continue;
             }
             if !job_desc.compaction_key_range.contains(&key) {

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

@@ -369,6 +369,13 @@ pub(super) async fn handle_walreceiver_connection(
                 // advances it to its end LSN. 0 is just an initialization placeholder.
                 let mut modification = timeline.begin_modification(Lsn(0));
 
+                if !records.is_empty() {
+                    timeline
+                        .metrics
+                        .wal_records_received
+                        .inc_by(records.len() as u64);
+                }
+
                 for interpreted in records {
                     if matches!(interpreted.flush_uncommitted, FlushUncommittedRecords::Yes)
                         && uncommitted_records > 0
@@ -510,6 +517,7 @@ pub(super) async fn handle_walreceiver_connection(
                         }
 
                         // Ingest the records without immediately committing them.
+                        timeline.metrics.wal_records_received.inc();
                         let ingested = walingest
                             .ingest_record(interpreted, &mut modification, &ctx)
                             .await

pageserver/src/walingest.rs

@@ -582,18 +582,21 @@ impl WalIngest {
                 forknum: FSM_FORKNUM,
             };
 
+            // Zero out the last remaining FSM page, if this shard owns it. We are not precise here,
+            // and instead of digging in the FSM bitmap format we just clear the whole page.
             let fsm_logical_page_no = blkno / pg_constants::SLOTS_PER_FSM_PAGE;
             let mut fsm_physical_page_no = fsm_logical_to_physical(fsm_logical_page_no);
-            if blkno % pg_constants::SLOTS_PER_FSM_PAGE != 0 {
-                // Tail of last remaining FSM page has to be zeroed.
-                // We are not precise here and instead of digging in FSM bitmap format just clear the whole page.
+            if blkno % pg_constants::SLOTS_PER_FSM_PAGE != 0
+                && self
+                    .shard
+                    .is_key_local(&rel_block_to_key(rel, fsm_physical_page_no))
+            {
                 modification.put_rel_page_image_zero(rel, fsm_physical_page_no)?;
                 fsm_physical_page_no += 1;
             }
-            // TODO: re-examine the None case here wrt. sharding; should we error?
+            // Truncate this shard's view of the FSM relation size, if it even has one.
             let nblocks = get_relsize(modification, rel, ctx).await?.unwrap_or(0);
             if nblocks > fsm_physical_page_no {
-                // check if something to do: FSM is larger than truncate position
                 self.put_rel_truncation(modification, rel, fsm_physical_page_no, ctx)
                     .await?;
             }
@@ -617,7 +620,7 @@ impl WalIngest {
             // tail bits in the last remaining map page, representing truncated heap
             // blocks, need to be cleared. This is not only tidy, but also necessary
             // because we don't get a chance to clear the bits if the heap is extended
-            // again.
+            // again. Only do this on the shard that owns the page.
             if (trunc_byte != 0 || trunc_offs != 0)
                 && self.shard.is_key_local(&rel_block_to_key(rel, vm_page_no))
             {
@@ -631,10 +634,9 @@ impl WalIngest {
                 )?;
                 vm_page_no += 1;
             }
-            // TODO: re-examine the None case here wrt. sharding; should we error?
+            // Truncate this shard's view of the VM relation size, if it even has one.
             let nblocks = get_relsize(modification, rel, ctx).await?.unwrap_or(0);
             if nblocks > vm_page_no {
-                // check if something to do: VM is larger than truncate position
                 self.put_rel_truncation(modification, rel, vm_page_no, ctx)
                     .await?;
             }

pgxn/neon/libpagestore.c

@@ -22,6 +22,7 @@
 #include "libpq/pqformat.h"
 #include "miscadmin.h"
 #include "pgstat.h"
+#include "portability/instr_time.h"
 #include "postmaster/interrupt.h"
 #include "storage/buf_internals.h"
 #include "storage/ipc.h"
@@ -118,6 +119,11 @@ typedef struct
 	 */
 	PSConnectionState state;
 	PGconn		   *conn;
+
+	/* request / response counters for debugging */
+	uint64			nrequests_sent;
+	uint64			nresponses_received;
+
 	/*---
 	 * WaitEventSet containing:
 	 *	- WL_SOCKET_READABLE on 'conn'
@@ -628,6 +634,8 @@ pageserver_connect(shardno_t shard_no, int elevel)
 		}
 
 		shard->state = PS_Connected;
+		shard->nrequests_sent = 0;
+		shard->nresponses_received = 0;
 	}
 	/* FALLTHROUGH */
 	case PS_Connected:
@@ -656,6 +664,27 @@ call_PQgetCopyData(shardno_t shard_no, char **buffer)
 	int			ret;
 	PageServer *shard = &page_servers[shard_no];
 	PGconn	   *pageserver_conn = shard->conn;
+	instr_time	now,
+				start_ts,
+				since_start,
+				last_log_ts,
+				since_last_log;
+	bool		logged = false;
+
+	/*
+	 * As a debugging aid, if we don't get a response for a long time, print a
+	 * log message.
+	 *
+	 * 10 s is a very generous threshold, normally we expect a response in a
+	 * few milliseconds. We have metrics to track latencies in normal ranges,
+	 * but in the cases that take exceptionally long, it's useful to log the
+	 * exact timestamps.
+	 */
+#define LOG_INTERVAL_US		UINT64CONST(10 * 1000000)
+
+	INSTR_TIME_SET_CURRENT(now);
+	start_ts = last_log_ts = now;
+	INSTR_TIME_SET_ZERO(since_last_log);
 
 retry:
 	ret = PQgetCopyData(pageserver_conn, buffer, 1 /* async */ );
@@ -663,9 +692,12 @@ retry:
 	if (ret == 0)
 	{
 		WaitEvent	event;
+		long		timeout;
+
+		timeout = Min(0, LOG_INTERVAL_US - INSTR_TIME_GET_MICROSEC(since_last_log));
 
 		/* Sleep until there's something to do */
-		(void) WaitEventSetWait(shard->wes_read, -1L, &event, 1,
+		(void) WaitEventSetWait(shard->wes_read, timeout, &event, 1,
 								WAIT_EVENT_NEON_PS_READ);
 		ResetLatch(MyLatch);
 
@@ -684,9 +716,40 @@ retry:
 			}
 		}
 
+		/*
+		 * Print a message to the log if a long time has passed with no
+		 * response.
+		 */
+		INSTR_TIME_SET_CURRENT(now);
+		since_last_log = now;
+		INSTR_TIME_SUBTRACT(since_last_log, last_log_ts);
+		if (INSTR_TIME_GET_MICROSEC(since_last_log) >= LOG_INTERVAL_US)
+		{
+			since_start = now;
+			INSTR_TIME_SUBTRACT(since_start, start_ts);
+			neon_shard_log(shard_no, LOG, "no response received from pageserver for %0.3f s, still waiting (sent " UINT64_FORMAT " requests, received " UINT64_FORMAT " responses)",
+						   INSTR_TIME_GET_DOUBLE(since_start),
+						   shard->nrequests_sent, shard->nresponses_received);
+			last_log_ts = now;
+			logged = true;
+		}
+
 		goto retry;
 	}
 
+	/*
+	 * If we logged earlier that the response is taking a long time, log
+	 * another message when the response is finally received.
+	 */
+	if (logged)
+	{
+		INSTR_TIME_SET_CURRENT(now);
+		since_start = now;
+		INSTR_TIME_SUBTRACT(since_start, start_ts);
+		neon_shard_log(shard_no, LOG, "received response from pageserver after %0.3f s",
+					   INSTR_TIME_GET_DOUBLE(since_start));
+	}
+
 	return ret;
 }
 
@@ -786,6 +849,7 @@ pageserver_send(shardno_t shard_no, NeonRequest *request)
 	 * PGRES_POLLING_WRITING state. It's kinda dirty to disconnect at this
 	 * point, but on the grand scheme of things it's only a small issue.
 	 */
+	shard->nrequests_sent++;
 	if (PQputCopyData(pageserver_conn, req_buff.data, req_buff.len) <= 0)
 	{
 		char	   *msg = pchomp(PQerrorMessage(pageserver_conn));
@@ -878,6 +942,7 @@ pageserver_receive(shardno_t shard_no)
 		neon_shard_log(shard_no, ERROR, "pageserver_receive disconnect: unexpected PQgetCopyData return value: %d", rc);
 	}
 
+	shard->nresponses_received++;
 	return (NeonResponse *) resp;
 }
 

pgxn/neon/pagestore_smgr.c

@@ -423,7 +423,11 @@ readahead_buffer_resize(int newsize, void *extra)
 	 * ensuring we have received all but the last n requests (n = newsize).
 	 */
 	if (MyPState->n_requests_inflight > newsize)
-		prefetch_wait_for(MyPState->ring_unused - newsize);
+	{
+		Assert(MyPState->ring_unused >= MyPState->n_requests_inflight - newsize);
+		prefetch_wait_for(MyPState->ring_unused - (MyPState->n_requests_inflight - newsize));
+		Assert(MyPState->n_requests_inflight <= newsize);
+	}
 
 	/* construct the new PrefetchState, and copy over the memory contexts */
 	newPState = MemoryContextAllocZero(TopMemoryContext, newprfs_size);
@@ -438,7 +442,6 @@ readahead_buffer_resize(int newsize, void *extra)
 	newPState->ring_last = newsize;
 	newPState->ring_unused = newsize;
 	newPState->ring_receive = newsize;
-	newPState->ring_flush = newsize;
 	newPState->max_shard_no = MyPState->max_shard_no;
 	memcpy(newPState->shard_bitmap, MyPState->shard_bitmap, sizeof(MyPState->shard_bitmap));
 
@@ -489,6 +492,7 @@ readahead_buffer_resize(int newsize, void *extra)
 		}
 		newPState->n_unused -= 1;
 	}
+	newPState->ring_flush = newPState->ring_receive;
 
 	MyNeonCounters->getpage_prefetches_buffered =
 		MyPState->n_responses_buffered;
@@ -498,6 +502,7 @@ readahead_buffer_resize(int newsize, void *extra)
 	for (; end >= MyPState->ring_last && end != UINT64_MAX; end -= 1)
 	{
 		PrefetchRequest *slot = GetPrfSlot(end);
+		Assert(slot->status != PRFS_REQUESTED);
 		if (slot->status == PRFS_RECEIVED)
 		{
 			pfree(slot->response);
@@ -610,6 +615,9 @@ prefetch_read(PrefetchRequest *slot)
 {
 	NeonResponse *response;
 	MemoryContext old;
+	BufferTag	buftag;
+	shardno_t	shard_no;
+	uint64		my_ring_index;
 
 	Assert(slot->status == PRFS_REQUESTED);
 	Assert(slot->response == NULL);
@@ -623,11 +631,29 @@ prefetch_read(PrefetchRequest *slot)
 					   slot->status, slot->response,
 					   (long)slot->my_ring_index, (long)MyPState->ring_receive);
 
+	/*
+	 * Copy the request info so that if an error happens and the prefetch
+	 * queue is flushed during the receive call, we can print the original
+	 * values in the error message
+	 */
+	buftag = slot->buftag;
+	shard_no = slot->shard_no;
+	my_ring_index = slot->my_ring_index;
+
 	old = MemoryContextSwitchTo(MyPState->errctx);
-	response = (NeonResponse *) page_server->receive(slot->shard_no);
+	response = (NeonResponse *) page_server->receive(shard_no);
 	MemoryContextSwitchTo(old);
 	if (response)
 	{
+		/* The slot should still be valid */
+		if (slot->status != PRFS_REQUESTED ||
+			slot->response != NULL ||
+			slot->my_ring_index != MyPState->ring_receive)
+			neon_shard_log(shard_no, ERROR,
+						   "Incorrect prefetch slot state after receive: status=%d response=%p my=%lu receive=%lu",
+						   slot->status, slot->response,
+						   (long) slot->my_ring_index, (long) MyPState->ring_receive);
+
 		/* update prefetch state */
 		MyPState->n_responses_buffered += 1;
 		MyPState->n_requests_inflight -= 1;
@@ -642,11 +668,15 @@ prefetch_read(PrefetchRequest *slot)
 	}
 	else
 	{
-		neon_shard_log(slot->shard_no, LOG,
+		/*
+		 * Note: The slot might no longer be valid, if the connection was lost
+		 * and the prefetch queue was flushed during the receive call
+		 */
+		neon_shard_log(shard_no, LOG,
 					   "No response from reading prefetch entry %lu: %u/%u/%u.%u block %u. This can be caused by a concurrent disconnect",
-					   (long)slot->my_ring_index,
-					   RelFileInfoFmt(BufTagGetNRelFileInfo(slot->buftag)),
-					   slot->buftag.forkNum, slot->buftag.blockNum);
+					   (long) my_ring_index,
+					   RelFileInfoFmt(BufTagGetNRelFileInfo(buftag)),
+					   buftag.forkNum, buftag.blockNum);
 		return false;
 	}
 }

proxy/src/cancellation.rs

@@ -115,7 +115,8 @@ impl<P: CancellationPublisher> CancellationHandler<P> {
                 IpAddr::V6(ip) => IpNet::V6(Ipv6Net::new_assert(ip, 64).trunc()),
             };
             if !self.limiter.lock().unwrap().check(subnet_key, 1) {
-                tracing::debug!("Rate limit exceeded. Skipping cancellation message");
+                // log only the subnet part of the IP address to know which subnet is rate limited
+                tracing::warn!("Rate limit exceeded. Skipping cancellation message, {subnet_key}");
                 Metrics::get()
                     .proxy
                     .cancellation_requests_total

proxy/src/console_redirect_proxy.rs

@@ -163,32 +163,36 @@ pub(crate) async fn handle_client<S: AsyncRead + AsyncWrite + Unpin>(
     let pause = ctx.latency_timer_pause(crate::metrics::Waiting::Client);
     let do_handshake = handshake(ctx, stream, tls, record_handshake_error);
 
-    let (mut stream, params) =
-        match tokio::time::timeout(config.handshake_timeout, do_handshake).await?? {
-            HandshakeData::Startup(stream, params) => (stream, params),
-            HandshakeData::Cancel(cancel_key_data) => {
-                // spawn a task to cancel the session, but don't wait for it
-                cancellations.spawn({
-                    let cancellation_handler_clone = Arc::clone(&cancellation_handler);
-                    let session_id = ctx.session_id();
-                    let peer_ip = ctx.peer_addr();
-                    async move {
-                        drop(
-                            cancellation_handler_clone
-                                .cancel_session(
-                                    cancel_key_data,
-                                    session_id,
-                                    peer_ip,
-                                    config.authentication_config.ip_allowlist_check_enabled,
-                                )
-                                .await,
-                        );
-                    }
-                });
+    let (mut stream, params) = match tokio::time::timeout(config.handshake_timeout, do_handshake)
+        .await??
+    {
+        HandshakeData::Startup(stream, params) => (stream, params),
+        HandshakeData::Cancel(cancel_key_data) => {
+            // spawn a task to cancel the session, but don't wait for it
+            cancellations.spawn({
+                let cancellation_handler_clone = Arc::clone(&cancellation_handler);
+                let session_id = ctx.session_id();
+                let peer_ip = ctx.peer_addr();
+                let cancel_span = tracing::span!(parent: None, tracing::Level::INFO, "cancel_session", session_id = ?session_id);
+                cancel_span.follows_from(tracing::Span::current());
+                async move {
+                    drop(
+                        cancellation_handler_clone
+                            .cancel_session(
+                                cancel_key_data,
+                                session_id,
+                                peer_ip,
+                                config.authentication_config.ip_allowlist_check_enabled,
+                            )
+                            .instrument(cancel_span)
+                            .await,
+                    );
+                }
+            });
 
-                return Ok(None);
-            }
-        };
+            return Ok(None);
+        }
+    };
     drop(pause);
 
     ctx.set_db_options(params.clone());

proxy/src/proxy/mod.rs

@@ -272,32 +272,36 @@ pub(crate) async fn handle_client<S: AsyncRead + AsyncWrite + Unpin>(
     let pause = ctx.latency_timer_pause(crate::metrics::Waiting::Client);
     let do_handshake = handshake(ctx, stream, mode.handshake_tls(tls), record_handshake_error);
 
-    let (mut stream, params) =
-        match tokio::time::timeout(config.handshake_timeout, do_handshake).await?? {
-            HandshakeData::Startup(stream, params) => (stream, params),
-            HandshakeData::Cancel(cancel_key_data) => {
-                // spawn a task to cancel the session, but don't wait for it
-                cancellations.spawn({
-                    let cancellation_handler_clone = Arc::clone(&cancellation_handler);
-                    let session_id = ctx.session_id();
-                    let peer_ip = ctx.peer_addr();
-                    async move {
-                        drop(
-                            cancellation_handler_clone
-                                .cancel_session(
-                                    cancel_key_data,
-                                    session_id,
-                                    peer_ip,
-                                    config.authentication_config.ip_allowlist_check_enabled,
-                                )
-                                .await,
-                        );
-                    }
-                });
+    let (mut stream, params) = match tokio::time::timeout(config.handshake_timeout, do_handshake)
+        .await??
+    {
+        HandshakeData::Startup(stream, params) => (stream, params),
+        HandshakeData::Cancel(cancel_key_data) => {
+            // spawn a task to cancel the session, but don't wait for it
+            cancellations.spawn({
+                let cancellation_handler_clone = Arc::clone(&cancellation_handler);
+                let session_id = ctx.session_id();
+                let peer_ip = ctx.peer_addr();
+                let cancel_span = tracing::span!(parent: None, tracing::Level::INFO, "cancel_session", session_id = ?session_id);
+                cancel_span.follows_from(tracing::Span::current());
+                async move {
+                    drop(
+                        cancellation_handler_clone
+                            .cancel_session(
+                                cancel_key_data,
+                                session_id,
+                                peer_ip,
+                                config.authentication_config.ip_allowlist_check_enabled,
+                            )
+                            .instrument(cancel_span)
+                            .await,
+                    );
+                }
+            });
 
-                return Ok(None);
-            }
-        };
+            return Ok(None);
+        }
+    };
     drop(pause);
 
     ctx.set_db_options(params.clone());

proxy/src/redis/notifications.rs

@@ -13,6 +13,7 @@ use crate::cache::project_info::ProjectInfoCache;
 use crate::cancellation::{CancelMap, CancellationHandler};
 use crate::intern::{ProjectIdInt, RoleNameInt};
 use crate::metrics::{Metrics, RedisErrors, RedisEventsCount};
+use tracing::Instrument;
 
 const CPLANE_CHANNEL_NAME: &str = "neondb-proxy-ws-updates";
 pub(crate) const PROXY_CHANNEL_NAME: &str = "neondb-proxy-to-proxy-updates";
@@ -143,6 +144,8 @@ impl<C: ProjectInfoCache + Send + Sync + 'static> MessageHandler<C> {
                 let peer_addr = cancel_session
                     .peer_addr
                     .unwrap_or(std::net::IpAddr::V4(std::net::Ipv4Addr::UNSPECIFIED));
+                let cancel_span = tracing::span!(parent: None, tracing::Level::INFO, "cancel_session", session_id = ?cancel_session.session_id);
+                cancel_span.follows_from(tracing::Span::current());
                 // This instance of cancellation_handler doesn't have a RedisPublisherClient so it can't publish the message.
                 match self
                     .cancellation_handler
@@ -152,6 +155,7 @@ impl<C: ProjectInfoCache + Send + Sync + 'static> MessageHandler<C> {
                         peer_addr,
                         cancel_session.peer_addr.is_some(),
                     )
+                    .instrument(cancel_span)
                     .await
                 {
                     Ok(()) => {}

safekeeper/benches/benchutils.rs

@@ -83,14 +83,20 @@ impl Env {
         node_id: NodeId,
         ttid: TenantTimelineId,
     ) -> anyhow::Result<Arc<Timeline>> {
-        let conf = self.make_conf(node_id);
+        let conf = Arc::new(self.make_conf(node_id));
         let timeline_dir = get_timeline_dir(&conf, &ttid);
         let remote_path = remote_timeline_path(&ttid)?;
 
         let safekeeper = self.make_safekeeper(node_id, ttid).await?;
         let shared_state = SharedState::new(StateSK::Loaded(safekeeper));
 
-        let timeline = Timeline::new(ttid, &timeline_dir, &remote_path, shared_state);
+        let timeline = Timeline::new(
+            ttid,
+            &timeline_dir,
+            &remote_path,
+            shared_state,
+            conf.clone(),
+        );
         timeline.bootstrap(
             &mut timeline.write_shared_state().await,
             &conf,

safekeeper/src/bin/safekeeper.rs

@@ -338,7 +338,7 @@ async fn main() -> anyhow::Result<()> {
         }
     };
 
-    let conf = SafeKeeperConf {
+    let conf = Arc::new(SafeKeeperConf {
         workdir,
         my_id: id,
         listen_pg_addr: args.listen_pg,
@@ -368,7 +368,7 @@ async fn main() -> anyhow::Result<()> {
         control_file_save_interval: args.control_file_save_interval,
         partial_backup_concurrency: args.partial_backup_concurrency,
         eviction_min_resident: args.eviction_min_resident,
-    };
+    });
 
     // initialize sentry if SENTRY_DSN is provided
     let _sentry_guard = init_sentry(
@@ -382,7 +382,7 @@ async fn main() -> anyhow::Result<()> {
 /// complete, e.g. panicked, inner is error produced by task itself.
 type JoinTaskRes = Result<anyhow::Result<()>, JoinError>;
 
-async fn start_safekeeper(conf: SafeKeeperConf) -> Result<()> {
+async fn start_safekeeper(conf: Arc<SafeKeeperConf>) -> Result<()> {
     // fsync the datadir to make sure we have a consistent state on disk.
     if !conf.no_sync {
         let dfd = File::open(&conf.workdir).context("open datadir for syncfs")?;
@@ -428,9 +428,11 @@ async fn start_safekeeper(conf: SafeKeeperConf) -> Result<()> {
         e
     })?;
 
+    let global_timelines = Arc::new(GlobalTimelines::new(conf.clone()));
+
     // Register metrics collector for active timelines. It's important to do this
     // after daemonizing, otherwise process collector will be upset.
-    let timeline_collector = safekeeper::metrics::TimelineCollector::new();
+    let timeline_collector = safekeeper::metrics::TimelineCollector::new(global_timelines.clone());
     metrics::register_internal(Box::new(timeline_collector))?;
 
     wal_backup::init_remote_storage(&conf).await;
@@ -447,9 +449,8 @@ async fn start_safekeeper(conf: SafeKeeperConf) -> Result<()> {
         .then(|| Handle::try_current().expect("no runtime in main"));
 
     // Load all timelines from disk to memory.
-    GlobalTimelines::init(conf.clone()).await?;
+    global_timelines.init().await?;
 
-    let conf_ = conf.clone();
     // Run everything in current thread rt, if asked.
     if conf.current_thread_runtime {
         info!("running in current thread runtime");
@@ -459,14 +460,16 @@ async fn start_safekeeper(conf: SafeKeeperConf) -> Result<()> {
         .as_ref()
         .unwrap_or_else(|| WAL_SERVICE_RUNTIME.handle())
         .spawn(wal_service::task_main(
-            conf_,
+            conf.clone(),
             pg_listener,
             Scope::SafekeeperData,
+            global_timelines.clone(),
         ))
         // wrap with task name for error reporting
         .map(|res| ("WAL service main".to_owned(), res));
     tasks_handles.push(Box::pin(wal_service_handle));
 
+    let global_timelines_ = global_timelines.clone();
     let timeline_housekeeping_handle = current_thread_rt
         .as_ref()
         .unwrap_or_else(|| WAL_SERVICE_RUNTIME.handle())
@@ -474,40 +477,45 @@ async fn start_safekeeper(conf: SafeKeeperConf) -> Result<()> {
             const TOMBSTONE_TTL: Duration = Duration::from_secs(3600 * 24);
             loop {
                 tokio::time::sleep(TOMBSTONE_TTL).await;
-                GlobalTimelines::housekeeping(&TOMBSTONE_TTL);
+                global_timelines_.housekeeping(&TOMBSTONE_TTL);
             }
         })
         .map(|res| ("Timeline map housekeeping".to_owned(), res));
     tasks_handles.push(Box::pin(timeline_housekeeping_handle));
 
     if let Some(pg_listener_tenant_only) = pg_listener_tenant_only {
-        let conf_ = conf.clone();
         let wal_service_handle = current_thread_rt
             .as_ref()
             .unwrap_or_else(|| WAL_SERVICE_RUNTIME.handle())
             .spawn(wal_service::task_main(
-                conf_,
+                conf.clone(),
                 pg_listener_tenant_only,
                 Scope::Tenant,
+                global_timelines.clone(),
             ))
             // wrap with task name for error reporting
             .map(|res| ("WAL service tenant only main".to_owned(), res));
         tasks_handles.push(Box::pin(wal_service_handle));
     }
 
-    let conf_ = conf.clone();
     let http_handle = current_thread_rt
         .as_ref()
         .unwrap_or_else(|| HTTP_RUNTIME.handle())
-        .spawn(http::task_main(conf_, http_listener))
+        .spawn(http::task_main(
+            conf.clone(),
+            http_listener,
+            global_timelines.clone(),
+        ))
         .map(|res| ("HTTP service main".to_owned(), res));
     tasks_handles.push(Box::pin(http_handle));
 
-    let conf_ = conf.clone();
     let broker_task_handle = current_thread_rt
         .as_ref()
         .unwrap_or_else(|| BROKER_RUNTIME.handle())
-        .spawn(broker::task_main(conf_).instrument(info_span!("broker")))
+        .spawn(
+            broker::task_main(conf.clone(), global_timelines.clone())
+                .instrument(info_span!("broker")),
+        )
         .map(|res| ("broker main".to_owned(), res));
     tasks_handles.push(Box::pin(broker_task_handle));
 

safekeeper/src/broker.rs

@@ -39,14 +39,17 @@ const RETRY_INTERVAL_MSEC: u64 = 1000;
 const PUSH_INTERVAL_MSEC: u64 = 1000;
 
 /// Push once in a while data about all active timelines to the broker.
-async fn push_loop(conf: SafeKeeperConf) -> anyhow::Result<()> {
+async fn push_loop(
+    conf: Arc<SafeKeeperConf>,
+    global_timelines: Arc<GlobalTimelines>,
+) -> anyhow::Result<()> {
     if conf.disable_periodic_broker_push {
         info!("broker push_loop is disabled, doing nothing...");
         futures::future::pending::<()>().await; // sleep forever
         return Ok(());
     }
 
-    let active_timelines_set = GlobalTimelines::get_global_broker_active_set();
+    let active_timelines_set = global_timelines.get_global_broker_active_set();
 
     let mut client =
         storage_broker::connect(conf.broker_endpoint.clone(), conf.broker_keepalive_interval)?;
@@ -87,8 +90,13 @@ async fn push_loop(conf: SafeKeeperConf) -> anyhow::Result<()> {
 
 /// Subscribe and fetch all the interesting data from the broker.
 #[instrument(name = "broker_pull", skip_all)]
-async fn pull_loop(conf: SafeKeeperConf, stats: Arc<BrokerStats>) -> Result<()> {
-    let mut client = storage_broker::connect(conf.broker_endpoint, conf.broker_keepalive_interval)?;
+async fn pull_loop(
+    conf: Arc<SafeKeeperConf>,
+    global_timelines: Arc<GlobalTimelines>,
+    stats: Arc<BrokerStats>,
+) -> Result<()> {
+    let mut client =
+        storage_broker::connect(conf.broker_endpoint.clone(), conf.broker_keepalive_interval)?;
 
     // TODO: subscribe only to local timelines instead of all
     let request = SubscribeSafekeeperInfoRequest {
@@ -113,7 +121,7 @@ async fn pull_loop(conf: SafeKeeperConf, stats: Arc<BrokerStats>) -> Result<()>
             .as_ref()
             .ok_or_else(|| anyhow!("missing tenant_timeline_id"))?;
         let ttid = parse_proto_ttid(proto_ttid)?;
-        if let Ok(tli) = GlobalTimelines::get(ttid) {
+        if let Ok(tli) = global_timelines.get(ttid) {
             // Note that we also receive *our own* info. That's
             // important, as it is used as an indication of live
             // connection to the broker.
@@ -135,7 +143,11 @@ async fn pull_loop(conf: SafeKeeperConf, stats: Arc<BrokerStats>) -> Result<()>
 
 /// Process incoming discover requests. This is done in a separate task to avoid
 /// interfering with the normal pull/push loops.
-async fn discover_loop(conf: SafeKeeperConf, stats: Arc<BrokerStats>) -> Result<()> {
+async fn discover_loop(
+    conf: Arc<SafeKeeperConf>,
+    global_timelines: Arc<GlobalTimelines>,
+    stats: Arc<BrokerStats>,
+) -> Result<()> {
     let mut client =
         storage_broker::connect(conf.broker_endpoint.clone(), conf.broker_keepalive_interval)?;
 
@@ -171,7 +183,7 @@ async fn discover_loop(conf: SafeKeeperConf, stats: Arc<BrokerStats>) -> Result<
                     .as_ref()
                     .ok_or_else(|| anyhow!("missing tenant_timeline_id"))?;
                 let ttid = parse_proto_ttid(proto_ttid)?;
-                if let Ok(tli) = GlobalTimelines::get(ttid) {
+                if let Ok(tli) = global_timelines.get(ttid) {
                     // we received a discovery request for a timeline we know about
                     discover_counter.inc();
 
@@ -210,7 +222,10 @@ async fn discover_loop(conf: SafeKeeperConf, stats: Arc<BrokerStats>) -> Result<
     bail!("end of stream");
 }
 
-pub async fn task_main(conf: SafeKeeperConf) -> anyhow::Result<()> {
+pub async fn task_main(
+    conf: Arc<SafeKeeperConf>,
+    global_timelines: Arc<GlobalTimelines>,
+) -> anyhow::Result<()> {
     info!("started, broker endpoint {:?}", conf.broker_endpoint);
 
     let mut ticker = tokio::time::interval(Duration::from_millis(RETRY_INTERVAL_MSEC));
@@ -261,13 +276,13 @@ pub async fn task_main(conf: SafeKeeperConf) -> anyhow::Result<()> {
                 },
                 _ = ticker.tick() => {
                     if push_handle.is_none() {
-                        push_handle = Some(tokio::spawn(push_loop(conf.clone())));
+                        push_handle = Some(tokio::spawn(push_loop(conf.clone(), global_timelines.clone())));
                     }
                     if pull_handle.is_none() {
-                        pull_handle = Some(tokio::spawn(pull_loop(conf.clone(), stats.clone())));
+                        pull_handle = Some(tokio::spawn(pull_loop(conf.clone(), global_timelines.clone(), stats.clone())));
                     }
                     if discover_handle.is_none() {
-                        discover_handle = Some(tokio::spawn(discover_loop(conf.clone(), stats.clone())));
+                        discover_handle = Some(tokio::spawn(discover_loop(conf.clone(), global_timelines.clone(), stats.clone())));
                     }
                 },
                 _ = &mut stats_task => {}

safekeeper/src/copy_timeline.rs

@@ -1,9 +1,7 @@
-use std::sync::Arc;
-
 use anyhow::{bail, Result};
 use camino::Utf8PathBuf;
-
 use postgres_ffi::{MAX_SEND_SIZE, WAL_SEGMENT_SIZE};
+use std::sync::Arc;
 use tokio::{
     fs::OpenOptions,
     io::{AsyncSeekExt, AsyncWriteExt},
@@ -14,7 +12,7 @@ use utils::{id::TenantTimelineId, lsn::Lsn};
 use crate::{
     control_file::FileStorage,
     state::TimelinePersistentState,
-    timeline::{Timeline, TimelineError, WalResidentTimeline},
+    timeline::{TimelineError, WalResidentTimeline},
     timelines_global_map::{create_temp_timeline_dir, validate_temp_timeline},
     wal_backup::copy_s3_segments,
     wal_storage::{wal_file_paths, WalReader},
@@ -25,16 +23,19 @@ use crate::{
 const MAX_BACKUP_LAG: u64 = 10 * WAL_SEGMENT_SIZE as u64;
 
 pub struct Request {
-    pub source: Arc<Timeline>,
+    pub source_ttid: TenantTimelineId,
     pub until_lsn: Lsn,
     pub destination_ttid: TenantTimelineId,
 }
 
-pub async fn handle_request(request: Request) -> Result<()> {
+pub async fn handle_request(
+    request: Request,
+    global_timelines: Arc<GlobalTimelines>,
+) -> Result<()> {
     // TODO: request.until_lsn MUST be a valid LSN, and we cannot check it :(
     //   if LSN will point to the middle of a WAL record, timeline will be in "broken" state
 
-    match GlobalTimelines::get(request.destination_ttid) {
+    match global_timelines.get(request.destination_ttid) {
         // timeline already exists. would be good to check that this timeline is the copy
         // of the source timeline, but it isn't obvious how to do that
         Ok(_) => return Ok(()),
@@ -46,9 +47,10 @@ pub async fn handle_request(request: Request) -> Result<()> {
         }
     }
 
-    let source_tli = request.source.wal_residence_guard().await?;
+    let source = global_timelines.get(request.source_ttid)?;
+    let source_tli = source.wal_residence_guard().await?;
 
-    let conf = &GlobalTimelines::get_global_config();
+    let conf = &global_timelines.get_global_config();
     let ttid = request.destination_ttid;
 
     let (_tmp_dir, tli_dir_path) = create_temp_timeline_dir(conf, ttid).await?;
@@ -127,7 +129,7 @@ pub async fn handle_request(request: Request) -> Result<()> {
 
     copy_s3_segments(
         wal_seg_size,
-        &request.source.ttid,
+        &request.source_ttid,
         &request.destination_ttid,
         first_segment,
         first_ondisk_segment,
@@ -158,7 +160,9 @@ pub async fn handle_request(request: Request) -> Result<()> {
 
     // now we have a ready timeline in a temp directory
     validate_temp_timeline(conf, request.destination_ttid, &tli_dir_path).await?;
-    GlobalTimelines::load_temp_timeline(request.destination_ttid, &tli_dir_path, true).await?;
+    global_timelines
+        .load_temp_timeline(request.destination_ttid, &tli_dir_path, true)
+        .await?;
 
     Ok(())
 }

safekeeper/src/debug_dump.rs

@@ -207,23 +207,23 @@ pub struct FileInfo {
 }
 
 /// Build debug dump response, using the provided [`Args`] filters.
-pub async fn build(args: Args) -> Result<Response> {
+pub async fn build(args: Args, global_timelines: Arc<GlobalTimelines>) -> Result<Response> {
     let start_time = Utc::now();
-    let timelines_count = GlobalTimelines::timelines_count();
-    let config = GlobalTimelines::get_global_config();
+    let timelines_count = global_timelines.timelines_count();
+    let config = global_timelines.get_global_config();
 
     let ptrs_snapshot = if args.tenant_id.is_some() && args.timeline_id.is_some() {
         // If both tenant_id and timeline_id are specified, we can just get the
         // timeline directly, without taking a snapshot of the whole list.
         let ttid = TenantTimelineId::new(args.tenant_id.unwrap(), args.timeline_id.unwrap());
-        if let Ok(tli) = GlobalTimelines::get(ttid) {
+        if let Ok(tli) = global_timelines.get(ttid) {
             vec![tli]
         } else {
             vec![]
         }
     } else {
         // Otherwise, take a snapshot of the whole list.
-        GlobalTimelines::get_all()
+        global_timelines.get_all()
     };
 
     let mut timelines = Vec::new();
@@ -344,12 +344,12 @@ fn get_wal_last_modified(path: &Utf8Path) -> Result<Option<DateTime<Utc>>> {
 
 /// Converts SafeKeeperConf to Config, filtering out the fields that are not
 /// supposed to be exposed.
-fn build_config(config: SafeKeeperConf) -> Config {
+fn build_config(config: Arc<SafeKeeperConf>) -> Config {
     Config {
         id: config.my_id,
-        workdir: config.workdir.into(),
-        listen_pg_addr: config.listen_pg_addr,
-        listen_http_addr: config.listen_http_addr,
+        workdir: config.workdir.clone().into(),
+        listen_pg_addr: config.listen_pg_addr.clone(),
+        listen_http_addr: config.listen_http_addr.clone(),
         no_sync: config.no_sync,
         max_offloader_lag_bytes: config.max_offloader_lag_bytes,
         wal_backup_enabled: config.wal_backup_enabled,

safekeeper/src/handler.rs

@@ -33,7 +33,7 @@ use utils::{
 
 /// Safekeeper handler of postgres commands
 pub struct SafekeeperPostgresHandler {
-    pub conf: SafeKeeperConf,
+    pub conf: Arc<SafeKeeperConf>,
     /// assigned application name
     pub appname: Option<String>,
     pub tenant_id: Option<TenantId>,
@@ -43,6 +43,7 @@ pub struct SafekeeperPostgresHandler {
     pub protocol: Option<PostgresClientProtocol>,
     /// Unique connection id is logged in spans for observability.
     pub conn_id: ConnectionId,
+    pub global_timelines: Arc<GlobalTimelines>,
     /// Auth scope allowed on the connections and public key used to check auth tokens. None if auth is not configured.
     auth: Option<(Scope, Arc<JwtAuth>)>,
     claims: Option<Claims>,
@@ -314,10 +315,11 @@ impl<IO: AsyncRead + AsyncWrite + Unpin + Send> postgres_backend::Handler<IO>
 
 impl SafekeeperPostgresHandler {
     pub fn new(
-        conf: SafeKeeperConf,
+        conf: Arc<SafeKeeperConf>,
         conn_id: u32,
         io_metrics: Option<TrafficMetrics>,
         auth: Option<(Scope, Arc<JwtAuth>)>,
+        global_timelines: Arc<GlobalTimelines>,
     ) -> Self {
         SafekeeperPostgresHandler {
             conf,
@@ -331,6 +333,7 @@ impl SafekeeperPostgresHandler {
             claims: None,
             auth,
             io_metrics,
+            global_timelines,
         }
     }
 
@@ -360,7 +363,7 @@ impl SafekeeperPostgresHandler {
         pgb: &mut PostgresBackend<IO>,
     ) -> Result<(), QueryError> {
         // Get timeline, handling "not found" error
-        let tli = match GlobalTimelines::get(self.ttid) {
+        let tli = match self.global_timelines.get(self.ttid) {
             Ok(tli) => Ok(Some(tli)),
             Err(TimelineError::NotFound(_)) => Ok(None),
             Err(e) => Err(QueryError::Other(e.into())),
@@ -394,7 +397,10 @@ impl SafekeeperPostgresHandler {
         &mut self,
         pgb: &mut PostgresBackend<IO>,
     ) -> Result<(), QueryError> {
-        let tli = GlobalTimelines::get(self.ttid).map_err(|e| QueryError::Other(e.into()))?;
+        let tli = self
+            .global_timelines
+            .get(self.ttid)
+            .map_err(|e| QueryError::Other(e.into()))?;
 
         let lsn = if self.is_walproposer_recovery() {
             // walproposer should get all local WAL until flush_lsn

safekeeper/src/http/mod.rs

@@ -3,14 +3,16 @@ pub mod routes;
 pub use routes::make_router;
 
 pub use safekeeper_api::models;
+use std::sync::Arc;
 
-use crate::SafeKeeperConf;
+use crate::{GlobalTimelines, SafeKeeperConf};
 
 pub async fn task_main(
-    conf: SafeKeeperConf,
+    conf: Arc<SafeKeeperConf>,
     http_listener: std::net::TcpListener,
+    global_timelines: Arc<GlobalTimelines>,
 ) -> anyhow::Result<()> {
-    let router = make_router(conf)
+    let router = make_router(conf, global_timelines)
         .build()
         .map_err(|err| anyhow::anyhow!(err))?;
     let service = utils::http::RouterService::new(router).unwrap();

safekeeper/src/http/routes.rs

@@ -66,6 +66,13 @@ fn get_conf(request: &Request<Body>) -> &SafeKeeperConf {
         .as_ref()
 }
 
+fn get_global_timelines(request: &Request<Body>) -> Arc<GlobalTimelines> {
+    request
+        .data::<Arc<GlobalTimelines>>()
+        .expect("unknown state type")
+        .clone()
+}
+
 /// Same as TermLsn, but serializes LSN using display serializer
 /// in Postgres format, i.e. 0/FFFFFFFF. Used only for the API response.
 #[derive(Debug, Clone, Copy, Serialize, Deserialize)]
@@ -123,9 +130,11 @@ async fn tenant_delete_handler(mut request: Request<Body>) -> Result<Response<Bo
     let only_local = parse_query_param(&request, "only_local")?.unwrap_or(false);
     check_permission(&request, Some(tenant_id))?;
     ensure_no_body(&mut request).await?;
+    let global_timelines = get_global_timelines(&request);
     // FIXME: `delete_force_all_for_tenant` can return an error for multiple different reasons;
     // Using an `InternalServerError` should be fixed when the types support it
-    let delete_info = GlobalTimelines::delete_force_all_for_tenant(&tenant_id, only_local)
+    let delete_info = global_timelines
+        .delete_force_all_for_tenant(&tenant_id, only_local)
         .await
         .map_err(ApiError::InternalServerError)?;
     json_response(
@@ -156,7 +165,9 @@ async fn timeline_create_handler(mut request: Request<Body>) -> Result<Response<
             .commit_lsn
             .segment_lsn(server_info.wal_seg_size as usize)
     });
-    GlobalTimelines::create(ttid, server_info, request_data.commit_lsn, local_start_lsn)
+    let global_timelines = get_global_timelines(&request);
+    global_timelines
+        .create(ttid, server_info, request_data.commit_lsn, local_start_lsn)
         .await
         .map_err(ApiError::InternalServerError)?;
 
@@ -167,7 +178,9 @@ async fn timeline_create_handler(mut request: Request<Body>) -> Result<Response<
 /// Note: it is possible to do the same with debug_dump.
 async fn timeline_list_handler(request: Request<Body>) -> Result<Response<Body>, ApiError> {
     check_permission(&request, None)?;
-    let res: Vec<TenantTimelineId> = GlobalTimelines::get_all()
+    let global_timelines = get_global_timelines(&request);
+    let res: Vec<TenantTimelineId> = global_timelines
+        .get_all()
         .iter()
         .map(|tli| tli.ttid)
         .collect();
@@ -182,7 +195,8 @@ async fn timeline_status_handler(request: Request<Body>) -> Result<Response<Body
     );
     check_permission(&request, Some(ttid.tenant_id))?;
 
-    let tli = GlobalTimelines::get(ttid).map_err(ApiError::from)?;
+    let global_timelines = get_global_timelines(&request);
+    let tli = global_timelines.get(ttid).map_err(ApiError::from)?;
     let (inmem, state) = tli.get_state().await;
     let flush_lsn = tli.get_flush_lsn().await;
 
@@ -233,9 +247,11 @@ async fn timeline_delete_handler(mut request: Request<Body>) -> Result<Response<
     let only_local = parse_query_param(&request, "only_local")?.unwrap_or(false);
     check_permission(&request, Some(ttid.tenant_id))?;
     ensure_no_body(&mut request).await?;
+    let global_timelines = get_global_timelines(&request);
     // FIXME: `delete_force` can fail from both internal errors and bad requests. Add better
     // error handling here when we're able to.
-    let resp = GlobalTimelines::delete(&ttid, only_local)
+    let resp = global_timelines
+        .delete(&ttid, only_local)
         .await
         .map_err(ApiError::InternalServerError)?;
     json_response(StatusCode::OK, resp)
@@ -247,8 +263,9 @@ async fn timeline_pull_handler(mut request: Request<Body>) -> Result<Response<Bo
 
     let data: pull_timeline::Request = json_request(&mut request).await?;
     let conf = get_conf(&request);
+    let global_timelines = get_global_timelines(&request);
 
-    let resp = pull_timeline::handle_request(data, conf.sk_auth_token.clone())
+    let resp = pull_timeline::handle_request(data, conf.sk_auth_token.clone(), global_timelines)
         .await
         .map_err(ApiError::InternalServerError)?;
     json_response(StatusCode::OK, resp)
@@ -263,7 +280,8 @@ async fn timeline_snapshot_handler(request: Request<Body>) -> Result<Response<Bo
     );
     check_permission(&request, Some(ttid.tenant_id))?;
 
-    let tli = GlobalTimelines::get(ttid).map_err(ApiError::from)?;
+    let global_timelines = get_global_timelines(&request);
+    let tli = global_timelines.get(ttid).map_err(ApiError::from)?;
 
     // To stream the body use wrap_stream which wants Stream of Result<Bytes>,
     // so create the chan and write to it in another task.
@@ -293,19 +311,19 @@ async fn timeline_copy_handler(mut request: Request<Body>) -> Result<Response<Bo
     check_permission(&request, None)?;
 
     let request_data: TimelineCopyRequest = json_request(&mut request).await?;
-    let ttid = TenantTimelineId::new(
+    let source_ttid = TenantTimelineId::new(
         parse_request_param(&request, "tenant_id")?,
         parse_request_param(&request, "source_timeline_id")?,
     );
 
-    let source = GlobalTimelines::get(ttid)?;
+    let global_timelines = get_global_timelines(&request);
 
     copy_timeline::handle_request(copy_timeline::Request{
-        source,
+        source_ttid,
         until_lsn: request_data.until_lsn,
-        destination_ttid: TenantTimelineId::new(ttid.tenant_id, request_data.target_timeline_id),
-    })
-        .instrument(info_span!("copy_timeline", from=%ttid, to=%request_data.target_timeline_id, until_lsn=%request_data.until_lsn))
+        destination_ttid: TenantTimelineId::new(source_ttid.tenant_id, request_data.target_timeline_id),
+    }, global_timelines)
+        .instrument(info_span!("copy_timeline", from=%source_ttid, to=%request_data.target_timeline_id, until_lsn=%request_data.until_lsn))
         .await
         .map_err(ApiError::InternalServerError)?;
 
@@ -322,7 +340,8 @@ async fn patch_control_file_handler(
         parse_request_param(&request, "timeline_id")?,
     );
 
-    let tli = GlobalTimelines::get(ttid).map_err(ApiError::from)?;
+    let global_timelines = get_global_timelines(&request);
+    let tli = global_timelines.get(ttid).map_err(ApiError::from)?;
 
     let patch_request: patch_control_file::Request = json_request(&mut request).await?;
     let response = patch_control_file::handle_request(tli, patch_request)
@@ -341,7 +360,8 @@ async fn timeline_checkpoint_handler(request: Request<Body>) -> Result<Response<
         parse_request_param(&request, "timeline_id")?,
     );
 
-    let tli = GlobalTimelines::get(ttid)?;
+    let global_timelines = get_global_timelines(&request);
+    let tli = global_timelines.get(ttid)?;
     tli.write_shared_state()
         .await
         .sk
@@ -359,6 +379,7 @@ async fn timeline_digest_handler(request: Request<Body>) -> Result<Response<Body
     );
     check_permission(&request, Some(ttid.tenant_id))?;
 
+    let global_timelines = get_global_timelines(&request);
     let from_lsn: Option<Lsn> = parse_query_param(&request, "from_lsn")?;
     let until_lsn: Option<Lsn> = parse_query_param(&request, "until_lsn")?;
 
@@ -371,7 +392,7 @@ async fn timeline_digest_handler(request: Request<Body>) -> Result<Response<Body
         )))?,
     };
 
-    let tli = GlobalTimelines::get(ttid).map_err(ApiError::from)?;
+    let tli = global_timelines.get(ttid).map_err(ApiError::from)?;
     let tli = tli
         .wal_residence_guard()
         .await
@@ -393,7 +414,8 @@ async fn timeline_backup_partial_reset(request: Request<Body>) -> Result<Respons
     );
     check_permission(&request, Some(ttid.tenant_id))?;
 
-    let tli = GlobalTimelines::get(ttid).map_err(ApiError::from)?;
+    let global_timelines = get_global_timelines(&request);
+    let tli = global_timelines.get(ttid).map_err(ApiError::from)?;
 
     let response = tli
         .backup_partial_reset()
@@ -415,7 +437,8 @@ async fn timeline_term_bump_handler(
 
     let request_data: TimelineTermBumpRequest = json_request(&mut request).await?;
 
-    let tli = GlobalTimelines::get(ttid).map_err(ApiError::from)?;
+    let global_timelines = get_global_timelines(&request);
+    let tli = global_timelines.get(ttid).map_err(ApiError::from)?;
     let response = tli
         .term_bump(request_data.term)
         .await
@@ -452,7 +475,8 @@ async fn record_safekeeper_info(mut request: Request<Body>) -> Result<Response<B
         standby_horizon: sk_info.standby_horizon.0,
     };
 
-    let tli = GlobalTimelines::get(ttid).map_err(ApiError::from)?;
+    let global_timelines = get_global_timelines(&request);
+    let tli = global_timelines.get(ttid).map_err(ApiError::from)?;
     tli.record_safekeeper_info(proto_sk_info)
         .await
         .map_err(ApiError::InternalServerError)?;
@@ -506,6 +530,8 @@ async fn dump_debug_handler(mut request: Request<Body>) -> Result<Response<Body>
     let dump_term_history = dump_term_history.unwrap_or(true);
     let dump_wal_last_modified = dump_wal_last_modified.unwrap_or(dump_all);
 
+    let global_timelines = get_global_timelines(&request);
+
     let args = debug_dump::Args {
         dump_all,
         dump_control_file,
@@ -517,7 +543,7 @@ async fn dump_debug_handler(mut request: Request<Body>) -> Result<Response<Body>
         timeline_id,
     };
 
-    let resp = debug_dump::build(args)
+    let resp = debug_dump::build(args, global_timelines)
         .await
         .map_err(ApiError::InternalServerError)?;
 
@@ -570,7 +596,10 @@ async fn dump_debug_handler(mut request: Request<Body>) -> Result<Response<Body>
 }
 
 /// Safekeeper http router.
-pub fn make_router(conf: SafeKeeperConf) -> RouterBuilder<hyper::Body, ApiError> {
+pub fn make_router(
+    conf: Arc<SafeKeeperConf>,
+    global_timelines: Arc<GlobalTimelines>,
+) -> RouterBuilder<hyper::Body, ApiError> {
     let mut router = endpoint::make_router();
     if conf.http_auth.is_some() {
         router = router.middleware(auth_middleware(|request| {
@@ -592,7 +621,8 @@ pub fn make_router(conf: SafeKeeperConf) -> RouterBuilder<hyper::Body, ApiError>
     // located nearby (/safekeeper/src/http/openapi_spec.yaml).
     let auth = conf.http_auth.clone();
     router
-        .data(Arc::new(conf))
+        .data(conf)
+        .data(global_timelines)
         .data(auth)
         .get("/metrics", |r| request_span(r, prometheus_metrics_handler))
         .get("/profile/cpu", |r| request_span(r, profile_cpu_handler))

safekeeper/src/json_ctrl.rs

@@ -11,7 +11,6 @@ use postgres_backend::QueryError;
 use serde::{Deserialize, Serialize};
 use tokio::io::{AsyncRead, AsyncWrite};
 use tracing::*;
-use utils::id::TenantTimelineId;
 
 use crate::handler::SafekeeperPostgresHandler;
 use crate::safekeeper::{AcceptorProposerMessage, AppendResponse, ServerInfo};
@@ -21,7 +20,6 @@ use crate::safekeeper::{
 use crate::safekeeper::{Term, TermHistory, TermLsn};
 use crate::state::TimelinePersistentState;
 use crate::timeline::WalResidentTimeline;
-use crate::GlobalTimelines;
 use postgres_backend::PostgresBackend;
 use postgres_ffi::encode_logical_message;
 use postgres_ffi::WAL_SEGMENT_SIZE;
@@ -70,7 +68,7 @@ pub async fn handle_json_ctrl<IO: AsyncRead + AsyncWrite + Unpin>(
     info!("JSON_CTRL request: {append_request:?}");
 
     // need to init safekeeper state before AppendRequest
-    let tli = prepare_safekeeper(spg.ttid, append_request.pg_version).await?;
+    let tli = prepare_safekeeper(spg, append_request.pg_version).await?;
 
     // if send_proposer_elected is true, we need to update local history
     if append_request.send_proposer_elected {
@@ -99,20 +97,22 @@ pub async fn handle_json_ctrl<IO: AsyncRead + AsyncWrite + Unpin>(
 /// Prepare safekeeper to process append requests without crashes,
 /// by sending ProposerGreeting with default server.wal_seg_size.
 async fn prepare_safekeeper(
-    ttid: TenantTimelineId,
+    spg: &SafekeeperPostgresHandler,
     pg_version: u32,
 ) -> anyhow::Result<WalResidentTimeline> {
-    let tli = GlobalTimelines::create(
-        ttid,
-        ServerInfo {
-            pg_version,
-            wal_seg_size: WAL_SEGMENT_SIZE as u32,
-            system_id: 0,
-        },
-        Lsn::INVALID,
-        Lsn::INVALID,
-    )
-    .await?;
+    let tli = spg
+        .global_timelines
+        .create(
+            spg.ttid,
+            ServerInfo {
+                pg_version,
+                wal_seg_size: WAL_SEGMENT_SIZE as u32,
+                system_id: 0,
+            },
+            Lsn::INVALID,
+            Lsn::INVALID,
+        )
+        .await?;
 
     tli.wal_residence_guard().await
 }

safekeeper/src/metrics.rs

@@ -455,6 +455,7 @@ pub struct FullTimelineInfo {
 
 /// Collects metrics for all active timelines.
 pub struct TimelineCollector {
+    global_timelines: Arc<GlobalTimelines>,
     descs: Vec<Desc>,
     commit_lsn: GenericGaugeVec<AtomicU64>,
     backup_lsn: GenericGaugeVec<AtomicU64>,
@@ -478,14 +479,8 @@ pub struct TimelineCollector {
     active_timelines_count: IntGauge,
 }
 
-impl Default for TimelineCollector {
-    fn default() -> Self {
-        Self::new()
-    }
-}
-
 impl TimelineCollector {
-    pub fn new() -> TimelineCollector {
+    pub fn new(global_timelines: Arc<GlobalTimelines>) -> TimelineCollector {
         let mut descs = Vec::new();
 
         let commit_lsn = GenericGaugeVec::new(
@@ -676,6 +671,7 @@ impl TimelineCollector {
         descs.extend(active_timelines_count.desc().into_iter().cloned());
 
         TimelineCollector {
+            global_timelines,
             descs,
             commit_lsn,
             backup_lsn,
@@ -728,17 +724,18 @@ impl Collector for TimelineCollector {
         self.written_wal_seconds.reset();
         self.flushed_wal_seconds.reset();
 
-        let timelines_count = GlobalTimelines::get_all().len();
+        let timelines_count = self.global_timelines.get_all().len();
         let mut active_timelines_count = 0;
 
         // Prometheus Collector is sync, and data is stored under async lock. To
         // bridge the gap with a crutch, collect data in spawned thread with
         // local tokio runtime.
+        let global_timelines = self.global_timelines.clone();
         let infos = std::thread::spawn(|| {
             let rt = tokio::runtime::Builder::new_current_thread()
                 .build()
                 .expect("failed to create rt");
-            rt.block_on(collect_timeline_metrics())
+            rt.block_on(collect_timeline_metrics(global_timelines))
         })
         .join()
         .expect("collect_timeline_metrics thread panicked");
@@ -857,9 +854,9 @@ impl Collector for TimelineCollector {
     }
 }
 
-async fn collect_timeline_metrics() -> Vec<FullTimelineInfo> {
+async fn collect_timeline_metrics(global_timelines: Arc<GlobalTimelines>) -> Vec<FullTimelineInfo> {
     let mut res = vec![];
-    let active_timelines = GlobalTimelines::get_global_broker_active_set().get_all();
+    let active_timelines = global_timelines.get_global_broker_active_set().get_all();
 
     for tli in active_timelines {
         if let Some(info) = tli.info_for_metrics().await {

safekeeper/src/pull_timeline.rs

@@ -409,8 +409,9 @@ pub struct DebugDumpResponse {
 pub async fn handle_request(
     request: Request,
     sk_auth_token: Option<SecretString>,
+    global_timelines: Arc<GlobalTimelines>,
 ) -> Result<Response> {
-    let existing_tli = GlobalTimelines::get(TenantTimelineId::new(
+    let existing_tli = global_timelines.get(TenantTimelineId::new(
         request.tenant_id,
         request.timeline_id,
     ));
@@ -453,13 +454,14 @@ pub async fn handle_request(
     assert!(status.tenant_id == request.tenant_id);
     assert!(status.timeline_id == request.timeline_id);
 
-    pull_timeline(status, safekeeper_host, sk_auth_token).await
+    pull_timeline(status, safekeeper_host, sk_auth_token, global_timelines).await
 }
 
 async fn pull_timeline(
     status: TimelineStatus,
     host: String,
     sk_auth_token: Option<SecretString>,
+    global_timelines: Arc<GlobalTimelines>,
 ) -> Result<Response> {
     let ttid = TenantTimelineId::new(status.tenant_id, status.timeline_id);
     info!(
@@ -472,7 +474,7 @@ async fn pull_timeline(
         status.acceptor_state.epoch
     );
 
-    let conf = &GlobalTimelines::get_global_config();
+    let conf = &global_timelines.get_global_config();
 
     let (_tmp_dir, tli_dir_path) = create_temp_timeline_dir(conf, ttid).await?;
 
@@ -531,7 +533,9 @@ async fn pull_timeline(
     assert!(status.commit_lsn <= status.flush_lsn);
 
     // Finally, load the timeline.
-    let _tli = GlobalTimelines::load_temp_timeline(ttid, &tli_dir_path, false).await?;
+    let _tli = global_timelines
+        .load_temp_timeline(ttid, &tli_dir_path, false)
+        .await?;
 
     Ok(Response {
         safekeeper_host: host,

safekeeper/src/receive_wal.rs

@@ -267,6 +267,7 @@ impl SafekeeperPostgresHandler {
             pgb_reader: &mut pgb_reader,
             peer_addr,
             acceptor_handle: &mut acceptor_handle,
+            global_timelines: self.global_timelines.clone(),
         };
 
         // Read first message and create timeline if needed.
@@ -331,6 +332,7 @@ struct NetworkReader<'a, IO> {
     // WalAcceptor is spawned when we learn server info from walproposer and
     // create timeline; handle is put here.
     acceptor_handle: &'a mut Option<JoinHandle<anyhow::Result<()>>>,
+    global_timelines: Arc<GlobalTimelines>,
 }
 
 impl<'a, IO: AsyncRead + AsyncWrite + Unpin> NetworkReader<'a, IO> {
@@ -350,10 +352,11 @@ impl<'a, IO: AsyncRead + AsyncWrite + Unpin> NetworkReader<'a, IO> {
                     system_id: greeting.system_id,
                     wal_seg_size: greeting.wal_seg_size,
                 };
-                let tli =
-                    GlobalTimelines::create(self.ttid, server_info, Lsn::INVALID, Lsn::INVALID)
-                        .await
-                        .context("create timeline")?;
+                let tli = self
+                    .global_timelines
+                    .create(self.ttid, server_info, Lsn::INVALID, Lsn::INVALID)
+                    .await
+                    .context("create timeline")?;
                 tli.wal_residence_guard().await?
             }
             _ => {

safekeeper/src/send_wal.rs

@@ -10,7 +10,6 @@ use crate::timeline::WalResidentTimeline;
 use crate::wal_reader_stream::WalReaderStreamBuilder;
 use crate::wal_service::ConnectionId;
 use crate::wal_storage::WalReader;
-use crate::GlobalTimelines;
 use anyhow::{bail, Context as AnyhowContext};
 use bytes::Bytes;
 use futures::future::Either;
@@ -400,7 +399,10 @@ impl SafekeeperPostgresHandler {
         start_pos: Lsn,
         term: Option<Term>,
     ) -> Result<(), QueryError> {
-        let tli = GlobalTimelines::get(self.ttid).map_err(|e| QueryError::Other(e.into()))?;
+        let tli = self
+            .global_timelines
+            .get(self.ttid)
+            .map_err(|e| QueryError::Other(e.into()))?;
         let residence_guard = tli.wal_residence_guard().await?;
 
         if let Err(end) = self

safekeeper/src/timeline.rs

@@ -44,8 +44,8 @@ use crate::wal_backup_partial::PartialRemoteSegment;
 
 use crate::metrics::{FullTimelineInfo, WalStorageMetrics, MISC_OPERATION_SECONDS};
 use crate::wal_storage::{Storage as wal_storage_iface, WalReader};
+use crate::SafeKeeperConf;
 use crate::{debug_dump, timeline_manager, wal_storage};
-use crate::{GlobalTimelines, SafeKeeperConf};
 
 /// Things safekeeper should know about timeline state on peers.
 #[derive(Debug, Clone, Serialize, Deserialize)]
@@ -467,6 +467,7 @@ pub struct Timeline {
     walreceivers: Arc<WalReceivers>,
     timeline_dir: Utf8PathBuf,
     manager_ctl: ManagerCtl,
+    conf: Arc<SafeKeeperConf>,
 
     /// Hold this gate from code that depends on the Timeline's non-shut-down state.  While holding
     /// this gate, you must respect [`Timeline::cancel`]
@@ -489,6 +490,7 @@ impl Timeline {
         timeline_dir: &Utf8Path,
         remote_path: &RemotePath,
         shared_state: SharedState,
+        conf: Arc<SafeKeeperConf>,
     ) -> Arc<Self> {
         let (commit_lsn_watch_tx, commit_lsn_watch_rx) =
             watch::channel(shared_state.sk.state().commit_lsn);
@@ -516,6 +518,7 @@ impl Timeline {
             gate: Default::default(),
             cancel: CancellationToken::default(),
             manager_ctl: ManagerCtl::new(),
+            conf,
             broker_active: AtomicBool::new(false),
             wal_backup_active: AtomicBool::new(false),
             last_removed_segno: AtomicU64::new(0),
@@ -524,11 +527,14 @@ impl Timeline {
     }
 
     /// Load existing timeline from disk.
-    pub fn load_timeline(conf: &SafeKeeperConf, ttid: TenantTimelineId) -> Result<Arc<Timeline>> {
+    pub fn load_timeline(
+        conf: Arc<SafeKeeperConf>,
+        ttid: TenantTimelineId,
+    ) -> Result<Arc<Timeline>> {
         let _enter = info_span!("load_timeline", timeline = %ttid.timeline_id).entered();
 
-        let shared_state = SharedState::restore(conf, &ttid)?;
-        let timeline_dir = get_timeline_dir(conf, &ttid);
+        let shared_state = SharedState::restore(conf.as_ref(), &ttid)?;
+        let timeline_dir = get_timeline_dir(conf.as_ref(), &ttid);
         let remote_path = remote_timeline_path(&ttid)?;
 
         Ok(Timeline::new(
@@ -536,6 +542,7 @@ impl Timeline {
             &timeline_dir,
             &remote_path,
             shared_state,
+            conf,
         ))
     }
 
@@ -604,8 +611,7 @@ impl Timeline {
         // it is cancelled, so WAL storage won't be opened again.
         shared_state.sk.close_wal_store();
 
-        let conf = GlobalTimelines::get_global_config();
-        if !only_local && conf.is_wal_backup_enabled() {
+        if !only_local && self.conf.is_wal_backup_enabled() {
             // Note: we concurrently delete remote storage data from multiple
             // safekeepers. That's ok, s3 replies 200 if object doesn't exist and we
             // do some retries anyway.
@@ -951,7 +957,7 @@ impl WalResidentTimeline {
 
     pub async fn get_walreader(&self, start_lsn: Lsn) -> Result<WalReader> {
         let (_, persisted_state) = self.get_state().await;
-        let enable_remote_read = GlobalTimelines::get_global_config().is_wal_backup_enabled();
+        let enable_remote_read = self.conf.is_wal_backup_enabled();
 
         WalReader::new(
             &self.ttid,
@@ -1061,7 +1067,6 @@ impl ManagerTimeline {
 
     /// Try to switch state Offloaded->Present.
     pub(crate) async fn switch_to_present(&self) -> anyhow::Result<()> {
-        let conf = GlobalTimelines::get_global_config();
         let mut shared = self.write_shared_state().await;
 
         // trying to restore WAL storage
@@ -1069,7 +1074,7 @@ impl ManagerTimeline {
             &self.ttid,
             &self.timeline_dir,
             shared.sk.state(),
-            conf.no_sync,
+            self.conf.no_sync,
         )?;
 
         // updating control file
@@ -1096,7 +1101,7 @@ impl ManagerTimeline {
         // now we can switch shared.sk to Present, shouldn't fail
         let prev_sk = std::mem::replace(&mut shared.sk, StateSK::Empty);
         let cfile_state = prev_sk.take_state();
-        shared.sk = StateSK::Loaded(SafeKeeper::new(cfile_state, wal_store, conf.my_id)?);
+        shared.sk = StateSK::Loaded(SafeKeeper::new(cfile_state, wal_store, self.conf.my_id)?);
 
         Ok(())
     }

safekeeper/src/timelines_global_map.rs

@@ -13,7 +13,6 @@ use crate::{control_file, wal_storage, SafeKeeperConf};
 use anyhow::{bail, Context, Result};
 use camino::Utf8PathBuf;
 use camino_tempfile::Utf8TempDir;
-use once_cell::sync::Lazy;
 use serde::Serialize;
 use std::collections::HashMap;
 use std::str::FromStr;
@@ -42,23 +41,16 @@ struct GlobalTimelinesState {
     // this map is dropped on restart.
     tombstones: HashMap<TenantTimelineId, Instant>,
 
-    conf: Option<SafeKeeperConf>,
+    conf: Arc<SafeKeeperConf>,
     broker_active_set: Arc<TimelinesSet>,
     global_rate_limiter: RateLimiter,
 }
 
 impl GlobalTimelinesState {
-    /// Get configuration, which must be set once during init.
-    fn get_conf(&self) -> &SafeKeeperConf {
-        self.conf
-            .as_ref()
-            .expect("GlobalTimelinesState conf is not initialized")
-    }
-
     /// Get dependencies for a timeline constructor.
-    fn get_dependencies(&self) -> (SafeKeeperConf, Arc<TimelinesSet>, RateLimiter) {
+    fn get_dependencies(&self) -> (Arc<SafeKeeperConf>, Arc<TimelinesSet>, RateLimiter) {
         (
-            self.get_conf().clone(),
+            self.conf.clone(),
             self.broker_active_set.clone(),
             self.global_rate_limiter.clone(),
         )
@@ -82,35 +74,39 @@ impl GlobalTimelinesState {
     }
 }
 
-static TIMELINES_STATE: Lazy<Mutex<GlobalTimelinesState>> = Lazy::new(|| {
-    Mutex::new(GlobalTimelinesState {
-        timelines: HashMap::new(),
-        tombstones: HashMap::new(),
-        conf: None,
-        broker_active_set: Arc::new(TimelinesSet::default()),
-        global_rate_limiter: RateLimiter::new(1, 1),
-    })
-});
-
-/// A zero-sized struct used to manage access to the global timelines map.
-pub struct GlobalTimelines;
+/// A struct used to manage access to the global timelines map.
+pub struct GlobalTimelines {
+    state: Mutex<GlobalTimelinesState>,
+}
 
 impl GlobalTimelines {
+    /// Create a new instance of the global timelines map.
+    pub fn new(conf: Arc<SafeKeeperConf>) -> Self {
+        Self {
+            state: Mutex::new(GlobalTimelinesState {
+                timelines: HashMap::new(),
+                tombstones: HashMap::new(),
+                conf,
+                broker_active_set: Arc::new(TimelinesSet::default()),
+                global_rate_limiter: RateLimiter::new(1, 1),
+            }),
+        }
+    }
+
     /// Inject dependencies needed for the timeline constructors and load all timelines to memory.
-    pub async fn init(conf: SafeKeeperConf) -> Result<()> {
+    pub async fn init(&self) -> Result<()> {
         // clippy isn't smart enough to understand that drop(state) releases the
         // lock, so use explicit block
         let tenants_dir = {
-            let mut state = TIMELINES_STATE.lock().unwrap();
+            let mut state = self.state.lock().unwrap();
             state.global_rate_limiter = RateLimiter::new(
-                conf.partial_backup_concurrency,
+                state.conf.partial_backup_concurrency,
                 DEFAULT_EVICTION_CONCURRENCY,
             );
-            state.conf = Some(conf);
 
             // Iterate through all directories and load tenants for all directories
             // named as a valid tenant_id.
-            state.get_conf().workdir.clone()
+            state.conf.workdir.clone()
         };
         let mut tenant_count = 0;
         for tenants_dir_entry in std::fs::read_dir(&tenants_dir)
@@ -122,7 +118,7 @@ impl GlobalTimelines {
                         TenantId::from_str(tenants_dir_entry.file_name().to_str().unwrap_or(""))
                     {
                         tenant_count += 1;
-                        GlobalTimelines::load_tenant_timelines(tenant_id).await?;
+                        self.load_tenant_timelines(tenant_id).await?;
                     }
                 }
                 Err(e) => error!(
@@ -135,7 +131,7 @@ impl GlobalTimelines {
         info!(
             "found {} tenants directories, successfully loaded {} timelines",
             tenant_count,
-            TIMELINES_STATE.lock().unwrap().timelines.len()
+            self.state.lock().unwrap().timelines.len()
         );
         Ok(())
     }
@@ -143,13 +139,13 @@ impl GlobalTimelines {
     /// Loads all timelines for the given tenant to memory. Returns fs::read_dir
     /// errors if any.
     ///
-    /// It is async, but TIMELINES_STATE lock is sync and there is no important
+    /// It is async, but self.state lock is sync and there is no important
     /// reason to make it async (it is always held for a short while), so we
     /// just lock and unlock it for each timeline -- this function is called
     /// during init when nothing else is running, so this is fine.
-    async fn load_tenant_timelines(tenant_id: TenantId) -> Result<()> {
+    async fn load_tenant_timelines(&self, tenant_id: TenantId) -> Result<()> {
         let (conf, broker_active_set, partial_backup_rate_limiter) = {
-            let state = TIMELINES_STATE.lock().unwrap();
+            let state = self.state.lock().unwrap();
             state.get_dependencies()
         };
 
@@ -163,10 +159,10 @@ impl GlobalTimelines {
                         TimelineId::from_str(timeline_dir_entry.file_name().to_str().unwrap_or(""))
                     {
                         let ttid = TenantTimelineId::new(tenant_id, timeline_id);
-                        match Timeline::load_timeline(&conf, ttid) {
+                        match Timeline::load_timeline(conf.clone(), ttid) {
                             Ok(tli) => {
                                 let mut shared_state = tli.write_shared_state().await;
-                                TIMELINES_STATE
+                                self.state
                                     .lock()
                                     .unwrap()
                                     .timelines
@@ -200,29 +196,30 @@ impl GlobalTimelines {
     }
 
     /// Get the number of timelines in the map.
-    pub fn timelines_count() -> usize {
-        TIMELINES_STATE.lock().unwrap().timelines.len()
+    pub fn timelines_count(&self) -> usize {
+        self.state.lock().unwrap().timelines.len()
     }
 
     /// Get the global safekeeper config.
-    pub fn get_global_config() -> SafeKeeperConf {
-        TIMELINES_STATE.lock().unwrap().get_conf().clone()
+    pub fn get_global_config(&self) -> Arc<SafeKeeperConf> {
+        self.state.lock().unwrap().conf.clone()
     }
 
-    pub fn get_global_broker_active_set() -> Arc<TimelinesSet> {
-        TIMELINES_STATE.lock().unwrap().broker_active_set.clone()
+    pub fn get_global_broker_active_set(&self) -> Arc<TimelinesSet> {
+        self.state.lock().unwrap().broker_active_set.clone()
     }
 
     /// Create a new timeline with the given id. If the timeline already exists, returns
     /// an existing timeline.
     pub(crate) async fn create(
+        &self,
         ttid: TenantTimelineId,
         server_info: ServerInfo,
         commit_lsn: Lsn,
         local_start_lsn: Lsn,
     ) -> Result<Arc<Timeline>> {
         let (conf, _, _) = {
-            let state = TIMELINES_STATE.lock().unwrap();
+            let state = self.state.lock().unwrap();
             if let Ok(timeline) = state.get(&ttid) {
                 // Timeline already exists, return it.
                 return Ok(timeline);
@@ -245,7 +242,7 @@ impl GlobalTimelines {
         let state =
             TimelinePersistentState::new(&ttid, server_info, vec![], commit_lsn, local_start_lsn)?;
         control_file::FileStorage::create_new(&tmp_dir_path, state, conf.no_sync).await?;
-        let timeline = GlobalTimelines::load_temp_timeline(ttid, &tmp_dir_path, true).await?;
+        let timeline = self.load_temp_timeline(ttid, &tmp_dir_path, true).await?;
         Ok(timeline)
     }
 
@@ -261,13 +258,14 @@ impl GlobalTimelines {
     /// 2) move the directory and load the timeline
     /// 3) take lock again and insert the timeline into the global map.
     pub async fn load_temp_timeline(
+        &self,
         ttid: TenantTimelineId,
         tmp_path: &Utf8PathBuf,
         check_tombstone: bool,
     ) -> Result<Arc<Timeline>> {
         // Check for existence and mark that we're creating it.
         let (conf, broker_active_set, partial_backup_rate_limiter) = {
-            let mut state = TIMELINES_STATE.lock().unwrap();
+            let mut state = self.state.lock().unwrap();
             match state.timelines.get(&ttid) {
                 Some(GlobalMapTimeline::CreationInProgress) => {
                     bail!(TimelineError::CreationInProgress(ttid));
@@ -295,10 +293,10 @@ impl GlobalTimelines {
         };
 
         // Do the actual move and reflect the result in the map.
-        match GlobalTimelines::install_temp_timeline(ttid, tmp_path, &conf).await {
+        match GlobalTimelines::install_temp_timeline(ttid, tmp_path, conf.clone()).await {
             Ok(timeline) => {
                 let mut timeline_shared_state = timeline.write_shared_state().await;
-                let mut state = TIMELINES_STATE.lock().unwrap();
+                let mut state = self.state.lock().unwrap();
                 assert!(matches!(
                     state.timelines.get(&ttid),
                     Some(GlobalMapTimeline::CreationInProgress)
@@ -319,7 +317,7 @@ impl GlobalTimelines {
             }
             Err(e) => {
                 // Init failed, remove the marker from the map
-                let mut state = TIMELINES_STATE.lock().unwrap();
+                let mut state = self.state.lock().unwrap();
                 assert!(matches!(
                     state.timelines.get(&ttid),
                     Some(GlobalMapTimeline::CreationInProgress)
@@ -334,10 +332,10 @@ impl GlobalTimelines {
     async fn install_temp_timeline(
         ttid: TenantTimelineId,
         tmp_path: &Utf8PathBuf,
-        conf: &SafeKeeperConf,
+        conf: Arc<SafeKeeperConf>,
     ) -> Result<Arc<Timeline>> {
-        let tenant_path = get_tenant_dir(conf, &ttid.tenant_id);
-        let timeline_path = get_timeline_dir(conf, &ttid);
+        let tenant_path = get_tenant_dir(conf.as_ref(), &ttid.tenant_id);
+        let timeline_path = get_timeline_dir(conf.as_ref(), &ttid);
 
         // We must have already checked that timeline doesn't exist in the map,
         // but there might be existing datadir: if timeline is corrupted it is
@@ -382,9 +380,9 @@ impl GlobalTimelines {
     /// Get a timeline from the global map. If it's not present, it doesn't exist on disk,
     /// or was corrupted and couldn't be loaded on startup. Returned timeline is always valid,
     /// i.e. loaded in memory and not cancelled.
-    pub(crate) fn get(ttid: TenantTimelineId) -> Result<Arc<Timeline>, TimelineError> {
+    pub(crate) fn get(&self, ttid: TenantTimelineId) -> Result<Arc<Timeline>, TimelineError> {
         let tli_res = {
-            let state = TIMELINES_STATE.lock().unwrap();
+            let state = self.state.lock().unwrap();
             state.get(&ttid)
         };
         match tli_res {
@@ -399,8 +397,8 @@ impl GlobalTimelines {
     }
 
     /// Returns all timelines. This is used for background timeline processes.
-    pub fn get_all() -> Vec<Arc<Timeline>> {
-        let global_lock = TIMELINES_STATE.lock().unwrap();
+    pub fn get_all(&self) -> Vec<Arc<Timeline>> {
+        let global_lock = self.state.lock().unwrap();
         global_lock
             .timelines
             .values()
@@ -419,8 +417,8 @@ impl GlobalTimelines {
 
     /// Returns all timelines belonging to a given tenant. Used for deleting all timelines of a tenant,
     /// and that's why it can return cancelled timelines, to retry deleting them.
-    fn get_all_for_tenant(tenant_id: TenantId) -> Vec<Arc<Timeline>> {
-        let global_lock = TIMELINES_STATE.lock().unwrap();
+    fn get_all_for_tenant(&self, tenant_id: TenantId) -> Vec<Arc<Timeline>> {
+        let global_lock = self.state.lock().unwrap();
         global_lock
             .timelines
             .values()
@@ -435,11 +433,12 @@ impl GlobalTimelines {
     /// Cancels timeline, then deletes the corresponding data directory.
     /// If only_local, doesn't remove WAL segments in remote storage.
     pub(crate) async fn delete(
+        &self,
         ttid: &TenantTimelineId,
         only_local: bool,
     ) -> Result<TimelineDeleteForceResult> {
         let tli_res = {
-            let state = TIMELINES_STATE.lock().unwrap();
+            let state = self.state.lock().unwrap();
 
             if state.tombstones.contains_key(ttid) {
                 // Presence of a tombstone guarantees that a previous deletion has completed and there is no work to do.
@@ -472,7 +471,7 @@ impl GlobalTimelines {
             }
             Err(_) => {
                 // Timeline is not memory, but it may still exist on disk in broken state.
-                let dir_path = get_timeline_dir(TIMELINES_STATE.lock().unwrap().get_conf(), ttid);
+                let dir_path = get_timeline_dir(self.state.lock().unwrap().conf.as_ref(), ttid);
                 let dir_existed = delete_dir(dir_path)?;
 
                 Ok(TimelineDeleteForceResult {
@@ -485,7 +484,7 @@ impl GlobalTimelines {
         // Finalize deletion, by dropping Timeline objects and storing smaller tombstones.  The tombstones
         // are used to prevent still-running computes from re-creating the same timeline when they send data,
         // and to speed up repeated deletion calls by avoiding re-listing objects.
-        TIMELINES_STATE.lock().unwrap().delete(*ttid);
+        self.state.lock().unwrap().delete(*ttid);
 
         result
     }
@@ -497,17 +496,18 @@ impl GlobalTimelines {
     ///
     /// If only_local, doesn't remove WAL segments in remote storage.
     pub async fn delete_force_all_for_tenant(
+        &self,
         tenant_id: &TenantId,
         only_local: bool,
     ) -> Result<HashMap<TenantTimelineId, TimelineDeleteForceResult>> {
         info!("deleting all timelines for tenant {}", tenant_id);
-        let to_delete = Self::get_all_for_tenant(*tenant_id);
+        let to_delete = self.get_all_for_tenant(*tenant_id);
 
         let mut err = None;
 
         let mut deleted = HashMap::new();
         for tli in &to_delete {
-            match Self::delete(&tli.ttid, only_local).await {
+            match self.delete(&tli.ttid, only_local).await {
                 Ok(result) => {
                     deleted.insert(tli.ttid, result);
                 }
@@ -529,15 +529,15 @@ impl GlobalTimelines {
         // so the directory may be not empty. In this case timelines will have bad state
         // and timeline background jobs can panic.
         delete_dir(get_tenant_dir(
-            TIMELINES_STATE.lock().unwrap().get_conf(),
+            self.state.lock().unwrap().conf.as_ref(),
             tenant_id,
         ))?;
 
         Ok(deleted)
     }
 
-    pub fn housekeeping(tombstone_ttl: &Duration) {
-        let mut state = TIMELINES_STATE.lock().unwrap();
+    pub fn housekeeping(&self, tombstone_ttl: &Duration) {
+        let mut state = self.state.lock().unwrap();
 
         // We keep tombstones long enough to have a good chance of preventing rogue computes from re-creating deleted
         // timelines.  If a compute kept running for longer than this TTL (or across a safekeeper restart) then they

safekeeper/src/wal_service.rs

@@ -4,6 +4,7 @@
 //!
 use anyhow::{Context, Result};
 use postgres_backend::QueryError;
+use std::sync::Arc;
 use std::time::Duration;
 use tokio::net::TcpStream;
 use tokio_io_timeout::TimeoutReader;
@@ -11,9 +12,9 @@ use tokio_util::sync::CancellationToken;
 use tracing::*;
 use utils::{auth::Scope, measured_stream::MeasuredStream};
 
-use crate::handler::SafekeeperPostgresHandler;
 use crate::metrics::TrafficMetrics;
 use crate::SafeKeeperConf;
+use crate::{handler::SafekeeperPostgresHandler, GlobalTimelines};
 use postgres_backend::{AuthType, PostgresBackend};
 
 /// Accept incoming TCP connections and spawn them into a background thread.
@@ -22,9 +23,10 @@ use postgres_backend::{AuthType, PostgresBackend};
 /// to any tenant are allowed) or Tenant (only tokens giving access to specific
 /// tenant are allowed). Doesn't matter if auth is disabled in conf.
 pub async fn task_main(
-    conf: SafeKeeperConf,
+    conf: Arc<SafeKeeperConf>,
     pg_listener: std::net::TcpListener,
     allowed_auth_scope: Scope,
+    global_timelines: Arc<GlobalTimelines>,
 ) -> anyhow::Result<()> {
     // Tokio's from_std won't do this for us, per its comment.
     pg_listener.set_nonblocking(true)?;
@@ -37,10 +39,10 @@ pub async fn task_main(
         debug!("accepted connection from {}", peer_addr);
         let conf = conf.clone();
         let conn_id = issue_connection_id(&mut connection_count);
-
+        let global_timelines = global_timelines.clone();
         tokio::spawn(
             async move {
-                if let Err(err) = handle_socket(socket, conf, conn_id, allowed_auth_scope).await {
+                if let Err(err) = handle_socket(socket, conf, conn_id, allowed_auth_scope, global_timelines).await {
                     error!("connection handler exited: {}", err);
                 }
             }
@@ -53,9 +55,10 @@ pub async fn task_main(
 ///
 async fn handle_socket(
     socket: TcpStream,
-    conf: SafeKeeperConf,
+    conf: Arc<SafeKeeperConf>,
     conn_id: ConnectionId,
     allowed_auth_scope: Scope,
+    global_timelines: Arc<GlobalTimelines>,
 ) -> Result<(), QueryError> {
     socket.set_nodelay(true)?;
     let peer_addr = socket.peer_addr()?;
@@ -96,8 +99,13 @@ async fn handle_socket(
         Some(_) => AuthType::NeonJWT,
     };
     let auth_pair = auth_key.map(|key| (allowed_auth_scope, key));
-    let mut conn_handler =
-        SafekeeperPostgresHandler::new(conf, conn_id, Some(traffic_metrics.clone()), auth_pair);
+    let mut conn_handler = SafekeeperPostgresHandler::new(
+        conf,
+        conn_id,
+        Some(traffic_metrics.clone()),
+        auth_pair,
+        global_timelines,
+    );
     let pgbackend = PostgresBackend::new_from_io(socket, peer_addr, auth_type, None)?;
     // libpq protocol between safekeeper and walproposer / pageserver
     // We don't use shutdown.

storage_controller/src/persistence.rs

@@ -636,6 +636,13 @@ impl Persistence {
                     .into_boxed(),
             };
 
+            // Clear generation_pageserver if we are moving into a state where we won't have
+            // any attached pageservers.
+            let input_generation_pageserver = match input_placement_policy {
+                None | Some(PlacementPolicy::Attached(_)) => None,
+                Some(PlacementPolicy::Detached | PlacementPolicy::Secondary) => Some(None),
+            };
+
             #[derive(AsChangeset)]
             #[diesel(table_name = crate::schema::tenant_shards)]
             struct ShardUpdate {
@@ -643,6 +650,7 @@ impl Persistence {
                 placement_policy: Option<String>,
                 config: Option<String>,
                 scheduling_policy: Option<String>,
+                generation_pageserver: Option<Option<i64>>,
             }
 
             let update = ShardUpdate {
@@ -655,6 +663,7 @@ impl Persistence {
                     .map(|c| serde_json::to_string(&c).unwrap()),
                 scheduling_policy: input_scheduling_policy
                     .map(|p| serde_json::to_string(&p).unwrap()),
+                generation_pageserver: input_generation_pageserver,
             };
 
             query.set(update).execute(conn)?;

storage_controller/src/service.rs

@@ -513,6 +513,9 @@ struct ShardUpdate {
 
     /// If this is None, generation is not updated.
     generation: Option<Generation>,
+
+    /// If this is None, scheduling policy is not updated.
+    scheduling_policy: Option<ShardSchedulingPolicy>,
 }
 
 enum StopReconciliationsReason {
@@ -789,7 +792,7 @@ impl Service {
             node_list_futs.push({
                 async move {
                     tracing::info!("Scanning shards on node {node}...");
-                    let timeout = Duration::from_secs(1);
+                    let timeout = Duration::from_secs(5);
                     let response = node
                         .with_client_retries(
                             |client| async move { client.list_location_config().await },
@@ -2376,6 +2379,23 @@ impl Service {
             }
         };
 
+        // Ordinarily we do not update scheduling policy, but when making major changes
+        // like detaching or demoting to secondary-only, we need to force the scheduling
+        // mode to Active, or the caller's expected outcome (detach it) will not happen.
+        let scheduling_policy = match req.config.mode {
+            LocationConfigMode::Detached | LocationConfigMode::Secondary => {
+                // Special case: when making major changes like detaching or demoting to secondary-only,
+                // we need to force the scheduling mode to Active, or nothing will happen.
+                Some(ShardSchedulingPolicy::Active)
+            }
+            LocationConfigMode::AttachedMulti
+            | LocationConfigMode::AttachedSingle
+            | LocationConfigMode::AttachedStale => {
+                // While attached, continue to respect whatever the existing scheduling mode is.
+                None
+            }
+        };
+
         let mut create = true;
         for (shard_id, shard) in tenants.range_mut(TenantShardId::tenant_range(tenant_id)) {
             // Saw an existing shard: this is not a creation
@@ -2401,6 +2421,7 @@ impl Service {
                 placement_policy: placement_policy.clone(),
                 tenant_config: req.config.tenant_conf.clone(),
                 generation: set_generation,
+                scheduling_policy,
             });
         }
 
@@ -2497,6 +2518,7 @@ impl Service {
                     placement_policy,
                     tenant_config,
                     generation,
+                    scheduling_policy,
                 } in &updates
                 {
                     self.persistence
@@ -2505,7 +2527,7 @@ impl Service {
                             Some(placement_policy.clone()),
                             Some(tenant_config.clone()),
                             *generation,
-                            None,
+                            *scheduling_policy,
                         )
                         .await?;
                 }
@@ -2521,6 +2543,7 @@ impl Service {
                         placement_policy,
                         tenant_config,
                         generation: update_generation,
+                        scheduling_policy,
                     } in updates
                     {
                         let Some(shard) = tenants.get_mut(&tenant_shard_id) else {
@@ -2539,6 +2562,10 @@ impl Service {
                             shard.generation = Some(generation);
                         }
 
+                        if let Some(scheduling_policy) = scheduling_policy {
+                            shard.set_scheduling_policy(scheduling_policy);
+                        }
+
                         shard.schedule(scheduler, &mut schedule_context)?;
 
                         let maybe_waiter = self.maybe_reconcile_shard(shard, nodes);
@@ -2992,9 +3019,17 @@ impl Service {
 
         let TenantPolicyRequest {
             placement,
-            scheduling,
+            mut scheduling,
         } = req;
 
+        if let Some(PlacementPolicy::Detached | PlacementPolicy::Secondary) = placement {
+            // When someone configures a tenant to detach, we force the scheduling policy to enable
+            // this to take effect.
+            if scheduling.is_none() {
+                scheduling = Some(ShardSchedulingPolicy::Active);
+            }
+        }
+
         self.persistence
             .update_tenant_shard(
                 TenantFilter::Tenant(tenant_id),

storage_scrubber/src/garbage.rs

@@ -459,12 +459,10 @@ pub async fn get_timeline_objects(
     Ok(list.keys)
 }
 
-const MAX_KEYS_PER_DELETE: usize = 1000;
-
 /// Drain a buffer of keys into DeleteObjects requests
 ///
 /// If `drain` is true, drains keys completely; otherwise stops when <
-/// MAX_KEYS_PER_DELETE keys are left.
+/// `max_keys_per_delete`` keys are left.
 /// `num_deleted` returns number of deleted keys.
 async fn do_delete(
     remote_client: &GenericRemoteStorage,
@@ -474,9 +472,10 @@ async fn do_delete(
     progress_tracker: &mut DeletionProgressTracker,
 ) -> anyhow::Result<()> {
     let cancel = CancellationToken::new();
-    while (!keys.is_empty() && drain) || (keys.len() >= MAX_KEYS_PER_DELETE) {
+    let max_keys_per_delete = remote_client.max_keys_per_delete();
+    while (!keys.is_empty() && drain) || (keys.len() >= max_keys_per_delete) {
         let request_keys =
-            keys.split_off(keys.len() - (std::cmp::min(MAX_KEYS_PER_DELETE, keys.len())));
+            keys.split_off(keys.len() - (std::cmp::min(max_keys_per_delete, keys.len())));
 
         let request_keys: Vec<RemotePath> = request_keys.into_iter().map(|o| o.key).collect();
 
@@ -617,7 +616,7 @@ pub async fn purge_garbage(
         }
 
         objects_to_delete.append(&mut object_list);
-        if objects_to_delete.len() >= MAX_KEYS_PER_DELETE {
+        if objects_to_delete.len() >= remote_client.max_keys_per_delete() {
             do_delete(
                 &remote_client,
                 &mut objects_to_delete,

storage_scrubber/src/lib.rs

@@ -268,7 +268,7 @@ impl BucketConfig {
                 config.bucket_name, config.bucket_region
             ),
             RemoteStorageKind::AzureContainer(config) => format!(
-                "bucket {}, storage account {:?}, region {}",
+                "container {}, storage account {:?}, region {}",
                 config.container_name, config.storage_account, config.container_region
             ),
         }

storage_scrubber/src/main.rs

@@ -86,6 +86,8 @@ enum Command {
         /// For safekeeper node_kind only, json list of timelines and their lsn info
         #[arg(long, default_value = None)]
         timeline_lsns: Option<String>,
+        #[arg(long, default_value_t = false)]
+        verbose: bool,
     },
     TenantSnapshot {
         #[arg(long = "tenant-id")]
@@ -166,6 +168,7 @@ async fn main() -> anyhow::Result<()> {
             dump_db_connstr,
             dump_db_table,
             timeline_lsns,
+            verbose,
         } => {
             if let NodeKind::Safekeeper = node_kind {
                 let db_or_list = match (timeline_lsns, dump_db_connstr) {
@@ -203,6 +206,7 @@ async fn main() -> anyhow::Result<()> {
                     tenant_ids,
                     json,
                     post_to_storcon,
+                    verbose,
                     cli.exit_code,
                 )
                 .await
@@ -313,6 +317,7 @@ pub async fn run_cron_job(
         Vec::new(),
         true,
         post_to_storcon,
+        false, // default to non-verbose mode
         exit_code,
     )
     .await?;
@@ -362,12 +367,13 @@ pub async fn scan_pageserver_metadata_cmd(
     tenant_shard_ids: Vec<TenantShardId>,
     json: bool,
     post_to_storcon: bool,
+    verbose: bool,
     exit_code: bool,
 ) -> anyhow::Result<()> {
     if controller_client.is_none() && post_to_storcon {
         return Err(anyhow!("Posting pageserver scan health status to storage controller requires `--controller-api` and `--controller-jwt` to run"));
     }
-    match scan_pageserver_metadata(bucket_config.clone(), tenant_shard_ids).await {
+    match scan_pageserver_metadata(bucket_config.clone(), tenant_shard_ids, verbose).await {
         Err(e) => {
             tracing::error!("Failed: {e}");
             Err(e)

storage_scrubber/src/scan_pageserver_metadata.rs

@@ -21,8 +21,12 @@ pub struct MetadataSummary {
     tenant_count: usize,
     timeline_count: usize,
     timeline_shard_count: usize,
-    with_errors: HashSet<TenantShardTimelineId>,
-    with_warnings: HashSet<TenantShardTimelineId>,
+    /// Tenant-shard timeline (key) mapping to errors. The key has to be a string because it will be serialized to a JSON.
+    /// The key is generated using `TenantShardTimelineId::to_string()`.
+    with_errors: HashMap<String, Vec<String>>,
+    /// Tenant-shard timeline (key) mapping to warnings. The key has to be a string because it will be serialized to a JSON.
+    /// The key is generated using `TenantShardTimelineId::to_string()`.
+    with_warnings: HashMap<String, Vec<String>>,
     with_orphans: HashSet<TenantShardTimelineId>,
     indices_by_version: HashMap<usize, usize>,
 
@@ -52,7 +56,12 @@ impl MetadataSummary {
         }
     }
 
-    fn update_analysis(&mut self, id: &TenantShardTimelineId, analysis: &TimelineAnalysis) {
+    fn update_analysis(
+        &mut self,
+        id: &TenantShardTimelineId,
+        analysis: &TimelineAnalysis,
+        verbose: bool,
+    ) {
         if analysis.is_healthy() {
             self.healthy_tenant_shards.insert(id.tenant_shard_id);
         } else {
@@ -61,11 +70,17 @@ impl MetadataSummary {
         }
 
         if !analysis.errors.is_empty() {
-            self.with_errors.insert(*id);
+            let entry = self.with_errors.entry(id.to_string()).or_default();
+            if verbose {
+                entry.extend(analysis.errors.iter().cloned());
+            }
         }
 
         if !analysis.warnings.is_empty() {
-            self.with_warnings.insert(*id);
+            let entry = self.with_warnings.entry(id.to_string()).or_default();
+            if verbose {
+                entry.extend(analysis.warnings.iter().cloned());
+            }
         }
     }
 
@@ -120,6 +135,7 @@ Index versions: {version_summary}
 pub async fn scan_pageserver_metadata(
     bucket_config: BucketConfig,
     tenant_ids: Vec<TenantShardId>,
+    verbose: bool,
 ) -> anyhow::Result<MetadataSummary> {
     let (remote_client, target) = init_remote(bucket_config, NodeKind::Pageserver).await?;
 
@@ -164,6 +180,7 @@ pub async fn scan_pageserver_metadata(
         mut tenant_objects: TenantObjectListing,
         timelines: Vec<(TenantShardTimelineId, RemoteTimelineBlobData)>,
         highest_shard_count: ShardCount,
+        verbose: bool,
     ) {
         summary.tenant_count += 1;
 
@@ -203,7 +220,7 @@ pub async fn scan_pageserver_metadata(
                         Some(data),
                     )
                     .await;
-                    summary.update_analysis(&ttid, &analysis);
+                    summary.update_analysis(&ttid, &analysis, verbose);
 
                     timeline_ids.insert(ttid.timeline_id);
                 } else {
@@ -271,10 +288,6 @@ pub async fn scan_pageserver_metadata(
         summary.update_data(&data);
 
         match tenant_id {
-            None => {
-                tenant_id = Some(ttid.tenant_shard_id.tenant_id);
-                highest_shard_count = highest_shard_count.max(ttid.tenant_shard_id.shard_count);
-            }
             Some(prev_tenant_id) => {
                 if prev_tenant_id != ttid.tenant_shard_id.tenant_id {
                     // New tenant: analyze this tenant's timelines, clear accumulated tenant_timeline_results
@@ -287,6 +300,7 @@ pub async fn scan_pageserver_metadata(
                         tenant_objects,
                         timelines,
                         highest_shard_count,
+                        verbose,
                     )
                     .instrument(info_span!("analyze-tenant", tenant = %prev_tenant_id))
                     .await;
@@ -296,6 +310,10 @@ pub async fn scan_pageserver_metadata(
                     highest_shard_count = highest_shard_count.max(ttid.tenant_shard_id.shard_count);
                 }
             }
+            None => {
+                tenant_id = Some(ttid.tenant_shard_id.tenant_id);
+                highest_shard_count = highest_shard_count.max(ttid.tenant_shard_id.shard_count);
+            }
         }
 
         match &data.blob_data {
@@ -326,6 +344,7 @@ pub async fn scan_pageserver_metadata(
             tenant_objects,
             tenant_timeline_results,
             highest_shard_count,
+            verbose,
         )
         .instrument(info_span!("analyze-tenant", tenant = %tenant_id))
         .await;

test_runner/cloud_regress/README.md

@@ -0,0 +1,21 @@
+# How to run the `pg_regress` tests on a cloud Neon instance.
+
+* Create a Neon project on staging.
+* Grant the superuser privileges to the DB user.
+* (Optional) create a branch for testing
+* Configure the endpoint by updating the control-plane database with the following settings:
+  * `Timeone`: `America/Los_Angeles`
+  * `DateStyle`: `Postgres,MDY`
+  * `compute_query_id`: `off`
+* Checkout the actual `Neon` sources
+* Patch the sql and expected files for the specific PostgreSQL version, e.g. for v17:
+```bash
+$ cd vendor/postgres-v17
+$ patch -p1 <../../compute/patches/cloud_regress_pg17.patch
+```
+* Set the environment variable `BENCHMARK_CONNSTR` to the connection URI of your project.
+* Set the environment variable `PG_VERSION` to the version of your project.
+* Run 
+```bash
+$ pytest -m remote_cluster -k cloud_regress
+```

test_runner/cloud_regress/test_cloud_regress.py

@@ -5,68 +5,15 @@ Run the regression tests on the cloud instance of Neon
 from __future__ import annotations
 
 from pathlib import Path
-from typing import Any
 
-import psycopg2
 import pytest
-from fixtures.log_helper import log
 from fixtures.neon_fixtures import RemotePostgres
 from fixtures.pg_version import PgVersion
 
 
-@pytest.fixture
-def setup(remote_pg: RemotePostgres):
-    """
-    Setup and teardown of the tests
-    """
-    with psycopg2.connect(remote_pg.connstr()) as conn:
-        with conn.cursor() as cur:
-            log.info("Creating the extension")
-            cur.execute("CREATE EXTENSION IF NOT EXISTS regress_so")
-            conn.commit()
-            # TODO: Migrate to branches and remove this code
-            log.info("Looking for subscriptions in the regress database")
-            cur.execute(
-                "SELECT subname FROM pg_catalog.pg_subscription WHERE "
-                "subdbid = (SELECT oid FROM pg_catalog.pg_database WHERE datname='regression');"
-            )
-            if cur.rowcount > 0:
-                with psycopg2.connect(
-                    dbname="regression",
-                    host=remote_pg.default_options["host"],
-                    user=remote_pg.default_options["user"],
-                    password=remote_pg.default_options["password"],
-                ) as regress_conn:
-                    with regress_conn.cursor() as regress_cur:
-                        for sub in cur:
-                            regress_cur.execute(f"ALTER SUBSCRIPTION {sub[0]} DISABLE")
-                            regress_cur.execute(
-                                f"ALTER SUBSCRIPTION {sub[0]} SET (slot_name = NONE)"
-                            )
-                            regress_cur.execute(f"DROP SUBSCRIPTION {sub[0]}")
-                        regress_conn.commit()
-
-    yield
-    # TODO: Migrate to branches and remove this code
-    log.info("Looking for extra roles...")
-    with psycopg2.connect(remote_pg.connstr()) as conn:
-        with conn.cursor() as cur:
-            cur.execute(
-                "SELECT rolname FROM pg_catalog.pg_roles WHERE oid > 16384 AND rolname <> 'neondb_owner'"
-            )
-            roles: list[Any] = []
-            for role in cur:
-                log.info("Role found: %s", role[0])
-                roles.append(role[0])
-            for role in roles:
-                cur.execute(f"DROP ROLE {role}")
-            conn.commit()
-
-
 @pytest.mark.timeout(7200)
 @pytest.mark.remote_cluster
 def test_cloud_regress(
-    setup,
     remote_pg: RemotePostgres,
     pg_version: PgVersion,
     pg_distrib_dir: Path,

test_runner/fixtures/metrics.py

@@ -152,6 +152,8 @@ PAGESERVER_PER_TENANT_METRICS: tuple[str, ...] = (
     "pageserver_resident_physical_size",
     "pageserver_io_operations_bytes_total",
     "pageserver_last_record_lsn",
+    "pageserver_disk_consistent_lsn",
+    "pageserver_projected_remote_consistent_lsn",
     "pageserver_standby_horizon",
     "pageserver_smgr_query_seconds_bucket",
     "pageserver_smgr_query_seconds_count",
@@ -173,6 +175,8 @@ PAGESERVER_PER_TENANT_METRICS: tuple[str, ...] = (
     counter("pageserver_tenant_throttling_count_accounted_finish"),
     counter("pageserver_tenant_throttling_wait_usecs_sum"),
     counter("pageserver_tenant_throttling_count"),
+    counter("pageserver_timeline_wal_records_received"),
+    counter("pageserver_page_service_pagestream_flush_in_progress_micros"),
     *histogram("pageserver_page_service_batch_size"),
     *PAGESERVER_PER_TENANT_REMOTE_TIMELINE_CLIENT_METRICS,
     # "pageserver_directory_entries_count", -- only used if above a certain threshold

test_runner/fixtures/neon_fixtures.py

@@ -4556,6 +4556,7 @@ class StorageScrubber:
     def __init__(self, env: NeonEnv, log_dir: Path):
         self.env = env
         self.log_dir = log_dir
+        self.allowed_errors: list[str] = []
 
     def scrubber_cli(
         self, args: list[str], timeout, extra_env: dict[str, str] | None = None
@@ -4633,19 +4634,70 @@ class StorageScrubber:
         if timeline_lsns is not None:
             args.append("--timeline-lsns")
             args.append(json.dumps(timeline_lsns))
+        if node_kind == NodeKind.PAGESERVER:
+            args.append("--verbose")
         stdout = self.scrubber_cli(args, timeout=30, extra_env=extra_env)
 
         try:
             summary = json.loads(stdout)
-            # summary does not contain "with_warnings" if node_kind is the safekeeper
-            no_warnings = "with_warnings" not in summary or not summary["with_warnings"]
-            healthy = not summary["with_errors"] and no_warnings
+            healthy = self._check_run_healthy(summary)
             return healthy, summary
         except:
             log.error("Failed to decode JSON output from `scan-metadata`.  Dumping stdout:")
             log.error(stdout)
             raise
 
+    def _check_line_allowed(self, line: str) -> bool:
+        for a in self.allowed_errors:
+            try:
+                if re.match(a, line):
+                    return True
+            except re.error:
+                log.error(f"Invalid regex: '{a}'")
+                raise
+        return False
+
+    def _check_line_list_allowed(self, lines: list[str]) -> bool:
+        for line in lines:
+            if not self._check_line_allowed(line):
+                return False
+        return True
+
+    def _check_run_healthy(self, summary: dict[str, Any]) -> bool:
+        # summary does not contain "with_warnings" if node_kind is the safekeeper
+        healthy = True
+        with_warnings = summary.get("with_warnings", None)
+        if with_warnings is not None:
+            if isinstance(with_warnings, list):
+                if len(with_warnings) > 0:
+                    # safekeeper scan_metadata output is a list of tenants
+                    healthy = False
+            else:
+                for _, warnings in with_warnings.items():
+                    assert (
+                        len(warnings) > 0
+                    ), "with_warnings value should not be empty, running without verbose mode?"
+                    if not self._check_line_list_allowed(warnings):
+                        healthy = False
+                        break
+        if not healthy:
+            return healthy
+        with_errors = summary.get("with_errors", None)
+        if with_errors is not None:
+            if isinstance(with_errors, list):
+                if len(with_errors) > 0:
+                    # safekeeper scan_metadata output is a list of tenants
+                    healthy = False
+            else:
+                for _, errors in with_errors.items():
+                    assert (
+                        len(errors) > 0
+                    ), "with_errors value should not be empty, running without verbose mode?"
+                    if not self._check_line_list_allowed(errors):
+                        healthy = False
+                        break
+        return healthy
+
     def tenant_snapshot(self, tenant_id: TenantId, output_path: Path):
         stdout = self.scrubber_cli(
             ["tenant-snapshot", "--tenant-id", str(tenant_id), "--output-path", str(output_path)],

test_runner/fixtures/pageserver/http.py

@@ -850,6 +850,7 @@ class PageserverHttpClient(requests.Session, MetricsGetter):
         force_repartition=False,
         force_image_layer_creation=False,
         force_l0_compaction=False,
+        wait_until_flushed=True,
         wait_until_uploaded=False,
         compact: bool | None = None,
         **kwargs,
@@ -862,6 +863,8 @@ class PageserverHttpClient(requests.Session, MetricsGetter):
             query["force_image_layer_creation"] = "true"
         if force_l0_compaction:
             query["force_l0_compaction"] = "true"
+        if not wait_until_flushed:
+            query["wait_until_flushed"] = "false"
         if wait_until_uploaded:
             query["wait_until_uploaded"] = "true"
 
@@ -869,7 +872,7 @@ class PageserverHttpClient(requests.Session, MetricsGetter):
             query["compact"] = "true" if compact else "false"
 
         log.info(
-            f"Requesting checkpoint: tenant {tenant_id}, timeline {timeline_id}, wait_until_uploaded={wait_until_uploaded}"
+            f"Requesting checkpoint: tenant={tenant_id} timeline={timeline_id} wait_until_flushed={wait_until_flushed} wait_until_uploaded={wait_until_uploaded} compact={compact}"
         )
         res = self.put(
             f"http://localhost:{self.port}/v1/tenant/{tenant_id}/timeline/{timeline_id}/checkpoint",

test_runner/fixtures/pageserver/utils.py

@@ -54,23 +54,15 @@ def wait_for_upload(
     tenant: TenantId | TenantShardId,
     timeline: TimelineId,
     lsn: Lsn,
+    timeout=20,
 ):
-    """waits for local timeline upload up to specified lsn"""
+    """Waits for local timeline upload up to specified LSN"""
 
-    current_lsn = Lsn(0)
-    for i in range(20):
-        current_lsn = remote_consistent_lsn(pageserver_http, tenant, timeline)
-        if current_lsn >= lsn:
-            log.info("wait finished")
-            return
-        lr_lsn = last_record_lsn(pageserver_http, tenant, timeline)
-        log.info(
-            f"waiting for remote_consistent_lsn to reach {lsn}, now {current_lsn}, last_record_lsn={lr_lsn}, iteration {i + 1}"
-        )
-        time.sleep(1)
-    raise Exception(
-        f"timed out while waiting for {tenant}/{timeline} remote_consistent_lsn to reach {lsn}, was {current_lsn}"
-    )
+    def is_uploaded():
+        remote_lsn = remote_consistent_lsn(pageserver_http, tenant, timeline)
+        assert remote_lsn >= lsn, f"remote_consistent_lsn at {remote_lsn}"
+
+    wait_until(is_uploaded, name=f"upload to {lsn}", timeout=timeout)
 
 
 def _tenant_in_expected_state(tenant_info: dict[str, Any], expected_state: str):

test_runner/performance/test_ingest_insert_bulk.py

@@ -0,0 +1,142 @@
+from __future__ import annotations
+
+import random
+from concurrent.futures import ThreadPoolExecutor
+
+import pytest
+from fixtures.benchmark_fixture import MetricReport, NeonBenchmarker
+from fixtures.common_types import Lsn
+from fixtures.log_helper import log
+from fixtures.neon_fixtures import (
+    NeonEnvBuilder,
+    wait_for_last_flush_lsn,
+)
+from fixtures.pageserver.utils import (
+    wait_for_last_record_lsn,
+    wait_for_upload,
+    wait_for_upload_queue_empty,
+)
+from fixtures.remote_storage import s3_storage
+
+
+@pytest.mark.timeout(900)
+@pytest.mark.parametrize("size", [8, 1024, 8192])
+@pytest.mark.parametrize("s3", [True, False], ids=["s3", "local"])
+@pytest.mark.parametrize("backpressure", [True, False], ids=["backpressure", "nobackpressure"])
+@pytest.mark.parametrize("fsync", [True, False], ids=["fsync", "nofsync"])
+def test_ingest_insert_bulk(
+    request: pytest.FixtureRequest,
+    neon_env_builder: NeonEnvBuilder,
+    zenbenchmark: NeonBenchmarker,
+    fsync: bool,
+    backpressure: bool,
+    s3: bool,
+    size: int,
+):
+    """
+    Benchmarks ingestion of 5 GB of sequential insert WAL. Measures ingestion and S3 upload
+    separately. Also does a Safekeeper→Pageserver re-ingestion to measure Pageserver ingestion in
+    isolation.
+    """
+
+    CONCURRENCY = 1  # 1 is optimal without fsync or backpressure
+    VOLUME = 5 * 1024**3
+    rows = VOLUME // (size + 64)  # +64 roughly accounts for per-row WAL overhead
+
+    neon_env_builder.safekeepers_enable_fsync = fsync
+
+    if s3:
+        neon_env_builder.enable_pageserver_remote_storage(s3_storage())
+        # NB: don't use S3 for Safekeeper. It doesn't affect throughput (no backpressure), but it
+        # would compete with Pageserver for bandwidth.
+        # neon_env_builder.enable_safekeeper_remote_storage(s3_storage())
+
+    neon_env_builder.disable_scrub_on_exit()  # immediate shutdown may leave stray layers
+    env = neon_env_builder.init_start()
+
+    endpoint = env.endpoints.create_start(
+        "main",
+        config_lines=[
+            f"fsync = {fsync}",
+            "max_replication_apply_lag = 0",
+            f"max_replication_flush_lag = {'10GB' if backpressure else '0'}",
+            # NB: neon_local defaults to 15MB, which is too slow -- production uses 500MB.
+            f"max_replication_write_lag = {'500MB' if backpressure else '0'}",
+        ],
+    )
+    endpoint.safe_psql("create extension neon")
+
+    # Wait for the timeline to be propagated to the pageserver.
+    wait_for_last_flush_lsn(env, endpoint, env.initial_tenant, env.initial_timeline)
+
+    # Ingest rows.
+    log.info("Ingesting data")
+    start_lsn = Lsn(endpoint.safe_psql("select pg_current_wal_lsn()")[0][0])
+
+    def insert_rows(endpoint, table, count, value):
+        with endpoint.connect().cursor() as cur:
+            cur.execute("set statement_timeout = 0")
+            cur.execute(f"create table {table} (id int, data bytea)")
+            cur.execute(f"insert into {table} values (generate_series(1, {count}), %s)", (value,))
+
+    with zenbenchmark.record_duration("upload"):
+        with zenbenchmark.record_duration("ingest"):
+            with ThreadPoolExecutor(max_workers=CONCURRENCY) as pool:
+                for i in range(CONCURRENCY):
+                    # Write a random value for all rows. This is sufficient to prevent compression,
+                    # e.g. in TOAST. Randomly generating every row is too slow.
+                    value = random.randbytes(size)
+                    worker_rows = rows / CONCURRENCY
+                    pool.submit(insert_rows, endpoint, f"table{i}", worker_rows, value)
+
+        end_lsn = Lsn(endpoint.safe_psql("select pg_current_wal_lsn()")[0][0])
+
+        # Wait for pageserver to ingest the WAL.
+        client = env.pageserver.http_client()
+        wait_for_last_record_lsn(client, env.initial_tenant, env.initial_timeline, end_lsn)
+
+        # Wait for pageserver S3 upload. Checkpoint to flush the last in-memory layer.
+        client.timeline_checkpoint(
+            env.initial_tenant,
+            env.initial_timeline,
+            compact=False,
+            wait_until_flushed=False,
+        )
+        wait_for_upload(client, env.initial_tenant, env.initial_timeline, end_lsn, timeout=600)
+
+    # Empty out upload queue for next benchmark.
+    wait_for_upload_queue_empty(client, env.initial_tenant, env.initial_timeline)
+
+    backpressure_time = endpoint.safe_psql("select backpressure_throttling_time()")[0][0]
+
+    # Now that all data is ingested, delete and recreate the tenant in the pageserver. This will
+    # reingest all the WAL directly from the safekeeper. This gives us a baseline of how fast the
+    # pageserver can ingest this WAL in isolation.
+    status = env.storage_controller.inspect(tenant_shard_id=env.initial_tenant)
+    assert status is not None
+
+    endpoint.stop()  # avoid spurious getpage errors
+    client.tenant_delete(env.initial_tenant)
+    env.pageserver.tenant_create(tenant_id=env.initial_tenant, generation=status[0])
+
+    with zenbenchmark.record_duration("recover"):
+        log.info("Recovering WAL into pageserver")
+        client.timeline_create(env.pg_version, env.initial_tenant, env.initial_timeline)
+        wait_for_last_record_lsn(client, env.initial_tenant, env.initial_timeline, end_lsn)
+
+    # Emit metrics.
+    wal_written_mb = round((end_lsn - start_lsn) / (1024 * 1024))
+    zenbenchmark.record("wal_written", wal_written_mb, "MB", MetricReport.TEST_PARAM)
+    zenbenchmark.record("row_count", rows, "rows", MetricReport.TEST_PARAM)
+    zenbenchmark.record("concurrency", CONCURRENCY, "clients", MetricReport.TEST_PARAM)
+    zenbenchmark.record(
+        "backpressure_time", backpressure_time // 1000, "ms", MetricReport.LOWER_IS_BETTER
+    )
+
+    props = {p["name"]: p["value"] for _, p in request.node.user_properties}
+    for name in ("ingest", "upload", "recover"):
+        throughput = int(wal_written_mb / props[name])
+        zenbenchmark.record(f"{name}_throughput", throughput, "MB/s", MetricReport.HIGHER_IS_BETTER)
+
+    # Pageserver shutdown will likely get stuck on the upload queue, just shut it down immediately.
+    env.stop(immediate=True)

test_runner/regress/test_compaction.py

@@ -15,7 +15,7 @@ from fixtures.pageserver.http import PageserverApiException
 from fixtures.utils import skip_in_debug_build, wait_until
 from fixtures.workload import Workload
 
-AGGRESIVE_COMPACTION_TENANT_CONF = {
+AGGRESSIVE_COMPACTION_TENANT_CONF = {
     # Disable gc and compaction. The test runs compaction manually.
     "gc_period": "0s",
     "compaction_period": "0s",
@@ -24,6 +24,7 @@ AGGRESIVE_COMPACTION_TENANT_CONF = {
     # Compact small layers
     "compaction_target_size": 1024**2,
     "image_creation_threshold": 2,
+    # "lsn_lease_length": "0s", -- TODO: would cause branch creation errors, should fix later
 }
 
 
@@ -51,7 +52,7 @@ def test_pageserver_compaction_smoke(
 page_cache_size=10
 """
 
-    env = neon_env_builder.init_start(initial_tenant_conf=AGGRESIVE_COMPACTION_TENANT_CONF)
+    env = neon_env_builder.init_start(initial_tenant_conf=AGGRESSIVE_COMPACTION_TENANT_CONF)
 
     tenant_id = env.initial_tenant
     timeline_id = env.initial_timeline
@@ -120,14 +121,25 @@ page_cache_size=10
     assert vectored_average < 8
 
 
+@skip_in_debug_build("only run with release build")
 def test_pageserver_gc_compaction_smoke(neon_env_builder: NeonEnvBuilder):
-    env = neon_env_builder.init_start(initial_tenant_conf=AGGRESIVE_COMPACTION_TENANT_CONF)
+    SMOKE_CONF = {
+        # Run both gc and gc-compaction.
+        "gc_period": "5s",
+        "compaction_period": "5s",
+        # No PiTR interval and small GC horizon
+        "pitr_interval": "0s",
+        "gc_horizon": f"{1024 ** 2}",
+        "lsn_lease_length": "0s",
+    }
+
+    env = neon_env_builder.init_start(initial_tenant_conf=SMOKE_CONF)
 
     tenant_id = env.initial_tenant
     timeline_id = env.initial_timeline
 
-    row_count = 1000
-    churn_rounds = 10
+    row_count = 10000
+    churn_rounds = 50
 
     ps_http = env.pageserver.http_client()
 
@@ -141,23 +153,35 @@ def test_pageserver_gc_compaction_smoke(neon_env_builder: NeonEnvBuilder):
         if i % 10 == 0:
             log.info(f"Running churn round {i}/{churn_rounds} ...")
 
+            # Run gc-compaction every 10 rounds to ensure the test doesn't take too long time.
+            ps_http.timeline_compact(
+                tenant_id,
+                timeline_id,
+                enhanced_gc_bottom_most_compaction=True,
+                body={
+                    "scheduled": True,
+                    "sub_compaction": True,
+                    "compact_range": {
+                        "start": "000000000000000000000000000000000000",
+                        "end": "030000000000000000000000000000000000",
+                    },
+                },
+            )
+
         workload.churn_rows(row_count, env.pageserver.id)
-        # Force L0 compaction to ensure the number of layers is within bounds, so that gc-compaction can run.
-        ps_http.timeline_compact(tenant_id, timeline_id, force_l0_compaction=True)
-        assert ps_http.perf_info(tenant_id, timeline_id)[0]["num_of_l0"] <= 1
-        ps_http.timeline_compact(
-            tenant_id,
-            timeline_id,
-            enhanced_gc_bottom_most_compaction=True,
-            body={
-                "start": "000000000000000000000000000000000000",
-                "end": "030000000000000000000000000000000000",
-            },
-        )
+
+    # ensure gc_compaction is scheduled and it's actually running (instead of skipping due to no layers picked)
+    env.pageserver.assert_log_contains(
+        "scheduled_compact_timeline.*picked .* layers for compaction"
+    )
 
     log.info("Validating at workload end ...")
     workload.validate(env.pageserver.id)
 
+    # Run a legacy compaction+gc to ensure gc-compaction can coexist with legacy compaction.
+    ps_http.timeline_checkpoint(tenant_id, timeline_id, wait_until_uploaded=True)
+    ps_http.timeline_gc(tenant_id, timeline_id, None)
+
 
 # Stripe sizes in number of pages.
 TINY_STRIPES = 16

test_runner/regress/test_compute_metrics.py

@@ -215,7 +215,7 @@ if SQL_EXPORTER is None:
             #
             # The "host" network mode allows sql_exporter to talk to the
             # endpoint which is running on the host.
-            super().__init__("docker.io/burningalchemist/sql_exporter:0.13.1", network_mode="host")
+            super().__init__("docker.io/burningalchemist/sql_exporter:0.16.0", network_mode="host")
 
             self.__logs_dir = logs_dir
             self.__port = port

test_runner/regress/test_nbtree_pagesplit_cycleid.py

@@ -0,0 +1,124 @@
+import threading
+import time
+
+from fixtures.neon_fixtures import NeonEnv
+
+BTREE_NUM_CYCLEID_PAGES = """
+    WITH raw_pages AS (
+        SELECT blkno, get_raw_page_at_lsn('t_uidx', 'main', blkno, NULL, NULL) page
+        FROM generate_series(1, pg_relation_size('t_uidx'::regclass) / 8192) blkno
+    ),
+    parsed_pages AS (
+        /* cycle ID is the last 2 bytes of the btree page */
+        SELECT blkno, SUBSTRING(page FROM 8191 FOR 2) as cycle_id
+        FROM raw_pages
+    )
+    SELECT count(*),
+           encode(cycle_id, 'hex')
+     FROM parsed_pages
+    WHERE encode(cycle_id, 'hex') != '0000'
+    GROUP BY encode(cycle_id, 'hex');
+    """
+
+
+def test_nbtree_pagesplit_cycleid(neon_simple_env: NeonEnv):
+    env = neon_simple_env
+    endpoint = env.endpoints.create_start("main")
+
+    ses1 = endpoint.connect().cursor()
+    ses1.execute("ALTER SYSTEM SET autovacuum = off;")
+    ses1.execute("ALTER SYSTEM SET enable_seqscan = off;")
+    ses1.execute("ALTER SYSTEM SET full_page_writes = off;")
+    ses1.execute("SELECT pg_reload_conf();")
+    ses1.execute("CREATE EXTENSION neon_test_utils;")
+    # prepare a large index
+    ses1.execute("CREATE TABLE t(id integer GENERATED ALWAYS AS IDENTITY, txt text);")
+    ses1.execute("CREATE UNIQUE INDEX t_uidx ON t(id);")
+    ses1.execute("INSERT INTO t (txt) SELECT i::text FROM generate_series(1, 2035) i;")
+
+    ses1.execute("SELECT neon_xlogflush();")
+    ses1.execute(BTREE_NUM_CYCLEID_PAGES)
+    pages = ses1.fetchall()
+    assert (
+        len(pages) == 0
+    ), f"0 back splits with cycle ID expected, real {len(pages)} first {pages[0]}"
+    # Delete enough tuples to clear the first index page.
+    # (there are up to 407 rows per 8KiB page; 406 for non-rightmost leafs.
+    ses1.execute("DELETE FROM t WHERE id <= 406;")
+    # Make sure the page is cleaned up
+    ses1.execute("VACUUM (FREEZE, INDEX_CLEANUP ON) t;")
+
+    # Do another delete-then-indexcleanup cycle, to move the pages from
+    # "dead" to "reusable"
+    ses1.execute("DELETE FROM t WHERE id <= 446;")
+    ses1.execute("VACUUM (FREEZE, INDEX_CLEANUP ON) t;")
+
+    # Make sure the vacuum we're about to trigger in s3 has cleanup work to do
+    ses1.execute("DELETE FROM t WHERE id <= 610;")
+
+    # Flush wal, for checking purposes
+    ses1.execute("SELECT neon_xlogflush();")
+    ses1.execute(BTREE_NUM_CYCLEID_PAGES)
+    pages = ses1.fetchall()
+    assert len(pages) == 0, f"No back splits with cycle ID expected, got batches of {pages} instead"
+
+    ses2 = endpoint.connect().cursor()
+    ses3 = endpoint.connect().cursor()
+
+    # Session 2 pins a btree page, which prevents vacuum from processing that
+    # page, thus allowing us to reliably split pages while a concurrent vacuum
+    # is running.
+    ses2.execute("BEGIN;")
+    ses2.execute(
+        "DECLARE foo NO SCROLL CURSOR FOR SELECT row_number() over () FROM t ORDER BY id ASC"
+    )
+    ses2.execute("FETCH FROM foo;")  # pins the leaf page with id 611
+    wait_evt = threading.Event()
+
+    # Session 3 runs the VACUUM command. Note that this will block, and
+    # therefore must run on another thread.
+    # We rely on this running quickly enough to hit the pinned page from
+    # session 2 by the time we start other work again in session 1, but
+    # technically there is a race where the thread (and/or PostgreSQL process)
+    # don't get to that pinned page with vacuum until >2s after evt.set() was
+    # called, and session 1 thus might already have split pages.
+    def vacuum_freeze_t(ses3, evt: threading.Event):
+        # Begin parallel vacuum that should hit the index
+        evt.set()
+        # this'll hang until s2 fetches enough new data from its cursor.
+        # this is technically a race with the time.sleep(2) below, but if this
+        # command doesn't hit
+        ses3.execute("VACUUM (FREEZE, INDEX_CLEANUP on, DISABLE_PAGE_SKIPPING on) t;")
+
+    ses3t = threading.Thread(target=vacuum_freeze_t, args=(ses3, wait_evt))
+    ses3t.start()
+    wait_evt.wait()
+    # Make extra sure we got the thread started and vacuum is stuck, by waiting
+    # some time even after wait_evt got set. This isn't truly reliable (it is
+    # possible
+    time.sleep(2)
+
+    # Insert 2 pages worth of new data.
+    # This should reuse the one empty page, plus another page at the end of
+    # the index relation; with split ordering
+    #    old_blk -> blkno=1 -> old_blk + 1.
+    # As this is run while vacuum in session 3 is happening, these splits
+    # should receive cycle IDs where applicable.
+    ses1.execute("INSERT INTO t (txt) SELECT i::text FROM generate_series(1, 812) i;")
+    # unpin the btree page, allowing s3's vacuum to complete
+    ses2.execute("FETCH ALL FROM foo;")
+    ses2.execute("ROLLBACK;")
+    # flush WAL to make sure PS is up-to-date
+    ses1.execute("SELECT neon_xlogflush();")
+    # check that our expectations are correct
+    ses1.execute(BTREE_NUM_CYCLEID_PAGES)
+    pages = ses1.fetchall()
+    assert (
+        len(pages) == 1 and pages[0][0] == 3
+    ), f"3 page splits with cycle ID expected; actual {pages}"
+
+    # final cleanup
+    ses3t.join()
+    ses1.close()
+    ses2.close()
+    ses3.close()

test_runner/regress/test_storage_controller.py

@@ -3230,3 +3230,55 @@ def test_multi_attached_timeline_creation(neon_env_builder: NeonEnvBuilder, migr
         # Always disable 'pause' failpoints, even on failure, to avoid hanging in shutdown
         env.storage_controller.configure_failpoints((migration_failpoint.value, "off"))
         raise
+
+
+@run_only_on_default_postgres("Postgres version makes no difference here")
+def test_storage_controller_detached_stopped(
+    neon_env_builder: NeonEnvBuilder,
+):
+    """
+    Test that detaching a tenant while it has scheduling policy set to Paused or Stop works
+    """
+
+    remote_storage_kind = s3_storage()
+    neon_env_builder.enable_pageserver_remote_storage(remote_storage_kind)
+
+    neon_env_builder.num_pageservers = 1
+
+    env = neon_env_builder.init_configs()
+    env.start()
+    virtual_ps_http = PageserverHttpClient(env.storage_controller_port, lambda: True)
+
+    tenant_id = TenantId.generate()
+    env.storage_controller.tenant_create(
+        tenant_id,
+        shard_count=1,
+    )
+
+    assert len(env.pageserver.http_client().tenant_list_locations()["tenant_shards"]) == 1
+
+    # Disable scheduling: ordinarily this would prevent the tenant's configuration being
+    # reconciled to pageservers, but this should be overridden when detaching.
+    env.storage_controller.allowed_errors.append(".*Scheduling is disabled by policy.*")
+    env.storage_controller.tenant_policy_update(
+        tenant_id,
+        {"scheduling": "Stop"},
+    )
+
+    env.storage_controller.consistency_check()
+
+    # Detach the tenant
+    virtual_ps_http.tenant_location_conf(
+        tenant_id,
+        {
+            "mode": "Detached",
+            "secondary_conf": None,
+            "tenant_conf": {},
+            "generation": None,
+        },
+    )
+
+    env.storage_controller.consistency_check()
+
+    # Confirm the detach happened
+    assert env.pageserver.http_client().tenant_list_locations()["tenant_shards"] == []

test_runner/regress/test_storage_scrubber.py

@@ -572,4 +572,10 @@ def test_scrubber_scan_pageserver_metadata(
     unhealthy = env.storage_controller.metadata_health_list_unhealthy()["unhealthy_tenant_shards"]
     assert len(unhealthy) == 1 and unhealthy[0] == str(tenant_shard_id)
 
-    neon_env_builder.disable_scrub_on_exit()
+    healthy, _ = env.storage_scrubber.scan_metadata()
+    assert not healthy
+    env.storage_scrubber.allowed_errors.append(".*not present in remote storage.*")
+    healthy, _ = env.storage_scrubber.scan_metadata()
+    assert healthy
+
+    neon_env_builder.disable_scrub_on_exit()  # We already ran scrubber, no need to do an extra run

test_runner/regress/test_vm_bits.py

@@ -4,7 +4,7 @@ import time
 from contextlib import closing
 
 from fixtures.log_helper import log
-from fixtures.neon_fixtures import NeonEnv, NeonEnvBuilder, fork_at_current_lsn
+from fixtures.neon_fixtures import NeonEnv, NeonEnvBuilder, PgBin, fork_at_current_lsn
 from fixtures.utils import query_scalar
 
 
@@ -292,3 +292,76 @@ def test_vm_bit_clear_on_heap_lock_blackbox(neon_env_builder: NeonEnvBuilder):
     tup = cur.fetchall()
     log.info(f"tuple = {tup}")
     cur.execute("commit transaction")
+
+
+def test_check_visibility_map(neon_env_builder: NeonEnvBuilder, pg_bin: PgBin):
+    """
+    Runs pgbench across a few databases on a sharded tenant, then performs a visibility map
+    consistency check. Regression test for https://github.com/neondatabase/neon/issues/9914.
+    """
+
+    # Use a large number of shards with small stripe sizes, to ensure the visibility
+    # map will end up on non-zero shards.
+    SHARD_COUNT = 8
+    STRIPE_SIZE = 32  # in 8KB pages
+    PGBENCH_RUNS = 4
+
+    env = neon_env_builder.init_start(
+        initial_tenant_shard_count=SHARD_COUNT, initial_tenant_shard_stripe_size=STRIPE_SIZE
+    )
+    endpoint = env.endpoints.create_start(
+        "main",
+        config_lines=[
+            "shared_buffers = 64MB",
+        ],
+    )
+
+    # Run pgbench in 4 different databases, to exercise different shards.
+    dbnames = [f"pgbench{i}" for i in range(PGBENCH_RUNS)]
+    for i, dbname in enumerate(dbnames):
+        log.info(f"pgbench run {i+1}/{PGBENCH_RUNS}")
+        endpoint.safe_psql(f"create database {dbname}")
+        connstr = endpoint.connstr(dbname=dbname)
+        # pgbench -i will automatically vacuum the tables. This creates the visibility map.
+        pg_bin.run(["pgbench", "-i", "-s", "10", connstr])
+        # Freeze the tuples to set the initial frozen bit.
+        endpoint.safe_psql("vacuum freeze", dbname=dbname)
+        # Run pgbench.
+        pg_bin.run(["pgbench", "-c", "32", "-j", "8", "-T", "10", connstr])
+
+    # Restart the endpoint to flush the compute page cache. We want to make sure we read VM pages
+    # from storage, not cache.
+    endpoint.stop()
+    endpoint.start()
+
+    # Check that the visibility map matches the heap contents for pg_accounts (the main table).
+    for dbname in dbnames:
+        log.info(f"Checking visibility map for {dbname}")
+        with endpoint.cursor(dbname=dbname) as cur:
+            cur.execute("create extension pg_visibility")
+
+            cur.execute("select count(*) from pg_check_visible('pgbench_accounts')")
+            row = cur.fetchone()
+            assert row is not None
+            assert row[0] == 0, f"{row[0]} inconsistent VM pages (visible)"
+
+            cur.execute("select count(*) from pg_check_frozen('pgbench_accounts')")
+            row = cur.fetchone()
+            assert row is not None
+            assert row[0] == 0, f"{row[0]} inconsistent VM pages (frozen)"
+
+    # Vacuum and freeze the tables, and check that the visibility map is still accurate.
+    for dbname in dbnames:
+        log.info(f"Vacuuming and checking visibility map for {dbname}")
+        with endpoint.cursor(dbname=dbname) as cur:
+            cur.execute("vacuum freeze")
+
+            cur.execute("select count(*) from pg_check_visible('pgbench_accounts')")
+            row = cur.fetchone()
+            assert row is not None
+            assert row[0] == 0, f"{row[0]} inconsistent VM pages (visible)"
+
+            cur.execute("select count(*) from pg_check_frozen('pgbench_accounts')")
+            row = cur.fetchone()
+            assert row is not None
+            assert row[0] == 0, f"{row[0]} inconsistent VM pages (frozen)"

vendor/revisions.json

@@ -1,18 +1,18 @@
 {
   "v17": [
     "17.2",
-    "a10d95be67265e0f10a422ba0457f5a7af01de71"
+    "01fa3c48664ca030cfb69bb4a350aa9df4691d88"
   ],
   "v16": [
     "16.6",
-    "dff6615a8e48a10bb17a03fa3c00635f1ace7a92"
+    "81428621f7c04aed03671cf80a928e0a36d92505"
   ],
   "v15": [
     "15.10",
-    "972e325e62b455957adbbdd8580e31275bb5b8c9"
+    "8736b10c1d93d11b9c0489872dd529c4c0f5338f"
   ],
   "v14": [
     "14.15",
-    "373f9decad933d2d46f321231032ae8b0da81acd"
+    "13ff324150fceaac72920e01742addc053db9462"
   ]
 }

workspace_hack/Cargo.toml

@@ -33,6 +33,7 @@ deranged = { version = "0.3", default-features = false, features = ["powerfmt",
 digest = { version = "0.10", features = ["mac", "oid", "std"] }
 either = { version = "1" }
 fail = { version = "0.5", default-features = false, features = ["failpoints"] }
+form_urlencoded = { version = "1" }
 futures-channel = { version = "0.3", features = ["sink"] }
 futures-executor = { version = "0.3" }
 futures-io = { version = "0.3" }
@@ -78,6 +79,7 @@ sha2 = { version = "0.10", features = ["asm", "oid"] }
 signature = { version = "2", default-features = false, features = ["digest", "rand_core", "std"] }
 smallvec = { version = "1", default-features = false, features = ["const_new", "write"] }
 spki = { version = "0.7", default-features = false, features = ["pem", "std"] }
+stable_deref_trait = { version = "1" }
 subtle = { version = "2" }
 sync_wrapper = { version = "0.1", default-features = false, features = ["futures"] }
 tikv-jemalloc-ctl = { version = "0.6", features = ["stats", "use_std"] }
@@ -105,6 +107,7 @@ anyhow = { version = "1", features = ["backtrace"] }
 bytes = { version = "1", features = ["serde"] }
 cc = { version = "1", default-features = false, features = ["parallel"] }
 chrono = { version = "0.4", default-features = false, features = ["clock", "serde", "wasmbind"] }
+displaydoc = { version = "0.2" }
 either = { version = "1" }
 getrandom = { version = "0.2", default-features = false, features = ["std"] }
 half = { version = "2", default-features = false, features = ["num-traits"] }