Merge with main

Resolve #1663.

## Changes

- ignore a "broken" [1] timeline on page server startup
- fix the race condition when creating multiple timelines in parallel for a tenant
- added tests for the above changes

[1]: a timeline is marked as "broken" if either
- failed to load the timeline's metadata or
- the timeline's disk consistent LSN is zero

.circleci/ansible/production.hosts

@@ -1,5 +1,6 @@
 [pageservers]
-zenith-1-ps-1 console_region_id=1
+#zenith-1-ps-1 console_region_id=1
+zenith-1-ps-2 console_region_id=1
 
 [safekeepers]
 zenith-1-sk-1 console_region_id=1
@@ -15,4 +16,3 @@ console_mgmt_base_url = http://console-release.local
 bucket_name           = zenith-storage-oregon
 bucket_region         = us-west-2
 etcd_endpoints        = etcd-release.local:2379
-safekeeper_enable_s3_offload = true

.circleci/ansible/staging.hosts

@@ -6,6 +6,7 @@ zenith-us-stage-ps-2 console_region_id=27
 zenith-us-stage-sk-1 console_region_id=27
 zenith-us-stage-sk-4 console_region_id=27
 zenith-us-stage-sk-5 console_region_id=27
+zenith-us-stage-sk-6 console_region_id=27
 
 [storage:children]
 pageservers
@@ -16,4 +17,3 @@ console_mgmt_base_url = http://console-staging.local
 bucket_name           = zenith-staging-storage-us-east-1
 bucket_region         = us-east-1
 etcd_endpoints        = etcd-staging.local:2379
-safekeeper_enable_s3_offload = false

.circleci/ansible/systemd/safekeeper.service

@@ -6,7 +6,7 @@ After=network.target auditd.service
 Type=simple
 User=safekeeper
 Environment=RUST_BACKTRACE=1 ZENITH_REPO_DIR=/storage/safekeeper/data LD_LIBRARY_PATH=/usr/local/lib
-ExecStart=/usr/local/bin/safekeeper -l {{ inventory_hostname }}.local:6500 --listen-http {{ inventory_hostname }}.local:7676 -p {{ first_pageserver }}:6400 -D /storage/safekeeper/data --broker-endpoints={{ etcd_endpoints }} --enable-s3-offload={{ safekeeper_enable_s3_offload }}
+ExecStart=/usr/local/bin/safekeeper -l {{ inventory_hostname }}.local:6500 --listen-http {{ inventory_hostname }}.local:7676 -p {{ first_pageserver }}:6400 -D /storage/safekeeper/data --broker-endpoints={{ etcd_endpoints }} --remote-storage='{bucket_name="{{bucket_name}}", bucket_region="{{bucket_region}}", prefix_in_bucket="wal"}'
 ExecReload=/bin/kill -HUP $MAINPID
 KillMode=mixed
 KillSignal=SIGINT

.circleci/config.yml

@@ -11,15 +11,6 @@ executors:
       - image: zimg/rust:1.58
 
 jobs:
-  check-codestyle-rust:
-    executor: neon-xlarge-executor
-    steps:
-      - checkout
-      - run:
-          name: rustfmt
-          when: always
-          command: cargo fmt --all -- --check
-
   # A job to build postgres
   build-postgres:
     executor: neon-xlarge-executor
@@ -685,7 +676,7 @@ jobs:
           name: Setup helm v3
           command: |
             curl -s https://raw.githubusercontent.com/helm/helm/main/scripts/get-helm-3 | bash
-            helm repo add zenithdb https://neondatabase.github.io/helm-charts
+            helm repo add neondatabase https://neondatabase.github.io/helm-charts
       - run:
           name: Re-deploy proxy
           command: |
@@ -740,7 +731,6 @@ jobs:
 workflows:
   build_and_test:
     jobs:
-      - check-codestyle-rust
       - check-codestyle-python
       - build-postgres:
           name: build-postgres-<< matrix.build_type >>

.github/workflows/testing.yml

@@ -1,8 +1,10 @@
 name: Build and Test
 
 on:
-  pull_request:
   push:
+    branches:
+    - main
+  pull_request:
 
 jobs:
   regression-check:
@@ -23,13 +25,17 @@ jobs:
           submodules: true
           fetch-depth: 2
 
-      - name: install rust toolchain ${{ matrix.rust_toolchain }}
+      - name: Install rust toolchain ${{ matrix.rust_toolchain }}
         uses: actions-rs/toolchain@v1
         with:
           profile: minimal
           toolchain: ${{ matrix.rust_toolchain }}
+          components: rustfmt, clippy
           override: true
 
+      - name: Check formatting
+        run: cargo fmt --all -- --check
+
       - name: Install Ubuntu postgres dependencies
         if: matrix.os == 'ubuntu-latest'
         run: |

Cargo.lock

@@ -292,6 +292,9 @@ name = "cc"
 version = "1.0.72"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "22a9137b95ea06864e018375b72adfb7db6e6f68cfc8df5a04d00288050485ee"
+dependencies = [
+ "jobserver",
+]
 
 [[package]]
 name = "cexpr"
@@ -1356,6 +1359,15 @@ version = "1.0.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "1aab8fc367588b89dcee83ab0fd66b72b50b72fa1904d7095045ace2b0c81c35"
 
+[[package]]
+name = "jobserver"
+version = "0.1.24"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "af25a77299a7f711a01975c35a6a424eb6862092cc2d6c72c4ed6cbc56dfc1fa"
+dependencies = [
+ "libc",
+]
+
 [[package]]
 name = "js-sys"
 version = "0.3.56"
@@ -1722,9 +1734,9 @@ dependencies = [
 
 [[package]]
 name = "once_cell"
-version = "1.9.0"
+version = "1.10.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "da32515d9f6e6e489d7bc9d84c71b060db7247dc035bbe44eac88cf87486d8d5"
+checksum = "87f3e037eac156d1775da914196f0f37741a274155e34a0b7e427c35d2a2ecb9"
 
 [[package]]
 name = "oorandom"
@@ -1831,6 +1843,7 @@ dependencies = [
  "url",
  "utils",
  "workspace_hack",
+ "zstd",
 ]
 
 [[package]]
@@ -2394,6 +2407,8 @@ version = "0.1.0"
 dependencies = [
  "anyhow",
  "async-trait",
+ "metrics",
+ "once_cell",
  "rusoto_core",
  "rusoto_s3",
  "serde",
@@ -2401,6 +2416,7 @@ dependencies = [
  "tempfile",
  "tokio",
  "tokio-util 0.7.0",
+ "toml_edit",
  "tracing",
  "workspace_hack",
 ]
@@ -2652,6 +2668,7 @@ name = "safekeeper"
 version = "0.1.0"
 dependencies = [
  "anyhow",
+ "async-trait",
  "byteorder",
  "bytes",
  "clap 3.0.14",
@@ -2660,12 +2677,14 @@ dependencies = [
  "daemonize",
  "etcd_broker",
  "fs2",
+ "futures",
  "git-version",
  "hex",
  "humantime",
  "hyper",
  "lazy_static",
  "metrics",
+ "once_cell",
  "postgres",
  "postgres-protocol",
  "postgres_ffi",
@@ -2679,6 +2698,7 @@ dependencies = [
  "tokio",
  "tokio-postgres",
  "tokio-util 0.7.0",
+ "toml_edit",
  "tracing",
  "url",
  "utils",
@@ -3931,3 +3951,32 @@ name = "zeroize"
 version = "1.5.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "7c88870063c39ee00ec285a2f8d6a966e5b6fb2becc4e8dac77ed0d370ed6006"
+
+[[package]]
+name = "zstd"
+version = "0.11.1+zstd.1.5.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "77a16b8414fde0414e90c612eba70985577451c4c504b99885ebed24762cb81a"
+dependencies = [
+ "zstd-safe",
+]
+
+[[package]]
+name = "zstd-safe"
+version = "5.0.1+zstd.1.5.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "7c12659121420dd6365c5c3de4901f97145b79651fb1d25814020ed2ed0585ae"
+dependencies = [
+ "libc",
+ "zstd-sys",
+]
+
+[[package]]
+name = "zstd-sys"
+version = "2.0.1+zstd.1.5.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "9fd07cbbc53846d9145dbffdf6dd09a7a0aa52be46741825f5c97bdd4f73f12b"
+dependencies = [
+ "cc",
+ "libc",
+]

compute_tools/src/compute.rs

@@ -146,8 +146,14 @@ impl ComputeNode {
             _ => format!("basebackup {} {} {}", &self.tenant, &self.timeline, lsn),
         };
         let copyreader = client.copy_out(basebackup_cmd.as_str())?;
-        let mut ar = tar::Archive::new(copyreader);
 
+        // Read the archive directly from the `CopyOutReader`
+        //
+        // Set `ignore_zeros` so that unpack() reads all the Copy data and
+        // doesn't stop at the end-of-archive marker. Otherwise, if the server
+        // sends an Error after finishing the tarball, we will not notice it.
+        let mut ar = tar::Archive::new(copyreader);
+        ar.set_ignore_zeros(true);
         ar.unpack(&self.pgdata)?;
 
         self.metrics.basebackup_ms.store(

control_plane/Cargo.toml

@@ -4,7 +4,7 @@ version = "0.1.0"
 edition = "2021"
 
 [dependencies]
-tar = "0.4.33"
+tar = "0.4.38"
 postgres = { git = "https://github.com/zenithdb/rust-postgres.git", rev="d052ee8b86fff9897c77b0fe89ea9daba0e1fa38" }
 serde = { version = "1.0", features = ["derive"] }
 serde_with = "1.12.0"

control_plane/src/compute.rs

@@ -231,8 +231,13 @@ impl PostgresNode {
             .context("page server 'basebackup' command failed")?;
 
         // Read the archive directly from the `CopyOutReader`
-        tar::Archive::new(copyreader)
+        //
-            .unpack(&self.pgdata())
+        // Set `ignore_zeros` so that unpack() reads all the Copy data and
+        // doesn't stop at the end-of-archive marker. Otherwise, if the server
+        // sends an Error after finishing the tarball, we will not notice it.
+        let mut ar = tar::Archive::new(copyreader);
+        ar.set_ignore_zeros(true);
+        ar.unpack(&self.pgdata())
             .context("extracting base backup failed")?;
 
         Ok(())
@@ -274,6 +279,8 @@ impl PostgresNode {
         conf.append("listen_addresses", &self.address.ip().to_string());
         conf.append("port", &self.address.port().to_string());
         conf.append("wal_keep_size", "0");
+        // walproposer panics when basebackup is invalid, it is pointless to restart in this case.
+        conf.append("restart_after_crash", "off");
 
         // Configure the node to fetch pages from pageserver
         let pageserver_connstr = {

control_plane/src/etcd.rs

@@ -48,6 +48,10 @@ pub fn start_etcd_process(env: &local_env::LocalEnv) -> anyhow::Result<()> {
             format!("--data-dir={}", etcd_data_dir.display()),
             format!("--listen-client-urls={client_urls}"),
             format!("--advertise-client-urls={client_urls}"),
+            // Set --quota-backend-bytes to keep the etcd virtual memory
+            // size smaller. Our test etcd clusters are very small.
+            // See https://github.com/etcd-io/etcd/issues/7910
+            "--quota-backend-bytes=100000000".to_string(),
         ])
         .stdout(Stdio::from(etcd_stdout_file))
         .stderr(Stdio::from(etcd_stderr_file))

control_plane/src/lib.rs

@@ -49,3 +49,12 @@ fn fill_rust_env_vars(cmd: &mut Command) -> &mut Command {
         cmd
     }
 }
+
+fn fill_aws_secrets_vars(mut cmd: &mut Command) -> &mut Command {
+    for env_key in ["AWS_ACCESS_KEY_ID", "AWS_SECRET_ACCESS_KEY"] {
+        if let Ok(value) = std::env::var(env_key) {
+            cmd = cmd.env(env_key, value);
+        }
+    }
+    cmd
+}

control_plane/src/local_env.rs

@@ -15,7 +15,7 @@ use std::process::{Command, Stdio};
 use utils::{
     auth::{encode_from_key_file, Claims, Scope},
     postgres_backend::AuthType,
-    zid::{ZNodeId, ZTenantId, ZTenantTimelineId, ZTimelineId},
+    zid::{NodeId, ZTenantId, ZTenantTimelineId, ZTimelineId},
 };
 
 use crate::safekeeper::SafekeeperNode;
@@ -136,7 +136,7 @@ impl EtcdBroker {
 #[serde(default)]
 pub struct PageServerConf {
     // node id
-    pub id: ZNodeId,
+    pub id: NodeId,
     // Pageserver connection settings
     pub listen_pg_addr: String,
     pub listen_http_addr: String,
@@ -151,7 +151,7 @@ pub struct PageServerConf {
 impl Default for PageServerConf {
     fn default() -> Self {
         Self {
-            id: ZNodeId(0),
+            id: NodeId(0),
             listen_pg_addr: String::new(),
             listen_http_addr: String::new(),
             auth_type: AuthType::Trust,
@@ -163,19 +163,23 @@ impl Default for PageServerConf {
 #[derive(Serialize, Deserialize, PartialEq, Eq, Clone, Debug)]
 #[serde(default)]
 pub struct SafekeeperConf {
-    pub id: ZNodeId,
+    pub id: NodeId,
     pub pg_port: u16,
     pub http_port: u16,
     pub sync: bool,
+    pub remote_storage: Option<String>,
+    pub backup_threads: Option<u32>,
 }
 
 impl Default for SafekeeperConf {
     fn default() -> Self {
         Self {
-            id: ZNodeId(0),
+            id: NodeId(0),
             pg_port: 0,
             http_port: 0,
             sync: true,
+            remote_storage: None,
+            backup_threads: None,
         }
     }
 }
@@ -377,6 +381,7 @@ impl LocalEnv {
             base_path != Path::new(""),
             "repository base path is missing"
         );
+
         ensure!(
             !base_path.exists(),
             "directory '{}' already exists. Perhaps already initialized?",

control_plane/src/safekeeper.rs

@@ -18,12 +18,12 @@ use thiserror::Error;
 use utils::{
     connstring::connection_address,
     http::error::HttpErrorBody,
-    zid::{ZNodeId, ZTenantId, ZTimelineId},
+    zid::{NodeId, ZTenantId, ZTimelineId},
 };
 
 use crate::local_env::{LocalEnv, SafekeeperConf};
 use crate::storage::PageServerNode;
-use crate::{fill_rust_env_vars, read_pidfile};
+use crate::{fill_aws_secrets_vars, fill_rust_env_vars, read_pidfile};
 
 #[derive(Error, Debug)]
 pub enum SafekeeperHttpError {
@@ -65,7 +65,7 @@ impl ResponseErrorMessageExt for Response {
 //
 #[derive(Debug)]
 pub struct SafekeeperNode {
-    pub id: ZNodeId,
+    pub id: NodeId,
 
     pub conf: SafekeeperConf,
 
@@ -100,7 +100,7 @@ impl SafekeeperNode {
             .unwrap()
     }
 
-    pub fn datadir_path_by_id(env: &LocalEnv, sk_id: ZNodeId) -> PathBuf {
+    pub fn datadir_path_by_id(env: &LocalEnv, sk_id: NodeId) -> PathBuf {
         env.safekeeper_data_dir(format!("sk{}", sk_id).as_ref())
     }
 
@@ -143,6 +143,14 @@ impl SafekeeperNode {
         if let Some(prefix) = self.env.etcd_broker.broker_etcd_prefix.as_deref() {
             cmd.args(&["--broker-etcd-prefix", prefix]);
         }
+        if let Some(threads) = self.conf.backup_threads {
+            cmd.args(&["--backup-threads", threads.to_string().as_ref()]);
+        }
+        if let Some(ref remote_storage) = self.conf.remote_storage {
+            cmd.args(&["--remote-storage", remote_storage]);
+        }
+
+        fill_aws_secrets_vars(&mut cmd);
 
         if !cmd.status()?.success() {
             bail!(
@@ -286,7 +294,7 @@ impl SafekeeperNode {
         &self,
         tenant_id: ZTenantId,
         timeline_id: ZTimelineId,
-        peer_ids: Vec<ZNodeId>,
+        peer_ids: Vec<NodeId>,
     ) -> Result<()> {
         Ok(self
             .http_request(

control_plane/src/storage.rs

@@ -25,7 +25,7 @@ use utils::{
 };
 
 use crate::local_env::LocalEnv;
-use crate::{fill_rust_env_vars, read_pidfile};
+use crate::{fill_aws_secrets_vars, fill_rust_env_vars, read_pidfile};
 use pageserver::tenant_mgr::TenantInfo;
 
 #[derive(Error, Debug)]
@@ -493,12 +493,3 @@ impl PageServerNode {
         Ok(timeline_info_response)
     }
 }
-
-fn fill_aws_secrets_vars(mut cmd: &mut Command) -> &mut Command {
-    for env_key in ["AWS_ACCESS_KEY_ID", "AWS_SECRET_ACCESS_KEY"] {
-        if let Ok(value) = std::env::var(env_key) {
-            cmd = cmd.env(env_key, value);
-        }
-    }
-    cmd
-}

libs/etcd_broker/src/lib.rs

@@ -16,7 +16,7 @@ use tokio::{sync::mpsc, task::JoinHandle};
 use tracing::*;
 use utils::{
     lsn::Lsn,
-    zid::{ZNodeId, ZTenantId, ZTenantTimelineId},
+    zid::{NodeId, ZTenantId, ZTenantTimelineId},
 };
 
 /// Default value to use for prefixing to all etcd keys with.
@@ -25,7 +25,7 @@ pub const DEFAULT_NEON_BROKER_ETCD_PREFIX: &str = "neon";
 
 #[derive(Debug, Deserialize, Serialize)]
 struct SafekeeperTimeline {
-    safekeeper_id: ZNodeId,
+    safekeeper_id: NodeId,
     info: SkTimelineInfo,
 }
 
@@ -43,10 +43,10 @@ pub struct SkTimelineInfo {
     #[serde_as(as = "Option<DisplayFromStr>")]
     #[serde(default)]
     pub commit_lsn: Option<Lsn>,
-    /// LSN up to which safekeeper offloaded WAL to s3.
+    /// LSN up to which safekeeper has backed WAL.
     #[serde_as(as = "Option<DisplayFromStr>")]
     #[serde(default)]
-    pub s3_wal_lsn: Option<Lsn>,
+    pub backup_lsn: Option<Lsn>,
     /// LSN of last checkpoint uploaded by pageserver.
     #[serde_as(as = "Option<DisplayFromStr>")]
     #[serde(default)]
@@ -71,7 +71,7 @@ pub enum BrokerError {
 /// A way to control the data retrieval from a certain subscription.
 pub struct SkTimelineSubscription {
     safekeeper_timeline_updates:
-        mpsc::UnboundedReceiver<HashMap<ZTenantTimelineId, HashMap<ZNodeId, SkTimelineInfo>>>,
+        mpsc::UnboundedReceiver<HashMap<ZTenantTimelineId, HashMap<NodeId, SkTimelineInfo>>>,
     kind: SkTimelineSubscriptionKind,
     watcher_handle: JoinHandle<Result<(), BrokerError>>,
     watcher: Watcher,
@@ -81,7 +81,7 @@ impl SkTimelineSubscription {
     /// Asynchronously polls for more data from the subscription, suspending the current future if there's no data sent yet.
     pub async fn fetch_data(
         &mut self,
-    ) -> Option<HashMap<ZTenantTimelineId, HashMap<ZNodeId, SkTimelineInfo>>> {
+    ) -> Option<HashMap<ZTenantTimelineId, HashMap<NodeId, SkTimelineInfo>>> {
         self.safekeeper_timeline_updates.recv().await
     }
 
@@ -221,7 +221,7 @@ pub async fn subscribe_to_safekeeper_timeline_updates(
                 break;
             }
 
-            let mut timeline_updates: HashMap<ZTenantTimelineId, HashMap<ZNodeId, SkTimelineInfo>> = HashMap::new();
+            let mut timeline_updates: HashMap<ZTenantTimelineId, HashMap<NodeId, SkTimelineInfo>> = HashMap::new();
             // Keep track that the timeline data updates from etcd arrive in the right order.
             // https://etcd.io/docs/v3.5/learning/api_guarantees/#isolation-level-and-consistency-of-replicas
             // > etcd does not ensure linearizability for watch operations. Users are expected to verify the revision of watch responses to ensure correct ordering.
@@ -299,18 +299,18 @@ fn parse_etcd_key_value(
                 parse_capture(&caps, 1).map_err(BrokerError::ParsingError)?,
                 parse_capture(&caps, 2).map_err(BrokerError::ParsingError)?,
             ),
-            ZNodeId(parse_capture(&caps, 3).map_err(BrokerError::ParsingError)?),
+            NodeId(parse_capture(&caps, 3).map_err(BrokerError::ParsingError)?),
         ),
         SubscriptionKind::Tenant(tenant_id) => (
             ZTenantTimelineId::new(
                 tenant_id,
                 parse_capture(&caps, 1).map_err(BrokerError::ParsingError)?,
             ),
-            ZNodeId(parse_capture(&caps, 2).map_err(BrokerError::ParsingError)?),
+            NodeId(parse_capture(&caps, 2).map_err(BrokerError::ParsingError)?),
         ),
         SubscriptionKind::Timeline(zttid) => (
             zttid,
-            ZNodeId(parse_capture(&caps, 1).map_err(BrokerError::ParsingError)?),
+            NodeId(parse_capture(&caps, 1).map_err(BrokerError::ParsingError)?),
         ),
     };
 

libs/remote_storage/Cargo.toml

@@ -5,14 +5,17 @@ edition = "2021"
 
 [dependencies]
 anyhow = { version = "1.0", features = ["backtrace"] }
-tokio = { version = "1.17", features = ["sync", "macros", "fs", "io-util"] }
+async-trait = "0.1"
-tokio-util = { version = "0.7", features = ["io"] }
+metrics = { version = "0.1", path = "../metrics" }
-tracing = "0.1.27"
+once_cell = "1.8.0"
 rusoto_core = "0.48"
 rusoto_s3 = "0.48"
 serde = { version = "1.0", features = ["derive"] }
 serde_json = "1"
-async-trait = "0.1"
+tokio = { version = "1.17", features = ["sync", "macros", "fs", "io-util"] }
+tokio-util = { version = "0.7", features = ["io"] }
+toml_edit = { version = "0.13", features = ["easy"] }
+tracing = "0.1.27"
 
 workspace_hack = { version = "0.1", path = "../../workspace_hack" }
 

libs/remote_storage/src/lib.rs

@@ -16,8 +16,10 @@ use std::{
     path::{Path, PathBuf},
 };
 
-use anyhow::Context;
+use anyhow::{bail, Context};
+
 use tokio::io;
+use toml_edit::Item;
 use tracing::info;
 
 pub use self::{
@@ -203,6 +205,90 @@ pub fn path_with_suffix_extension(original_path: impl AsRef<Path>, suffix: &str)
         .with_extension(new_extension.as_ref())
 }
 
+impl RemoteStorageConfig {
+    pub fn from_toml(toml: &toml_edit::Item) -> anyhow::Result<RemoteStorageConfig> {
+        let local_path = toml.get("local_path");
+        let bucket_name = toml.get("bucket_name");
+        let bucket_region = toml.get("bucket_region");
+
+        let max_concurrent_syncs = NonZeroUsize::new(
+            parse_optional_integer("max_concurrent_syncs", toml)?
+                .unwrap_or(DEFAULT_REMOTE_STORAGE_MAX_CONCURRENT_SYNCS),
+        )
+        .context("Failed to parse 'max_concurrent_syncs' as a positive integer")?;
+
+        let max_sync_errors = NonZeroU32::new(
+            parse_optional_integer("max_sync_errors", toml)?
+                .unwrap_or(DEFAULT_REMOTE_STORAGE_MAX_SYNC_ERRORS),
+        )
+        .context("Failed to parse 'max_sync_errors' as a positive integer")?;
+
+        let concurrency_limit = NonZeroUsize::new(
+            parse_optional_integer("concurrency_limit", toml)?
+                .unwrap_or(DEFAULT_REMOTE_STORAGE_S3_CONCURRENCY_LIMIT),
+        )
+        .context("Failed to parse 'concurrency_limit' as a positive integer")?;
+
+        let storage = match (local_path, bucket_name, bucket_region) {
+            (None, None, None) => bail!("no 'local_path' nor 'bucket_name' option"),
+            (_, Some(_), None) => {
+                bail!("'bucket_region' option is mandatory if 'bucket_name' is given ")
+            }
+            (_, None, Some(_)) => {
+                bail!("'bucket_name' option is mandatory if 'bucket_region' is given ")
+            }
+            (None, Some(bucket_name), Some(bucket_region)) => RemoteStorageKind::AwsS3(S3Config {
+                bucket_name: parse_toml_string("bucket_name", bucket_name)?,
+                bucket_region: parse_toml_string("bucket_region", bucket_region)?,
+                prefix_in_bucket: toml
+                    .get("prefix_in_bucket")
+                    .map(|prefix_in_bucket| parse_toml_string("prefix_in_bucket", prefix_in_bucket))
+                    .transpose()?,
+                endpoint: toml
+                    .get("endpoint")
+                    .map(|endpoint| parse_toml_string("endpoint", endpoint))
+                    .transpose()?,
+                concurrency_limit,
+            }),
+            (Some(local_path), None, None) => RemoteStorageKind::LocalFs(PathBuf::from(
+                parse_toml_string("local_path", local_path)?,
+            )),
+            (Some(_), Some(_), _) => bail!("local_path and bucket_name are mutually exclusive"),
+        };
+
+        Ok(RemoteStorageConfig {
+            max_concurrent_syncs,
+            max_sync_errors,
+            storage,
+        })
+    }
+}
+
+// Helper functions to parse a toml Item
+fn parse_optional_integer<I, E>(name: &str, item: &toml_edit::Item) -> anyhow::Result<Option<I>>
+where
+    I: TryFrom<i64, Error = E>,
+    E: std::error::Error + Send + Sync + 'static,
+{
+    let toml_integer = match item.get(name) {
+        Some(item) => item
+            .as_integer()
+            .with_context(|| format!("configure option {name} is not an integer"))?,
+        None => return Ok(None),
+    };
+
+    I::try_from(toml_integer)
+        .map(Some)
+        .with_context(|| format!("configure option {name} is too large"))
+}
+
+fn parse_toml_string(name: &str, item: &Item) -> anyhow::Result<String> {
+    let s = item
+        .as_str()
+        .with_context(|| format!("configure option {name} is not a string"))?;
+    Ok(s.to_string())
+}
+
 #[cfg(test)]
 mod tests {
     use super::*;

libs/remote_storage/src/s3_bucket.rs

@@ -23,6 +23,71 @@ use crate::{strip_path_prefix, RemoteStorage, S3Config};
 
 use super::StorageMetadata;
 
+pub(super) mod metrics {
+    use metrics::{register_int_counter_vec, IntCounterVec};
+    use once_cell::sync::Lazy;
+
+    static S3_REQUESTS_COUNT: Lazy<IntCounterVec> = Lazy::new(|| {
+        register_int_counter_vec!(
+            "remote_storage_s3_requests_count",
+            "Number of s3 requests of particular type",
+            &["request_type"],
+        )
+        .expect("failed to define a metric")
+    });
+
+    static S3_REQUESTS_FAIL_COUNT: Lazy<IntCounterVec> = Lazy::new(|| {
+        register_int_counter_vec!(
+            "remote_storage_s3_failures_count",
+            "Number of failed s3 requests of particular type",
+            &["request_type"],
+        )
+        .expect("failed to define a metric")
+    });
+
+    pub fn inc_get_object() {
+        S3_REQUESTS_COUNT.with_label_values(&["get_object"]).inc();
+    }
+
+    pub fn inc_get_object_fail() {
+        S3_REQUESTS_FAIL_COUNT
+            .with_label_values(&["get_object"])
+            .inc();
+    }
+
+    pub fn inc_put_object() {
+        S3_REQUESTS_COUNT.with_label_values(&["put_object"]).inc();
+    }
+
+    pub fn inc_put_object_fail() {
+        S3_REQUESTS_FAIL_COUNT
+            .with_label_values(&["put_object"])
+            .inc();
+    }
+
+    pub fn inc_delete_object() {
+        S3_REQUESTS_COUNT
+            .with_label_values(&["delete_object"])
+            .inc();
+    }
+
+    pub fn inc_delete_object_fail() {
+        S3_REQUESTS_FAIL_COUNT
+            .with_label_values(&["delete_object"])
+            .inc();
+    }
+
+    pub fn inc_list_objects() {
+        S3_REQUESTS_COUNT.with_label_values(&["list_objects"]).inc();
+    }
+
+    pub fn inc_list_objects_fail() {
+        S3_REQUESTS_FAIL_COUNT
+            .with_label_values(&["list_objects"])
+            .inc();
+    }
+}
+
 const S3_PREFIX_SEPARATOR: char = '/';
 
 #[derive(Debug, Eq, PartialEq, PartialOrd, Ord, Hash)]
@@ -152,6 +217,9 @@ impl RemoteStorage for S3Bucket {
                 .acquire()
                 .await
                 .context("Concurrency limiter semaphore got closed during S3 list")?;
+
+            metrics::inc_list_objects();
+
             let fetch_response = self
                 .client
                 .list_objects_v2(ListObjectsV2Request {
@@ -160,7 +228,11 @@ impl RemoteStorage for S3Bucket {
                     continuation_token,
                     ..ListObjectsV2Request::default()
                 })
-                .await?;
+                .await
+                .map_err(|e| {
+                    metrics::inc_list_objects_fail();
+                    e
+                })?;
             document_keys.extend(
                 fetch_response
                     .contents
@@ -190,6 +262,8 @@ impl RemoteStorage for S3Bucket {
             .acquire()
             .await
             .context("Concurrency limiter semaphore got closed during S3 upload")?;
+
+        metrics::inc_put_object();
         self.client
             .put_object(PutObjectRequest {
                 body: Some(StreamingBody::new_with_size(
@@ -201,7 +275,11 @@ impl RemoteStorage for S3Bucket {
                 metadata: metadata.map(|m| m.0),
                 ..PutObjectRequest::default()
             })
-            .await?;
+            .await
+            .map_err(|e| {
+                metrics::inc_put_object_fail();
+                e
+            })?;
         Ok(())
     }
 
@@ -215,6 +293,9 @@ impl RemoteStorage for S3Bucket {
             .acquire()
             .await
             .context("Concurrency limiter semaphore got closed during S3 download")?;
+
+        metrics::inc_get_object();
+
         let object_output = self
             .client
             .get_object(GetObjectRequest {
@@ -222,7 +303,11 @@ impl RemoteStorage for S3Bucket {
                 key: from.key().to_owned(),
                 ..GetObjectRequest::default()
             })
-            .await?;
+            .await
+            .map_err(|e| {
+                metrics::inc_get_object_fail();
+                e
+            })?;
 
         if let Some(body) = object_output.body {
             let mut from = io::BufReader::new(body.into_async_read());
@@ -251,6 +336,9 @@ impl RemoteStorage for S3Bucket {
             .acquire()
             .await
             .context("Concurrency limiter semaphore got closed during S3 range download")?;
+
+        metrics::inc_get_object();
+
         let object_output = self
             .client
             .get_object(GetObjectRequest {
@@ -259,7 +347,11 @@ impl RemoteStorage for S3Bucket {
                 range,
                 ..GetObjectRequest::default()
             })
-            .await?;
+            .await
+            .map_err(|e| {
+                metrics::inc_get_object_fail();
+                e
+            })?;
 
         if let Some(body) = object_output.body {
             let mut from = io::BufReader::new(body.into_async_read());
@@ -275,13 +367,20 @@ impl RemoteStorage for S3Bucket {
             .acquire()
             .await
             .context("Concurrency limiter semaphore got closed during S3 delete")?;
+
+        metrics::inc_delete_object();
+
         self.client
             .delete_object(DeleteObjectRequest {
                 bucket: self.bucket_name.clone(),
                 key: path.key().to_owned(),
                 ..DeleteObjectRequest::default()
             })
-            .await?;
+            .await
+            .map_err(|e| {
+                metrics::inc_delete_object_fail();
+                e
+            })?;
         Ok(())
     }
 }

libs/utils/src/lsn.rs

@@ -26,6 +26,9 @@ impl Lsn {
     /// Maximum possible value for an LSN
     pub const MAX: Lsn = Lsn(u64::MAX);
 
+    /// Invalid value for InvalidXLogRecPtr, as defined in xlogdefs.h
+    pub const INVALID: Lsn = Lsn(0);
+
     /// Subtract a number, returning None on overflow.
     pub fn checked_sub<T: Into<u64>>(self, other: T) -> Option<Lsn> {
         let other: u64 = other.into();
@@ -103,6 +106,12 @@ impl Lsn {
     pub fn is_aligned(&self) -> bool {
         *self == self.align()
     }
+
+    /// Return if the LSN is valid
+    /// mimics postgres XLogRecPtrIsInvalid macro
+    pub fn is_valid(self) -> bool {
+        self != Lsn::INVALID
+    }
 }
 
 impl From<u64> for Lsn {

libs/utils/src/zid.rs

@@ -218,7 +218,7 @@ impl ZTenantTimelineId {
 
 impl fmt::Display for ZTenantTimelineId {
     fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
-        write!(f, "{}-{}", self.tenant_id, self.timeline_id)
+        write!(f, "{}/{}", self.tenant_id, self.timeline_id)
     }
 }
 
@@ -226,9 +226,9 @@ impl fmt::Display for ZTenantTimelineId {
 // by the console.
 #[derive(Clone, Copy, Eq, Ord, PartialEq, PartialOrd, Hash, Debug, Serialize, Deserialize)]
 #[serde(transparent)]
-pub struct ZNodeId(pub u64);
+pub struct NodeId(pub u64);
 
-impl fmt::Display for ZNodeId {
+impl fmt::Display for NodeId {
     fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
         write!(f, "{}", self.0)
     }

neon_local/src/main.rs

@@ -22,14 +22,14 @@ use utils::{
     lsn::Lsn,
     postgres_backend::AuthType,
     project_git_version,
-    zid::{ZNodeId, ZTenantId, ZTenantTimelineId, ZTimelineId},
+    zid::{NodeId, ZTenantId, ZTenantTimelineId, ZTimelineId},
 };
 
 use pageserver::timelines::TimelineInfo;
 
 // Default id of a safekeeper node, if not specified on the command line.
-const DEFAULT_SAFEKEEPER_ID: ZNodeId = ZNodeId(1);
+const DEFAULT_SAFEKEEPER_ID: NodeId = NodeId(1);
-const DEFAULT_PAGESERVER_ID: ZNodeId = ZNodeId(1);
+const DEFAULT_PAGESERVER_ID: NodeId = NodeId(1);
 const DEFAULT_BRANCH_NAME: &str = "main";
 project_git_version!(GIT_VERSION);
 
@@ -860,7 +860,7 @@ fn handle_pageserver(sub_match: &ArgMatches, env: &local_env::LocalEnv) -> Resul
     Ok(())
 }
 
-fn get_safekeeper(env: &local_env::LocalEnv, id: ZNodeId) -> Result<SafekeeperNode> {
+fn get_safekeeper(env: &local_env::LocalEnv, id: NodeId) -> Result<SafekeeperNode> {
     if let Some(node) = env.safekeepers.iter().find(|node| node.id == id) {
         Ok(SafekeeperNode::from_env(env, node))
     } else {
@@ -876,7 +876,7 @@ fn handle_safekeeper(sub_match: &ArgMatches, env: &local_env::LocalEnv) -> Resul
 
     // All the commands take an optional safekeeper name argument
     let sk_id = if let Some(id_str) = sub_args.value_of("id") {
-        ZNodeId(id_str.parse().context("while parsing safekeeper id")?)
+        NodeId(id_str.parse().context("while parsing safekeeper id")?)
     } else {
         DEFAULT_SAFEKEEPER_ID
     };

pageserver/Cargo.toml

@@ -54,6 +54,9 @@ crossbeam-utils = "0.8.5"
 fail = "0.5.0"
 git-version = "0.3.5"
 
+# 'experimental' is needed for the `zstd::bulk::Decompressor::upper_bound` function.
+zstd = { version = "0.11.1", features = ["experimental"] }
+
 postgres_ffi = { path = "../libs/postgres_ffi" }
 etcd_broker = { path = "../libs/etcd_broker" }
 metrics = { path = "../libs/metrics" }

pageserver/src/basebackup.rs

@@ -10,8 +10,9 @@
 //! This module is responsible for creation of such tarball
 //! from data stored in object storage.
 //!
-use anyhow::{anyhow, ensure, Context, Result};
+use anyhow::{anyhow, bail, ensure, Context, Result};
 use bytes::{BufMut, BytesMut};
+use fail::fail_point;
 use std::fmt::Write as FmtWrite;
 use std::io;
 use std::io::Write;
@@ -30,11 +31,16 @@ use utils::lsn::Lsn;
 /// This is short-living object only for the time of tarball creation,
 /// created mostly to avoid passing a lot of parameters between various functions
 /// used for constructing tarball.
-pub struct Basebackup<'a> {
+pub struct Basebackup<'a, W>
-    ar: Builder<&'a mut dyn Write>,
+where
+    W: Write,
+{
+    ar: Builder<AbortableWrite<W>>,
     timeline: &'a Arc<DatadirTimelineImpl>,
     pub lsn: Lsn,
     prev_record_lsn: Lsn,
+
+    finished: bool,
 }
 
 // Create basebackup with non-rel data in it. Omit relational data.
@@ -44,12 +50,15 @@ pub struct Basebackup<'a> {
 //  * When working without safekeepers. In this situation it is important to match the lsn
 //    we are taking basebackup on with the lsn that is used in pageserver's walreceiver
 //    to start the replication.
-impl<'a> Basebackup<'a> {
+impl<'a, W> Basebackup<'a, W>
+where
+    W: Write,
+{
     pub fn new(
-        write: &'a mut dyn Write,
+        write: W,
         timeline: &'a Arc<DatadirTimelineImpl>,
         req_lsn: Option<Lsn>,
-    ) -> Result<Basebackup<'a>> {
+    ) -> Result<Basebackup<'a, W>> {
         // Compute postgres doesn't have any previous WAL files, but the first
         // record that it's going to write needs to include the LSN of the
         // previous record (xl_prev). We include prev_record_lsn in the
@@ -90,14 +99,15 @@ impl<'a> Basebackup<'a> {
         );
 
         Ok(Basebackup {
-            ar: Builder::new(write),
+            ar: Builder::new(AbortableWrite::new(write)),
             timeline,
             lsn: backup_lsn,
             prev_record_lsn: backup_prev,
+            finished: false,
         })
     }
 
-    pub fn send_tarball(&mut self) -> anyhow::Result<()> {
+    pub fn send_tarball(mut self) -> anyhow::Result<()> {
         // Create pgdata subdirs structure
         for dir in pg_constants::PGDATA_SUBDIRS.iter() {
             let header = new_tar_header_dir(*dir)?;
@@ -135,9 +145,14 @@ impl<'a> Basebackup<'a> {
             self.add_twophase_file(xid)?;
         }
 
+        fail_point!("basebackup-before-control-file", |_| {
+            bail!("failpoint basebackup-before-control-file")
+        });
+
         // Generate pg_control and bootstrap WAL segment.
         self.add_pgcontrol_file()?;
         self.ar.finish()?;
+        self.finished = true;
         debug!("all tarred up!");
         Ok(())
     }
@@ -331,6 +346,19 @@ impl<'a> Basebackup<'a> {
     }
 }
 
+impl<'a, W> Drop for Basebackup<'a, W>
+where
+    W: Write,
+{
+    /// If the basebackup was not finished, prevent the Archive::drop() from
+    /// writing the end-of-archive marker.
+    fn drop(&mut self) {
+        if !self.finished {
+            self.ar.get_mut().abort();
+        }
+    }
+}
+
 //
 // Create new tarball entry header
 //
@@ -366,3 +394,49 @@ fn new_tar_header_dir(path: &str) -> anyhow::Result<Header> {
     header.set_cksum();
     Ok(header)
 }
+
+/// A wrapper that passes through all data to the underlying Write,
+/// until abort() is called.
+///
+/// tar::Builder has an annoying habit of finishing the archive with
+/// a valid tar end-of-archive marker (two 512-byte sectors of zeros),
+/// even if an error occurs and we don't finish building the archive.
+/// We'd rather abort writing the tarball immediately than construct
+/// a seemingly valid but incomplete archive. This wrapper allows us
+/// to swallow the end-of-archive marker that Builder::drop() emits,
+/// without writing it to the underlying sink.
+///
+struct AbortableWrite<W> {
+    w: W,
+    aborted: bool,
+}
+
+impl<W> AbortableWrite<W> {
+    pub fn new(w: W) -> Self {
+        AbortableWrite { w, aborted: false }
+    }
+
+    pub fn abort(&mut self) {
+        self.aborted = true;
+    }
+}
+
+impl<W> Write for AbortableWrite<W>
+where
+    W: Write,
+{
+    fn write(&mut self, data: &[u8]) -> io::Result<usize> {
+        if self.aborted {
+            Ok(data.len())
+        } else {
+            self.w.write(data)
+        }
+    }
+    fn flush(&mut self) -> io::Result<()> {
+        if self.aborted {
+            Ok(())
+        } else {
+            self.w.flush()
+        }
+    }
+}

pageserver/src/config.rs

@@ -5,9 +5,9 @@
 //! See also `settings.md` for better description on every parameter.
 
 use anyhow::{anyhow, bail, ensure, Context, Result};
-use remote_storage::{RemoteStorageConfig, RemoteStorageKind, S3Config};
+use remote_storage::RemoteStorageConfig;
 use std::env;
-use std::num::{NonZeroU32, NonZeroUsize};
+
 use std::path::{Path, PathBuf};
 use std::str::FromStr;
 use std::time::Duration;
@@ -16,12 +16,18 @@ use toml_edit::{Document, Item};
 use url::Url;
 use utils::{
     postgres_backend::AuthType,
-    zid::{ZNodeId, ZTenantId, ZTimelineId},
+    zid::{NodeId, ZTenantId, ZTimelineId},
 };
 
 use crate::layered_repository::TIMELINES_SEGMENT_NAME;
 use crate::tenant_config::{TenantConf, TenantConfOpt};
 
+pub const ZSTD_MAX_SAMPLES: usize = 1024;
+pub const ZSTD_MIN_SAMPLES: usize = 8; // magic requirement of zstd
+pub const ZSTD_MAX_SAMPLE_BYTES: usize = 10 * 1024 * 1024; // max memory size for holding samples
+pub const ZSTD_MAX_DICTIONARY_SIZE: usize = 8 * 1024 - 4; // make dictionary + BLOB length fit in first page
+pub const ZSTD_COMPRESSION_LEVEL: i32 = 0; // default compression level
+
 pub mod defaults {
     use crate::tenant_config::defaults::*;
     use const_format::formatcp;
@@ -78,7 +84,7 @@ pub mod defaults {
 pub struct PageServerConf {
     // Identifier of that particular pageserver so e g safekeepers
     // can safely distinguish different pageservers
-    pub id: ZNodeId,
+    pub id: NodeId,
 
     /// Example (default): 127.0.0.1:64000
     pub listen_pg_addr: String,
@@ -180,7 +186,7 @@ struct PageServerConfigBuilder {
     auth_validation_public_key_path: BuilderValue<Option<PathBuf>>,
     remote_storage_config: BuilderValue<Option<RemoteStorageConfig>>,
 
-    id: BuilderValue<ZNodeId>,
+    id: BuilderValue<NodeId>,
 
     profiling: BuilderValue<ProfilingConfig>,
     broker_etcd_prefix: BuilderValue<String>,
@@ -276,7 +282,7 @@ impl PageServerConfigBuilder {
         self.broker_etcd_prefix = BuilderValue::Set(broker_etcd_prefix)
     }
 
-    pub fn id(&mut self, node_id: ZNodeId) {
+    pub fn id(&mut self, node_id: NodeId) {
         self.id = BuilderValue::Set(node_id)
     }
 
@@ -394,12 +400,12 @@ impl PageServerConf {
                 )),
                 "auth_type" => builder.auth_type(parse_toml_from_str(key, item)?),
                 "remote_storage" => {
-                    builder.remote_storage_config(Some(Self::parse_remote_storage_config(item)?))
+                    builder.remote_storage_config(Some(RemoteStorageConfig::from_toml(item)?))
                 }
                 "tenant_config" => {
                     t_conf = Self::parse_toml_tenant_conf(item)?;
                 }
-                "id" => builder.id(ZNodeId(parse_toml_u64(key, item)?)),
+                "id" => builder.id(NodeId(parse_toml_u64(key, item)?)),
                 "profiling" => builder.profiling(parse_toml_from_str(key, item)?),
                 "broker_etcd_prefix" => builder.broker_etcd_prefix(parse_toml_string(key, item)?),
                 "broker_endpoints" => builder.broker_endpoints(
@@ -484,64 +490,6 @@ impl PageServerConf {
         Ok(t_conf)
     }
 
-    /// subroutine of parse_config(), to parse the `[remote_storage]` table.
-    fn parse_remote_storage_config(toml: &toml_edit::Item) -> anyhow::Result<RemoteStorageConfig> {
-        let local_path = toml.get("local_path");
-        let bucket_name = toml.get("bucket_name");
-        let bucket_region = toml.get("bucket_region");
-
-        let max_concurrent_syncs = NonZeroUsize::new(
-            parse_optional_integer("max_concurrent_syncs", toml)?
-                .unwrap_or(remote_storage::DEFAULT_REMOTE_STORAGE_MAX_CONCURRENT_SYNCS),
-        )
-        .context("Failed to parse 'max_concurrent_syncs' as a positive integer")?;
-
-        let max_sync_errors = NonZeroU32::new(
-            parse_optional_integer("max_sync_errors", toml)?
-                .unwrap_or(remote_storage::DEFAULT_REMOTE_STORAGE_MAX_SYNC_ERRORS),
-        )
-        .context("Failed to parse 'max_sync_errors' as a positive integer")?;
-
-        let concurrency_limit = NonZeroUsize::new(
-            parse_optional_integer("concurrency_limit", toml)?
-                .unwrap_or(remote_storage::DEFAULT_REMOTE_STORAGE_S3_CONCURRENCY_LIMIT),
-        )
-        .context("Failed to parse 'concurrency_limit' as a positive integer")?;
-
-        let storage = match (local_path, bucket_name, bucket_region) {
-            (None, None, None) => bail!("no 'local_path' nor 'bucket_name' option"),
-            (_, Some(_), None) => {
-                bail!("'bucket_region' option is mandatory if 'bucket_name' is given ")
-            }
-            (_, None, Some(_)) => {
-                bail!("'bucket_name' option is mandatory if 'bucket_region' is given ")
-            }
-            (None, Some(bucket_name), Some(bucket_region)) => RemoteStorageKind::AwsS3(S3Config {
-                bucket_name: parse_toml_string("bucket_name", bucket_name)?,
-                bucket_region: parse_toml_string("bucket_region", bucket_region)?,
-                prefix_in_bucket: toml
-                    .get("prefix_in_bucket")
-                    .map(|prefix_in_bucket| parse_toml_string("prefix_in_bucket", prefix_in_bucket))
-                    .transpose()?,
-                endpoint: toml
-                    .get("endpoint")
-                    .map(|endpoint| parse_toml_string("endpoint", endpoint))
-                    .transpose()?,
-                concurrency_limit,
-            }),
-            (Some(local_path), None, None) => RemoteStorageKind::LocalFs(PathBuf::from(
-                parse_toml_string("local_path", local_path)?,
-            )),
-            (Some(_), Some(_), _) => bail!("local_path and bucket_name are mutually exclusive"),
-        };
-
-        Ok(RemoteStorageConfig {
-            max_concurrent_syncs,
-            max_sync_errors,
-            storage,
-        })
-    }
-
     #[cfg(test)]
     pub fn test_repo_dir(test_name: &str) -> PathBuf {
         PathBuf::from(format!("../tmp_check/test_{test_name}"))
@@ -550,7 +498,7 @@ impl PageServerConf {
     #[cfg(test)]
     pub fn dummy_conf(repo_dir: PathBuf) -> Self {
         PageServerConf {
-            id: ZNodeId(0),
+            id: NodeId(0),
             wait_lsn_timeout: Duration::from_secs(60),
             wal_redo_timeout: Duration::from_secs(60),
             page_cache_size: defaults::DEFAULT_PAGE_CACHE_SIZE,
@@ -592,23 +540,6 @@ fn parse_toml_u64(name: &str, item: &Item) -> Result<u64> {
     Ok(i as u64)
 }
 
-fn parse_optional_integer<I, E>(name: &str, item: &toml_edit::Item) -> anyhow::Result<Option<I>>
-where
-    I: TryFrom<i64, Error = E>,
-    E: std::error::Error + Send + Sync + 'static,
-{
-    let toml_integer = match item.get(name) {
-        Some(item) => item
-            .as_integer()
-            .with_context(|| format!("configure option {name} is not an integer"))?,
-        None => return Ok(None),
-    };
-
-    I::try_from(toml_integer)
-        .map(Some)
-        .with_context(|| format!("configure option {name} is too large"))
-}
-
 fn parse_toml_duration(name: &str, item: &Item) -> Result<Duration> {
     let s = item
         .as_str()
@@ -651,8 +582,12 @@ fn parse_toml_array(name: &str, item: &Item) -> anyhow::Result<Vec<String>> {
 
 #[cfg(test)]
 mod tests {
-    use std::fs;
+    use std::{
+        fs,
+        num::{NonZeroU32, NonZeroUsize},
+    };
 
+    use remote_storage::{RemoteStorageKind, S3Config};
     use tempfile::{tempdir, TempDir};
 
     use super::*;
@@ -693,7 +628,7 @@ id = 10
         assert_eq!(
             parsed_config,
             PageServerConf {
-                id: ZNodeId(10),
+                id: NodeId(10),
                 listen_pg_addr: defaults::DEFAULT_PG_LISTEN_ADDR.to_string(),
                 listen_http_addr: defaults::DEFAULT_HTTP_LISTEN_ADDR.to_string(),
                 wait_lsn_timeout: humantime::parse_duration(defaults::DEFAULT_WAIT_LSN_TIMEOUT)?,
@@ -737,7 +672,7 @@ id = 10
         assert_eq!(
             parsed_config,
             PageServerConf {
-                id: ZNodeId(10),
+                id: NodeId(10),
                 listen_pg_addr: "127.0.0.1:64000".to_string(),
                 listen_http_addr: "127.0.0.1:9898".to_string(),
                 wait_lsn_timeout: Duration::from_secs(111),

pageserver/src/http/models.rs

@@ -2,7 +2,7 @@ use serde::{Deserialize, Serialize};
 use serde_with::{serde_as, DisplayFromStr};
 use utils::{
     lsn::Lsn,
-    zid::{ZNodeId, ZTenantId, ZTimelineId},
+    zid::{NodeId, ZTenantId, ZTimelineId},
 };
 
 #[serde_as]
@@ -42,7 +42,7 @@ pub struct TenantCreateResponse(#[serde_as(as = "DisplayFromStr")] pub ZTenantId
 
 #[derive(Serialize)]
 pub struct StatusResponse {
-    pub id: ZNodeId,
+    pub id: NodeId,
 }
 
 impl TenantCreateRequest {

pageserver/src/layered_repository.rs

@@ -1230,7 +1230,7 @@ impl LayeredTimeline {
             }),
             disk_consistent_lsn: AtomicLsn::new(metadata.disk_consistent_lsn().0),
 
-            last_freeze_at: AtomicLsn::new(0),
+            last_freeze_at: AtomicLsn::new(metadata.disk_consistent_lsn().0),
 
             ancestor_timeline: ancestor,
             ancestor_lsn: metadata.ancestor_lsn(),
@@ -2518,7 +2518,7 @@ fn rename_to_backup(path: PathBuf) -> anyhow::Result<()> {
     bail!("couldn't find an unused backup number for {:?}", path)
 }
 
-fn load_metadata(
+pub fn load_metadata(
     conf: &'static PageServerConf,
     timeline_id: ZTimelineId,
     tenant_id: ZTenantId,

pageserver/src/layered_repository/delta_layer.rs

@@ -23,6 +23,25 @@
 //! "values" part.  The actual page images and WAL records are stored in the
 //! "values" part.
 //!
+//! # Compression
+//!
+//! Each value is stored as a Blob, which can optionally be compressed. Compression
+//! is done by ZStandard, in dictionary mode, which gives pretty good compression
+//! ratio even for small inputs like WAL records.
+//!
+//! The dictionary is built separately for each delta layer file, and stored in
+//! the file itself.
+//!
+//! TODO: The ZStandard format includes constant 4-byte "magic bytes" in the beginning
+//! of each compressed block. With small values like WAL records, that's pretty wasteful.
+//! We could disable those bytes by setting the `include_magibytes' flag to false,
+//! but as of this writing that's considered experimental in the zstd crate, and the
+//! zstd::bulk::Decompressor::upper_bound() function doesn't work without the magic bytes
+//! so we would have to find a different way of allocating the decompression buffer if
+//! we did that.
+//!
+
+use crate::config;
 use crate::config::PageServerConf;
 use crate::layered_repository::blob_io::{BlobCursor, BlobWriter, WriteBlobWriter};
 use crate::layered_repository::block_io::{BlockBuf, BlockCursor, BlockReader, FileBlockReader};
@@ -36,7 +55,7 @@ use crate::repository::{Key, Value, KEY_SIZE};
 use crate::virtual_file::VirtualFile;
 use crate::walrecord;
 use crate::{DELTA_FILE_MAGIC, STORAGE_FORMAT_VERSION};
-use anyhow::{bail, ensure, Context, Result};
+use anyhow::{anyhow, bail, ensure, Context, Result};
 use rand::{distributions::Alphanumeric, Rng};
 use serde::{Deserialize, Serialize};
 use std::fs;
@@ -75,6 +94,9 @@ struct Summary {
     index_start_blk: u32,
     /// Block within the 'index', where the B-tree root page is stored
     index_root_blk: u32,
+
+    /// Byte offset of the compression dictionary, or 0 if no compression
+    dictionary_offset: u64,
 }
 
 impl From<&DeltaLayer> for Summary {
@@ -90,33 +112,46 @@ impl From<&DeltaLayer> for Summary {
 
             index_start_blk: 0,
             index_root_blk: 0,
+
+            dictionary_offset: 0,
         }
     }
 }
 
-// Flag indicating that this version initialize the page
-const WILL_INIT: u64 = 1;
-
 ///
-/// Struct representing reference to BLOB in layers. Reference contains BLOB
+/// Struct representing reference to BLOB in the file. The reference contains
-/// offset, and for WAL records it also contains `will_init` flag. The flag
+/// the offset to the BLOB within the file, a flag indicating if it's
+/// compressed or not, and also the `will_init` flag. The `will_init` flag
 /// helps to determine the range of records that needs to be applied, without
 /// reading/deserializing records themselves.
 ///
 #[derive(Debug, Serialize, Deserialize, Copy, Clone)]
 struct BlobRef(u64);
 
+/// Flag indicating that this blob is compressed
+const BLOB_COMPRESSED: u64 = 1;
+
+/// Flag indicating that this version initializes the page
+const WILL_INIT: u64 = 2;
+
 impl BlobRef {
+    pub fn compressed(&self) -> bool {
+        (self.0 & BLOB_COMPRESSED) != 0
+    }
+
     pub fn will_init(&self) -> bool {
         (self.0 & WILL_INIT) != 0
     }
 
     pub fn pos(&self) -> u64 {
-        self.0 >> 1
+        self.0 >> 2
     }
 
-    pub fn new(pos: u64, will_init: bool) -> BlobRef {
+    pub fn new(pos: u64, compressed: bool, will_init: bool) -> BlobRef {
-        let mut blob_ref = pos << 1;
+        let mut blob_ref = pos << 2;
+        if compressed {
+            blob_ref |= BLOB_COMPRESSED;
+        }
         if will_init {
             blob_ref |= WILL_INIT;
         }
@@ -193,6 +228,37 @@ pub struct DeltaLayerInner {
 
     /// Reader object for reading blocks from the file. (None if not loaded yet)
     file: Option<FileBlockReader<VirtualFile>>,
+
+    /// Compression dictionary, as raw bytes, and in prepared format ready for use
+    /// for decompression. None if there is no dictionary, or if 'loaded' is false.
+    dictionary: Option<(Vec<u8>, zstd::dict::DecoderDictionary<'static>)>,
+}
+
+impl DeltaLayerInner {
+    // Create a new Decompressor, using the prepared dictionary
+    fn create_decompressor(&self) -> Result<Option<zstd::bulk::Decompressor<'_>>> {
+        if let Some((_, dict)) = &self.dictionary {
+            let decompressor = zstd::bulk::Decompressor::with_prepared_dictionary(dict)?;
+            Ok(Some(decompressor))
+        } else {
+            Ok(None)
+        }
+    }
+
+    // Create a new Decompressor, without using the prepared dictionary.
+    //
+    // For the cases that you cannot use 'create_decompressor', if the
+    // Decompressor needs to outlive 'self'.
+    fn create_decompressor_not_prepared(
+        &self,
+    ) -> Result<Option<zstd::bulk::Decompressor<'static>>> {
+        if let Some((dict, _)) = &self.dictionary {
+            let decompressor = zstd::bulk::Decompressor::with_dictionary(dict)?;
+            Ok(Some(decompressor))
+        } else {
+            Ok(None)
+        }
+    }
 }
 
 impl Layer for DeltaLayer {
@@ -234,6 +300,8 @@ impl Layer for DeltaLayer {
         {
             // Open the file and lock the metadata in memory
             let inner = self.load()?;
+            let mut decompressor = inner.create_decompressor()?;
+            let mut decompress_buf = Vec::new();
 
             // Scan the page versions backwards, starting from `lsn`.
             let file = inner.file.as_ref().unwrap();
@@ -244,7 +312,7 @@ impl Layer for DeltaLayer {
             );
             let search_key = DeltaKey::from_key_lsn(&key, Lsn(lsn_range.end.0 - 1));
 
-            let mut offsets: Vec<(Lsn, u64)> = Vec::new();
+            let mut blob_refs: Vec<(Lsn, BlobRef)> = Vec::new();
 
             tree_reader.visit(&search_key.0, VisitDirection::Backwards, |key, value| {
                 let blob_ref = BlobRef(value);
@@ -255,21 +323,36 @@ impl Layer for DeltaLayer {
                 if entry_lsn < lsn_range.start {
                     return false;
                 }
-                offsets.push((entry_lsn, blob_ref.pos()));
+                blob_refs.push((entry_lsn, blob_ref));
 
                 !blob_ref.will_init()
             })?;
 
             // Ok, 'offsets' now contains the offsets of all the entries we need to read
             let mut cursor = file.block_cursor();
-            for (entry_lsn, pos) in offsets {
+            for (entry_lsn, blob_ref) in blob_refs {
-                let buf = cursor.read_blob(pos).with_context(|| {
+                let buf = cursor.read_blob(blob_ref.pos()).with_context(|| {
                     format!(
                         "Failed to read blob from virtual file {}",
                         file.file.path.display()
                     )
                 })?;
-                let val = Value::des(&buf).with_context(|| {
+                let uncompressed_bytes = if blob_ref.compressed() {
+                    if let Some(ref mut decompressor) = decompressor {
+                        let decompressed_max_len = zstd::bulk::Decompressor::upper_bound(&buf)
+                            .ok_or_else(|| anyhow!("could not get decompressed length"))?;
+                        decompress_buf.clear();
+                        decompress_buf.reserve(decompressed_max_len);
+                        let _ = decompressor.decompress_to_buffer(&buf, &mut decompress_buf)?;
+                        &decompress_buf
+                    } else {
+                        bail!("blob is compressed, but there was no dictionary");
+                    }
+                } else {
+                    &buf
+                };
+
+                let val = Value::des(uncompressed_bytes).with_context(|| {
                     format!(
                         "Failed to deserialize file blob from virtual file {}",
                         file.file.path.display()
@@ -347,7 +430,6 @@ impl Layer for DeltaLayer {
         }
 
         let inner = self.load()?;
-
         println!(
             "index_start_blk: {}, root {}",
             inner.index_start_blk, inner.index_root_blk
@@ -363,19 +445,49 @@ impl Layer for DeltaLayer {
         tree_reader.dump()?;
 
         let mut cursor = file.block_cursor();
+        let mut decompressor = inner.create_decompressor()?;
+        let mut decompress_buf = Vec::new();
 
         // A subroutine to dump a single blob
         let mut dump_blob = |blob_ref: BlobRef| -> anyhow::Result<String> {
-            let buf = cursor.read_blob(blob_ref.pos())?;
+            let buf = cursor.read_blob(blob_ref.pos()).with_context(|| {
-            let val = Value::des(&buf)?;
+                format!(
+                    "Failed to read blob from virtual file {}",
+                    file.file.path.display()
+                )
+            })?;
+
+            let uncompressed_bytes = if blob_ref.compressed() {
+                if let Some(ref mut decompressor) = decompressor {
+                    let decompressed_max_len = zstd::bulk::Decompressor::upper_bound(&buf)
+                        .ok_or_else(|| anyhow!("could not get decompressed length"))?;
+                    decompress_buf.clear();
+                    decompress_buf.reserve(decompressed_max_len);
+                    let _ = decompressor.decompress_to_buffer(&buf, &mut decompress_buf)?;
+                    &decompress_buf
+                } else {
+                    bail!("blob is compressed, but there was no dictionary");
+                }
+            } else {
+                &buf
+            };
+
+            let val = Value::des(uncompressed_bytes).with_context(|| {
+                format!(
+                    "Failed to deserialize file blob from virtual file {}",
+                    file.file.path.display()
+                )
+            })?;
+
             let desc = match val {
                 Value::Image(img) => {
-                    format!(" img {} bytes", img.len())
+                    format!("img {} bytes, {} compressed", img.len(), buf.len())
                 }
                 Value::WalRecord(rec) => {
                     let wal_desc = walrecord::describe_wal_record(&rec)?;
                     format!(
-                        " rec {} bytes will_init: {} {}",
+                        "rec {} bytes, {} compressed, will_init {}: {}",
+                        uncompressed_bytes.len(),
                         buf.len(),
                         rec.will_init(),
                         wal_desc
@@ -494,6 +606,7 @@ impl DeltaLayer {
                 let mut expected_summary = Summary::from(self);
                 expected_summary.index_start_blk = actual_summary.index_start_blk;
                 expected_summary.index_root_blk = actual_summary.index_root_blk;
+                expected_summary.dictionary_offset = actual_summary.dictionary_offset;
                 if actual_summary != expected_summary {
                     bail!("in-file summary does not match expected summary. actual = {:?} expected = {:?}", actual_summary, expected_summary);
                 }
@@ -512,6 +625,13 @@ impl DeltaLayer {
             }
         }
 
+        // Load and prepare the dictionary, if any
+        if actual_summary.dictionary_offset != 0 {
+            let mut cursor = file.block_cursor();
+            let dict = cursor.read_blob(actual_summary.dictionary_offset)?;
+            let prepared_dict = zstd::dict::DecoderDictionary::copy(&dict);
+            inner.dictionary = Some((dict, prepared_dict));
+        }
         inner.index_start_blk = actual_summary.index_start_blk;
         inner.index_root_blk = actual_summary.index_root_blk;
 
@@ -537,6 +657,7 @@ impl DeltaLayer {
             inner: RwLock::new(DeltaLayerInner {
                 loaded: false,
                 file: None,
+                dictionary: None,
                 index_start_blk: 0,
                 index_root_blk: 0,
             }),
@@ -564,6 +685,7 @@ impl DeltaLayer {
             inner: RwLock::new(DeltaLayerInner {
                 loaded: false,
                 file: None,
+                dictionary: None,
                 index_start_blk: 0,
                 index_root_blk: 0,
             }),
@@ -599,6 +721,16 @@ impl DeltaLayer {
 ///
 /// 3. Call `finish`.
 ///
+///
+/// To train the dictionary for compression, the first ZSTD_MAX_SAMPLES values
+/// (or up ZSTD_MAX_SAMPLE_BYTES) are buffered in memory, before writing them
+/// to disk. When the "sample buffer" fills up, the buffered values are used
+/// to train a zstandard dictionary, which is then used to compress all the
+/// buffered values, and all subsequent values. So the dictionary is built
+/// based on just the first values, but in practice that usually gives pretty
+/// good compression for all subsequent data as well. Things like page and
+/// tuple headers are similar across all pages of the same relation.
+///
 pub struct DeltaLayerWriter {
     conf: &'static PageServerConf,
     path: PathBuf,
@@ -611,6 +743,13 @@ pub struct DeltaLayerWriter {
     tree: DiskBtreeBuilder<BlockBuf, DELTA_KEY_SIZE>,
 
     blob_writer: WriteBlobWriter<BufWriter<VirtualFile>>,
+    compressor: Option<zstd::bulk::Compressor<'static>>,
+    dictionary_offset: u64,
+
+    training: bool,
+    sample_key_lsn_willinit: Vec<(Key, Lsn, bool)>,
+    sample_sizes: Vec<usize>,
+    sample_data: Vec<u8>,
 }
 
 impl DeltaLayerWriter {
@@ -641,7 +780,6 @@ impl DeltaLayerWriter {
         // Initialize the b-tree index builder
         let block_buf = BlockBuf::new();
         let tree_builder = DiskBtreeBuilder::new(block_buf);
-
         Ok(DeltaLayerWriter {
             conf,
             path,
@@ -651,6 +789,13 @@ impl DeltaLayerWriter {
             lsn_range,
             tree: tree_builder,
             blob_writer,
+            compressor: None,
+            dictionary_offset: 0,
+
+            training: true,
+            sample_key_lsn_willinit: Vec::new(),
+            sample_sizes: Vec::new(),
+            sample_data: Vec::new(),
         })
     }
 
@@ -660,18 +805,122 @@ impl DeltaLayerWriter {
     /// The values must be appended in key, lsn order.
     ///
     pub fn put_value(&mut self, key: Key, lsn: Lsn, val: Value) -> Result<()> {
+        let blob_content = &Value::ser(&val)?;
+
+        // Are we still accumulating values for training the compression dictionary?
+        if self.training {
+            self.put_value_train(key, lsn, val.will_init(), blob_content)?;
+
+            if self.sample_sizes.len() >= config::ZSTD_MAX_SAMPLES
+                || self.sample_data.len() >= config::ZSTD_MAX_SAMPLE_BYTES
+            {
+                self.finish_training()?;
+            }
+        } else {
+            self.put_value_flush(key, lsn, val.will_init(), blob_content)?;
+        }
+        Ok(())
+    }
+
+    /// Accumulate one key-value pair in the samples buffer
+    fn put_value_train(&mut self, key: Key, lsn: Lsn, will_init: bool, bytes: &[u8]) -> Result<()> {
+        assert!(self.training);
+        self.sample_key_lsn_willinit.push((key, lsn, will_init));
+        self.sample_sizes.push(bytes.len());
+        self.sample_data.extend_from_slice(bytes);
+        Ok(())
+    }
+
+    /// Train the compression dictionary, and flush out all the accumulated
+    /// key-value pairs to disk.
+    fn finish_training(&mut self) -> Result<()> {
+        assert!(self.training);
+        assert!(self.sample_sizes.len() == self.sample_key_lsn_willinit.len());
+
+        // Create the dictionary, if we had enough samples for it.
+        //
+        // If there weren't enough samples, we don't do any compression at
+        // all. Possibly we could still benefit from compression; for example
+        // if you have only one gigantic value in a single layer, it would
+        // still be good to compress that, without a dictionary. But we don't
+        // do that currently.
+        if self.sample_sizes.len() >= config::ZSTD_MIN_SAMPLES {
+            let dictionary = zstd::dict::from_continuous(
+                &self.sample_data,
+                &self.sample_sizes,
+                config::ZSTD_MAX_DICTIONARY_SIZE,
+            )?;
+
+            let off = self.blob_writer.write_blob(&dictionary)?;
+            self.dictionary_offset = off;
+
+            let compressor = zstd::bulk::Compressor::with_dictionary(
+                config::ZSTD_COMPRESSION_LEVEL,
+                &dictionary,
+            )?;
+            self.compressor = Some(compressor);
+        };
+        self.training = false;
+
+        // release the memory used by the sample buffers
+        let sample_key_lsn_willinit = std::mem::take(&mut self.sample_key_lsn_willinit);
+        let sample_sizes = std::mem::take(&mut self.sample_sizes);
+        let sample_data = std::mem::take(&mut self.sample_data);
+
+        // Compress and write out all the buffered key-value pairs
+        let mut buf_idx: usize = 0;
+        for ((key, lsn, will_init), len) in
+            itertools::izip!(sample_key_lsn_willinit.iter(), sample_sizes.iter())
+        {
+            let end = buf_idx + len;
+            self.put_value_flush(*key, *lsn, *will_init, &sample_data[buf_idx..end])?;
+            buf_idx = end;
+        }
+        assert!(buf_idx == sample_data.len());
+
+        Ok(())
+    }
+
+    /// Write a key-value pair to the file, compressing it if applicable.
+    pub fn put_value_flush(
+        &mut self,
+        key: Key,
+        lsn: Lsn,
+        will_init: bool,
+        bytes: &[u8],
+    ) -> Result<()> {
+        assert!(!self.training);
         assert!(self.lsn_range.start <= lsn);
 
-        let off = self.blob_writer.write_blob(&Value::ser(&val)?)?;
+        let mut blob_content = bytes;
+        let mut compressed = false;
 
-        let blob_ref = BlobRef::new(off, val.will_init());
+        // Try to compress the blob
+        let compressed_bytes;
+        if let Some(ref mut compressor) = self.compressor {
+            compressed_bytes = compressor.compress(blob_content)?;
+            // If compressed version is not any smaller than the original,
+            // store it uncompressed.
+            if compressed_bytes.len() < blob_content.len() {
+                blob_content = &compressed_bytes;
+                compressed = true;
+            }
+        }
 
+        // Write it to the file
+        let off = self.blob_writer.write_blob(blob_content)?;
+        let blob_ref = BlobRef::new(off, compressed, will_init);
+
+        // And store the reference in the B-tree
         let delta_key = DeltaKey::from_key_lsn(&key, lsn);
         self.tree.append(&delta_key.0, blob_ref.0)?;
 
         Ok(())
     }
 
+    ///
+    /// Return an estimate of the file, if it was finished now.
+    ///
     pub fn size(&self) -> u64 {
         self.blob_writer.size() + self.tree.borrow_writer().size()
     }
@@ -679,7 +928,11 @@ impl DeltaLayerWriter {
     ///
     /// Finish writing the delta layer.
     ///
-    pub fn finish(self, key_end: Key) -> anyhow::Result<DeltaLayer> {
+    pub fn finish(mut self, key_end: Key) -> anyhow::Result<DeltaLayer> {
+        if self.training {
+            self.finish_training()?;
+        }
+
         let index_start_blk =
             ((self.blob_writer.size() + PAGE_SZ as u64 - 1) / PAGE_SZ as u64) as u32;
 
@@ -703,6 +956,7 @@ impl DeltaLayerWriter {
             lsn_range: self.lsn_range.clone(),
             index_start_blk,
             index_root_blk,
+            dictionary_offset: self.dictionary_offset,
         };
         file.seek(SeekFrom::Start(0))?;
         Summary::ser_into(&summary, &mut file)?;
@@ -719,6 +973,7 @@ impl DeltaLayerWriter {
             inner: RwLock::new(DeltaLayerInner {
                 loaded: false,
                 file: None,
+                dictionary: None,
                 index_start_blk,
                 index_root_blk,
             }),
@@ -758,6 +1013,9 @@ struct DeltaValueIter<'a> {
     all_offsets: Vec<(DeltaKey, BlobRef)>,
     next_idx: usize,
     reader: BlockCursor<Adapter<'a>>,
+    decompressor: Option<zstd::bulk::Decompressor<'a>>,
+
+    decompress_buf: Vec<u8>,
 }
 
 struct Adapter<'a>(RwLockReadGuard<'a, DeltaLayerInner>);
@@ -797,10 +1055,20 @@ impl<'a> DeltaValueIter<'a> {
             },
         )?;
 
+        // We cannot use inner.create_decompressor() here, because it returns
+        // a Decompressor with lifetime that depends on 'inner', and that
+        // doesn't live long enough here. Cannot use the prepared dictionary
+        // for that reason either. Doesn't matter too much in practice because
+        // this Iterator is used for bulk operations, and loading the dictionary
+        // isn't that expensive in comparison.
+        let decompressor = inner.create_decompressor_not_prepared()?;
+
         let iter = DeltaValueIter {
             all_offsets,
             next_idx: 0,
             reader: BlockCursor::new(Adapter(inner)),
+            decompressor,
+            decompress_buf: Vec::new(),
         };
 
         Ok(iter)
@@ -814,7 +1082,31 @@ impl<'a> DeltaValueIter<'a> {
             let lsn = delta_key.lsn();
 
             let buf = self.reader.read_blob(blob_ref.pos())?;
-            let val = Value::des(&buf)?;
+            let uncompressed_bytes = if blob_ref.compressed() {
+                if let Some(decompressor) = &mut self.decompressor {
+                    let decompressed_max_len = zstd::bulk::Decompressor::upper_bound(&buf)
+                        .ok_or_else(|| {
+                            anyhow!(
+                                "could not get decompressed length at offset {}",
+                                blob_ref.pos()
+                            )
+                        })?;
+                    self.decompress_buf.clear();
+                    self.decompress_buf.reserve(decompressed_max_len);
+                    let _ = decompressor.decompress_to_buffer(&buf, &mut self.decompress_buf)?;
+                    &self.decompress_buf
+                } else {
+                    bail!("blob is compressed, but there was no dictionary");
+                }
+            } else {
+                &buf
+            };
+            let val = Value::des(uncompressed_bytes).with_context(|| {
+                format!(
+                    "Failed to deserialize file blob at offset {}",
+                    blob_ref.pos()
+                )
+            })?;
             self.next_idx += 1;
             Ok(Some((key, lsn, val)))
         } else {

pageserver/src/layered_repository/image_layer.rs

@@ -19,6 +19,11 @@
 //! layer, and offsets to the other parts. The "index" is a B-tree,
 //! mapping from Key to an offset in the "values" part.  The
 //! actual page images are stored in the "values" part.
+//!
+//! Each page image is compressed with ZStandard. See Compression section
+//! in the delta_layer.rs for more discussion. Difference from a delta
+//! layer is that we don't currently use a dictionary for image layers.
+use crate::config;
 use crate::config::PageServerConf;
 use crate::layered_repository::blob_io::{BlobCursor, BlobWriter, WriteBlobWriter};
 use crate::layered_repository::block_io::{BlockBuf, BlockReader, FileBlockReader};
@@ -90,6 +95,35 @@ impl From<&ImageLayer> for Summary {
     }
 }
 
+///
+/// Struct representing reference to BLOB in the file. In an image layer,
+/// each blob is an image of the page. It can be compressed or not, and
+/// that is stored in low bit of the BlobRef.
+///
+#[derive(Debug, Serialize, Deserialize, Copy, Clone)]
+struct BlobRef(u64);
+
+/// Flag indicating that this blob is compressed
+const BLOB_COMPRESSED: u64 = 1;
+
+impl BlobRef {
+    pub fn compressed(&self) -> bool {
+        (self.0 & BLOB_COMPRESSED) != 0
+    }
+
+    pub fn pos(&self) -> u64 {
+        self.0 >> 1
+    }
+
+    pub fn new(pos: u64, compressed: bool) -> BlobRef {
+        let mut blob_ref = pos << 1;
+        if compressed {
+            blob_ref |= BLOB_COMPRESSED;
+        }
+        BlobRef(blob_ref)
+    }
+}
+
 ///
 /// ImageLayer is the in-memory data structure associated with an on-disk image
 /// file.  We keep an ImageLayer in memory for each file, in the LayerMap. If a
@@ -121,6 +155,13 @@ pub struct ImageLayerInner {
     file: Option<FileBlockReader<VirtualFile>>,
 }
 
+impl ImageLayerInner {
+    fn create_decompressor(&self) -> Result<zstd::bulk::Decompressor<'_>> {
+        let decompressor = zstd::bulk::Decompressor::new()?;
+        Ok(decompressor)
+    }
+}
+
 impl Layer for ImageLayer {
     fn filename(&self) -> PathBuf {
         PathBuf::from(self.layer_name().to_string())
@@ -160,20 +201,33 @@ impl Layer for ImageLayer {
 
         let inner = self.load()?;
 
+        let mut decompressor = inner.create_decompressor()?;
+
         let file = inner.file.as_ref().unwrap();
         let tree_reader = DiskBtreeReader::new(inner.index_start_blk, inner.index_root_blk, file);
 
         let mut keybuf: [u8; KEY_SIZE] = [0u8; KEY_SIZE];
         key.write_to_byte_slice(&mut keybuf);
-        if let Some(offset) = tree_reader.get(&keybuf)? {
+        if let Some(value) = tree_reader.get(&keybuf)? {
-            let blob = file.block_cursor().read_blob(offset).with_context(|| {
+            let blob_ref = BlobRef(value);
-                format!(
+            let blob_content =
-                    "failed to read value from data file {} at offset {}",
+                file.block_cursor()
-                    self.filename().display(),
+                    .read_blob(blob_ref.pos())
-                    offset
+                    .with_context(|| {
-                )
+                        format!(
-            })?;
+                            "failed to read value from data file {} at offset {}",
-            let value = Bytes::from(blob);
+                            self.filename().display(),
+                            blob_ref.pos()
+                        )
+                    })?;
+
+            let uncompressed_bytes = if blob_ref.compressed() {
+                decompressor.decompress(&blob_content, PAGE_SZ)?
+            } else {
+                blob_content
+            };
+
+            let value = Bytes::from(uncompressed_bytes);
 
             reconstruct_state.img = Some((self.lsn, value));
             Ok(ValueReconstructResult::Complete)
@@ -219,7 +273,17 @@ impl Layer for ImageLayer {
         tree_reader.dump()?;
 
         tree_reader.visit(&[0u8; KEY_SIZE], VisitDirection::Forwards, |key, value| {
-            println!("key: {} offset {}", hex::encode(key), value);
+            let blob_ref = BlobRef(value);
+            println!(
+                "key: {} offset {}{}",
+                hex::encode(key),
+                blob_ref.pos(),
+                if blob_ref.compressed() {
+                    " (compressed)"
+                } else {
+                    ""
+                }
+            );
             true
         })?;
 
@@ -423,6 +487,8 @@ pub struct ImageLayerWriter {
 
     blob_writer: WriteBlobWriter<VirtualFile>,
     tree: DiskBtreeBuilder<BlockBuf, KEY_SIZE>,
+
+    compressor: Option<zstd::bulk::Compressor<'static>>,
 }
 
 impl ImageLayerWriter {
@@ -454,6 +520,12 @@ impl ImageLayerWriter {
         let block_buf = BlockBuf::new();
         let tree_builder = DiskBtreeBuilder::new(block_buf);
 
+        // TODO: use a dictionary
+        let compressor = {
+            let compressor = zstd::bulk::Compressor::new(config::ZSTD_COMPRESSION_LEVEL)?;
+            Some(compressor)
+        };
+
         let writer = ImageLayerWriter {
             conf,
             path,
@@ -463,6 +535,7 @@ impl ImageLayerWriter {
             lsn,
             tree: tree_builder,
             blob_writer,
+            compressor,
         };
 
         Ok(writer)
@@ -475,11 +548,37 @@ impl ImageLayerWriter {
     ///
     pub fn put_image(&mut self, key: Key, img: &[u8]) -> Result<()> {
         ensure!(self.key_range.contains(&key));
-        let off = self.blob_writer.write_blob(img)?;
 
+        let mut blob_content = img;
+        let mut compressed = false;
+
+        // Try to compress the blob
+        let compressed_bytes;
+        if blob_content.len() <= PAGE_SZ {
+            if let Some(ref mut compressor) = self.compressor {
+                compressed_bytes = compressor.compress(blob_content)?;
+
+                // If compressed version is not any smaller than the original,
+                // store it uncompressed. This not just an optimization, the
+                // the decompression assumes that too. That simplifies the
+                // decompression, because you don't need to jump through any
+                // hoops to determine how large a buffer you need to hold the
+                // decompression result.
+                if compressed_bytes.len() < blob_content.len() {
+                    blob_content = &compressed_bytes;
+                    compressed = true;
+                }
+            }
+        }
+
+        // Write it to the file
+        let off = self.blob_writer.write_blob(blob_content)?;
+        let blob_ref = BlobRef::new(off, compressed);
+
+        // And store the reference in the B-tree
         let mut keybuf: [u8; KEY_SIZE] = [0u8; KEY_SIZE];
         key.write_to_byte_slice(&mut keybuf);
-        self.tree.append(&keybuf, off)?;
+        self.tree.append(&keybuf, blob_ref.0)?;
 
         Ok(())
     }

pageserver/src/lib.rs

@@ -37,7 +37,7 @@ use pgdatadir_mapping::DatadirTimeline;
 /// This is embedded in the metadata file, and also in the header of all the
 /// layer files. If you make any backwards-incompatible changes to the storage
 /// format, bump this!
-pub const STORAGE_FORMAT_VERSION: u16 = 3;
+pub const STORAGE_FORMAT_VERSION: u16 = 4;
 
 // Magic constants used to identify different kinds of files
 pub const IMAGE_FILE_MAGIC: u16 = 0x5A60;

pageserver/src/page_service.rs

@@ -305,7 +305,29 @@ fn page_service_conn_main(
 
     let mut conn_handler = PageServerHandler::new(conf, auth);
     let pgbackend = PostgresBackend::new(socket, auth_type, None, true)?;
-    pgbackend.run(&mut conn_handler)
+    match pgbackend.run(&mut conn_handler) {
+        Ok(()) => {
+            // we've been requested to shut down
+            Ok(())
+        }
+        Err(err) => {
+            let root_cause_io_err_kind = err
+                .root_cause()
+                .downcast_ref::<io::Error>()
+                .map(|e| e.kind());
+
+            // `ConnectionReset` error happens when the Postgres client closes the connection.
+            // As this disconnection happens quite often and is expected,
+            // we decided to downgrade the logging level to `INFO`.
+            // See: https://github.com/neondatabase/neon/issues/1683.
+            if root_cause_io_err_kind == Some(io::ErrorKind::ConnectionReset) {
+                info!("Postgres client disconnected");
+                Ok(())
+            } else {
+                Err(err)
+            }
+        }
+    }
 }
 
 #[derive(Debug)]
@@ -593,7 +615,8 @@ impl PageServerHandler {
         /* Send a tarball of the latest layer on the timeline */
         {
             let mut writer = CopyDataSink { pgb };
-            let mut basebackup = basebackup::Basebackup::new(&mut writer, &timeline, lsn)?;
+
+            let basebackup = basebackup::Basebackup::new(&mut writer, &timeline, lsn)?;
             span.record("lsn", &basebackup.lsn.to_string().as_str());
             basebackup.send_tarball()?;
         }

pageserver/src/tenant_mgr.rs

@@ -2,7 +2,7 @@
 //! page server.
 
 use crate::config::PageServerConf;
-use crate::layered_repository::LayeredRepository;
+use crate::layered_repository::{load_metadata, LayeredRepository};
 use crate::pgdatadir_mapping::DatadirTimeline;
 use crate::repository::{Repository, TimelineSyncStatusUpdate};
 use crate::storage_sync::index::RemoteIndex;
@@ -22,6 +22,7 @@ use std::collections::HashMap;
 use std::fmt;
 use std::sync::Arc;
 use tracing::*;
+use utils::lsn::Lsn;
 
 use utils::zid::{ZTenantId, ZTimelineId};
 
@@ -327,8 +328,8 @@ pub fn get_local_timeline_with_load(
         return Ok(Arc::clone(page_tline));
     }
 
-    let page_tline = new_local_timeline(&tenant.repo, timeline_id)
+    let page_tline = load_local_timeline(&tenant.repo, timeline_id)
-        .with_context(|| format!("Failed to create new local timeline for tenant {tenant_id}"))?;
+        .with_context(|| format!("Failed to load local timeline for tenant {tenant_id}"))?;
     tenant
         .local_timelines
         .insert(timeline_id, Arc::clone(&page_tline));
@@ -365,7 +366,7 @@ pub fn detach_timeline(
     Ok(())
 }
 
-fn new_local_timeline(
+fn load_local_timeline(
     repo: &RepositoryImpl,
     timeline_id: ZTimelineId,
 ) -> anyhow::Result<Arc<DatadirTimeline<LayeredRepository>>> {
@@ -399,6 +400,26 @@ pub fn list_tenants() -> Vec<TenantInfo> {
         .collect()
 }
 
+/// Check if a given timeline is "broken" \[1\].
+/// The function returns an error if the timeline is "broken".
+///
+/// \[1\]: it's not clear now how should we classify a timeline as broken.
+/// A timeline is categorized as broken when any of following conditions is true:
+/// - failed to load the timeline's metadata
+/// - the timeline's disk consistent LSN is zero
+fn check_broken_timeline(repo: &LayeredRepository, timeline_id: ZTimelineId) -> anyhow::Result<()> {
+    let metadata = load_metadata(repo.conf, timeline_id, repo.tenant_id())
+        .context("failed to load metadata")?;
+
+    // A timeline with zero disk consistent LSN can happen when the page server
+    // failed to checkpoint the timeline import data when creating that timeline.
+    if metadata.disk_consistent_lsn() == Lsn::INVALID {
+        bail!("Timeline {timeline_id} has a zero disk consistent LSN.");
+    }
+
+    Ok(())
+}
+
 fn init_local_repository(
     conf: &'static PageServerConf,
     tenant_id: ZTenantId,
@@ -414,7 +435,13 @@ fn init_local_repository(
         match init_status {
             LocalTimelineInitStatus::LocallyComplete => {
                 debug!("timeline {timeline_id} for tenant {tenant_id} is locally complete, registering it in repository");
-                status_updates.insert(timeline_id, TimelineSyncStatusUpdate::Downloaded);
+                if let Err(err) = check_broken_timeline(&repo, timeline_id) {
+                    info!(
+                        "Found a broken timeline {timeline_id} (err={err:?}), skip registering it in repository"
+                    );
+                } else {
+                    status_updates.insert(timeline_id, TimelineSyncStatusUpdate::Downloaded);
+                }
             }
             LocalTimelineInitStatus::NeedsSync => {
                 debug!(
@@ -458,8 +485,8 @@ fn apply_timeline_remote_sync_status_updates(
                     bail!("Local timeline {timeline_id} already registered")
                 }
                 Entry::Vacant(v) => {
-                    v.insert(new_local_timeline(repo, timeline_id).with_context(|| {
+                    v.insert(load_local_timeline(repo, timeline_id).with_context(|| {
-                        format!("Failed to register new local timeline for tenant {tenant_id}")
+                        format!("Failed to register add local timeline for tenant {tenant_id}")
                     })?);
                 }
             },

pageserver/src/timelines.rs

@@ -285,7 +285,9 @@ fn bootstrap_timeline<R: Repository>(
 ) -> Result<()> {
     let _enter = info_span!("bootstrapping", timeline = %tli, tenant = %tenantid).entered();
 
-    let initdb_path = conf.tenant_path(&tenantid).join("tmp");
+    let initdb_path = conf
+        .tenant_path(&tenantid)
+        .join(format!("tmp-timeline-{}", tli));
 
     // Init temporarily repo to get bootstrap data
     run_initdb(conf, &initdb_path)?;
@@ -300,10 +302,15 @@ fn bootstrap_timeline<R: Repository>(
     let timeline = repo.create_empty_timeline(tli, lsn)?;
     let mut page_tline: DatadirTimeline<R> = DatadirTimeline::new(timeline, u64::MAX);
     import_datadir::import_timeline_from_postgres_datadir(&pgdata_path, &mut page_tline, lsn)?;
+
+    fail::fail_point!("before-checkpoint-new-timeline", |_| {
+        bail!("failpoint before-checkpoint-new-timeline");
+    });
+
     page_tline.tline.checkpoint(CheckpointConfig::Forced)?;
 
-    println!(
+    info!(
-        "created initial timeline {} timeline.lsn {}",
+        "created root timeline {} timeline.lsn {}",
         tli,
         page_tline.tline.get_last_record_lsn()
     );

proxy/src/auth_backend/link.rs

@@ -5,12 +5,9 @@ use utils::pq_proto::{BeMessage as Be, BeParameterStatusMessage};
 fn hello_message(redirect_uri: &str, session_id: &str) -> String {
     format!(
         concat![
-            "☀️  Welcome to Neon!\n",
+            "Welcome to Neon!\n",
-            "To proceed with database creation, open the following link:\n\n",
+            "Authenticate by visiting:\n",
             "    {redirect_uri}{session_id}\n\n",
-            "It needs to be done once and we will send you '.pgpass' file,\n",
-            "which will allow you to access or create ",
-            "databases without opening your web browser."
         ],
         redirect_uri = redirect_uri,
         session_id = session_id,

proxy/src/config.rs

@@ -61,7 +61,8 @@ pub fn configure_tls(key_path: &str, cert_path: &str) -> anyhow::Result<TlsConfi
     let config = rustls::ServerConfig::builder()
         .with_safe_default_cipher_suites()
         .with_safe_default_kx_groups()
-        .with_protocol_versions(&[&rustls::version::TLS13])?
+        // allow TLS 1.2 to be compatible with older client libraries
+        .with_protocol_versions(&[&rustls::version::TLS13, &rustls::version::TLS12])?
         .with_no_client_auth()
         .with_single_cert(cert_chain, key)?;
 

safekeeper/Cargo.toml

@@ -30,6 +30,10 @@ const_format = "0.2.21"
 tokio-postgres = { git = "https://github.com/zenithdb/rust-postgres.git", rev="d052ee8b86fff9897c77b0fe89ea9daba0e1fa38" }
 tokio-util = { version = "0.7", features = ["io"] }
 git-version = "0.3.5"
+async-trait = "0.1"
+once_cell = "1.10.0"
+futures = "0.3.13"
+toml_edit = { version = "0.13", features = ["easy"] }
 
 postgres_ffi = { path = "../libs/postgres_ffi" }
 metrics = { path = "../libs/metrics" }

safekeeper/src/bin/safekeeper.rs

@@ -6,25 +6,30 @@ use clap::{App, Arg};
 use const_format::formatcp;
 use daemonize::Daemonize;
 use fs2::FileExt;
+use remote_storage::RemoteStorageConfig;
 use std::fs::{self, File};
 use std::io::{ErrorKind, Write};
 use std::path::{Path, PathBuf};
 use std::thread;
 use tokio::sync::mpsc;
+use toml_edit::Document;
 use tracing::*;
 use url::{ParseError, Url};
 
 use safekeeper::control_file::{self};
-use safekeeper::defaults::{DEFAULT_HTTP_LISTEN_ADDR, DEFAULT_PG_LISTEN_ADDR};
+use safekeeper::defaults::{
+    DEFAULT_HTTP_LISTEN_ADDR, DEFAULT_PG_LISTEN_ADDR, DEFAULT_WAL_BACKUP_RUNTIME_THREADS,
+};
+use safekeeper::http;
 use safekeeper::remove_wal;
 use safekeeper::timeline::GlobalTimelines;
+use safekeeper::wal_backup;
 use safekeeper::wal_service;
 use safekeeper::SafeKeeperConf;
 use safekeeper::{broker, callmemaybe};
-use safekeeper::{http, s3_offload};
 use utils::{
     http::endpoint, logging, project_git_version, shutdown::exit_now, signals, tcp_listener,
-    zid::ZNodeId,
+    zid::NodeId,
 };
 
 const LOCK_FILE_NAME: &str = "safekeeper.lock";
@@ -71,12 +76,6 @@ fn main() -> anyhow::Result<()> {
                 .long("pageserver")
                 .takes_value(true),
         )
-        .arg(
-            Arg::new("ttl")
-                .long("ttl")
-                .takes_value(true)
-                .help("interval for keeping WAL at safekeeper node, after which them will be uploaded to S3 and removed locally"),
-        )
         .arg(
             Arg::new("recall")
                 .long("recall")
@@ -118,12 +117,20 @@ fn main() -> anyhow::Result<()> {
             .help("a prefix to always use when polling/pusing data in etcd from this safekeeper"),
         )
         .arg(
-            Arg::new("enable-s3-offload")
+            Arg::new("wal-backup-threads").long("backup-threads").takes_value(true).help(formatcp!("number of threads for wal backup (default {DEFAULT_WAL_BACKUP_RUNTIME_THREADS}")),
-                .long("enable-s3-offload")
+        ).arg(
+            Arg::new("remote-storage")
+                .long("remote-storage")
+                .takes_value(true)
+                .help("Remote storage configuration for WAL backup (offloading to s3) as TOML inline table, e.g. {\"max_concurrent_syncs\" = 17, \"max_sync_errors\": 13, \"bucket_name\": \"<BUCKETNAME>\", \"bucket_region\":\"<REGION>\", \"concurrency_limit\": 119}.\nSafekeeper offloads WAL to [prefix_in_bucket/]<tenant_id>/<timeline_id>/<segment_file>, mirroring structure on the file system.")
+        )
+        .arg(
+            Arg::new("enable-wal-backup")
+                .long("enable-wal-backup")
                 .takes_value(true)
                 .default_value("true")
                 .default_missing_value("true")
-                .help("Enable/disable s3 offloading. When disabled, safekeeper removes WAL ignoring s3 WAL horizon."),
+                .help("Enable/disable WAL backup to s3. When disabled, safekeeper removes WAL ignoring WAL backup horizon."),
         )
         .get_matches();
 
@@ -157,17 +164,13 @@ fn main() -> anyhow::Result<()> {
         conf.listen_http_addr = addr.to_owned();
     }
 
-    if let Some(ttl) = arg_matches.value_of("ttl") {
-        conf.ttl = Some(humantime::parse_duration(ttl)?);
-    }
-
     if let Some(recall) = arg_matches.value_of("recall") {
         conf.recall_period = humantime::parse_duration(recall)?;
     }
 
     let mut given_id = None;
     if let Some(given_id_str) = arg_matches.value_of("id") {
-        given_id = Some(ZNodeId(
+        given_id = Some(NodeId(
             given_id_str
                 .parse()
                 .context("failed to parse safekeeper id")?,
@@ -182,9 +185,21 @@ fn main() -> anyhow::Result<()> {
         conf.broker_etcd_prefix = prefix.to_string();
     }
 
+    if let Some(backup_threads) = arg_matches.value_of("wal-backup-threads") {
+        conf.backup_runtime_threads = backup_threads
+            .parse()
+            .with_context(|| format!("Failed to parse backup threads {}", backup_threads))?;
+    }
+    if let Some(storage_conf) = arg_matches.value_of("remote-storage") {
+        // funny toml doesn't consider plain inline table as valid document, so wrap in a key to parse
+        let storage_conf_toml = format!("remote_storage = {}", storage_conf);
+        let parsed_toml = storage_conf_toml.parse::<Document>()?; // parse
+        let (_, storage_conf_parsed_toml) = parsed_toml.iter().next().unwrap(); // and strip key off again
+        conf.remote_storage = Some(RemoteStorageConfig::from_toml(storage_conf_parsed_toml)?);
+    }
     // Seems like there is no better way to accept bool values explicitly in clap.
-    conf.s3_offload_enabled = arg_matches
+    conf.wal_backup_enabled = arg_matches
-        .value_of("enable-s3-offload")
+        .value_of("enable-wal-backup")
         .unwrap()
         .parse()
         .context("failed to parse bool enable-s3-offload bool")?;
@@ -192,7 +207,7 @@ fn main() -> anyhow::Result<()> {
     start_safekeeper(conf, given_id, arg_matches.is_present("init"))
 }
 
-fn start_safekeeper(mut conf: SafeKeeperConf, given_id: Option<ZNodeId>, init: bool) -> Result<()> {
+fn start_safekeeper(mut conf: SafeKeeperConf, given_id: Option<NodeId>, init: bool) -> Result<()> {
     let log_file = logging::init("safekeeper.log", conf.daemonize)?;
 
     info!("version: {GIT_VERSION}");
@@ -252,7 +267,8 @@ fn start_safekeeper(mut conf: SafeKeeperConf, given_id: Option<ZNodeId>, init: b
     let signals = signals::install_shutdown_handlers()?;
     let mut threads = vec![];
     let (callmemaybe_tx, callmemaybe_rx) = mpsc::unbounded_channel();
-    GlobalTimelines::set_callmemaybe_tx(callmemaybe_tx);
+    let (wal_backup_launcher_tx, wal_backup_launcher_rx) = mpsc::channel(100);
+    GlobalTimelines::init(callmemaybe_tx, wal_backup_launcher_tx);
 
     let conf_ = conf.clone();
     threads.push(
@@ -270,17 +286,6 @@ fn start_safekeeper(mut conf: SafeKeeperConf, given_id: Option<ZNodeId>, init: b
             })?,
     );
 
-    if conf.ttl.is_some() {
-        let conf_ = conf.clone();
-        threads.push(
-            thread::Builder::new()
-                .name("S3 offload thread".into())
-                .spawn(|| {
-                    s3_offload::thread_main(conf_);
-                })?,
-        );
-    }
-
     let conf_cloned = conf.clone();
     let safekeeper_thread = thread::Builder::new()
         .name("Safekeeper thread".into())
@@ -330,6 +335,15 @@ fn start_safekeeper(mut conf: SafeKeeperConf, given_id: Option<ZNodeId>, init: b
             })?,
     );
 
+    let conf_ = conf.clone();
+    threads.push(
+        thread::Builder::new()
+            .name("wal backup launcher thread".into())
+            .spawn(move || {
+                wal_backup::wal_backup_launcher_thread_main(conf_, wal_backup_launcher_rx);
+            })?,
+    );
+
     // TODO: put more thoughts into handling of failed threads
     // We probably should restart them.
 
@@ -345,14 +359,14 @@ fn start_safekeeper(mut conf: SafeKeeperConf, given_id: Option<ZNodeId>, init: b
 }
 
 /// Determine safekeeper id and set it in config.
-fn set_id(conf: &mut SafeKeeperConf, given_id: Option<ZNodeId>) -> Result<()> {
+fn set_id(conf: &mut SafeKeeperConf, given_id: Option<NodeId>) -> Result<()> {
     let id_file_path = conf.workdir.join(ID_FILE_NAME);
 
-    let my_id: ZNodeId;
+    let my_id: NodeId;
     // If ID exists, read it in; otherwise set one passed
     match fs::read(&id_file_path) {
         Ok(id_serialized) => {
-            my_id = ZNodeId(
+            my_id = NodeId(
                 std::str::from_utf8(&id_serialized)
                     .context("failed to parse safekeeper id")?
                     .parse()

safekeeper/src/broker.rs

@@ -1,5 +1,6 @@
 //! Communication with etcd, providing safekeeper peers and pageserver coordination.
 
+use anyhow::anyhow;
 use anyhow::Context;
 use anyhow::Error;
 use anyhow::Result;
@@ -7,12 +8,14 @@ use etcd_broker::Client;
 use etcd_broker::PutOptions;
 use etcd_broker::SkTimelineSubscriptionKind;
 use std::time::Duration;
+use tokio::spawn;
 use tokio::task::JoinHandle;
 use tokio::{runtime, time::sleep};
 use tracing::*;
+use url::Url;
 
 use crate::{timeline::GlobalTimelines, SafeKeeperConf};
-use utils::zid::{ZNodeId, ZTenantTimelineId};
+use utils::zid::{NodeId, ZTenantTimelineId};
 
 const RETRY_INTERVAL_MSEC: u64 = 1000;
 const PUSH_INTERVAL_MSEC: u64 = 1000;
@@ -36,7 +39,7 @@ pub fn thread_main(conf: SafeKeeperConf) {
 fn timeline_safekeeper_path(
     broker_etcd_prefix: String,
     zttid: ZTenantTimelineId,
-    sk_id: ZNodeId,
+    sk_id: NodeId,
 ) -> String {
     format!(
         "{}/{sk_id}",
@@ -44,6 +47,118 @@ fn timeline_safekeeper_path(
     )
 }
 
+pub struct Election {
+    pub election_name: String,
+    pub candidate_name: String,
+    pub broker_endpoints: Vec<Url>,
+}
+
+impl Election {
+    pub fn new(election_name: String, candidate_name: String, broker_endpoints: Vec<Url>) -> Self {
+        Self {
+            election_name,
+            candidate_name,
+            broker_endpoints,
+        }
+    }
+}
+
+pub struct ElectionLeader {
+    client: Client,
+    keep_alive: JoinHandle<Result<()>>,
+}
+
+impl ElectionLeader {
+    pub async fn check_am_i(
+        &mut self,
+        election_name: String,
+        candidate_name: String,
+    ) -> Result<bool> {
+        let resp = self.client.leader(election_name).await?;
+
+        let kv = resp.kv().ok_or(anyhow!("failed to get leader response"))?;
+        let leader = kv.value_str()?;
+
+        Ok(leader == candidate_name)
+    }
+
+    pub async fn give_up(self) {
+        self.keep_alive.abort();
+        // TODO: it'll be wise to resign here but it'll happen after lease expiration anyway
+        // should we await for keep alive termination?
+        let _ = self.keep_alive.await;
+    }
+}
+
+pub async fn get_leader(req: &Election) -> Result<ElectionLeader> {
+    let mut client = Client::connect(req.broker_endpoints.clone(), None)
+        .await
+        .context("Could not connect to etcd")?;
+
+    let lease = client
+        .lease_grant(LEASE_TTL_SEC, None)
+        .await
+        .context("Could not acquire a lease");
+
+    let lease_id = lease.map(|l| l.id()).unwrap();
+
+    let keep_alive = spawn::<_>(lease_keep_alive(client.clone(), lease_id));
+
+    if let Err(e) = client
+        .campaign(
+            req.election_name.clone(),
+            req.candidate_name.clone(),
+            lease_id,
+        )
+        .await
+    {
+        keep_alive.abort();
+        let _ = keep_alive.await;
+        return Err(e.into());
+    }
+
+    Ok(ElectionLeader { client, keep_alive })
+}
+
+async fn lease_keep_alive(mut client: Client, lease_id: i64) -> Result<()> {
+    let (mut keeper, mut ka_stream) = client
+        .lease_keep_alive(lease_id)
+        .await
+        .context("failed to create keepalive stream")?;
+
+    loop {
+        let push_interval = Duration::from_millis(PUSH_INTERVAL_MSEC);
+
+        keeper
+            .keep_alive()
+            .await
+            .context("failed to send LeaseKeepAliveRequest")?;
+
+        ka_stream
+            .message()
+            .await
+            .context("failed to receive LeaseKeepAliveResponse")?;
+
+        sleep(push_interval).await;
+    }
+}
+
+pub fn get_campaign_name(
+    election_name: String,
+    broker_prefix: String,
+    timeline_id: &ZTenantTimelineId,
+) -> String {
+    return format!(
+        "{}/{}",
+        SkTimelineSubscriptionKind::timeline(broker_prefix, *timeline_id).watch_key(),
+        election_name
+    );
+}
+
+pub fn get_candiate_name(system_id: NodeId) -> String {
+    format!("id_{}", system_id)
+}
+
 /// Push once in a while data about all active timelines to the broker.
 async fn push_loop(conf: SafeKeeperConf) -> anyhow::Result<()> {
     let mut client = Client::connect(&conf.broker_endpoints, None).await?;
@@ -59,7 +174,7 @@ async fn push_loop(conf: SafeKeeperConf) -> anyhow::Result<()> {
         // sensitive and there is no risk of deadlock as we don't await while
         // lock is held.
         for zttid in GlobalTimelines::get_active_timelines() {
-            if let Ok(tli) = GlobalTimelines::get(&conf, zttid, false) {
+            if let Some(tli) = GlobalTimelines::get_loaded(zttid) {
                 let sk_info = tli.get_public_info(&conf)?;
                 let put_opts = PutOptions::new().with_lease(lease.id());
                 client
@@ -106,12 +221,13 @@ async fn pull_loop(conf: SafeKeeperConf) -> Result<()> {
                     // note: there are blocking operations below, but it's considered fine for now
                     if let Ok(tli) = GlobalTimelines::get(&conf, zttid, false) {
                         for (safekeeper_id, info) in sk_info {
-                            tli.record_safekeeper_info(&info, safekeeper_id)?
+                            tli.record_safekeeper_info(&info, safekeeper_id).await?
                         }
                     }
                 }
             }
             None => {
+                // XXX it means we lost connection with etcd, error is consumed inside sub object
                 debug!("timeline updates sender closed, aborting the pull loop");
                 return Ok(());
             }
@@ -142,11 +258,12 @@ async fn main_loop(conf: SafeKeeperConf) {
                 },
                 res = async { pull_handle.as_mut().unwrap().await }, if pull_handle.is_some() => {
                     // was it panic or normal error?
-                    let err = match res {
+                    match res {
-                        Ok(res_internal) => res_internal.unwrap_err(),
+                        Ok(res_internal) => if let Err(err_inner) = res_internal {
-                        Err(err_outer) => err_outer.into(),
+                            warn!("pull task failed: {:?}", err_inner);
+                        }
+                        Err(err_outer) => { warn!("pull task panicked: {:?}", err_outer) }
                     };
-                    warn!("pull task failed: {:?}", err);
                     pull_handle = None;
                 },
                 _ = ticker.tick() => {

safekeeper/src/control_file_upgrade.rs

@@ -165,7 +165,7 @@ pub fn upgrade_control_file(buf: &[u8], version: u32) -> Result<SafeKeeperState>
             timeline_start_lsn: Lsn(0),
             local_start_lsn: Lsn(0),
             commit_lsn: oldstate.commit_lsn,
-            s3_wal_lsn: Lsn(0),
+            backup_lsn: Lsn(0),
             peer_horizon_lsn: oldstate.truncate_lsn,
             remote_consistent_lsn: Lsn(0),
             peers: Peers(vec![]),
@@ -188,7 +188,7 @@ pub fn upgrade_control_file(buf: &[u8], version: u32) -> Result<SafeKeeperState>
             timeline_start_lsn: Lsn(0),
             local_start_lsn: Lsn(0),
             commit_lsn: oldstate.commit_lsn,
-            s3_wal_lsn: Lsn(0),
+            backup_lsn: Lsn(0),
             peer_horizon_lsn: oldstate.truncate_lsn,
             remote_consistent_lsn: Lsn(0),
             peers: Peers(vec![]),
@@ -211,7 +211,7 @@ pub fn upgrade_control_file(buf: &[u8], version: u32) -> Result<SafeKeeperState>
             timeline_start_lsn: Lsn(0),
             local_start_lsn: Lsn(0),
             commit_lsn: oldstate.commit_lsn,
-            s3_wal_lsn: Lsn(0),
+            backup_lsn: Lsn(0),
             peer_horizon_lsn: oldstate.truncate_lsn,
             remote_consistent_lsn: Lsn(0),
             peers: Peers(vec![]),
@@ -234,7 +234,7 @@ pub fn upgrade_control_file(buf: &[u8], version: u32) -> Result<SafeKeeperState>
             timeline_start_lsn: Lsn(0),
             local_start_lsn: Lsn(0),
             commit_lsn: oldstate.commit_lsn,
-            s3_wal_lsn: Lsn(0),
+            backup_lsn: Lsn::INVALID,
             peer_horizon_lsn: oldstate.peer_horizon_lsn,
             remote_consistent_lsn: Lsn(0),
             peers: Peers(vec![]),

safekeeper/src/http/models.rs

@@ -1,9 +1,9 @@
 use serde::{Deserialize, Serialize};
-use utils::zid::{ZNodeId, ZTenantId, ZTimelineId};
+use utils::zid::{NodeId, ZTenantId, ZTimelineId};
 
 #[derive(Serialize, Deserialize)]
 pub struct TimelineCreateRequest {
     pub tenant_id: ZTenantId,
     pub timeline_id: ZTimelineId,
-    pub peer_ids: Vec<ZNodeId>,
+    pub peer_ids: Vec<NodeId>,
 }

safekeeper/src/http/routes.rs

@@ -20,14 +20,14 @@ use utils::{
         RequestExt, RouterBuilder,
     },
     lsn::Lsn,
-    zid::{ZNodeId, ZTenantId, ZTenantTimelineId, ZTimelineId},
+    zid::{NodeId, ZTenantId, ZTenantTimelineId, ZTimelineId},
 };
 
 use super::models::TimelineCreateRequest;
 
 #[derive(Debug, Serialize)]
 struct SafekeeperStatus {
-    id: ZNodeId,
+    id: NodeId,
 }
 
 /// Healthcheck handler.
@@ -70,19 +70,19 @@ struct TimelineStatus {
     timeline_id: ZTimelineId,
     acceptor_state: AcceptorStateStatus,
     #[serde(serialize_with = "display_serialize")]
+    flush_lsn: Lsn,
+    #[serde(serialize_with = "display_serialize")]
     timeline_start_lsn: Lsn,
     #[serde(serialize_with = "display_serialize")]
     local_start_lsn: Lsn,
     #[serde(serialize_with = "display_serialize")]
     commit_lsn: Lsn,
     #[serde(serialize_with = "display_serialize")]
-    s3_wal_lsn: Lsn,
+    backup_lsn: Lsn,
     #[serde(serialize_with = "display_serialize")]
     peer_horizon_lsn: Lsn,
     #[serde(serialize_with = "display_serialize")]
     remote_consistent_lsn: Lsn,
-    #[serde(serialize_with = "display_serialize")]
-    flush_lsn: Lsn,
 }
 
 /// Report info about timeline.
@@ -107,13 +107,13 @@ async fn timeline_status_handler(request: Request<Body>) -> Result<Response<Body
         tenant_id: zttid.tenant_id,
         timeline_id: zttid.timeline_id,
         acceptor_state: acc_state,
+        flush_lsn,
         timeline_start_lsn: state.timeline_start_lsn,
         local_start_lsn: state.local_start_lsn,
         commit_lsn: inmem.commit_lsn,
-        s3_wal_lsn: inmem.s3_wal_lsn,
+        backup_lsn: inmem.backup_lsn,
         peer_horizon_lsn: inmem.peer_horizon_lsn,
         remote_consistent_lsn: inmem.remote_consistent_lsn,
-        flush_lsn,
     };
     json_response(StatusCode::OK, status)
 }
@@ -148,7 +148,9 @@ async fn timeline_delete_force_handler(
     ensure_no_body(&mut request).await?;
     json_response(
         StatusCode::OK,
-        GlobalTimelines::delete_force(get_conf(&request), &zttid).map_err(ApiError::from_err)?,
+        GlobalTimelines::delete_force(get_conf(&request), &zttid)
+            .await
+            .map_err(ApiError::from_err)?,
     )
 }
 
@@ -162,6 +164,7 @@ async fn tenant_delete_force_handler(
     json_response(
         StatusCode::OK,
         GlobalTimelines::delete_force_all_for_tenant(get_conf(&request), &tenant_id)
+            .await
             .map_err(ApiError::from_err)?
             .iter()
             .map(|(zttid, resp)| (format!("{}", zttid.timeline_id), *resp))
@@ -178,7 +181,8 @@ async fn record_safekeeper_info(mut request: Request<Body>) -> Result<Response<B
     let safekeeper_info: SkTimelineInfo = json_request(&mut request).await?;
 
     let tli = GlobalTimelines::get(get_conf(&request), zttid, false).map_err(ApiError::from_err)?;
-    tli.record_safekeeper_info(&safekeeper_info, ZNodeId(1))?;
+    tli.record_safekeeper_info(&safekeeper_info, NodeId(1))
+        .await?;
 
     json_response(StatusCode::OK, ())
 }

safekeeper/src/lib.rs

@@ -1,9 +1,11 @@
+use defaults::DEFAULT_WAL_BACKUP_RUNTIME_THREADS;
 //
+use remote_storage::RemoteStorageConfig;
 use std::path::PathBuf;
 use std::time::Duration;
 use url::Url;
 
-use utils::zid::{ZNodeId, ZTenantId, ZTenantTimelineId};
+use utils::zid::{NodeId, ZTenantId, ZTenantTimelineId};
 
 pub mod broker;
 pub mod callmemaybe;
@@ -14,10 +16,10 @@ pub mod http;
 pub mod json_ctrl;
 pub mod receive_wal;
 pub mod remove_wal;
-pub mod s3_offload;
 pub mod safekeeper;
 pub mod send_wal;
 pub mod timeline;
+pub mod wal_backup;
 pub mod wal_service;
 pub mod wal_storage;
 
@@ -31,6 +33,7 @@ pub mod defaults {
     pub const DEFAULT_HTTP_LISTEN_PORT: u16 = 7676;
     pub const DEFAULT_HTTP_LISTEN_ADDR: &str = formatcp!("127.0.0.1:{DEFAULT_HTTP_LISTEN_PORT}");
     pub const DEFAULT_RECALL_PERIOD: Duration = Duration::from_secs(10);
+    pub const DEFAULT_WAL_BACKUP_RUNTIME_THREADS: usize = 8;
 }
 
 #[derive(Debug, Clone)]
@@ -47,12 +50,13 @@ pub struct SafeKeeperConf {
     pub no_sync: bool,
     pub listen_pg_addr: String,
     pub listen_http_addr: String,
-    pub ttl: Option<Duration>,
     pub recall_period: Duration,
-    pub my_id: ZNodeId,
+    pub remote_storage: Option<RemoteStorageConfig>,
+    pub backup_runtime_threads: usize,
+    pub wal_backup_enabled: bool,
+    pub my_id: NodeId,
     pub broker_endpoints: Vec<Url>,
     pub broker_etcd_prefix: String,
-    pub s3_offload_enabled: bool,
 }
 
 impl SafeKeeperConf {
@@ -77,12 +81,13 @@ impl Default for SafeKeeperConf {
             no_sync: false,
             listen_pg_addr: defaults::DEFAULT_PG_LISTEN_ADDR.to_string(),
             listen_http_addr: defaults::DEFAULT_HTTP_LISTEN_ADDR.to_string(),
-            ttl: None,
+            remote_storage: None,
             recall_period: defaults::DEFAULT_RECALL_PERIOD,
-            my_id: ZNodeId(0),
+            my_id: NodeId(0),
             broker_endpoints: Vec::new(),
             broker_etcd_prefix: etcd_broker::DEFAULT_NEON_BROKER_ETCD_PREFIX.to_string(),
-            s3_offload_enabled: true,
+            backup_runtime_threads: DEFAULT_WAL_BACKUP_RUNTIME_THREADS,
+            wal_backup_enabled: true,
         }
     }
 }

safekeeper/src/receive_wal.rs

@@ -85,16 +85,10 @@ impl<'pg> ReceiveWalConn<'pg> {
             _ => bail!("unexpected message {:?} instead of greeting", next_msg),
         }
 
-        // Register the connection and defer unregister.
-        spg.timeline
-            .get()
-            .on_compute_connect(self.pageserver_connstr.as_ref())?;
-        let _guard = ComputeConnectionGuard {
-            timeline: Arc::clone(spg.timeline.get()),
-        };
-
         let mut next_msg = Some(next_msg);
 
+        let mut first_time_through = true;
+        let mut _guard: Option<ComputeConnectionGuard> = None;
         loop {
             if matches!(next_msg, Some(ProposerAcceptorMessage::AppendRequest(_))) {
                 // poll AppendRequest's without blocking and write WAL to disk without flushing,
@@ -122,6 +116,18 @@ impl<'pg> ReceiveWalConn<'pg> {
                     self.write_msg(&reply)?;
                 }
             }
+            if first_time_through {
+                // Register the connection and defer unregister. Do that only
+                // after processing first message, as it sets wal_seg_size,
+                // wanted by many.
+                spg.timeline
+                    .get()
+                    .on_compute_connect(self.pageserver_connstr.as_ref())?;
+                _guard = Some(ComputeConnectionGuard {
+                    timeline: Arc::clone(spg.timeline.get()),
+                });
+                first_time_through = false;
+            }
 
             // blocking wait for the next message
             if next_msg.is_none() {

safekeeper/src/remove_wal.rs

@@ -12,7 +12,7 @@ pub fn thread_main(conf: SafeKeeperConf) {
         let active_tlis = GlobalTimelines::get_active_timelines();
         for zttid in &active_tlis {
             if let Ok(tli) = GlobalTimelines::get(&conf, *zttid, false) {
-                if let Err(e) = tli.remove_old_wal(conf.s3_offload_enabled) {
+                if let Err(e) = tli.remove_old_wal(conf.wal_backup_enabled) {
                     warn!(
                         "failed to remove WAL for tenant {} timeline {}: {}",
                         tli.zttid.tenant_id, tli.zttid.timeline_id, e

safekeeper/src/s3_offload.rs

@@ -1,107 +0,0 @@
-//
-// Offload old WAL segments to S3 and remove them locally
-// Needs `AWS_ACCESS_KEY_ID` and `AWS_SECRET_ACCESS_KEY` environment variables to be set
-// if no IAM bucket access is used.
-//
-
-use anyhow::{bail, Context};
-use postgres_ffi::xlog_utils::*;
-use remote_storage::{
-    GenericRemoteStorage, RemoteStorage, RemoteStorageConfig, S3Bucket, S3Config, S3ObjectKey,
-};
-use std::collections::HashSet;
-use std::env;
-use std::num::{NonZeroU32, NonZeroUsize};
-use std::path::Path;
-use std::time::SystemTime;
-use tokio::fs::{self, File};
-use tokio::io::BufReader;
-use tokio::runtime;
-use tokio::time::sleep;
-use tracing::*;
-use walkdir::WalkDir;
-
-use crate::SafeKeeperConf;
-
-pub fn thread_main(conf: SafeKeeperConf) {
-    // Create a new thread pool
-    //
-    // FIXME: keep it single-threaded for now, make it easier to debug with gdb,
-    // and we're not concerned with performance yet.
-    //let runtime = runtime::Runtime::new().unwrap();
-    let runtime = runtime::Builder::new_current_thread()
-        .enable_all()
-        .build()
-        .unwrap();
-
-    info!("Starting S3 offload task");
-
-    runtime.block_on(async {
-        main_loop(&conf).await.unwrap();
-    });
-}
-
-async fn offload_files(
-    remote_storage: &S3Bucket,
-    listing: &HashSet<S3ObjectKey>,
-    dir_path: &Path,
-    conf: &SafeKeeperConf,
-) -> anyhow::Result<u64> {
-    let horizon = SystemTime::now() - conf.ttl.unwrap();
-    let mut n: u64 = 0;
-    for entry in WalkDir::new(dir_path) {
-        let entry = entry?;
-        let path = entry.path();
-
-        if path.is_file()
-            && IsXLogFileName(entry.file_name().to_str().unwrap())
-            && entry.metadata().unwrap().created().unwrap() <= horizon
-        {
-            let remote_path = remote_storage.remote_object_id(path)?;
-            if !listing.contains(&remote_path) {
-                let file = File::open(&path).await?;
-                let file_length = file.metadata().await?.len() as usize;
-                remote_storage
-                    .upload(BufReader::new(file), file_length, &remote_path, None)
-                    .await?;
-
-                fs::remove_file(&path).await?;
-                n += 1;
-            }
-        }
-    }
-    Ok(n)
-}
-
-async fn main_loop(conf: &SafeKeeperConf) -> anyhow::Result<()> {
-    let remote_storage = match GenericRemoteStorage::new(
-        conf.workdir.clone(),
-        &RemoteStorageConfig {
-            max_concurrent_syncs: NonZeroUsize::new(10).unwrap(),
-            max_sync_errors: NonZeroU32::new(1).unwrap(),
-            storage: remote_storage::RemoteStorageKind::AwsS3(S3Config {
-                bucket_name: "zenith-testbucket".to_string(),
-                bucket_region: env::var("S3_REGION").context("S3_REGION env var is not set")?,
-                prefix_in_bucket: Some("walarchive/".to_string()),
-                endpoint: Some(env::var("S3_ENDPOINT").context("S3_ENDPOINT env var is not set")?),
-                concurrency_limit: NonZeroUsize::new(20).unwrap(),
-            }),
-        },
-    )? {
-        GenericRemoteStorage::Local(_) => {
-            bail!("Unexpected: got local storage for the remote config")
-        }
-        GenericRemoteStorage::S3(remote_storage) => remote_storage,
-    };
-
-    loop {
-        let listing = remote_storage
-            .list()
-            .await?
-            .into_iter()
-            .collect::<HashSet<_>>();
-        let n = offload_files(&remote_storage, &listing, &conf.workdir, conf).await?;
-        info!("Offload {n} files to S3");
-        sleep(conf.ttl.unwrap()).await;
-    }
-}

safekeeper/src/safekeeper.rs

@@ -19,6 +19,7 @@ use lazy_static::lazy_static;
 
 use crate::control_file;
 use crate::send_wal::HotStandbyFeedback;
+
 use crate::wal_storage;
 use metrics::{register_gauge_vec, Gauge, GaugeVec};
 use postgres_ffi::xlog_utils::MAX_SEND_SIZE;
@@ -26,7 +27,7 @@ use utils::{
     bin_ser::LeSer,
     lsn::Lsn,
     pq_proto::{SystemId, ZenithFeedback},
-    zid::{ZNodeId, ZTenantId, ZTenantTimelineId, ZTimelineId},
+    zid::{NodeId, ZTenantId, ZTenantTimelineId, ZTimelineId},
 };
 
 pub const SK_MAGIC: u32 = 0xcafeceefu32;
@@ -141,7 +142,7 @@ pub struct ServerInfo {
 #[derive(Debug, Clone, Serialize, Deserialize)]
 pub struct PeerInfo {
     /// LSN up to which safekeeper offloaded WAL to s3.
-    s3_wal_lsn: Lsn,
+    backup_lsn: Lsn,
     /// Term of the last entry.
     term: Term,
     /// LSN of the last record.
@@ -153,7 +154,7 @@ pub struct PeerInfo {
 impl PeerInfo {
     fn new() -> Self {
         Self {
-            s3_wal_lsn: Lsn(0),
+            backup_lsn: Lsn::INVALID,
             term: INVALID_TERM,
             flush_lsn: Lsn(0),
             commit_lsn: Lsn(0),
@@ -164,7 +165,7 @@ impl PeerInfo {
 // vector-based node id -> peer state map with very limited functionality we
 // need/
 #[derive(Debug, Clone, Serialize, Deserialize)]
-pub struct Peers(pub Vec<(ZNodeId, PeerInfo)>);
+pub struct Peers(pub Vec<(NodeId, PeerInfo)>);
 
 /// Persistent information stored on safekeeper node
 /// On disk data is prefixed by magic and format version and followed by checksum.
@@ -193,9 +194,9 @@ pub struct SafeKeeperState {
     /// Part of WAL acknowledged by quorum and available locally. Always points
     /// to record boundary.
     pub commit_lsn: Lsn,
-    /// First LSN not yet offloaded to s3. Useful to persist to avoid finding
+    /// LSN that points to the end of the last backed up segment. Useful to
-    /// out offloading progress on boot.
+    /// persist to avoid finding out offloading progress on boot.
-    pub s3_wal_lsn: Lsn,
+    pub backup_lsn: Lsn,
     /// Minimal LSN which may be needed for recovery of some safekeeper (end_lsn
     /// of last record streamed to everyone). Persisting it helps skipping
     /// recovery in walproposer, generally we compute it from peers. In
@@ -217,14 +218,14 @@ pub struct SafeKeeperState {
 // are not flushed yet.
 pub struct SafekeeperMemState {
     pub commit_lsn: Lsn,
-    pub s3_wal_lsn: Lsn, // TODO: keep only persistent version
+    pub backup_lsn: Lsn,
     pub peer_horizon_lsn: Lsn,
     pub remote_consistent_lsn: Lsn,
     pub proposer_uuid: PgUuid,
 }
 
 impl SafeKeeperState {
-    pub fn new(zttid: &ZTenantTimelineId, peers: Vec<ZNodeId>) -> SafeKeeperState {
+    pub fn new(zttid: &ZTenantTimelineId, peers: Vec<NodeId>) -> SafeKeeperState {
         SafeKeeperState {
             tenant_id: zttid.tenant_id,
             timeline_id: zttid.timeline_id,
@@ -241,7 +242,7 @@ impl SafeKeeperState {
             timeline_start_lsn: Lsn(0),
             local_start_lsn: Lsn(0),
             commit_lsn: Lsn(0),
-            s3_wal_lsn: Lsn(0),
+            backup_lsn: Lsn::INVALID,
             peer_horizon_lsn: Lsn(0),
             remote_consistent_lsn: Lsn(0),
             peers: Peers(peers.iter().map(|p| (*p, PeerInfo::new())).collect()),
@@ -277,7 +278,7 @@ pub struct ProposerGreeting {
 #[derive(Debug, Serialize)]
 pub struct AcceptorGreeting {
     term: u64,
-    node_id: ZNodeId,
+    node_id: NodeId,
 }
 
 /// Vote request sent from proposer to safekeepers
@@ -531,7 +532,7 @@ pub struct SafeKeeper<CTRL: control_file::Storage, WAL: wal_storage::Storage> {
 
     pub wal_store: WAL,
 
-    node_id: ZNodeId, // safekeeper's node id
+    node_id: NodeId, // safekeeper's node id
 }
 
 impl<CTRL, WAL> SafeKeeper<CTRL, WAL>
@@ -544,7 +545,7 @@ where
         ztli: ZTimelineId,
         state: CTRL,
         mut wal_store: WAL,
-        node_id: ZNodeId,
+        node_id: NodeId,
     ) -> Result<SafeKeeper<CTRL, WAL>> {
         if state.timeline_id != ZTimelineId::from([0u8; 16]) && ztli != state.timeline_id {
             bail!("Calling SafeKeeper::new with inconsistent ztli ({}) and SafeKeeperState.server.timeline_id ({})", ztli, state.timeline_id);
@@ -559,7 +560,7 @@ where
             epoch_start_lsn: Lsn(0),
             inmem: SafekeeperMemState {
                 commit_lsn: state.commit_lsn,
-                s3_wal_lsn: state.s3_wal_lsn,
+                backup_lsn: state.backup_lsn,
                 peer_horizon_lsn: state.peer_horizon_lsn,
                 remote_consistent_lsn: state.remote_consistent_lsn,
                 proposer_uuid: state.proposer_uuid,
@@ -649,7 +650,6 @@ where
             self.state.persist(&state)?;
         }
 
-        // pass wal_seg_size to read WAL and find flush_lsn
         self.wal_store.init_storage(&self.state)?;
 
         info!(
@@ -731,24 +731,36 @@ where
         {
             let mut state = self.state.clone();
 
-            // Remeber point where WAL begins globally, if not yet.
+            // Here we learn initial LSN for the first time, set fields
+            // interested in that.
+
             if state.timeline_start_lsn == Lsn(0) {
+                // Remember point where WAL begins globally.
                 state.timeline_start_lsn = msg.timeline_start_lsn;
                 info!(
                     "setting timeline_start_lsn to {:?}",
                     state.timeline_start_lsn
                 );
-            }
 
-            // Remember point where WAL begins locally, if not yet. (I doubt the
-            // second condition is ever possible)
-            if state.local_start_lsn == Lsn(0) || state.local_start_lsn >= msg.start_streaming_at {
                 state.local_start_lsn = msg.start_streaming_at;
                 info!("setting local_start_lsn to {:?}", state.local_start_lsn);
             }
+            // Initializing commit_lsn before acking first flushed record is
+            // important to let find_end_of_wal skip the whole in the beginning
+            // of the first segment.
+            //
+            // NB: on new clusters, this happens at the same time as
+            // timeline_start_lsn initialization, it is taken outside to provide
+            // upgrade.
+            self.global_commit_lsn = max(self.global_commit_lsn, state.timeline_start_lsn);
+            self.inmem.commit_lsn = max(self.inmem.commit_lsn, state.timeline_start_lsn);
+            self.metrics.commit_lsn.set(self.inmem.commit_lsn.0 as f64);
+
+            // Initalizing backup_lsn is useful to avoid making backup think it should upload 0 segment.
+            self.inmem.backup_lsn = max(self.inmem.backup_lsn, state.timeline_start_lsn);
 
             state.acceptor_state.term_history = msg.term_history.clone();
-            self.state.persist(&state)?;
+            self.persist_control_file(state)?;
         }
 
         info!("start receiving WAL since {:?}", msg.start_streaming_at);
@@ -772,25 +784,16 @@ where
         // that we receive new epoch_start_lsn, and we still need to sync
         // control file in this case.
         if commit_lsn == self.epoch_start_lsn && self.state.commit_lsn != commit_lsn {
-            self.persist_control_file()?;
+            self.persist_control_file(self.state.clone())?;
-        }
-
-        // We got our first commit_lsn, which means we should sync
-        // everything to disk, to initialize the state.
-        if self.state.commit_lsn == Lsn(0) && commit_lsn > Lsn(0) {
-            self.wal_store.flush_wal()?;
-            self.persist_control_file()?;
         }
 
         Ok(())
     }
 
-    /// Persist in-memory state to the disk.
+    /// Persist in-memory state to the disk, taking other data from state.
-    fn persist_control_file(&mut self) -> Result<()> {
+    fn persist_control_file(&mut self, mut state: SafeKeeperState) -> Result<()> {
-        let mut state = self.state.clone();
-
         state.commit_lsn = self.inmem.commit_lsn;
-        state.s3_wal_lsn = self.inmem.s3_wal_lsn;
+        state.backup_lsn = self.inmem.backup_lsn;
         state.peer_horizon_lsn = self.inmem.peer_horizon_lsn;
         state.remote_consistent_lsn = self.inmem.remote_consistent_lsn;
         state.proposer_uuid = self.inmem.proposer_uuid;
@@ -823,13 +826,6 @@ where
         // do the job
         if !msg.wal_data.is_empty() {
             self.wal_store.write_wal(msg.h.begin_lsn, &msg.wal_data)?;
-
-            // If this was the first record we ever received, initialize
-            // commit_lsn to help find_end_of_wal skip the hole in the
-            // beginning.
-            if self.global_commit_lsn == Lsn(0) {
-                self.global_commit_lsn = msg.h.begin_lsn;
-            }
         }
 
         // flush wal to the disk, if required
@@ -852,7 +848,7 @@ where
         if self.state.peer_horizon_lsn + (self.state.server.wal_seg_size as u64)
             < self.inmem.peer_horizon_lsn
         {
-            self.persist_control_file()?;
+            self.persist_control_file(self.state.clone())?;
         }
 
         trace!(
@@ -898,11 +894,11 @@ where
                 self.update_commit_lsn()?;
             }
         }
-        if let Some(s3_wal_lsn) = sk_info.s3_wal_lsn {
+        if let Some(backup_lsn) = sk_info.backup_lsn {
-            let new_s3_wal_lsn = max(s3_wal_lsn, self.inmem.s3_wal_lsn);
+            let new_backup_lsn = max(backup_lsn, self.inmem.backup_lsn);
             sync_control_file |=
-                self.state.s3_wal_lsn + (self.state.server.wal_seg_size as u64) < new_s3_wal_lsn;
+                self.state.backup_lsn + (self.state.server.wal_seg_size as u64) < new_backup_lsn;
-            self.inmem.s3_wal_lsn = new_s3_wal_lsn;
+            self.inmem.backup_lsn = new_backup_lsn;
         }
         if let Some(remote_consistent_lsn) = sk_info.remote_consistent_lsn {
             let new_remote_consistent_lsn =
@@ -920,7 +916,7 @@ where
             self.inmem.peer_horizon_lsn = new_peer_horizon_lsn;
         }
         if sync_control_file {
-            self.persist_control_file()?;
+            self.persist_control_file(self.state.clone())?;
         }
         Ok(())
     }
@@ -930,29 +926,23 @@ where
     /// offloading.
     /// While it is safe to use inmem values for determining horizon,
     /// we use persistent to make possible normal states less surprising.
-    pub fn get_horizon_segno(&self, s3_offload_enabled: bool) -> XLogSegNo {
+    pub fn get_horizon_segno(&self, wal_backup_enabled: bool) -> XLogSegNo {
-        let s3_offload_horizon = if s3_offload_enabled {
+        let mut horizon_lsn = min(
-            self.state.s3_wal_lsn
+            self.state.remote_consistent_lsn,
-        } else {
+            self.state.peer_horizon_lsn,
-            Lsn(u64::MAX)
-        };
-        let horizon_lsn = min(
-            min(
-                self.state.remote_consistent_lsn,
-                self.state.peer_horizon_lsn,
-            ),
-            s3_offload_horizon,
         );
+        if wal_backup_enabled {
+            horizon_lsn = min(horizon_lsn, self.state.backup_lsn);
+        }
         horizon_lsn.segment_number(self.state.server.wal_seg_size as usize)
     }
 }
 
 #[cfg(test)]
 mod tests {
-    use std::ops::Deref;
-
     use super::*;
     use crate::wal_storage::Storage;
+    use std::ops::Deref;
 
     // fake storage for tests
     struct InMemoryState {
@@ -1013,7 +1003,8 @@ mod tests {
         };
         let wal_store = DummyWalStore { lsn: Lsn(0) };
         let ztli = ZTimelineId::from([0u8; 16]);
-        let mut sk = SafeKeeper::new(ztli, storage, wal_store, ZNodeId(0)).unwrap();
+
+        let mut sk = SafeKeeper::new(ztli, storage, wal_store, NodeId(0)).unwrap();
 
         // check voting for 1 is ok
         let vote_request = ProposerAcceptorMessage::VoteRequest(VoteRequest { term: 1 });
@@ -1028,7 +1019,8 @@ mod tests {
         let storage = InMemoryState {
             persisted_state: state,
         };
-        sk = SafeKeeper::new(ztli, storage, sk.wal_store, ZNodeId(0)).unwrap();
+
+        sk = SafeKeeper::new(ztli, storage, sk.wal_store, NodeId(0)).unwrap();
 
         // and ensure voting second time for 1 is not ok
         vote_resp = sk.process_msg(&vote_request);
@@ -1045,7 +1037,8 @@ mod tests {
         };
         let wal_store = DummyWalStore { lsn: Lsn(0) };
         let ztli = ZTimelineId::from([0u8; 16]);
-        let mut sk = SafeKeeper::new(ztli, storage, wal_store, ZNodeId(0)).unwrap();
+
+        let mut sk = SafeKeeper::new(ztli, storage, wal_store, NodeId(0)).unwrap();
 
         let mut ar_hdr = AppendRequestHeader {
             term: 1,

safekeeper/src/send_wal.rs

@@ -315,7 +315,7 @@ impl ReplicationConn {
                 } else {
                     // TODO: also check once in a while whether we are walsender
                     // to right pageserver.
-                    if spg.timeline.get().check_deactivate(replica_id)? {
+                    if spg.timeline.get().stop_walsender(replica_id)? {
                         // Shut down, timeline is suspended.
                         // TODO create proper error type for this
                         bail!("end streaming to {:?}", spg.appname);

safekeeper/src/timeline.rs

@@ -8,6 +8,7 @@ use lazy_static::lazy_static;
 use postgres_ffi::xlog_utils::XLogSegNo;
 
 use serde::Serialize;
+use tokio::sync::watch;
 
 use std::cmp::{max, min};
 use std::collections::HashMap;
@@ -15,23 +16,23 @@ use std::fs::{self};
 
 use std::sync::{Arc, Condvar, Mutex, MutexGuard};
 use std::time::Duration;
-use tokio::sync::mpsc::UnboundedSender;
+use tokio::sync::mpsc::{Sender, UnboundedSender};
 use tracing::*;
 
 use utils::{
     lsn::Lsn,
     pq_proto::ZenithFeedback,
-    zid::{ZNodeId, ZTenantId, ZTenantTimelineId},
+    zid::{NodeId, ZTenantId, ZTenantTimelineId},
 };
 
 use crate::callmemaybe::{CallmeEvent, SubscriptionStateKey};
-
 use crate::control_file;
 use crate::safekeeper::{
     AcceptorProposerMessage, ProposerAcceptorMessage, SafeKeeper, SafeKeeperState,
     SafekeeperMemState,
 };
 use crate::send_wal::HotStandbyFeedback;
+
 use crate::wal_storage;
 use crate::wal_storage::Storage as wal_storage_iface;
 use crate::SafeKeeperConf;
@@ -81,10 +82,14 @@ struct SharedState {
     notified_commit_lsn: Lsn,
     /// State of replicas
     replicas: Vec<Option<ReplicaState>>,
-    /// Inactive clusters shouldn't occupy any resources, so timeline is
+    /// True when WAL backup launcher oversees the timeline, making sure WAL is
-    /// activated whenever there is a compute connection or pageserver is not
+    /// offloaded, allows to bother launcher less.
-    /// caughtup (it must have latest WAL for new compute start) and suspended
+    wal_backup_active: bool,
-    /// otherwise.
+    /// True whenever there is at least some pending activity on timeline: live
+    /// compute connection, pageserver is not caughtup (it must have latest WAL
+    /// for new compute start) or WAL backuping is not finished. Practically it
+    /// means safekeepers broadcast info to peers about the timeline, old WAL is
+    /// trimmed.
     ///
     /// TODO: it might be better to remove tli completely from GlobalTimelines
     /// when tli is inactive instead of having this flag.
@@ -99,10 +104,11 @@ impl SharedState {
     fn create(
         conf: &SafeKeeperConf,
         zttid: &ZTenantTimelineId,
-        peer_ids: Vec<ZNodeId>,
+        peer_ids: Vec<NodeId>,
     ) -> Result<Self> {
         let state = SafeKeeperState::new(zttid, peer_ids);
         let control_store = control_file::FileStorage::create_new(zttid, conf, state)?;
+
         let wal_store = wal_storage::PhysicalStorage::new(zttid, conf);
         let sk = SafeKeeper::new(zttid.timeline_id, control_store, wal_store, conf.my_id)?;
 
@@ -110,6 +116,7 @@ impl SharedState {
             notified_commit_lsn: Lsn(0),
             sk,
             replicas: Vec::new(),
+            wal_backup_active: false,
             active: false,
             num_computes: 0,
             pageserver_connstr: None,
@@ -129,15 +136,62 @@ impl SharedState {
             notified_commit_lsn: Lsn(0),
             sk: SafeKeeper::new(zttid.timeline_id, control_store, wal_store, conf.my_id)?,
             replicas: Vec::new(),
+            wal_backup_active: false,
             active: false,
             num_computes: 0,
             pageserver_connstr: None,
             last_removed_segno: 0,
         })
     }
+    fn is_active(&self) -> bool {
+        self.is_wal_backup_required()
+            // FIXME: add tracking of relevant pageservers and check them here individually,
+            // otherwise migration won't work (we suspend too early).
+            || self.sk.inmem.remote_consistent_lsn <= self.sk.inmem.commit_lsn
+    }
 
-    /// Activate the timeline: start/change walsender (via callmemaybe).
+    /// Mark timeline active/inactive and return whether s3 offloading requires
-    fn activate(
+    /// start/stop action.
+    fn update_status(&mut self) -> bool {
+        self.active = self.is_active();
+        self.is_wal_backup_action_pending()
+    }
+
+    /// Should we run s3 offloading in current state?
+    fn is_wal_backup_required(&self) -> bool {
+        let seg_size = self.get_wal_seg_size();
+        self.num_computes > 0 ||
+        // Currently only the whole segment is offloaded, so compare segment numbers.
+               (self.sk.inmem.commit_lsn.segment_number(seg_size) >
+                self.sk.inmem.backup_lsn.segment_number(seg_size))
+    }
+
+    /// Is current state of s3 offloading is not what it ought to be?
+    fn is_wal_backup_action_pending(&self) -> bool {
+        let res = self.wal_backup_active != self.is_wal_backup_required();
+        if res {
+            let action_pending = if self.is_wal_backup_required() {
+                "start"
+            } else {
+                "stop"
+            };
+            trace!(
+                "timeline {} s3 offloading action {} pending: num_computes={}, commit_lsn={}, backup_lsn={}",
+                self.sk.state.timeline_id, action_pending, self.num_computes, self.sk.inmem.commit_lsn, self.sk.inmem.backup_lsn
+            );
+        }
+        res
+    }
+
+    /// Returns whether s3 offloading is required and sets current status as
+    /// matching.
+    fn wal_backup_attend(&mut self) -> bool {
+        self.wal_backup_active = self.is_wal_backup_required();
+        self.wal_backup_active
+    }
+
+    /// start/change walsender (via callmemaybe).
+    fn callmemaybe_sub(
         &mut self,
         zttid: &ZTenantTimelineId,
         pageserver_connstr: Option<&String>,
@@ -179,42 +233,42 @@ impl SharedState {
             );
         }
         self.pageserver_connstr = pageserver_connstr.map(|c| c.to_owned());
-        self.active = true;
         Ok(())
     }
 
     /// Deactivate the timeline: stop callmemaybe.
-    fn deactivate(
+    fn callmemaybe_unsub(
         &mut self,
         zttid: &ZTenantTimelineId,
         callmemaybe_tx: &UnboundedSender<CallmeEvent>,
     ) -> Result<()> {
-        if self.active {
+        if let Some(ref pageserver_connstr) = self.pageserver_connstr {
-            if let Some(ref pageserver_connstr) = self.pageserver_connstr {
+            let subscription_key = SubscriptionStateKey::new(
-                let subscription_key = SubscriptionStateKey::new(
+                zttid.tenant_id,
-                    zttid.tenant_id,
+                zttid.timeline_id,
-                    zttid.timeline_id,
+                pageserver_connstr.to_owned(),
-                    pageserver_connstr.to_owned(),
+            );
-                );
+            callmemaybe_tx
-                callmemaybe_tx
+                .send(CallmeEvent::Unsubscribe(subscription_key))
-                    .send(CallmeEvent::Unsubscribe(subscription_key))
+                .unwrap_or_else(|e| {
-                    .unwrap_or_else(|e| {
+                    error!(
-                        error!(
+                        "failed to send Unsubscribe request to callmemaybe thread {}",
-                            "failed to send Unsubscribe request to callmemaybe thread {}",
+                        e
-                            e
+                    );
-                        );
+                });
-                    });
+            info!(
-                info!(
+                "timeline {} is unsubscribed from callmemaybe to {}",
-                    "timeline {} is unsubscribed from callmemaybe to {}",
+                zttid.timeline_id,
-                    zttid.timeline_id,
+                self.pageserver_connstr.as_ref().unwrap()
-                    self.pageserver_connstr.as_ref().unwrap()
+            );
-                );
-            }
-            self.active = false;
         }
         Ok(())
     }
 
+    fn get_wal_seg_size(&self) -> usize {
+        self.sk.state.server.wal_seg_size as usize
+    }
+
     /// Get combined state of all alive replicas
     pub fn get_replicas_state(&self) -> ReplicaState {
         let mut acc = ReplicaState::new();
@@ -278,6 +332,13 @@ impl SharedState {
 pub struct Timeline {
     pub zttid: ZTenantTimelineId,
     pub callmemaybe_tx: UnboundedSender<CallmeEvent>,
+    /// Sending here asks for wal backup launcher attention (start/stop
+    /// offloading). Sending zttid instead of concrete command allows to do
+    /// sending without timeline lock.
+    wal_backup_launcher_tx: Sender<ZTenantTimelineId>,
+    commit_lsn_watch_tx: watch::Sender<Lsn>,
+    /// For breeding receivers.
+    commit_lsn_watch_rx: watch::Receiver<Lsn>,
     mutex: Mutex<SharedState>,
     /// conditional variable used to notify wal senders
     cond: Condvar,
@@ -287,11 +348,17 @@ impl Timeline {
     fn new(
         zttid: ZTenantTimelineId,
         callmemaybe_tx: UnboundedSender<CallmeEvent>,
+        wal_backup_launcher_tx: Sender<ZTenantTimelineId>,
         shared_state: SharedState,
     ) -> Timeline {
+        let (commit_lsn_watch_tx, commit_lsn_watch_rx) =
+            watch::channel(shared_state.sk.inmem.commit_lsn);
         Timeline {
             zttid,
             callmemaybe_tx,
+            wal_backup_launcher_tx,
+            commit_lsn_watch_tx,
+            commit_lsn_watch_rx,
             mutex: Mutex::new(shared_state),
             cond: Condvar::new(),
         }
@@ -301,13 +368,21 @@ impl Timeline {
     /// not running yet.
     /// Can fail only if channel to a static thread got closed, which is not normal at all.
     pub fn on_compute_connect(&self, pageserver_connstr: Option<&String>) -> Result<()> {
-        let mut shared_state = self.mutex.lock().unwrap();
+        let is_wal_backup_action_pending: bool;
-        shared_state.num_computes += 1;
+        {
-        // FIXME: currently we always adopt latest pageserver connstr, but we
+            let mut shared_state = self.mutex.lock().unwrap();
-        // should have kind of generations assigned by compute to distinguish
+            shared_state.num_computes += 1;
-        // the latest one or even pass it through consensus to reliably deliver
+            is_wal_backup_action_pending = shared_state.update_status();
-        // to all safekeepers.
+            // FIXME: currently we always adopt latest pageserver connstr, but we
-        shared_state.activate(&self.zttid, pageserver_connstr, &self.callmemaybe_tx)?;
+            // should have kind of generations assigned by compute to distinguish
+            // the latest one or even pass it through consensus to reliably deliver
+            // to all safekeepers.
+            shared_state.callmemaybe_sub(&self.zttid, pageserver_connstr, &self.callmemaybe_tx)?;
+        }
+        // Wake up wal backup launcher, if offloading not started yet.
+        if is_wal_backup_action_pending {
+            self.wal_backup_launcher_tx.blocking_send(self.zttid)?;
+        }
         Ok(())
     }
 
@@ -315,38 +390,43 @@ impl Timeline {
     /// pageserver doesn't need catchup.
     /// Can fail only if channel to a static thread got closed, which is not normal at all.
     pub fn on_compute_disconnect(&self) -> Result<()> {
-        let mut shared_state = self.mutex.lock().unwrap();
+        let is_wal_backup_action_pending: bool;
-        shared_state.num_computes -= 1;
+        {
-        // If there is no pageserver, can suspend right away; otherwise let
+            let mut shared_state = self.mutex.lock().unwrap();
-        // walsender do that.
+            shared_state.num_computes -= 1;
-        if shared_state.num_computes == 0 && shared_state.pageserver_connstr.is_none() {
+            is_wal_backup_action_pending = shared_state.update_status();
-            shared_state.deactivate(&self.zttid, &self.callmemaybe_tx)?;
+        }
+        // Wake up wal backup launcher, if it is time to stop the offloading.
+        if is_wal_backup_action_pending {
+            self.wal_backup_launcher_tx.blocking_send(self.zttid)?;
         }
         Ok(())
     }
 
-    /// Deactivate tenant if there is no computes and pageserver is caughtup,
+    /// Whether we still need this walsender running?
-    /// assuming the pageserver status is in replica_id.
+    /// TODO: check this pageserver is actually interested in this timeline.
-    /// Returns true if deactivated.
+    pub fn stop_walsender(&self, replica_id: usize) -> Result<bool> {
-    pub fn check_deactivate(&self, replica_id: usize) -> Result<bool> {
         let mut shared_state = self.mutex.lock().unwrap();
-        if !shared_state.active {
-            // already suspended
-            return Ok(true);
-        }
         if shared_state.num_computes == 0 {
             let replica_state = shared_state.replicas[replica_id].unwrap();
-            let deactivate = shared_state.notified_commit_lsn == Lsn(0) || // no data at all yet
+            let stop = shared_state.notified_commit_lsn == Lsn(0) || // no data at all yet
-            (replica_state.last_received_lsn != Lsn::MAX && // Lsn::MAX means that we don't know the latest LSN yet.
+            (replica_state.remote_consistent_lsn != Lsn::MAX && // Lsn::MAX means that we don't know the latest LSN yet.
-             replica_state.last_received_lsn >= shared_state.sk.inmem.commit_lsn);
+             replica_state.remote_consistent_lsn >= shared_state.sk.inmem.commit_lsn);
-            if deactivate {
+            if stop {
-                shared_state.deactivate(&self.zttid, &self.callmemaybe_tx)?;
+                shared_state.callmemaybe_unsub(&self.zttid, &self.callmemaybe_tx)?;
                 return Ok(true);
             }
         }
         Ok(false)
     }
 
+    /// Returns whether s3 offloading is required and sets current status as
+    /// matching it.
+    pub fn wal_backup_attend(&self) -> bool {
+        let mut shared_state = self.mutex.lock().unwrap();
+        shared_state.wal_backup_attend()
+    }
+
     /// Deactivates the timeline, assuming it is being deleted.
     /// Returns whether the timeline was already active.
     ///
@@ -354,10 +434,14 @@ impl Timeline {
     /// will stop by themselves eventually (possibly with errors, but no panics). There should be no
     /// compute threads (as we're deleting the timeline), actually. Some WAL may be left unsent, but
     /// we're deleting the timeline anyway.
-    pub fn deactivate_for_delete(&self) -> Result<bool> {
+    pub async fn deactivate_for_delete(&self) -> Result<bool> {
-        let mut shared_state = self.mutex.lock().unwrap();
+        let was_active: bool;
-        let was_active = shared_state.active;
+        {
-        shared_state.deactivate(&self.zttid, &self.callmemaybe_tx)?;
+            let mut shared_state = self.mutex.lock().unwrap();
+            was_active = shared_state.active;
+            shared_state.callmemaybe_unsub(&self.zttid, &self.callmemaybe_tx)?;
+        }
+        self.wal_backup_launcher_tx.send(self.zttid).await?;
         Ok(was_active)
     }
 
@@ -391,6 +475,7 @@ impl Timeline {
     }
 
     // Notify caught-up WAL senders about new WAL data received
+    // TODO: replace-unify it with commit_lsn_watch.
     fn notify_wal_senders(&self, shared_state: &mut MutexGuard<SharedState>) {
         if shared_state.notified_commit_lsn < shared_state.sk.inmem.commit_lsn {
             shared_state.notified_commit_lsn = shared_state.sk.inmem.commit_lsn;
@@ -398,12 +483,17 @@ impl Timeline {
         }
     }
 
+    pub fn get_commit_lsn_watch_rx(&self) -> watch::Receiver<Lsn> {
+        self.commit_lsn_watch_rx.clone()
+    }
+
     /// Pass arrived message to the safekeeper.
     pub fn process_msg(
         &self,
         msg: &ProposerAcceptorMessage,
     ) -> Result<Option<AcceptorProposerMessage>> {
         let mut rmsg: Option<AcceptorProposerMessage>;
+        let commit_lsn: Lsn;
         {
             let mut shared_state = self.mutex.lock().unwrap();
             rmsg = shared_state.sk.process_msg(msg)?;
@@ -419,15 +509,31 @@ impl Timeline {
 
             // Ping wal sender that new data might be available.
             self.notify_wal_senders(&mut shared_state);
+            commit_lsn = shared_state.sk.inmem.commit_lsn;
         }
+        self.commit_lsn_watch_tx.send(commit_lsn)?;
         Ok(rmsg)
     }
 
+    pub fn get_wal_seg_size(&self) -> usize {
+        self.mutex.lock().unwrap().get_wal_seg_size()
+    }
+
     pub fn get_state(&self) -> (SafekeeperMemState, SafeKeeperState) {
         let shared_state = self.mutex.lock().unwrap();
         (shared_state.sk.inmem.clone(), shared_state.sk.state.clone())
     }
 
+    pub fn get_wal_backup_lsn(&self) -> Lsn {
+        self.mutex.lock().unwrap().sk.inmem.backup_lsn
+    }
+
+    pub fn set_wal_backup_lsn(&self, backup_lsn: Lsn) {
+        self.mutex.lock().unwrap().sk.inmem.backup_lsn = backup_lsn;
+        // we should check whether to shut down offloader, but this will be done
+        // soon by peer communication anyway.
+    }
+
     /// Prepare public safekeeper info for reporting.
     pub fn get_public_info(&self, conf: &SafeKeeperConf) -> anyhow::Result<SkTimelineInfo> {
         let shared_state = self.mutex.lock().unwrap();
@@ -436,7 +542,6 @@ impl Timeline {
             flush_lsn: Some(shared_state.sk.wal_store.flush_lsn()),
             // note: this value is not flushed to control file yet and can be lost
             commit_lsn: Some(shared_state.sk.inmem.commit_lsn),
-            s3_wal_lsn: Some(shared_state.sk.inmem.s3_wal_lsn),
             // TODO: rework feedbacks to avoid max here
             remote_consistent_lsn: Some(max(
                 shared_state.get_replicas_state().remote_consistent_lsn,
@@ -444,14 +549,35 @@ impl Timeline {
             )),
             peer_horizon_lsn: Some(shared_state.sk.inmem.peer_horizon_lsn),
             safekeeper_connection_string: Some(conf.listen_pg_addr.clone()),
+            backup_lsn: Some(shared_state.sk.inmem.backup_lsn),
         })
     }
 
     /// Update timeline state with peer safekeeper data.
-    pub fn record_safekeeper_info(&self, sk_info: &SkTimelineInfo, _sk_id: ZNodeId) -> Result<()> {
+    pub async fn record_safekeeper_info(
-        let mut shared_state = self.mutex.lock().unwrap();
+        &self,
-        shared_state.sk.record_safekeeper_info(sk_info)?;
+        sk_info: &SkTimelineInfo,
-        self.notify_wal_senders(&mut shared_state);
+        _sk_id: NodeId,
+    ) -> Result<()> {
+        let is_wal_backup_action_pending: bool;
+        let commit_lsn: Lsn;
+        {
+            let mut shared_state = self.mutex.lock().unwrap();
+            // WAL seg size not initialized yet (no message from compute ever
+            // received), can't do much without it.
+            if shared_state.get_wal_seg_size() == 0 {
+                return Ok(());
+            }
+            shared_state.sk.record_safekeeper_info(sk_info)?;
+            self.notify_wal_senders(&mut shared_state);
+            is_wal_backup_action_pending = shared_state.update_status();
+            commit_lsn = shared_state.sk.inmem.commit_lsn;
+        }
+        self.commit_lsn_watch_tx.send(commit_lsn)?;
+        // Wake up wal backup launcher, if it is time to stop the offloading.
+        if is_wal_backup_action_pending {
+            self.wal_backup_launcher_tx.send(self.zttid).await?;
+        }
         Ok(())
     }
 
@@ -476,16 +602,16 @@ impl Timeline {
         shared_state.sk.wal_store.flush_lsn()
     }
 
-    pub fn remove_old_wal(&self, s3_offload_enabled: bool) -> Result<()> {
+    pub fn remove_old_wal(&self, wal_backup_enabled: bool) -> Result<()> {
         let horizon_segno: XLogSegNo;
         let remover: Box<dyn Fn(u64) -> Result<(), anyhow::Error>>;
         {
             let shared_state = self.mutex.lock().unwrap();
             // WAL seg size not initialized yet, no WAL exists.
-            if shared_state.sk.state.server.wal_seg_size == 0 {
+            if shared_state.get_wal_seg_size() == 0 {
                 return Ok(());
             }
-            horizon_segno = shared_state.sk.get_horizon_segno(s3_offload_enabled);
+            horizon_segno = shared_state.sk.get_horizon_segno(wal_backup_enabled);
             remover = shared_state.sk.wal_store.remove_up_to();
             if horizon_segno <= 1 || horizon_segno <= shared_state.last_removed_segno {
                 return Ok(());
@@ -522,12 +648,14 @@ impl TimelineTools for Option<Arc<Timeline>> {
 struct GlobalTimelinesState {
     timelines: HashMap<ZTenantTimelineId, Arc<Timeline>>,
     callmemaybe_tx: Option<UnboundedSender<CallmeEvent>>,
+    wal_backup_launcher_tx: Option<Sender<ZTenantTimelineId>>,
 }
 
 lazy_static! {
     static ref TIMELINES_STATE: Mutex<GlobalTimelinesState> = Mutex::new(GlobalTimelinesState {
         timelines: HashMap::new(),
-        callmemaybe_tx: None
+        callmemaybe_tx: None,
+        wal_backup_launcher_tx: None,
     });
 }
 
@@ -541,17 +669,22 @@ pub struct TimelineDeleteForceResult {
 pub struct GlobalTimelines;
 
 impl GlobalTimelines {
-    pub fn set_callmemaybe_tx(callmemaybe_tx: UnboundedSender<CallmeEvent>) {
+    pub fn init(
+        callmemaybe_tx: UnboundedSender<CallmeEvent>,
+        wal_backup_launcher_tx: Sender<ZTenantTimelineId>,
+    ) {
         let mut state = TIMELINES_STATE.lock().unwrap();
         assert!(state.callmemaybe_tx.is_none());
         state.callmemaybe_tx = Some(callmemaybe_tx);
+        assert!(state.wal_backup_launcher_tx.is_none());
+        state.wal_backup_launcher_tx = Some(wal_backup_launcher_tx);
     }
 
     fn create_internal(
         mut state: MutexGuard<GlobalTimelinesState>,
         conf: &SafeKeeperConf,
         zttid: ZTenantTimelineId,
-        peer_ids: Vec<ZNodeId>,
+        peer_ids: Vec<NodeId>,
     ) -> Result<Arc<Timeline>> {
         match state.timelines.get(&zttid) {
             Some(_) => bail!("timeline {} already exists", zttid),
@@ -559,12 +692,14 @@ impl GlobalTimelines {
                 // TODO: check directory existence
                 let dir = conf.timeline_dir(&zttid);
                 fs::create_dir_all(dir)?;
+
                 let shared_state = SharedState::create(conf, &zttid, peer_ids)
                     .context("failed to create shared state")?;
 
                 let new_tli = Arc::new(Timeline::new(
                     zttid,
                     state.callmemaybe_tx.as_ref().unwrap().clone(),
+                    state.wal_backup_launcher_tx.as_ref().unwrap().clone(),
                     shared_state,
                 ));
                 state.timelines.insert(zttid, Arc::clone(&new_tli));
@@ -576,7 +711,7 @@ impl GlobalTimelines {
     pub fn create(
         conf: &SafeKeeperConf,
         zttid: ZTenantTimelineId,
-        peer_ids: Vec<ZNodeId>,
+        peer_ids: Vec<NodeId>,
     ) -> Result<Arc<Timeline>> {
         let state = TIMELINES_STATE.lock().unwrap();
         GlobalTimelines::create_internal(state, conf, zttid, peer_ids)
@@ -594,8 +729,7 @@ impl GlobalTimelines {
         match state.timelines.get(&zttid) {
             Some(result) => Ok(Arc::clone(result)),
             None => {
-                let shared_state =
+                let shared_state = SharedState::restore(conf, &zttid);
-                    SharedState::restore(conf, &zttid).context("failed to restore shared state");
 
                 let shared_state = match shared_state {
                     Ok(shared_state) => shared_state,
@@ -617,6 +751,7 @@ impl GlobalTimelines {
                 let new_tli = Arc::new(Timeline::new(
                     zttid,
                     state.callmemaybe_tx.as_ref().unwrap().clone(),
+                    state.wal_backup_launcher_tx.as_ref().unwrap().clone(),
                     shared_state,
                 ));
                 state.timelines.insert(zttid, Arc::clone(&new_tli));
@@ -625,6 +760,12 @@ impl GlobalTimelines {
         }
     }
 
+    /// Get loaded timeline, if it exists.
+    pub fn get_loaded(zttid: ZTenantTimelineId) -> Option<Arc<Timeline>> {
+        let state = TIMELINES_STATE.lock().unwrap();
+        state.timelines.get(&zttid).map(Arc::clone)
+    }
+
     /// Get ZTenantTimelineIDs of all active timelines.
     pub fn get_active_timelines() -> Vec<ZTenantTimelineId> {
         let state = TIMELINES_STATE.lock().unwrap();
@@ -665,22 +806,23 @@ impl GlobalTimelines {
     /// b) an HTTP GET request about the timeline is made and it's able to restore the current state, or
     /// c) an HTTP POST request for timeline creation is made after the timeline is already deleted.
     /// TODO: ensure all of the above never happens.
-    pub fn delete_force(
+    pub async fn delete_force(
         conf: &SafeKeeperConf,
         zttid: &ZTenantTimelineId,
     ) -> Result<TimelineDeleteForceResult> {
         info!("deleting timeline {}", zttid);
-        let was_active = match TIMELINES_STATE.lock().unwrap().timelines.remove(zttid) {
+        let timeline = TIMELINES_STATE.lock().unwrap().timelines.remove(zttid);
-            None => false,
+        let mut was_active = false;
-            Some(tli) => tli.deactivate_for_delete()?,
+        if let Some(tli) = timeline {
-        };
+            was_active = tli.deactivate_for_delete().await?;
+        }
         GlobalTimelines::delete_force_internal(conf, zttid, was_active)
     }
 
     /// Deactivates and deletes all timelines for the tenant, see `delete()`.
     /// Returns map of all timelines which the tenant had, `true` if a timeline was active.
     /// There may be a race if new timelines are created simultaneously.
-    pub fn delete_force_all_for_tenant(
+    pub async fn delete_force_all_for_tenant(
         conf: &SafeKeeperConf,
         tenant_id: &ZTenantId,
     ) -> Result<HashMap<ZTenantTimelineId, TimelineDeleteForceResult>> {
@@ -691,14 +833,15 @@ impl GlobalTimelines {
             let timelines = &mut TIMELINES_STATE.lock().unwrap().timelines;
             for (&zttid, tli) in timelines.iter() {
                 if zttid.tenant_id == *tenant_id {
-                    to_delete.insert(zttid, tli.deactivate_for_delete()?);
+                    to_delete.insert(zttid, tli.clone());
                 }
             }
             // TODO: test that the correct subset of timelines is removed. It's complicated because they are implicitly created currently.
             timelines.retain(|zttid, _| !to_delete.contains_key(zttid));
         }
         let mut deleted = HashMap::new();
-        for (zttid, was_active) in to_delete {
+        for (zttid, timeline) in to_delete {
+            let was_active = timeline.deactivate_for_delete().await?;
             deleted.insert(
                 zttid,
                 GlobalTimelines::delete_force_internal(conf, &zttid, was_active)?,

safekeeper/src/wal_backup.rs

@@ -0,0 +1,417 @@
+use anyhow::{Context, Result};
+use tokio::task::JoinHandle;
+
+use std::cmp::min;
+use std::collections::HashMap;
+use std::path::{Path, PathBuf};
+use std::sync::Arc;
+use std::time::Duration;
+
+use postgres_ffi::xlog_utils::{XLogFileName, XLogSegNo, XLogSegNoOffsetToRecPtr, PG_TLI};
+use remote_storage::{GenericRemoteStorage, RemoteStorage};
+use tokio::fs::File;
+use tokio::runtime::Builder;
+
+use tokio::select;
+use tokio::sync::mpsc::{self, Receiver, Sender};
+use tokio::sync::watch;
+use tokio::time::sleep;
+use tracing::*;
+
+use utils::{lsn::Lsn, zid::ZTenantTimelineId};
+
+use crate::broker::{Election, ElectionLeader};
+use crate::timeline::{GlobalTimelines, Timeline};
+use crate::{broker, SafeKeeperConf};
+
+use once_cell::sync::OnceCell;
+
+const BACKUP_ELECTION_NAME: &str = "WAL_BACKUP";
+
+const BROKER_CONNECTION_RETRY_DELAY_MS: u64 = 1000;
+
+const UPLOAD_FAILURE_RETRY_MIN_MS: u64 = 10;
+const UPLOAD_FAILURE_RETRY_MAX_MS: u64 = 5000;
+
+pub fn wal_backup_launcher_thread_main(
+    conf: SafeKeeperConf,
+    wal_backup_launcher_rx: Receiver<ZTenantTimelineId>,
+) {
+    let rt = Builder::new_multi_thread()
+        .worker_threads(conf.backup_runtime_threads)
+        .enable_all()
+        .build()
+        .expect("failed to create wal backup runtime");
+
+    rt.block_on(async {
+        wal_backup_launcher_main_loop(conf, wal_backup_launcher_rx).await;
+    });
+}
+
+/// Check whether wal backup is required for timeline and mark that launcher is
+/// aware of current status (if timeline exists).
+fn is_wal_backup_required(zttid: ZTenantTimelineId) -> bool {
+    if let Some(tli) = GlobalTimelines::get_loaded(zttid) {
+        tli.wal_backup_attend()
+    } else {
+        false
+    }
+}
+
+struct WalBackupTaskHandle {
+    shutdown_tx: Sender<()>,
+    handle: JoinHandle<()>,
+}
+
+/// Sits on wal_backup_launcher_rx and starts/stops per timeline wal backup
+/// tasks. Having this in separate task simplifies locking, allows to reap
+/// panics and separate elections from offloading itself.
+async fn wal_backup_launcher_main_loop(
+    conf: SafeKeeperConf,
+    mut wal_backup_launcher_rx: Receiver<ZTenantTimelineId>,
+) {
+    info!(
+        "WAL backup launcher: started, remote config {:?}",
+        conf.remote_storage
+    );
+
+    let conf_ = conf.clone();
+    REMOTE_STORAGE.get_or_init(|| {
+        conf_.remote_storage.as_ref().map(|c| {
+            GenericRemoteStorage::new(conf_.workdir, c).expect("failed to create remote storage")
+        })
+    });
+
+    let mut tasks: HashMap<ZTenantTimelineId, WalBackupTaskHandle> = HashMap::new();
+
+    loop {
+        // channel is never expected to get closed
+        let zttid = wal_backup_launcher_rx.recv().await.unwrap();
+        let is_wal_backup_required = is_wal_backup_required(zttid);
+        if conf.remote_storage.is_none() || !conf.wal_backup_enabled {
+            continue; /* just drain the channel and do nothing */
+        }
+        // do we need to do anything at all?
+        if is_wal_backup_required != tasks.contains_key(&zttid) {
+            if is_wal_backup_required {
+                // need to start the task
+                info!("starting WAL backup task for {}", zttid);
+
+                // TODO: decide who should offload in launcher itself by simply checking current state
+                let election_name = broker::get_campaign_name(
+                    BACKUP_ELECTION_NAME.to_string(),
+                    conf.broker_etcd_prefix.clone(),
+                    &zttid,
+                );
+                let my_candidate_name = broker::get_candiate_name(conf.my_id);
+                let election = broker::Election::new(
+                    election_name,
+                    my_candidate_name,
+                    conf.broker_endpoints.clone(),
+                );
+
+                let (shutdown_tx, shutdown_rx) = mpsc::channel(1);
+                let timeline_dir = conf.timeline_dir(&zttid);
+
+                let handle = tokio::spawn(
+                    backup_task_main(zttid, timeline_dir, shutdown_rx, election)
+                        .instrument(info_span!("WAL backup task", zttid = %zttid)),
+                );
+
+                tasks.insert(
+                    zttid,
+                    WalBackupTaskHandle {
+                        shutdown_tx,
+                        handle,
+                    },
+                );
+            } else {
+                // need to stop the task
+                info!("stopping WAL backup task for {}", zttid);
+
+                let wb_handle = tasks.remove(&zttid).unwrap();
+                // Tell the task to shutdown. Error means task exited earlier, that's ok.
+                let _ = wb_handle.shutdown_tx.send(()).await;
+                // Await the task itself. TODO: restart panicked tasks earlier.
+                // Hm, why I can't await on reference to handle?
+                if let Err(e) = wb_handle.handle.await {
+                    warn!("WAL backup task for {} panicked: {}", zttid, e);
+                }
+            }
+        }
+    }
+}
+
+struct WalBackupTask {
+    timeline: Arc<Timeline>,
+    timeline_dir: PathBuf,
+    wal_seg_size: usize,
+    commit_lsn_watch_rx: watch::Receiver<Lsn>,
+    leader: Option<ElectionLeader>,
+    election: Election,
+}
+
+/// Offload single timeline.
+async fn backup_task_main(
+    zttid: ZTenantTimelineId,
+    timeline_dir: PathBuf,
+    mut shutdown_rx: Receiver<()>,
+    election: Election,
+) {
+    info!("started");
+    let timeline: Arc<Timeline> = if let Some(tli) = GlobalTimelines::get_loaded(zttid) {
+        tli
+    } else {
+        /* Timeline could get deleted while task was starting, just exit then. */
+        info!("no timeline, exiting");
+        return;
+    };
+
+    let mut wb = WalBackupTask {
+        wal_seg_size: timeline.get_wal_seg_size(),
+        commit_lsn_watch_rx: timeline.get_commit_lsn_watch_rx(),
+        timeline,
+        timeline_dir,
+        leader: None,
+        election,
+    };
+
+    // task is spinned up only when wal_seg_size already initialized
+    assert!(wb.wal_seg_size > 0);
+
+    let mut canceled = false;
+    select! {
+        _ = wb.run() => {}
+        _ = shutdown_rx.recv() => {
+            canceled = true;
+        }
+    }
+    if let Some(l) = wb.leader {
+        l.give_up().await;
+    }
+    info!("task {}", if canceled { "canceled" } else { "terminated" });
+}
+
+impl WalBackupTask {
+    async fn run(&mut self) {
+        let mut backup_lsn = Lsn(0);
+
+        // election loop
+        loop {
+            let mut retry_attempt = 0u32;
+
+            if let Some(l) = self.leader.take() {
+                l.give_up().await;
+            }
+
+            match broker::get_leader(&self.election).await {
+                Ok(l) => {
+                    self.leader = Some(l);
+                }
+                Err(e) => {
+                    error!("error during leader election {:?}", e);
+                    sleep(Duration::from_millis(BROKER_CONNECTION_RETRY_DELAY_MS)).await;
+                    continue;
+                }
+            }
+
+            // offload loop
+            loop {
+                if retry_attempt == 0 {
+                    // wait for new WAL to arrive
+                    if let Err(e) = self.commit_lsn_watch_rx.changed().await {
+                        // should never happen, as we hold Arc to timeline.
+                        error!("commit_lsn watch shut down: {:?}", e);
+                        return;
+                    }
+                } else {
+                    // or just sleep if we errored previously
+                    let mut retry_delay = UPLOAD_FAILURE_RETRY_MAX_MS;
+                    if let Some(backoff_delay) =
+                        UPLOAD_FAILURE_RETRY_MIN_MS.checked_shl(retry_attempt)
+                    {
+                        retry_delay = min(retry_delay, backoff_delay);
+                    }
+                    sleep(Duration::from_millis(retry_delay)).await;
+                }
+
+                let commit_lsn = *self.commit_lsn_watch_rx.borrow();
+
+                // Note that backup_lsn can be higher than commit_lsn if we
+                // don't have much local WAL and others already uploaded
+                // segments we don't even have.
+                if backup_lsn.segment_number(self.wal_seg_size)
+                    >= commit_lsn.segment_number(self.wal_seg_size)
+                {
+                    continue; /* nothing to do, common case as we wake up on every commit_lsn bump */
+                }
+                // Perhaps peers advanced the position, check shmem value.
+                backup_lsn = self.timeline.get_wal_backup_lsn();
+                if backup_lsn.segment_number(self.wal_seg_size)
+                    >= commit_lsn.segment_number(self.wal_seg_size)
+                {
+                    continue;
+                }
+
+                if let Some(l) = self.leader.as_mut() {
+                    // Optimization idea for later:
+                    //  Avoid checking election leader every time by returning current lease grant expiration time
+                    //  Re-check leadership only after expiration time,
+                    //  such approach woud reduce overhead on write-intensive workloads
+
+                    match l
+                        .check_am_i(
+                            self.election.election_name.clone(),
+                            self.election.candidate_name.clone(),
+                        )
+                        .await
+                    {
+                        Ok(leader) => {
+                            if !leader {
+                                info!("leader has changed");
+                                break;
+                            }
+                        }
+                        Err(e) => {
+                            warn!("error validating leader, {:?}", e);
+                            break;
+                        }
+                    }
+                }
+
+                match backup_lsn_range(
+                    backup_lsn,
+                    commit_lsn,
+                    self.wal_seg_size,
+                    &self.timeline_dir,
+                )
+                .await
+                {
+                    Ok(backup_lsn_result) => {
+                        backup_lsn = backup_lsn_result;
+                        self.timeline.set_wal_backup_lsn(backup_lsn_result);
+                        retry_attempt = 0;
+                    }
+                    Err(e) => {
+                        error!(
+                            "failed while offloading range {}-{}: {:?}",
+                            backup_lsn, commit_lsn, e
+                        );
+
+                        retry_attempt = min(retry_attempt + 1, u32::MAX);
+                    }
+                }
+            }
+        }
+    }
+}
+
+pub async fn backup_lsn_range(
+    start_lsn: Lsn,
+    end_lsn: Lsn,
+    wal_seg_size: usize,
+    timeline_dir: &Path,
+) -> Result<Lsn> {
+    let mut res = start_lsn;
+    let segments = get_segments(start_lsn, end_lsn, wal_seg_size);
+    for s in &segments {
+        backup_single_segment(s, timeline_dir)
+            .await
+            .with_context(|| format!("offloading segno {}", s.seg_no))?;
+
+        res = s.end_lsn;
+    }
+    info!(
+        "offloaded segnos {:?} up to {}, previous backup_lsn {}",
+        segments.iter().map(|&s| s.seg_no).collect::<Vec<_>>(),
+        end_lsn,
+        start_lsn,
+    );
+    Ok(res)
+}
+
+async fn backup_single_segment(seg: &Segment, timeline_dir: &Path) -> Result<()> {
+    let segment_file_name = seg.file_path(timeline_dir)?;
+
+    backup_object(&segment_file_name, seg.size()).await?;
+    debug!("Backup of {} done", segment_file_name.display());
+
+    Ok(())
+}
+
+#[derive(Debug, Copy, Clone)]
+pub struct Segment {
+    seg_no: XLogSegNo,
+    start_lsn: Lsn,
+    end_lsn: Lsn,
+}
+
+impl Segment {
+    pub fn new(seg_no: u64, start_lsn: Lsn, end_lsn: Lsn) -> Self {
+        Self {
+            seg_no,
+            start_lsn,
+            end_lsn,
+        }
+    }
+
+    pub fn object_name(self) -> String {
+        XLogFileName(PG_TLI, self.seg_no, self.size())
+    }
+
+    pub fn file_path(self, timeline_dir: &Path) -> Result<PathBuf> {
+        Ok(timeline_dir.join(self.object_name()))
+    }
+
+    pub fn size(self) -> usize {
+        (u64::from(self.end_lsn) - u64::from(self.start_lsn)) as usize
+    }
+}
+
+fn get_segments(start: Lsn, end: Lsn, seg_size: usize) -> Vec<Segment> {
+    let first_seg = start.segment_number(seg_size);
+    let last_seg = end.segment_number(seg_size);
+
+    let res: Vec<Segment> = (first_seg..last_seg)
+        .map(|s| {
+            let start_lsn = XLogSegNoOffsetToRecPtr(s, 0, seg_size);
+            let end_lsn = XLogSegNoOffsetToRecPtr(s + 1, 0, seg_size);
+            Segment::new(s, Lsn::from(start_lsn), Lsn::from(end_lsn))
+        })
+        .collect();
+    res
+}
+
+static REMOTE_STORAGE: OnceCell<Option<GenericRemoteStorage>> = OnceCell::new();
+
+async fn backup_object(source_file: &Path, size: usize) -> Result<()> {
+    let storage = REMOTE_STORAGE.get().expect("failed to get remote storage");
+
+    let file = File::open(&source_file).await?;
+
+    // Storage is initialized by launcher at ths point.
+    match storage.as_ref().unwrap() {
+        GenericRemoteStorage::Local(local_storage) => {
+            let destination = local_storage.remote_object_id(source_file)?;
+
+            debug!(
+                "local upload about to start from {} to {}",
+                source_file.display(),
+                destination.display()
+            );
+            local_storage.upload(file, size, &destination, None).await
+        }
+        GenericRemoteStorage::S3(s3_storage) => {
+            let s3key = s3_storage.remote_object_id(source_file)?;
+
+            debug!(
+                "S3 upload about to start from {} to {:?}",
+                source_file.display(),
+                s3key
+            );
+            s3_storage.upload(file, size, &s3key, None).await
+        }
+    }?;
+
+    Ok(())
+}

test_runner/batch_others/test_basebackup_error.py

@@ -0,0 +1,20 @@
+import pytest
+from contextlib import closing
+
+from fixtures.zenith_fixtures import ZenithEnv
+from fixtures.log_helper import log
+
+
+#
+# Test error handling, if the 'basebackup' command fails in the middle
+# of building the tar archive.
+#
+def test_basebackup_error(zenith_simple_env: ZenithEnv):
+    env = zenith_simple_env
+    env.zenith_cli.create_branch("test_basebackup_error", "empty")
+
+    # Introduce failpoint
+    env.pageserver.safe_psql(f"failpoints basebackup-before-control-file=return")
+
+    with pytest.raises(Exception, match="basebackup-before-control-file"):
+        pg = env.postgres.create_start('test_basebackup_error')

test_runner/batch_others/test_broken_timeline.py

@@ -1,6 +1,7 @@
 import pytest
+import concurrent.futures
 from contextlib import closing
-from fixtures.zenith_fixtures import ZenithEnvBuilder
+from fixtures.zenith_fixtures import ZenithEnvBuilder, ZenithEnv
 from fixtures.log_helper import log
 import os
 
@@ -78,3 +79,37 @@ def test_broken_timeline(zenith_env_builder: ZenithEnvBuilder):
         with pytest.raises(Exception, match="Cannot load local timeline") as err:
             pg.start()
         log.info(f'compute startup failed as expected: {err}')
+
+
+def test_create_multiple_timelines_parallel(zenith_simple_env: ZenithEnv):
+    env = zenith_simple_env
+
+    tenant_id, _ = env.zenith_cli.create_tenant()
+
+    with concurrent.futures.ThreadPoolExecutor(max_workers=4) as executor:
+        futures = [
+            executor.submit(env.zenith_cli.create_timeline,
+                            f"test-create-multiple-timelines-{i}",
+                            tenant_id) for i in range(4)
+        ]
+        for future in futures:
+            future.result()
+
+
+def test_fix_broken_timelines_on_startup(zenith_simple_env: ZenithEnv):
+    env = zenith_simple_env
+
+    tenant_id, _ = env.zenith_cli.create_tenant()
+
+    # Introduce failpoint when creating a new timeline
+    env.pageserver.safe_psql(f"failpoints before-checkpoint-new-timeline=return")
+    with pytest.raises(Exception, match="before-checkpoint-new-timeline"):
+        _ = env.zenith_cli.create_timeline("test_fix_broken_timelines", tenant_id)
+
+    # Restart the page server
+    env.zenith_cli.pageserver_stop(immediate=True)
+    env.zenith_cli.pageserver_start()
+
+    # Check that the "broken" timeline is not loaded
+    timelines = env.zenith_cli.list_timelines(tenant_id)
+    assert len(timelines) == 1

test_runner/batch_others/test_wal_acceptor.py

@@ -12,7 +12,7 @@ from contextlib import closing
 from dataclasses import dataclass, field
 from multiprocessing import Process, Value
 from pathlib import Path
-from fixtures.zenith_fixtures import PgBin, Etcd, Postgres, Safekeeper, ZenithEnv, ZenithEnvBuilder, PortDistributor, SafekeeperPort, zenith_binpath, PgProtocol
+from fixtures.zenith_fixtures import PgBin, Etcd, Postgres, RemoteStorageUsers, Safekeeper, ZenithEnv, ZenithEnvBuilder, PortDistributor, SafekeeperPort, zenith_binpath, PgProtocol
 from fixtures.utils import get_dir_size, lsn_to_hex, mkdir_if_needed, lsn_from_hex
 from fixtures.log_helper import log
 from typing import List, Optional, Any
@@ -401,7 +401,7 @@ def test_wal_removal(zenith_env_builder: ZenithEnvBuilder):
 
     http_cli = env.safekeepers[0].http_client()
     # Pretend WAL is offloaded to s3.
-    http_cli.record_safekeeper_info(tenant_id, timeline_id, {'s3_wal_lsn': 'FFFFFFFF/FEFFFFFF'})
+    http_cli.record_safekeeper_info(tenant_id, timeline_id, {'backup_lsn': 'FFFFFFFF/FEFFFFFF'})
 
     # wait till first segment is removed on all safekeepers
     started_at = time.time()
@@ -414,6 +414,56 @@ def test_wal_removal(zenith_env_builder: ZenithEnvBuilder):
         time.sleep(0.5)
 
 
+@pytest.mark.parametrize('storage_type', ['mock_s3', 'local_fs'])
+def test_wal_backup(zenith_env_builder: ZenithEnvBuilder, storage_type: str):
+    zenith_env_builder.num_safekeepers = 3
+    if storage_type == 'local_fs':
+        zenith_env_builder.enable_local_fs_remote_storage()
+    elif storage_type == 'mock_s3':
+        zenith_env_builder.enable_s3_mock_remote_storage('test_safekeepers_wal_backup')
+    else:
+        raise RuntimeError(f'Unknown storage type: {storage_type}')
+    zenith_env_builder.remote_storage_users = RemoteStorageUsers.SAFEKEEPER
+
+    env = zenith_env_builder.init_start()
+
+    env.zenith_cli.create_branch('test_safekeepers_wal_backup')
+    pg = env.postgres.create_start('test_safekeepers_wal_backup')
+
+    # learn zenith timeline from compute
+    tenant_id = pg.safe_psql("show zenith.zenith_tenant")[0][0]
+    timeline_id = pg.safe_psql("show zenith.zenith_timeline")[0][0]
+
+    pg_conn = pg.connect()
+    cur = pg_conn.cursor()
+    cur.execute('create table t(key int, value text)')
+
+    # Shut down subsequently each of safekeepers and fill a segment while sk is
+    # down; ensure segment gets offloaded by others.
+    offloaded_seg_end = ['0/2000000', '0/3000000', '0/4000000']
+    for victim, seg_end in zip(env.safekeepers, offloaded_seg_end):
+        victim.stop()
+        # roughly fills one segment
+        cur.execute("insert into t select generate_series(1,250000), 'payload'")
+        live_sk = [sk for sk in env.safekeepers if sk != victim][0]
+        http_cli = live_sk.http_client()
+
+        started_at = time.time()
+        while True:
+            tli_status = http_cli.timeline_status(tenant_id, timeline_id)
+            log.info(f"live sk status is {tli_status}")
+
+            if lsn_from_hex(tli_status.backup_lsn) >= lsn_from_hex(seg_end):
+                break
+            elapsed = time.time() - started_at
+            if elapsed > 20:
+                raise RuntimeError(
+                    f"timed out waiting {elapsed:.0f}s segment ending at {seg_end} get offloaded")
+            time.sleep(0.5)
+
+        victim.start()
+
+
 class ProposerPostgres(PgProtocol):
     """Object for running postgres without ZenithEnv"""
     def __init__(self,

test_runner/fixtures/zenith_fixtures.py

@@ -1,6 +1,7 @@
 from __future__ import annotations
 
 from dataclasses import field
+from enum import Flag, auto
 import textwrap
 from cached_property import cached_property
 import asyncpg
@@ -421,10 +422,51 @@ class MockS3Server:
     def secret_key(self) -> str:
         return 'test'
 
+    def access_env_vars(self) -> Dict[Any, Any]:
+        return {
+            'AWS_ACCESS_KEY_ID': self.access_key(),
+            'AWS_SECRET_ACCESS_KEY': self.secret_key(),
+        }
+
     def kill(self):
         self.subprocess.kill()
 
 
+@dataclass
+class LocalFsStorage:
+    local_path: Path
+
+
+@dataclass
+class S3Storage:
+    bucket_name: str
+    bucket_region: str
+    endpoint: Optional[str]
+
+
+RemoteStorage = Union[LocalFsStorage, S3Storage]
+
+
+# serialize as toml inline table
+def remote_storage_to_toml_inline_table(remote_storage):
+    if isinstance(remote_storage, LocalFsStorage):
+        res = f"local_path='{remote_storage.local_path}'"
+    elif isinstance(remote_storage, S3Storage):
+        res = f"bucket_name='{remote_storage.bucket_name}', bucket_region='{remote_storage.bucket_region}'"
+        if remote_storage.endpoint is not None:
+            res += f", endpoint='{remote_storage.endpoint}'"
+        else:
+            raise Exception(f'Unknown storage configuration {remote_storage}')
+    else:
+        raise Exception("invalid remote storage type")
+    return f"{{{res}}}"
+
+
+class RemoteStorageUsers(Flag):
+    PAGESERVER = auto()
+    SAFEKEEPER = auto()
+
+
 class ZenithEnvBuilder:
     """
     Builder object to create a Zenith runtime environment
@@ -440,6 +482,7 @@ class ZenithEnvBuilder:
                  broker: Etcd,
                  mock_s3_server: MockS3Server,
                  remote_storage: Optional[RemoteStorage] = None,
+                 remote_storage_users: RemoteStorageUsers = RemoteStorageUsers.PAGESERVER,
                  pageserver_config_override: Optional[str] = None,
                  num_safekeepers: int = 1,
                  pageserver_auth_enabled: bool = False,
@@ -449,6 +492,7 @@ class ZenithEnvBuilder:
         self.rust_log_override = rust_log_override
         self.port_distributor = port_distributor
         self.remote_storage = remote_storage
+        self.remote_storage_users = remote_storage_users
         self.broker = broker
         self.mock_s3_server = mock_s3_server
         self.pageserver_config_override = pageserver_config_override
@@ -497,9 +541,9 @@ class ZenithEnvBuilder:
             aws_access_key_id=self.mock_s3_server.access_key(),
             aws_secret_access_key=self.mock_s3_server.secret_key(),
         ).create_bucket(Bucket=bucket_name)
-        self.remote_storage = S3Storage(bucket=bucket_name,
+        self.remote_storage = S3Storage(bucket_name=bucket_name,
                                         endpoint=mock_endpoint,
-                                        region=mock_region)
+                                        bucket_region=mock_region)
 
     def __enter__(self):
         return self
@@ -557,6 +601,7 @@ class ZenithEnv:
         self.safekeepers: List[Safekeeper] = []
         self.broker = config.broker
         self.remote_storage = config.remote_storage
+        self.remote_storage_users = config.remote_storage_users
 
         # generate initial tenant ID here instead of letting 'zenith init' generate it,
         # so that we don't need to dig it out of the config file afterwards.
@@ -605,8 +650,12 @@ class ZenithEnv:
                 id = {id}
                 pg_port = {port.pg}
                 http_port = {port.http}
-                sync = false # Disable fsyncs to make the tests go faster
+                sync = false # Disable fsyncs to make the tests go faster""")
-            """)
+            if bool(self.remote_storage_users
+                    & RemoteStorageUsers.SAFEKEEPER) and self.remote_storage is not None:
+                toml += textwrap.dedent(f"""
+                remote_storage = "{remote_storage_to_toml_inline_table(self.remote_storage)}"
+                """)
             safekeeper = Safekeeper(env=self, id=id, port=port)
             self.safekeepers.append(safekeeper)
 
@@ -638,7 +687,7 @@ def _shared_simple_env(request: Any,
                        mock_s3_server: MockS3Server,
                        default_broker: Etcd) -> Iterator[ZenithEnv]:
     """
-    Internal fixture backing the `zenith_simple_env` fixture. If TEST_SHARED_FIXTURES
+   # Internal fixture backing the `zenith_simple_env` fixture. If TEST_SHARED_FIXTURES
     is set, this is shared by all tests using `zenith_simple_env`.
     """
 
@@ -822,20 +871,6 @@ class PageserverPort:
     http: int
 
 
-@dataclass
-class LocalFsStorage:
-    root: Path
-
-
-@dataclass
-class S3Storage:
-    bucket: str
-    region: str
-    endpoint: Optional[str]
-
-
-RemoteStorage = Union[LocalFsStorage, S3Storage]
-
 CREATE_TIMELINE_ID_EXTRACTOR = re.compile(r"^Created timeline '(?P<timeline_id>[^']+)'",
                                           re.MULTILINE)
 CREATE_TIMELINE_ID_EXTRACTOR = re.compile(r"^Created timeline '(?P<timeline_id>[^']+)'",
@@ -998,6 +1033,7 @@ class ZenithCli:
             append_pageserver_param_overrides(
                 params_to_update=cmd,
                 remote_storage=self.env.remote_storage,
+                remote_storage_users=self.env.remote_storage_users,
                 pageserver_config_override=self.env.pageserver.config_override)
 
             res = self.raw_cli(cmd)
@@ -1022,14 +1058,10 @@ class ZenithCli:
         append_pageserver_param_overrides(
             params_to_update=start_args,
             remote_storage=self.env.remote_storage,
+            remote_storage_users=self.env.remote_storage_users,
             pageserver_config_override=self.env.pageserver.config_override)
 
-        s3_env_vars = None
+        s3_env_vars = self.env.s3_mock_server.access_env_vars() if self.env.s3_mock_server else None
-        if self.env.s3_mock_server:
-            s3_env_vars = {
-                'AWS_ACCESS_KEY_ID': self.env.s3_mock_server.access_key(),
-                'AWS_SECRET_ACCESS_KEY': self.env.s3_mock_server.secret_key(),
-            }
         return self.raw_cli(start_args, extra_env_vars=s3_env_vars)
 
     def pageserver_stop(self, immediate=False) -> 'subprocess.CompletedProcess[str]':
@@ -1041,7 +1073,8 @@ class ZenithCli:
         return self.raw_cli(cmd)
 
     def safekeeper_start(self, id: int) -> 'subprocess.CompletedProcess[str]':
-        return self.raw_cli(['safekeeper', 'start', str(id)])
+        s3_env_vars = self.env.s3_mock_server.access_env_vars() if self.env.s3_mock_server else None
+        return self.raw_cli(['safekeeper', 'start', str(id)], extra_env_vars=s3_env_vars)
 
     def safekeeper_stop(self,
                         id: Optional[int] = None,
@@ -1237,22 +1270,13 @@ class ZenithPageserver(PgProtocol):
 def append_pageserver_param_overrides(
     params_to_update: List[str],
     remote_storage: Optional[RemoteStorage],
+    remote_storage_users: RemoteStorageUsers,
     pageserver_config_override: Optional[str] = None,
 ):
-    if remote_storage is not None:
+    if bool(remote_storage_users & RemoteStorageUsers.PAGESERVER) and remote_storage is not None:
-        if isinstance(remote_storage, LocalFsStorage):
+        remote_storage_toml_table = remote_storage_to_toml_inline_table(remote_storage)
-            pageserver_storage_override = f"local_path='{remote_storage.root}'"
-        elif isinstance(remote_storage, S3Storage):
-            pageserver_storage_override = f"bucket_name='{remote_storage.bucket}',\
-                bucket_region='{remote_storage.region}'"
-
-            if remote_storage.endpoint is not None:
-                pageserver_storage_override += f",endpoint='{remote_storage.endpoint}'"
-
-        else:
-            raise Exception(f'Unknown storage configuration {remote_storage}')
         params_to_update.append(
-            f'--pageserver-config-override=remote_storage={{{pageserver_storage_override}}}')
+            f'--pageserver-config-override=remote_storage={remote_storage_toml_table}')
 
     env_overrides = os.getenv('ZENITH_PAGESERVER_OVERRIDES')
     if env_overrides is not None:
@@ -1786,8 +1810,9 @@ class Safekeeper:
 class SafekeeperTimelineStatus:
     acceptor_epoch: int
     flush_lsn: str
-    remote_consistent_lsn: str
     timeline_start_lsn: str
+    backup_lsn: str
+    remote_consistent_lsn: str
 
 
 @dataclass
@@ -1812,8 +1837,9 @@ class SafekeeperHttpClient(requests.Session):
         resj = res.json()
         return SafekeeperTimelineStatus(acceptor_epoch=resj['acceptor_state']['epoch'],
                                         flush_lsn=resj['flush_lsn'],
-                                        remote_consistent_lsn=resj['remote_consistent_lsn'],
+                                        timeline_start_lsn=resj['timeline_start_lsn'],
-                                        timeline_start_lsn=resj['timeline_start_lsn'])
+                                        backup_lsn=resj['backup_lsn'],
+                                        remote_consistent_lsn=resj['remote_consistent_lsn'])
 
     def record_safekeeper_info(self, tenant_id: str, timeline_id: str, body):
         res = self.post(
@@ -1893,7 +1919,11 @@ class Etcd:
                 f"--data-dir={self.datadir}",
                 f"--listen-client-urls={client_url}",
                 f"--advertise-client-urls={client_url}",
-                f"--listen-peer-urls=http://127.0.0.1:{self.peer_port}"
+                f"--listen-peer-urls=http://127.0.0.1:{self.peer_port}",
+                # Set --quota-backend-bytes to keep the etcd virtual memory
+                # size smaller. Our test etcd clusters are very small.
+                # See https://github.com/etcd-io/etcd/issues/7910
+                f"--quota-backend-bytes=100000000"
             ]
             self.handle = subprocess.Popen(args, stdout=log_file, stderr=log_file)
 

test_runner/performance/test_compression.py

@@ -0,0 +1,24 @@
+# Test sequential scan speed
+#
+from contextlib import closing
+from dataclasses import dataclass
+from fixtures.zenith_fixtures import ZenithEnv
+from fixtures.log_helper import log
+from fixtures.benchmark_fixture import MetricReport, ZenithBenchmarker
+from fixtures.compare_fixtures import PgCompare
+import pytest
+
+
+@pytest.mark.parametrize('rows', [pytest.param(10000000)])
+def test_compression(zenith_with_baseline: PgCompare, rows: int):
+    env = zenith_with_baseline
+
+    with closing(env.pg.connect()) as conn:
+        with conn.cursor() as cur:
+            with env.record_duration('insert'):
+                cur.execute(
+                    f'create table t as select generate_series(1,{rows}) as pk,(random()*10)::bigint as r10,(random()*100)::bigint as r100,(random()*1000)::bigint as r1000,(random()*10000)::bigint as r10000'
+                )
+            cur.execute("vacuum t")
+            with env.record_duration('select'):
+                cur.execute('select sum(r100) from t')