pageserver: enable setting a target disk range

## Problem

There was an edge case where initial logical size calculation can be
downloading a layer that wasn't hit by the test's `SELECT`, and it's
on-disk but still marked as remote in the pageserver's internal state,
so evicting it fails.


https://neon-github-public-dev.s3.amazonaws.com/reports/pr-5648/6630099807/index.html#categories/dee044ec96f666edb90a77c01099a941/e38e97a2735ffa8c/

## Summary of changes

Use pageserver API to learn about layers, instead of inspecting local
disk, so that we will always agree with the pageserver about which layer
are local.

.github/workflows/build_and_test.yml

@@ -320,6 +320,9 @@ jobs:
       - name: Build neon extensions
         run: mold -run make neon-pg-ext -j$(nproc)
 
+      - name: Build walproposer-lib
+        run: mold -run make walproposer-lib -j$(nproc)
+
       - name: Run cargo build
         run: |
           ${cov_prefix} mold -run cargo build $CARGO_FLAGS $CARGO_FEATURES --bins --tests
@@ -335,6 +338,16 @@ jobs:
           # Avoid `$CARGO_FEATURES` since there's no `testing` feature in the e2e tests now
           ${cov_prefix} cargo test $CARGO_FLAGS --package remote_storage --test test_real_s3
 
+          # Run separate tests for real Azure Blob Storage
+          # XXX: replace region with `eu-central-1`-like region
+          export AZURE_STORAGE_ACCOUNT="${{ secrets.AZURE_STORAGE_ACCOUNT_DEV }}"
+          export AZURE_STORAGE_ACCESS_KEY="${{ secrets.AZURE_STORAGE_ACCESS_KEY_DEV }}"
+          export ENABLE_REAL_AZURE_REMOTE_STORAGE=y
+          export REMOTE_STORAGE_AZURE_CONTAINER=neon-github-sandbox
+          export REMOTE_STORAGE_AZURE_REGION=eastus2
+          # Avoid `$CARGO_FEATURES` since there's no `testing` feature in the e2e tests now
+          ${cov_prefix} cargo test $CARGO_FLAGS --package remote_storage --test test_real_azure
+
       - name: Install rust binaries
         run: |
           # Install target binaries
@@ -834,7 +847,7 @@ jobs:
       run:
         shell: sh -eu {0}
     env:
-      VM_BUILDER_VERSION: v0.17.12
+      VM_BUILDER_VERSION: v0.18.2
 
     steps:
       - name: Checkout

.github/workflows/neon_extra_builds.yml

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

Cargo.toml

@@ -26,6 +26,7 @@ members = [
     "libs/tracing-utils",
     "libs/postgres_ffi/wal_craft",
     "libs/vm_monitor",
+    "libs/walproposer",
 ]
 
 [workspace.package]
@@ -36,6 +37,10 @@ license = "Apache-2.0"
 [workspace.dependencies]
 anyhow = { version = "1.0", features = ["backtrace"] }
 async-compression = { version = "0.4.0", features = ["tokio", "gzip"] }
+azure_core = "0.16"
+azure_identity = "0.16"
+azure_storage = "0.16"
+azure_storage_blobs = "0.16"
 flate2 = "1.0.26"
 async-stream = "0.3"
 async-trait = "0.1"
@@ -76,6 +81,7 @@ hex = "0.4"
 hex-literal = "0.4"
 hmac = "0.12.1"
 hostname = "0.3.1"
+http-types = "2"
 humantime = "2.1"
 humantime-serde = "1.1.1"
 hyper = "0.14"
@@ -155,11 +161,11 @@ env_logger = "0.10"
 log = "0.4"
 
 ## Libraries from neondatabase/ git forks, ideally with changes to be upstreamed
-postgres = { git = "https://github.com/neondatabase/rust-postgres.git", rev="9011f7110db12b5e15afaf98f8ac834501d50ddc" }
-postgres-native-tls = { git = "https://github.com/neondatabase/rust-postgres.git", rev="9011f7110db12b5e15afaf98f8ac834501d50ddc" }
-postgres-protocol = { git = "https://github.com/neondatabase/rust-postgres.git", rev="9011f7110db12b5e15afaf98f8ac834501d50ddc" }
-postgres-types = { git = "https://github.com/neondatabase/rust-postgres.git", rev="9011f7110db12b5e15afaf98f8ac834501d50ddc" }
-tokio-postgres = { git = "https://github.com/neondatabase/rust-postgres.git", rev="9011f7110db12b5e15afaf98f8ac834501d50ddc" }
+postgres = { git = "https://github.com/neondatabase/rust-postgres.git", rev="7434d9388965a17a6d113e5dfc0e65666a03b4c2" }
+postgres-native-tls = { git = "https://github.com/neondatabase/rust-postgres.git", rev="7434d9388965a17a6d113e5dfc0e65666a03b4c2" }
+postgres-protocol = { git = "https://github.com/neondatabase/rust-postgres.git", rev="7434d9388965a17a6d113e5dfc0e65666a03b4c2" }
+postgres-types = { git = "https://github.com/neondatabase/rust-postgres.git", rev="7434d9388965a17a6d113e5dfc0e65666a03b4c2" }
+tokio-postgres = { git = "https://github.com/neondatabase/rust-postgres.git", rev="7434d9388965a17a6d113e5dfc0e65666a03b4c2" }
 
 ## Other git libraries
 heapless = { default-features=false, features=[], git = "https://github.com/japaric/heapless.git", rev = "644653bf3b831c6bb4963be2de24804acf5e5001" } # upstream release pending
@@ -180,6 +186,7 @@ tenant_size_model = { version = "0.1", path = "./libs/tenant_size_model/" }
 tracing-utils = { version = "0.1", path = "./libs/tracing-utils/" }
 utils = { version = "0.1", path = "./libs/utils/" }
 vm_monitor = { version = "0.1", path = "./libs/vm_monitor/" }
+walproposer = { version = "0.1", path = "./libs/walproposer/" }
 
 ## Common library dependency
 workspace_hack = { version = "0.1", path = "./workspace_hack/" }
@@ -195,7 +202,7 @@ tonic-build = "0.9"
 
 # This is only needed for proxy's tests.
 # TODO: we should probably fork `tokio-postgres-rustls` instead.
-tokio-postgres = { git = "https://github.com/neondatabase/rust-postgres.git", rev="9011f7110db12b5e15afaf98f8ac834501d50ddc" }
+tokio-postgres = { git = "https://github.com/neondatabase/rust-postgres.git", rev="7434d9388965a17a6d113e5dfc0e65666a03b4c2" }
 
 ################# Binary contents sections
 

Makefile

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

NOTICE

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

compute_tools/src/compute.rs

@@ -252,7 +252,7 @@ fn create_neon_superuser(spec: &ComputeSpec, client: &mut Client) -> Result<()>
                     IF NOT EXISTS (
                         SELECT FROM pg_catalog.pg_roles WHERE rolname = 'neon_superuser')
                     THEN
-                        CREATE ROLE neon_superuser CREATEDB CREATEROLE NOLOGIN IN ROLE pg_read_all_data, pg_write_all_data;
+                        CREATE ROLE neon_superuser CREATEDB CREATEROLE NOLOGIN REPLICATION IN ROLE pg_read_all_data, pg_write_all_data;
                         IF array_length(roles, 1) IS NOT NULL THEN
                             EXECUTE format('GRANT neon_superuser TO %s',
                                            array_to_string(ARRAY(SELECT quote_ident(x) FROM unnest(roles) as x), ', '));
@@ -692,10 +692,11 @@ impl ComputeNode {
         // Proceed with post-startup configuration. Note, that order of operations is important.
         let spec = &compute_state.pspec.as_ref().expect("spec must be set").spec;
         create_neon_superuser(spec, &mut client)?;
+        cleanup_instance(&mut client)?;
         handle_roles(spec, &mut client)?;
         handle_databases(spec, &mut client)?;
         handle_role_deletions(spec, self.connstr.as_str(), &mut client)?;
-        handle_grants(spec, self.connstr.as_str())?;
+        handle_grants(spec, &mut client, self.connstr.as_str())?;
         handle_extensions(spec, &mut client)?;
         create_availability_check_data(&mut client)?;
 
@@ -731,10 +732,11 @@ impl ComputeNode {
         // Disable DDL forwarding because control plane already knows about these roles/databases.
         if spec.mode == ComputeMode::Primary {
             client.simple_query("SET neon.forward_ddl = false")?;
+            cleanup_instance(&mut client)?;
             handle_roles(&spec, &mut client)?;
             handle_databases(&spec, &mut client)?;
             handle_role_deletions(&spec, self.connstr.as_str(), &mut client)?;
-            handle_grants(&spec, self.connstr.as_str())?;
+            handle_grants(&spec, &mut client, self.connstr.as_str())?;
             handle_extensions(&spec, &mut client)?;
         }
 

compute_tools/src/pg_helpers.rs

@@ -1,3 +1,4 @@
+use std::collections::HashMap;
 use std::fmt::Write;
 use std::fs;
 use std::fs::File;
@@ -205,22 +206,37 @@ pub fn get_existing_roles(xact: &mut Transaction<'_>) -> Result<Vec<Role>> {
 }
 
 /// Build a list of existing Postgres databases
-pub fn get_existing_dbs(client: &mut Client) -> Result<Vec<Database>> {
-    let postgres_dbs = client
+pub fn get_existing_dbs(client: &mut Client) -> Result<HashMap<String, Database>> {
+    // `pg_database.datconnlimit = -2` means that the database is in the
+    // invalid state. See:
+    //   https://github.com/postgres/postgres/commit/a4b4cc1d60f7e8ccfcc8ff8cb80c28ee411ad9a9
+    let postgres_dbs: Vec<Database> = client
         .query(
-            "SELECT datname, datdba::regrole::text as owner
-               FROM pg_catalog.pg_database;",
+            "SELECT
+                datname AS name,
+                datdba::regrole::text AS owner,
+                NOT datallowconn AS restrict_conn,
+                datconnlimit = - 2 AS invalid
+            FROM
+                pg_catalog.pg_database;",
             &[],
         )?
         .iter()
         .map(|row| Database {
-            name: row.get("datname"),
+            name: row.get("name"),
             owner: row.get("owner"),
+            restrict_conn: row.get("restrict_conn"),
+            invalid: row.get("invalid"),
             options: None,
         })
         .collect();
 
-    Ok(postgres_dbs)
+    let dbs_map = postgres_dbs
+        .iter()
+        .map(|db| (db.name.clone(), db.clone()))
+        .collect::<HashMap<_, _>>();
+
+    Ok(dbs_map)
 }
 
 /// Wait for Postgres to become ready to accept connections. It's ready to

compute_tools/src/spec.rs

@@ -13,7 +13,7 @@ use crate::params::PG_HBA_ALL_MD5;
 use crate::pg_helpers::*;
 
 use compute_api::responses::{ControlPlaneComputeStatus, ControlPlaneSpecResponse};
-use compute_api::spec::{ComputeSpec, Database, PgIdent, Role};
+use compute_api::spec::{ComputeSpec, PgIdent, Role};
 
 // Do control plane request and return response if any. In case of error it
 // returns a bool flag indicating whether it makes sense to retry the request
@@ -161,6 +161,38 @@ pub fn add_standby_signal(pgdata_path: &Path) -> Result<()> {
     Ok(())
 }
 
+/// Compute could be unexpectedly shut down, for example, during the
+/// database dropping. This leaves the database in the invalid state,
+/// which prevents new db creation with the same name. This function
+/// will clean it up before proceeding with catalog updates. All
+/// possible future cleanup operations may go here too.
+#[instrument(skip_all)]
+pub fn cleanup_instance(client: &mut Client) -> Result<()> {
+    let existing_dbs = get_existing_dbs(client)?;
+
+    for (_, db) in existing_dbs {
+        if db.invalid {
+            // After recent commit in Postgres, interrupted DROP DATABASE
+            // leaves the database in the invalid state. According to the
+            // commit message, the only option for user is to drop it again.
+            // See:
+            //   https://github.com/postgres/postgres/commit/a4b4cc1d60f7e8ccfcc8ff8cb80c28ee411ad9a9
+            //
+            // Postgres Neon extension is done the way, that db is de-registered
+            // in the control plane metadata only after it is dropped. So there is
+            // a chance that it still thinks that db should exist. This means
+            // that it will be re-created by `handle_databases()`. Yet, it's fine
+            // as user can just repeat drop (in vanilla Postgres they would need
+            // to do the same, btw).
+            let query = format!("DROP DATABASE IF EXISTS {}", db.name.pg_quote());
+            info!("dropping invalid database {}", db.name);
+            client.execute(query.as_str(), &[])?;
+        }
+    }
+
+    Ok(())
+}
+
 /// Given a cluster spec json and open transaction it handles roles creation,
 /// deletion and update.
 #[instrument(skip_all)]
@@ -270,7 +302,7 @@ pub fn handle_roles(spec: &ComputeSpec, client: &mut Client) -> Result<()> {
             }
             RoleAction::Create => {
                 let mut query: String = format!(
-                    "CREATE ROLE {} CREATEROLE CREATEDB BYPASSRLS IN ROLE neon_superuser",
+                    "CREATE ROLE {} CREATEROLE CREATEDB BYPASSRLS REPLICATION IN ROLE neon_superuser",
                     name.pg_quote()
                 );
                 info!("role create query: '{}'", &query);
@@ -379,13 +411,13 @@ fn reassign_owned_objects(spec: &ComputeSpec, connstr: &str, role_name: &PgIdent
 /// which together provide us idempotency.
 #[instrument(skip_all)]
 pub fn handle_databases(spec: &ComputeSpec, client: &mut Client) -> Result<()> {
-    let existing_dbs: Vec<Database> = get_existing_dbs(client)?;
+    let existing_dbs = get_existing_dbs(client)?;
 
     // Print a list of existing Postgres databases (only in debug mode)
     if span_enabled!(Level::INFO) {
         info!("postgres databases:");
-        for r in &existing_dbs {
-            info!("    {}:{}", r.name, r.owner);
+        for (dbname, db) in &existing_dbs {
+            info!("    {}:{}", dbname, db.owner);
         }
     }
 
@@ -439,8 +471,7 @@ pub fn handle_databases(spec: &ComputeSpec, client: &mut Client) -> Result<()> {
                 "rename_db" => {
                     let new_name = op.new_name.as_ref().unwrap();
 
-                    // XXX: with a limited number of roles it is fine, but consider making it a HashMap
-                    if existing_dbs.iter().any(|r| r.name == op.name) {
+                    if existing_dbs.get(&op.name).is_some() {
                         let query: String = format!(
                             "ALTER DATABASE {} RENAME TO {}",
                             op.name.pg_quote(),
@@ -457,14 +488,12 @@ pub fn handle_databases(spec: &ComputeSpec, client: &mut Client) -> Result<()> {
     }
 
     // Refresh Postgres databases info to handle possible renames
-    let existing_dbs: Vec<Database> = get_existing_dbs(client)?;
+    let existing_dbs = get_existing_dbs(client)?;
 
     info!("cluster spec databases:");
     for db in &spec.cluster.databases {
         let name = &db.name;
-
-        // XXX: with a limited number of databases it is fine, but consider making it a HashMap
-        let pg_db = existing_dbs.iter().find(|r| r.name == *name);
+        let pg_db = existing_dbs.get(name);
 
         enum DatabaseAction {
             None,
@@ -530,13 +559,32 @@ pub fn handle_databases(spec: &ComputeSpec, client: &mut Client) -> Result<()> {
 /// Grant CREATE ON DATABASE to the database owner and do some other alters and grants
 /// to allow users creating trusted extensions and re-creating `public` schema, for example.
 #[instrument(skip_all)]
-pub fn handle_grants(spec: &ComputeSpec, connstr: &str) -> Result<()> {
-    info!("cluster spec grants:");
+pub fn handle_grants(spec: &ComputeSpec, client: &mut Client, connstr: &str) -> Result<()> {
+    info!("modifying database permissions");
+    let existing_dbs = get_existing_dbs(client)?;
 
     // Do some per-database access adjustments. We'd better do this at db creation time,
     // but CREATE DATABASE isn't transactional. So we cannot create db + do some grants
     // atomically.
     for db in &spec.cluster.databases {
+        match existing_dbs.get(&db.name) {
+            Some(pg_db) => {
+                if pg_db.restrict_conn || pg_db.invalid {
+                    info!(
+                        "skipping grants for db {} (invalid: {}, connections not allowed: {})",
+                        db.name, pg_db.invalid, pg_db.restrict_conn
+                    );
+                    continue;
+                }
+            }
+            None => {
+                bail!(
+                    "database {} doesn't exist in Postgres after handle_databases()",
+                    db.name
+                );
+            }
+        }
+
         let mut conf = Config::from_str(connstr)?;
         conf.dbname(&db.name);
 
@@ -575,6 +623,11 @@ pub fn handle_grants(spec: &ComputeSpec, connstr: &str) -> Result<()> {
 
         // Explicitly grant CREATE ON SCHEMA PUBLIC to the web_access user.
         // This is needed because since postgres 15 this privilege is removed by default.
+        // TODO: web_access isn't created for almost 1 year. It could be that we have
+        // active users of 1 year old projects, but hopefully not, so check it and
+        // remove this code if possible. The worst thing that could happen is that
+        // user won't be able to use public schema in NEW databases created in the
+        // very OLD project.
         let grant_query = "DO $$\n\
                 BEGIN\n\
                     IF EXISTS(\n\

compute_tools/tests/pg_helpers_tests.rs

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

control_plane/src/background_process.rs

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

control_plane/src/bin/attachment_service.rs

@@ -13,6 +13,7 @@ use serde::{Deserialize, Serialize};
 use std::path::{Path, PathBuf};
 use std::{collections::HashMap, sync::Arc};
 use utils::logging::{self, LogFormat};
+use utils::signals::{ShutdownSignals, Signal};
 
 use utils::{
     http::{
@@ -268,7 +269,16 @@ async fn main() -> anyhow::Result<()> {
     let server = hyper::Server::from_tcp(http_listener)?.serve(service);
 
     tracing::info!("Serving on {0}", args.listen);
-    server.await?;
+
+    tokio::task::spawn(server);
+
+    ShutdownSignals::handle(|signal| match signal {
+        Signal::Interrupt | Signal::Terminate | Signal::Quit => {
+            tracing::info!("Got {}. Terminating", signal.name());
+            // We're just a test helper: no graceful shutdown.
+            std::process::exit(0);
+        }
+    })?;
 
     Ok(())
 }

control_plane/src/endpoint.rs

@@ -253,7 +253,7 @@ impl Endpoint {
         conf.append("shared_buffers", "1MB");
         conf.append("fsync", "off");
         conf.append("max_connections", "100");
-        conf.append("wal_level", "replica");
+        conf.append("wal_level", "logical");
         // wal_sender_timeout is the maximum time to wait for WAL replication.
         // It also defines how often the walreciever will send a feedback message to the wal sender.
         conf.append("wal_sender_timeout", "5s");

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

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

docs/error-handling.md

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

docs/pageserver-services.md

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

libs/compute_api/src/spec.rs

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

libs/compute_api/tests/cluster_spec.json

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

libs/metrics/src/wrappers.rs

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

libs/pageserver_api/src/reltag.rs

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

libs/postgres_backend/src/lib.rs

@@ -19,8 +19,8 @@ use tracing::{debug, error, info, trace};
 
 use pq_proto::framed::{ConnectionError, Framed, FramedReader, FramedWriter};
 use pq_proto::{
-    BeMessage, FeMessage, FeStartupPacket, ProtocolError, SQLSTATE_INTERNAL_ERROR,
-    SQLSTATE_SUCCESSFUL_COMPLETION,
+    BeMessage, FeMessage, FeStartupPacket, ProtocolError, SQLSTATE_ADMIN_SHUTDOWN,
+    SQLSTATE_INTERNAL_ERROR, SQLSTATE_SUCCESSFUL_COMPLETION,
 };
 
 /// An error, occurred during query processing:
@@ -30,6 +30,9 @@ pub enum QueryError {
     /// The connection was lost while processing the query.
     #[error(transparent)]
     Disconnected(#[from] ConnectionError),
+    /// We were instructed to shutdown while processing the query
+    #[error("Shutting down")]
+    Shutdown,
     /// Some other error
     #[error(transparent)]
     Other(#[from] anyhow::Error),
@@ -44,7 +47,8 @@ impl From<io::Error> for QueryError {
 impl QueryError {
     pub fn pg_error_code(&self) -> &'static [u8; 5] {
         match self {
-            Self::Disconnected(_) => b"08006",         // connection failure
+            Self::Disconnected(_) => b"08006", // connection failure
+            Self::Shutdown => SQLSTATE_ADMIN_SHUTDOWN,
             Self::Other(_) => SQLSTATE_INTERNAL_ERROR, // internal error
         }
     }
@@ -396,7 +400,20 @@ impl<IO: AsyncRead + AsyncWrite + Unpin> PostgresBackend<IO> {
         // socket might be already closed, e.g. if previously received error,
         // so ignore result.
         self.framed.shutdown().await.ok();
-        ret
+        match ret {
+            Ok(()) => Ok(()),
+            Err(QueryError::Shutdown) => {
+                info!("Stopped due to shutdown");
+                Ok(())
+            }
+            Err(QueryError::Disconnected(e)) => {
+                info!("Disconnected ({e:#})");
+                // Disconnection is not an error: we just use it that way internally to drop
+                // out of loops.
+                Ok(())
+            }
+            e => e,
+        }
     }
 
     async fn run_message_loop<F, S>(
@@ -416,15 +433,11 @@ impl<IO: AsyncRead + AsyncWrite + Unpin> PostgresBackend<IO> {
             _ = shutdown_watcher() => {
                 // We were requested to shut down.
                 tracing::info!("shutdown request received during handshake");
-                return Ok(())
+                return Err(QueryError::Shutdown)
             },
 
-            result = self.handshake(handler) => {
-                // Handshake complete.
-                result?;
-                if self.state == ProtoState::Closed {
-                    return Ok(()); // EOF during handshake
-                }
+            handshake_r = self.handshake(handler) => {
+                handshake_r?;
             }
         );
 
@@ -435,7 +448,7 @@ impl<IO: AsyncRead + AsyncWrite + Unpin> PostgresBackend<IO> {
             _ = shutdown_watcher() => {
                 // We were requested to shut down.
                 tracing::info!("shutdown request received in run_message_loop");
-                Ok(None)
+                return Err(QueryError::Shutdown)
             },
             msg = self.read_message() => { msg },
         )? {
@@ -447,7 +460,14 @@ impl<IO: AsyncRead + AsyncWrite + Unpin> PostgresBackend<IO> {
                 _ = shutdown_watcher() => {
                     // We were requested to shut down.
                     tracing::info!("shutdown request received during response flush");
-                    return Ok(())
+
+                    // If we exited process_message with a shutdown error, there may be
+                    // some valid response content on in our transmit buffer: permit sending
+                    // this within a short timeout.  This is a best effort thing so we don't
+                    // care about the result.
+                    tokio::time::timeout(std::time::Duration::from_millis(500), self.flush()).await.ok();
+
+                    return Err(QueryError::Shutdown)
                 },
                 flush_r = self.flush() => {
                     flush_r?;
@@ -560,7 +580,9 @@ impl<IO: AsyncRead + AsyncWrite + Unpin> PostgresBackend<IO> {
                         self.peer_addr
                     );
                     self.state = ProtoState::Closed;
-                    return Ok(());
+                    return Err(QueryError::Disconnected(ConnectionError::Protocol(
+                        ProtocolError::Protocol("EOF during handshake".to_string()),
+                    )));
                 }
             }
         }
@@ -599,7 +621,9 @@ impl<IO: AsyncRead + AsyncWrite + Unpin> PostgresBackend<IO> {
                         self.peer_addr
                     );
                     self.state = ProtoState::Closed;
-                    return Ok(());
+                    return Err(QueryError::Disconnected(ConnectionError::Protocol(
+                        ProtocolError::Protocol("EOF during auth".to_string()),
+                    )));
                 }
             }
         }
@@ -923,6 +947,7 @@ impl<'a, IO: AsyncRead + AsyncWrite + Unpin> AsyncWrite for CopyDataWriter<'a, I
 pub fn short_error(e: &QueryError) -> String {
     match e {
         QueryError::Disconnected(connection_error) => connection_error.to_string(),
+        QueryError::Shutdown => "shutdown".to_string(),
         QueryError::Other(e) => format!("{e:#}"),
     }
 }
@@ -939,6 +964,9 @@ fn log_query_error(query: &str, e: &QueryError) {
         QueryError::Disconnected(other_connection_error) => {
             error!("query handler for '{query}' failed with connection error: {other_connection_error:?}")
         }
+        QueryError::Shutdown => {
+            info!("query handler for '{query}' cancelled during tenant shutdown")
+        }
         QueryError::Other(e) => {
             error!("query handler for '{query}' failed: {e:?}");
         }

libs/postgres_ffi/src/lib.rs

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

libs/postgres_ffi/src/pg_constants.rs

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

libs/postgres_ffi/src/xlog_utils.rs

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

libs/pq_proto/src/lib.rs

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

libs/remote_storage/Cargo.toml

@@ -13,6 +13,7 @@ aws-types.workspace = true
 aws-config.workspace = true
 aws-sdk-s3.workspace = true
 aws-credential-types.workspace = true
+bytes.workspace = true
 camino.workspace = true
 hyper = { workspace = true, features = ["stream"] }
 serde.workspace = true
@@ -26,6 +27,13 @@ metrics.workspace = true
 utils.workspace = true
 pin-project-lite.workspace = true
 workspace_hack.workspace = true
+azure_core.workspace = true
+azure_identity.workspace = true
+azure_storage.workspace = true
+azure_storage_blobs.workspace = true
+futures-util.workspace = true
+http-types.workspace = true
+itertools.workspace = true
 
 [dev-dependencies]
 camino-tempfile.workspace = true

libs/remote_storage/src/azure_blob.rs

@@ -0,0 +1,356 @@
+//! Azure Blob Storage wrapper
+
+use std::env;
+use std::num::NonZeroU32;
+use std::sync::Arc;
+use std::{borrow::Cow, collections::HashMap, io::Cursor};
+
+use super::REMOTE_STORAGE_PREFIX_SEPARATOR;
+use anyhow::Result;
+use azure_core::request_options::{MaxResults, Metadata, Range};
+use azure_core::Header;
+use azure_identity::DefaultAzureCredential;
+use azure_storage::StorageCredentials;
+use azure_storage_blobs::prelude::ClientBuilder;
+use azure_storage_blobs::{
+    blob::operations::GetBlobBuilder,
+    prelude::{BlobClient, ContainerClient},
+};
+use futures_util::StreamExt;
+use http_types::StatusCode;
+use tokio::io::AsyncRead;
+use tracing::debug;
+
+use crate::s3_bucket::RequestKind;
+use crate::{
+    AzureConfig, ConcurrencyLimiter, Download, DownloadError, RemotePath, RemoteStorage,
+    StorageMetadata,
+};
+
+pub struct AzureBlobStorage {
+    client: ContainerClient,
+    prefix_in_container: Option<String>,
+    max_keys_per_list_response: Option<NonZeroU32>,
+    concurrency_limiter: ConcurrencyLimiter,
+}
+
+impl AzureBlobStorage {
+    pub fn new(azure_config: &AzureConfig) -> Result<Self> {
+        debug!(
+            "Creating azure remote storage for azure container {}",
+            azure_config.container_name
+        );
+
+        let account = env::var("AZURE_STORAGE_ACCOUNT").expect("missing AZURE_STORAGE_ACCOUNT");
+
+        // If the `AZURE_STORAGE_ACCESS_KEY` env var has an access key, use that,
+        // otherwise try the token based credentials.
+        let credentials = if let Ok(access_key) = env::var("AZURE_STORAGE_ACCESS_KEY") {
+            StorageCredentials::access_key(account.clone(), access_key)
+        } else {
+            let token_credential = DefaultAzureCredential::default();
+            StorageCredentials::token_credential(Arc::new(token_credential))
+        };
+
+        let builder = ClientBuilder::new(account, credentials);
+
+        let client = builder.container_client(azure_config.container_name.to_owned());
+
+        let max_keys_per_list_response =
+            if let Some(limit) = azure_config.max_keys_per_list_response {
+                Some(
+                    NonZeroU32::new(limit as u32)
+                        .ok_or_else(|| anyhow::anyhow!("max_keys_per_list_response can't be 0"))?,
+                )
+            } else {
+                None
+            };
+
+        Ok(AzureBlobStorage {
+            client,
+            prefix_in_container: azure_config.prefix_in_container.to_owned(),
+            max_keys_per_list_response,
+            concurrency_limiter: ConcurrencyLimiter::new(azure_config.concurrency_limit.get()),
+        })
+    }
+
+    pub fn relative_path_to_name(&self, path: &RemotePath) -> String {
+        assert_eq!(std::path::MAIN_SEPARATOR, REMOTE_STORAGE_PREFIX_SEPARATOR);
+        let path_string = path
+            .get_path()
+            .as_str()
+            .trim_end_matches(REMOTE_STORAGE_PREFIX_SEPARATOR);
+        match &self.prefix_in_container {
+            Some(prefix) => {
+                if prefix.ends_with(REMOTE_STORAGE_PREFIX_SEPARATOR) {
+                    prefix.clone() + path_string
+                } else {
+                    format!("{prefix}{REMOTE_STORAGE_PREFIX_SEPARATOR}{path_string}")
+                }
+            }
+            None => path_string.to_string(),
+        }
+    }
+
+    fn name_to_relative_path(&self, key: &str) -> RemotePath {
+        let relative_path =
+            match key.strip_prefix(self.prefix_in_container.as_deref().unwrap_or_default()) {
+                Some(stripped) => stripped,
+                // we rely on Azure to return properly prefixed paths
+                // for requests with a certain prefix
+                None => panic!(
+                    "Key {key} does not start with container prefix {:?}",
+                    self.prefix_in_container
+                ),
+            };
+        RemotePath(
+            relative_path
+                .split(REMOTE_STORAGE_PREFIX_SEPARATOR)
+                .collect(),
+        )
+    }
+
+    async fn download_for_builder(
+        &self,
+        metadata: StorageMetadata,
+        builder: GetBlobBuilder,
+    ) -> Result<Download, DownloadError> {
+        let mut response = builder.into_stream();
+
+        // TODO give proper streaming response instead of buffering into RAM
+        // https://github.com/neondatabase/neon/issues/5563
+        let mut buf = Vec::new();
+        while let Some(part) = response.next().await {
+            let part = part.map_err(to_download_error)?;
+            let data = part
+                .data
+                .collect()
+                .await
+                .map_err(|e| DownloadError::Other(e.into()))?;
+            buf.extend_from_slice(&data.slice(..));
+        }
+        Ok(Download {
+            download_stream: Box::pin(Cursor::new(buf)),
+            metadata: Some(metadata),
+        })
+    }
+    // TODO get rid of this function once we have metadata included in the response
+    // https://github.com/Azure/azure-sdk-for-rust/issues/1439
+    async fn get_metadata(
+        &self,
+        blob_client: &BlobClient,
+    ) -> Result<StorageMetadata, DownloadError> {
+        let builder = blob_client.get_metadata();
+
+        let response = builder.into_future().await.map_err(to_download_error)?;
+        let mut map = HashMap::new();
+
+        for md in response.metadata.iter() {
+            map.insert(
+                md.name().as_str().to_string(),
+                md.value().as_str().to_string(),
+            );
+        }
+        Ok(StorageMetadata(map))
+    }
+
+    async fn permit(&self, kind: RequestKind) -> tokio::sync::SemaphorePermit<'_> {
+        self.concurrency_limiter
+            .acquire(kind)
+            .await
+            .expect("semaphore is never closed")
+    }
+}
+
+fn to_azure_metadata(metadata: StorageMetadata) -> Metadata {
+    let mut res = Metadata::new();
+    for (k, v) in metadata.0.into_iter() {
+        res.insert(k, v);
+    }
+    res
+}
+
+fn to_download_error(error: azure_core::Error) -> DownloadError {
+    if let Some(http_err) = error.as_http_error() {
+        match http_err.status() {
+            StatusCode::NotFound => DownloadError::NotFound,
+            StatusCode::BadRequest => DownloadError::BadInput(anyhow::Error::new(error)),
+            _ => DownloadError::Other(anyhow::Error::new(error)),
+        }
+    } else {
+        DownloadError::Other(error.into())
+    }
+}
+
+#[async_trait::async_trait]
+impl RemoteStorage for AzureBlobStorage {
+    async fn list_prefixes(
+        &self,
+        prefix: Option<&RemotePath>,
+    ) -> Result<Vec<RemotePath>, DownloadError> {
+        // get the passed prefix or if it is not set use prefix_in_bucket value
+        let list_prefix = prefix
+            .map(|p| self.relative_path_to_name(p))
+            .or_else(|| self.prefix_in_container.clone())
+            .map(|mut p| {
+                // required to end with a separator
+                // otherwise request will return only the entry of a prefix
+                if !p.ends_with(REMOTE_STORAGE_PREFIX_SEPARATOR) {
+                    p.push(REMOTE_STORAGE_PREFIX_SEPARATOR);
+                }
+                p
+            });
+
+        let mut builder = self
+            .client
+            .list_blobs()
+            .delimiter(REMOTE_STORAGE_PREFIX_SEPARATOR.to_string());
+
+        if let Some(prefix) = list_prefix {
+            builder = builder.prefix(Cow::from(prefix.to_owned()));
+        }
+
+        if let Some(limit) = self.max_keys_per_list_response {
+            builder = builder.max_results(MaxResults::new(limit));
+        }
+
+        let mut response = builder.into_stream();
+        let mut res = Vec::new();
+        while let Some(entry) = response.next().await {
+            let entry = entry.map_err(to_download_error)?;
+            let name_iter = entry
+                .blobs
+                .prefixes()
+                .map(|prefix| self.name_to_relative_path(&prefix.name));
+            res.extend(name_iter);
+        }
+        Ok(res)
+    }
+
+    async fn list_files(&self, folder: Option<&RemotePath>) -> anyhow::Result<Vec<RemotePath>> {
+        let folder_name = folder
+            .map(|p| self.relative_path_to_name(p))
+            .or_else(|| self.prefix_in_container.clone());
+
+        let mut builder = self.client.list_blobs();
+
+        if let Some(folder_name) = folder_name {
+            builder = builder.prefix(Cow::from(folder_name.to_owned()));
+        }
+
+        if let Some(limit) = self.max_keys_per_list_response {
+            builder = builder.max_results(MaxResults::new(limit));
+        }
+
+        let mut response = builder.into_stream();
+        let mut res = Vec::new();
+        while let Some(l) = response.next().await {
+            let entry = l.map_err(anyhow::Error::new)?;
+            let name_iter = entry
+                .blobs
+                .blobs()
+                .map(|bl| self.name_to_relative_path(&bl.name));
+            res.extend(name_iter);
+        }
+        Ok(res)
+    }
+
+    async fn upload(
+        &self,
+        mut from: impl AsyncRead + Unpin + Send + Sync + 'static,
+        data_size_bytes: usize,
+        to: &RemotePath,
+        metadata: Option<StorageMetadata>,
+    ) -> anyhow::Result<()> {
+        let _permit = self.permit(RequestKind::Put).await;
+        let blob_client = self.client.blob_client(self.relative_path_to_name(to));
+
+        // TODO FIX THIS UGLY HACK and don't buffer the entire object
+        // into RAM here, but use the streaming interface. For that,
+        // we'd have to change the interface though...
+        // https://github.com/neondatabase/neon/issues/5563
+        let mut buf = Vec::with_capacity(data_size_bytes);
+        tokio::io::copy(&mut from, &mut buf).await?;
+        let body = azure_core::Body::Bytes(buf.into());
+
+        let mut builder = blob_client.put_block_blob(body);
+
+        if let Some(metadata) = metadata {
+            builder = builder.metadata(to_azure_metadata(metadata));
+        }
+
+        let _response = builder.into_future().await?;
+
+        Ok(())
+    }
+
+    async fn download(&self, from: &RemotePath) -> Result<Download, DownloadError> {
+        let _permit = self.permit(RequestKind::Get).await;
+        let blob_client = self.client.blob_client(self.relative_path_to_name(from));
+
+        let metadata = self.get_metadata(&blob_client).await?;
+
+        let builder = blob_client.get();
+
+        self.download_for_builder(metadata, builder).await
+    }
+
+    async fn download_byte_range(
+        &self,
+        from: &RemotePath,
+        start_inclusive: u64,
+        end_exclusive: Option<u64>,
+    ) -> Result<Download, DownloadError> {
+        let _permit = self.permit(RequestKind::Get).await;
+        let blob_client = self.client.blob_client(self.relative_path_to_name(from));
+
+        let metadata = self.get_metadata(&blob_client).await?;
+
+        let mut builder = blob_client.get();
+
+        if let Some(end_exclusive) = end_exclusive {
+            builder = builder.range(Range::new(start_inclusive, end_exclusive));
+        } else {
+            // Open ranges are not supported by the SDK so we work around
+            // by setting the upper limit extremely high (but high enough
+            // to still be representable by signed 64 bit integers).
+            // TODO remove workaround once the SDK adds open range support
+            // https://github.com/Azure/azure-sdk-for-rust/issues/1438
+            let end_exclusive = u64::MAX / 4;
+            builder = builder.range(Range::new(start_inclusive, end_exclusive));
+        }
+
+        self.download_for_builder(metadata, builder).await
+    }
+
+    async fn delete(&self, path: &RemotePath) -> anyhow::Result<()> {
+        let _permit = self.permit(RequestKind::Delete).await;
+        let blob_client = self.client.blob_client(self.relative_path_to_name(path));
+
+        let builder = blob_client.delete();
+
+        match builder.into_future().await {
+            Ok(_response) => Ok(()),
+            Err(e) => {
+                if let Some(http_err) = e.as_http_error() {
+                    if http_err.status() == StatusCode::NotFound {
+                        return Ok(());
+                    }
+                }
+                Err(anyhow::Error::new(e))
+            }
+        }
+    }
+
+    async fn delete_objects<'a>(&self, paths: &'a [RemotePath]) -> anyhow::Result<()> {
+        // Permit is already obtained by inner delete function
+
+        // TODO batch requests are also not supported by the SDK
+        // https://github.com/Azure/azure-sdk-for-rust/issues/1068
+        // https://github.com/Azure/azure-sdk-for-rust/issues/1249
+        for path in paths {
+            self.delete(path).await?;
+        }
+        Ok(())
+    }
+}

libs/remote_storage/src/lib.rs

@@ -4,7 +4,10 @@
 //! [`RemoteStorage`] trait a CRUD-like generic abstraction to use for adapting external storages with a few implementations:
 //!   * [`local_fs`] allows to use local file system as an external storage
 //!   * [`s3_bucket`] uses AWS S3 bucket as an external storage
+//!   * [`azure_blob`] allows to use Azure Blob storage as an external storage
 //!
+
+mod azure_blob;
 mod local_fs;
 mod s3_bucket;
 mod simulate_failures;
@@ -21,11 +24,15 @@ use anyhow::{bail, Context};
 use camino::{Utf8Path, Utf8PathBuf};
 
 use serde::{Deserialize, Serialize};
-use tokio::io;
+use tokio::{io, sync::Semaphore};
 use toml_edit::Item;
 use tracing::info;
 
-pub use self::{local_fs::LocalFs, s3_bucket::S3Bucket, simulate_failures::UnreliableWrapper};
+pub use self::{
+    azure_blob::AzureBlobStorage, local_fs::LocalFs, s3_bucket::S3Bucket,
+    simulate_failures::UnreliableWrapper,
+};
+use s3_bucket::RequestKind;
 
 /// How many different timelines can be processed simultaneously when synchronizing layers with the remote storage.
 /// During regular work, pageserver produces one layer file per timeline checkpoint, with bursts of concurrency
@@ -39,6 +46,11 @@ pub const DEFAULT_REMOTE_STORAGE_MAX_SYNC_ERRORS: u32 = 10;
 /// ~3500 PUT/COPY/POST/DELETE or 5500 GET/HEAD S3 requests
 /// <https://aws.amazon.com/premiumsupport/knowledge-center/s3-request-limit-avoid-throttling/>
 pub const DEFAULT_REMOTE_STORAGE_S3_CONCURRENCY_LIMIT: usize = 100;
+/// We set this a little bit low as we currently buffer the entire file into RAM
+///
+/// Here, a limit of max 20k concurrent connections was noted.
+/// <https://learn.microsoft.com/en-us/answers/questions/1301863/is-there-any-limitation-to-concurrent-connections>
+pub const DEFAULT_REMOTE_STORAGE_AZURE_CONCURRENCY_LIMIT: usize = 30;
 /// No limits on the client side, which currenltly means 1000 for AWS S3.
 /// <https://docs.aws.amazon.com/AmazonS3/latest/API/API_ListObjectsV2.html#API_ListObjectsV2_RequestSyntax>
 pub const DEFAULT_MAX_KEYS_PER_LIST_RESPONSE: Option<i32> = None;
@@ -217,6 +229,7 @@ impl std::error::Error for DownloadError {}
 pub enum GenericRemoteStorage {
     LocalFs(LocalFs),
     AwsS3(Arc<S3Bucket>),
+    AzureBlob(Arc<AzureBlobStorage>),
     Unreliable(Arc<UnreliableWrapper>),
 }
 
@@ -228,6 +241,7 @@ impl GenericRemoteStorage {
         match self {
             Self::LocalFs(s) => s.list_files(folder).await,
             Self::AwsS3(s) => s.list_files(folder).await,
+            Self::AzureBlob(s) => s.list_files(folder).await,
             Self::Unreliable(s) => s.list_files(folder).await,
         }
     }
@@ -242,6 +256,7 @@ impl GenericRemoteStorage {
         match self {
             Self::LocalFs(s) => s.list_prefixes(prefix).await,
             Self::AwsS3(s) => s.list_prefixes(prefix).await,
+            Self::AzureBlob(s) => s.list_prefixes(prefix).await,
             Self::Unreliable(s) => s.list_prefixes(prefix).await,
         }
     }
@@ -256,6 +271,7 @@ impl GenericRemoteStorage {
         match self {
             Self::LocalFs(s) => s.upload(from, data_size_bytes, to, metadata).await,
             Self::AwsS3(s) => s.upload(from, data_size_bytes, to, metadata).await,
+            Self::AzureBlob(s) => s.upload(from, data_size_bytes, to, metadata).await,
             Self::Unreliable(s) => s.upload(from, data_size_bytes, to, metadata).await,
         }
     }
@@ -264,6 +280,7 @@ impl GenericRemoteStorage {
         match self {
             Self::LocalFs(s) => s.download(from).await,
             Self::AwsS3(s) => s.download(from).await,
+            Self::AzureBlob(s) => s.download(from).await,
             Self::Unreliable(s) => s.download(from).await,
         }
     }
@@ -283,6 +300,10 @@ impl GenericRemoteStorage {
                 s.download_byte_range(from, start_inclusive, end_exclusive)
                     .await
             }
+            Self::AzureBlob(s) => {
+                s.download_byte_range(from, start_inclusive, end_exclusive)
+                    .await
+            }
             Self::Unreliable(s) => {
                 s.download_byte_range(from, start_inclusive, end_exclusive)
                     .await
@@ -294,6 +315,7 @@ impl GenericRemoteStorage {
         match self {
             Self::LocalFs(s) => s.delete(path).await,
             Self::AwsS3(s) => s.delete(path).await,
+            Self::AzureBlob(s) => s.delete(path).await,
             Self::Unreliable(s) => s.delete(path).await,
         }
     }
@@ -302,6 +324,7 @@ impl GenericRemoteStorage {
         match self {
             Self::LocalFs(s) => s.delete_objects(paths).await,
             Self::AwsS3(s) => s.delete_objects(paths).await,
+            Self::AzureBlob(s) => s.delete_objects(paths).await,
             Self::Unreliable(s) => s.delete_objects(paths).await,
         }
     }
@@ -319,6 +342,11 @@ impl GenericRemoteStorage {
                       s3_config.bucket_name, s3_config.bucket_region, s3_config.prefix_in_bucket, s3_config.endpoint);
                 Self::AwsS3(Arc::new(S3Bucket::new(s3_config)?))
             }
+            RemoteStorageKind::AzureContainer(azure_config) => {
+                info!("Using azure container '{}' in region '{}' as a remote storage, prefix in container: '{:?}'",
+                      azure_config.container_name, azure_config.container_region, azure_config.prefix_in_container);
+                Self::AzureBlob(Arc::new(AzureBlobStorage::new(azure_config)?))
+            }
         })
     }
 
@@ -383,6 +411,9 @@ pub enum RemoteStorageKind {
     /// AWS S3 based storage, storing all files in the S3 bucket
     /// specified by the config
     AwsS3(S3Config),
+    /// Azure Blob based storage, storing all files in the container
+    /// specified by the config
+    AzureContainer(AzureConfig),
 }
 
 /// AWS S3 bucket coordinates and access credentials to manage the bucket contents (read and write).
@@ -422,11 +453,45 @@ impl Debug for S3Config {
     }
 }
 
+/// Azure  bucket coordinates and access credentials to manage the bucket contents (read and write).
+#[derive(Clone, PartialEq, Eq)]
+pub struct AzureConfig {
+    /// Name of the container to connect to.
+    pub container_name: String,
+    /// The region where the bucket is located at.
+    pub container_region: String,
+    /// A "subfolder" in the container, to use the same container separately by multiple remote storage users at once.
+    pub prefix_in_container: Option<String>,
+    /// Azure has various limits on its API calls, we need not to exceed those.
+    /// See [`DEFAULT_REMOTE_STORAGE_AZURE_CONCURRENCY_LIMIT`] for more details.
+    pub concurrency_limit: NonZeroUsize,
+    pub max_keys_per_list_response: Option<i32>,
+}
+
+impl Debug for AzureConfig {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        f.debug_struct("AzureConfig")
+            .field("bucket_name", &self.container_name)
+            .field("bucket_region", &self.container_region)
+            .field("prefix_in_bucket", &self.prefix_in_container)
+            .field("concurrency_limit", &self.concurrency_limit)
+            .field(
+                "max_keys_per_list_response",
+                &self.max_keys_per_list_response,
+            )
+            .finish()
+    }
+}
+
 impl RemoteStorageConfig {
     pub fn from_toml(toml: &toml_edit::Item) -> anyhow::Result<Option<RemoteStorageConfig>> {
         let local_path = toml.get("local_path");
         let bucket_name = toml.get("bucket_name");
         let bucket_region = toml.get("bucket_region");
+        let container_name = toml.get("container_name");
+        let container_region = toml.get("container_region");
+
+        let use_azure = container_name.is_some() && container_region.is_some();
 
         let max_concurrent_syncs = NonZeroUsize::new(
             parse_optional_integer("max_concurrent_syncs", toml)?
@@ -440,9 +505,13 @@ impl RemoteStorageConfig {
         )
         .context("Failed to parse 'max_sync_errors' as a positive integer")?;
 
+        let default_concurrency_limit = if use_azure {
+            DEFAULT_REMOTE_STORAGE_AZURE_CONCURRENCY_LIMIT
+        } else {
+            DEFAULT_REMOTE_STORAGE_S3_CONCURRENCY_LIMIT
+        };
         let concurrency_limit = NonZeroUsize::new(
-            parse_optional_integer("concurrency_limit", toml)?
-                .unwrap_or(DEFAULT_REMOTE_STORAGE_S3_CONCURRENCY_LIMIT),
+            parse_optional_integer("concurrency_limit", toml)?.unwrap_or(default_concurrency_limit),
         )
         .context("Failed to parse 'concurrency_limit' as a positive integer")?;
 
@@ -451,33 +520,70 @@ impl RemoteStorageConfig {
                 .context("Failed to parse 'max_keys_per_list_response' as a positive integer")?
                 .or(DEFAULT_MAX_KEYS_PER_LIST_RESPONSE);
 
-        let storage = match (local_path, bucket_name, bucket_region) {
+        let endpoint = toml
+            .get("endpoint")
+            .map(|endpoint| parse_toml_string("endpoint", endpoint))
+            .transpose()?;
+
+        let storage = match (
+            local_path,
+            bucket_name,
+            bucket_region,
+            container_name,
+            container_region,
+        ) {
             // no 'local_path' nor 'bucket_name' options are provided, consider this remote storage disabled
-            (None, None, None) => return Ok(None),
-            (_, Some(_), None) => {
+            (None, None, None, None, None) => return Ok(None),
+            (_, Some(_), None, ..) => {
                 bail!("'bucket_region' option is mandatory if 'bucket_name' is given ")
             }
-            (_, None, Some(_)) => {
+            (_, None, Some(_), ..) => {
                 bail!("'bucket_name' option is mandatory if 'bucket_region' is given ")
             }
-            (None, Some(bucket_name), Some(bucket_region)) => RemoteStorageKind::AwsS3(S3Config {
-                bucket_name: parse_toml_string("bucket_name", bucket_name)?,
-                bucket_region: parse_toml_string("bucket_region", bucket_region)?,
-                prefix_in_bucket: toml
-                    .get("prefix_in_bucket")
-                    .map(|prefix_in_bucket| parse_toml_string("prefix_in_bucket", prefix_in_bucket))
-                    .transpose()?,
-                endpoint: toml
-                    .get("endpoint")
-                    .map(|endpoint| parse_toml_string("endpoint", endpoint))
-                    .transpose()?,
-                concurrency_limit,
-                max_keys_per_list_response,
-            }),
-            (Some(local_path), None, None) => RemoteStorageKind::LocalFs(Utf8PathBuf::from(
-                parse_toml_string("local_path", local_path)?,
-            )),
-            (Some(_), Some(_), _) => bail!("local_path and bucket_name are mutually exclusive"),
+            (None, Some(bucket_name), Some(bucket_region), ..) => {
+                RemoteStorageKind::AwsS3(S3Config {
+                    bucket_name: parse_toml_string("bucket_name", bucket_name)?,
+                    bucket_region: parse_toml_string("bucket_region", bucket_region)?,
+                    prefix_in_bucket: toml
+                        .get("prefix_in_bucket")
+                        .map(|prefix_in_bucket| {
+                            parse_toml_string("prefix_in_bucket", prefix_in_bucket)
+                        })
+                        .transpose()?,
+                    endpoint,
+                    concurrency_limit,
+                    max_keys_per_list_response,
+                })
+            }
+            (_, _, _, Some(_), None) => {
+                bail!("'container_name' option is mandatory if 'container_region' is given ")
+            }
+            (_, _, _, None, Some(_)) => {
+                bail!("'container_name' option is mandatory if 'container_region' is given ")
+            }
+            (None, None, None, Some(container_name), Some(container_region)) => {
+                RemoteStorageKind::AzureContainer(AzureConfig {
+                    container_name: parse_toml_string("container_name", container_name)?,
+                    container_region: parse_toml_string("container_region", container_region)?,
+                    prefix_in_container: toml
+                        .get("prefix_in_container")
+                        .map(|prefix_in_container| {
+                            parse_toml_string("prefix_in_container", prefix_in_container)
+                        })
+                        .transpose()?,
+                    concurrency_limit,
+                    max_keys_per_list_response,
+                })
+            }
+            (Some(local_path), None, None, None, None) => RemoteStorageKind::LocalFs(
+                Utf8PathBuf::from(parse_toml_string("local_path", local_path)?),
+            ),
+            (Some(_), Some(_), ..) => {
+                bail!("'local_path' and 'bucket_name' are mutually exclusive")
+            }
+            (Some(_), _, _, Some(_), Some(_)) => {
+                bail!("local_path and 'container_name' are mutually exclusive")
+            }
         };
 
         Ok(Some(RemoteStorageConfig {
@@ -513,6 +619,46 @@ fn parse_toml_string(name: &str, item: &Item) -> anyhow::Result<String> {
     Ok(s.to_string())
 }
 
+struct ConcurrencyLimiter {
+    // Every request to S3 can be throttled or cancelled, if a certain number of requests per second is exceeded.
+    // Same goes to IAM, which is queried before every S3 request, if enabled. IAM has even lower RPS threshold.
+    // The helps to ensure we don't exceed the thresholds.
+    write: Arc<Semaphore>,
+    read: Arc<Semaphore>,
+}
+
+impl ConcurrencyLimiter {
+    fn for_kind(&self, kind: RequestKind) -> &Arc<Semaphore> {
+        match kind {
+            RequestKind::Get => &self.read,
+            RequestKind::Put => &self.write,
+            RequestKind::List => &self.read,
+            RequestKind::Delete => &self.write,
+        }
+    }
+
+    async fn acquire(
+        &self,
+        kind: RequestKind,
+    ) -> Result<tokio::sync::SemaphorePermit<'_>, tokio::sync::AcquireError> {
+        self.for_kind(kind).acquire().await
+    }
+
+    async fn acquire_owned(
+        &self,
+        kind: RequestKind,
+    ) -> Result<tokio::sync::OwnedSemaphorePermit, tokio::sync::AcquireError> {
+        Arc::clone(self.for_kind(kind)).acquire_owned().await
+    }
+
+    fn new(limit: usize) -> ConcurrencyLimiter {
+        Self {
+            read: Arc::new(Semaphore::new(limit)),
+            write: Arc::new(Semaphore::new(limit)),
+        }
+    }
+}
+
 #[cfg(test)]
 mod tests {
     use super::*;

libs/remote_storage/src/s3_bucket.rs

@@ -4,7 +4,7 @@
 //! allowing multiple api users to independently work with the same S3 bucket, if
 //! their bucket prefixes are both specified and different.
 
-use std::{borrow::Cow, sync::Arc};
+use std::borrow::Cow;
 
 use anyhow::Context;
 use aws_config::{
@@ -24,22 +24,20 @@ use aws_sdk_s3::{
 use aws_smithy_http::body::SdkBody;
 use hyper::Body;
 use scopeguard::ScopeGuard;
-use tokio::{
-    io::{self, AsyncRead},
-    sync::Semaphore,
-};
+use tokio::io::{self, AsyncRead};
 use tokio_util::io::ReaderStream;
 use tracing::debug;
 
 use super::StorageMetadata;
 use crate::{
-    Download, DownloadError, RemotePath, RemoteStorage, S3Config, MAX_KEYS_PER_DELETE,
-    REMOTE_STORAGE_PREFIX_SEPARATOR,
+    ConcurrencyLimiter, Download, DownloadError, RemotePath, RemoteStorage, S3Config,
+    MAX_KEYS_PER_DELETE, REMOTE_STORAGE_PREFIX_SEPARATOR,
 };
 
 pub(super) mod metrics;
 
-use self::metrics::{AttemptOutcome, RequestKind};
+use self::metrics::AttemptOutcome;
+pub(super) use self::metrics::RequestKind;
 
 /// AWS S3 storage.
 pub struct S3Bucket {
@@ -50,46 +48,6 @@ pub struct S3Bucket {
     concurrency_limiter: ConcurrencyLimiter,
 }
 
-struct ConcurrencyLimiter {
-    // Every request to S3 can be throttled or cancelled, if a certain number of requests per second is exceeded.
-    // Same goes to IAM, which is queried before every S3 request, if enabled. IAM has even lower RPS threshold.
-    // The helps to ensure we don't exceed the thresholds.
-    write: Arc<Semaphore>,
-    read: Arc<Semaphore>,
-}
-
-impl ConcurrencyLimiter {
-    fn for_kind(&self, kind: RequestKind) -> &Arc<Semaphore> {
-        match kind {
-            RequestKind::Get => &self.read,
-            RequestKind::Put => &self.write,
-            RequestKind::List => &self.read,
-            RequestKind::Delete => &self.write,
-        }
-    }
-
-    async fn acquire(
-        &self,
-        kind: RequestKind,
-    ) -> Result<tokio::sync::SemaphorePermit<'_>, tokio::sync::AcquireError> {
-        self.for_kind(kind).acquire().await
-    }
-
-    async fn acquire_owned(
-        &self,
-        kind: RequestKind,
-    ) -> Result<tokio::sync::OwnedSemaphorePermit, tokio::sync::AcquireError> {
-        Arc::clone(self.for_kind(kind)).acquire_owned().await
-    }
-
-    fn new(limit: usize) -> ConcurrencyLimiter {
-        Self {
-            read: Arc::new(Semaphore::new(limit)),
-            write: Arc::new(Semaphore::new(limit)),
-        }
-    }
-}
-
 #[derive(Default)]
 struct GetObjectRequest {
     bucket: String,

libs/remote_storage/src/s3_bucket/metrics.rs

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

libs/remote_storage/tests/test_real_azure.rs

@@ -0,0 +1,625 @@
+use std::collections::HashSet;
+use std::env;
+use std::num::{NonZeroU32, NonZeroUsize};
+use std::ops::ControlFlow;
+use std::path::PathBuf;
+use std::sync::Arc;
+use std::time::UNIX_EPOCH;
+
+use anyhow::Context;
+use camino::Utf8Path;
+use once_cell::sync::OnceCell;
+use remote_storage::{
+    AzureConfig, Download, GenericRemoteStorage, RemotePath, RemoteStorageConfig, RemoteStorageKind,
+};
+use test_context::{test_context, AsyncTestContext};
+use tokio::task::JoinSet;
+use tracing::{debug, error, info};
+
+static LOGGING_DONE: OnceCell<()> = OnceCell::new();
+
+const ENABLE_REAL_AZURE_REMOTE_STORAGE_ENV_VAR_NAME: &str = "ENABLE_REAL_AZURE_REMOTE_STORAGE";
+
+const BASE_PREFIX: &str = "test";
+
+/// Tests that the Azure client can list all prefixes, even if the response comes paginated and requires multiple HTTP queries.
+/// Uses real Azure and requires [`ENABLE_REAL_AZURE_REMOTE_STORAGE_ENV_VAR_NAME`] and related Azure cred env vars specified.
+/// See the client creation in [`create_azure_client`] for details on the required env vars.
+/// If real Azure tests are disabled, the test passes, skipping any real test run: currently, there's no way to mark the test ignored in runtime with the
+/// deafult test framework, see https://github.com/rust-lang/rust/issues/68007 for details.
+///
+/// First, the test creates a set of Azure blobs with keys `/${random_prefix_part}/${base_prefix_str}/sub_prefix_${i}/blob_${i}` in [`upload_azure_data`]
+/// where
+/// * `random_prefix_part` is set for the entire Azure client during the Azure client creation in [`create_azure_client`], to avoid multiple test runs interference
+/// * `base_prefix_str` is a common prefix to use in the client requests: we would want to ensure that the client is able to list nested prefixes inside the bucket
+///
+/// Then, verifies that the client does return correct prefixes when queried:
+/// * with no prefix, it lists everything after its `${random_prefix_part}/` — that should be `${base_prefix_str}` value only
+/// * with `${base_prefix_str}/` prefix, it lists every `sub_prefix_${i}`
+///
+/// With the real Azure enabled and `#[cfg(test)]` Rust configuration used, the Azure client test adds a `max-keys` param to limit the response keys.
+/// This way, we are able to test the pagination implicitly, by ensuring all results are returned from the remote storage and avoid uploading too many blobs to Azure.
+///
+/// Lastly, the test attempts to clean up and remove all uploaded Azure files.
+/// If any errors appear during the clean up, they get logged, but the test is not failed or stopped until clean up is finished.
+#[test_context(MaybeEnabledAzureWithTestBlobs)]
+#[tokio::test]
+async fn azure_pagination_should_work(
+    ctx: &mut MaybeEnabledAzureWithTestBlobs,
+) -> anyhow::Result<()> {
+    let ctx = match ctx {
+        MaybeEnabledAzureWithTestBlobs::Enabled(ctx) => ctx,
+        MaybeEnabledAzureWithTestBlobs::Disabled => return Ok(()),
+        MaybeEnabledAzureWithTestBlobs::UploadsFailed(e, _) => {
+            anyhow::bail!("Azure init failed: {e:?}")
+        }
+    };
+
+    let test_client = Arc::clone(&ctx.enabled.client);
+    let expected_remote_prefixes = ctx.remote_prefixes.clone();
+
+    let base_prefix = RemotePath::new(Utf8Path::new(ctx.enabled.base_prefix))
+        .context("common_prefix construction")?;
+    let root_remote_prefixes = test_client
+        .list_prefixes(None)
+        .await
+        .context("client list root prefixes failure")?
+        .into_iter()
+        .collect::<HashSet<_>>();
+    assert_eq!(
+        root_remote_prefixes, HashSet::from([base_prefix.clone()]),
+        "remote storage root prefixes list mismatches with the uploads. Returned prefixes: {root_remote_prefixes:?}"
+    );
+
+    let nested_remote_prefixes = test_client
+        .list_prefixes(Some(&base_prefix))
+        .await
+        .context("client list nested prefixes failure")?
+        .into_iter()
+        .collect::<HashSet<_>>();
+    let remote_only_prefixes = nested_remote_prefixes
+        .difference(&expected_remote_prefixes)
+        .collect::<HashSet<_>>();
+    let missing_uploaded_prefixes = expected_remote_prefixes
+        .difference(&nested_remote_prefixes)
+        .collect::<HashSet<_>>();
+    assert_eq!(
+        remote_only_prefixes.len() + missing_uploaded_prefixes.len(), 0,
+        "remote storage nested prefixes list mismatches with the uploads. Remote only prefixes: {remote_only_prefixes:?}, missing uploaded prefixes: {missing_uploaded_prefixes:?}",
+    );
+
+    Ok(())
+}
+
+/// Tests that Azure client can list all files in a folder, even if the response comes paginated and requirees multiple Azure queries.
+/// Uses real Azure and requires [`ENABLE_REAL_AZURE_REMOTE_STORAGE_ENV_VAR_NAME`] and related Azure cred env vars specified. Test will skip real code and pass if env vars not set.
+/// See `Azure_pagination_should_work` for more information.
+///
+/// First, create a set of Azure objects with keys `random_prefix/folder{j}/blob_{i}.txt` in [`upload_azure_data`]
+/// Then performs the following queries:
+///    1. `list_files(None)`. This should return all files `random_prefix/folder{j}/blob_{i}.txt`
+///    2. `list_files("folder1")`.  This  should return all files `random_prefix/folder1/blob_{i}.txt`
+#[test_context(MaybeEnabledAzureWithSimpleTestBlobs)]
+#[tokio::test]
+async fn azure_list_files_works(
+    ctx: &mut MaybeEnabledAzureWithSimpleTestBlobs,
+) -> anyhow::Result<()> {
+    let ctx = match ctx {
+        MaybeEnabledAzureWithSimpleTestBlobs::Enabled(ctx) => ctx,
+        MaybeEnabledAzureWithSimpleTestBlobs::Disabled => return Ok(()),
+        MaybeEnabledAzureWithSimpleTestBlobs::UploadsFailed(e, _) => {
+            anyhow::bail!("Azure init failed: {e:?}")
+        }
+    };
+    let test_client = Arc::clone(&ctx.enabled.client);
+    let base_prefix =
+        RemotePath::new(Utf8Path::new("folder1")).context("common_prefix construction")?;
+    let root_files = test_client
+        .list_files(None)
+        .await
+        .context("client list root files failure")?
+        .into_iter()
+        .collect::<HashSet<_>>();
+    assert_eq!(
+        root_files,
+        ctx.remote_blobs.clone(),
+        "remote storage list_files on root mismatches with the uploads."
+    );
+    let nested_remote_files = test_client
+        .list_files(Some(&base_prefix))
+        .await
+        .context("client list nested files failure")?
+        .into_iter()
+        .collect::<HashSet<_>>();
+    let trim_remote_blobs: HashSet<_> = ctx
+        .remote_blobs
+        .iter()
+        .map(|x| x.get_path())
+        .filter(|x| x.starts_with("folder1"))
+        .map(|x| RemotePath::new(x).expect("must be valid path"))
+        .collect();
+    assert_eq!(
+        nested_remote_files, trim_remote_blobs,
+        "remote storage list_files on subdirrectory mismatches with the uploads."
+    );
+    Ok(())
+}
+
+#[test_context(MaybeEnabledAzure)]
+#[tokio::test]
+async fn azure_delete_non_exising_works(ctx: &mut MaybeEnabledAzure) -> anyhow::Result<()> {
+    let ctx = match ctx {
+        MaybeEnabledAzure::Enabled(ctx) => ctx,
+        MaybeEnabledAzure::Disabled => return Ok(()),
+    };
+
+    let path = RemotePath::new(Utf8Path::new(
+        format!("{}/for_sure_there_is_nothing_there_really", ctx.base_prefix).as_str(),
+    ))
+    .with_context(|| "RemotePath conversion")?;
+
+    ctx.client.delete(&path).await.expect("should succeed");
+
+    Ok(())
+}
+
+#[test_context(MaybeEnabledAzure)]
+#[tokio::test]
+async fn azure_delete_objects_works(ctx: &mut MaybeEnabledAzure) -> anyhow::Result<()> {
+    let ctx = match ctx {
+        MaybeEnabledAzure::Enabled(ctx) => ctx,
+        MaybeEnabledAzure::Disabled => return Ok(()),
+    };
+
+    let path1 = RemotePath::new(Utf8Path::new(format!("{}/path1", ctx.base_prefix).as_str()))
+        .with_context(|| "RemotePath conversion")?;
+
+    let path2 = RemotePath::new(Utf8Path::new(format!("{}/path2", ctx.base_prefix).as_str()))
+        .with_context(|| "RemotePath conversion")?;
+
+    let path3 = RemotePath::new(Utf8Path::new(format!("{}/path3", ctx.base_prefix).as_str()))
+        .with_context(|| "RemotePath conversion")?;
+
+    let data1 = "remote blob data1".as_bytes();
+    let data1_len = data1.len();
+    let data2 = "remote blob data2".as_bytes();
+    let data2_len = data2.len();
+    let data3 = "remote blob data3".as_bytes();
+    let data3_len = data3.len();
+    ctx.client
+        .upload(std::io::Cursor::new(data1), data1_len, &path1, None)
+        .await?;
+
+    ctx.client
+        .upload(std::io::Cursor::new(data2), data2_len, &path2, None)
+        .await?;
+
+    ctx.client
+        .upload(std::io::Cursor::new(data3), data3_len, &path3, None)
+        .await?;
+
+    ctx.client.delete_objects(&[path1, path2]).await?;
+
+    let prefixes = ctx.client.list_prefixes(None).await?;
+
+    assert_eq!(prefixes.len(), 1);
+
+    ctx.client.delete_objects(&[path3]).await?;
+
+    Ok(())
+}
+
+#[test_context(MaybeEnabledAzure)]
+#[tokio::test]
+async fn azure_upload_download_works(ctx: &mut MaybeEnabledAzure) -> anyhow::Result<()> {
+    let MaybeEnabledAzure::Enabled(ctx) = ctx else {
+        return Ok(());
+    };
+
+    let path = RemotePath::new(Utf8Path::new(format!("{}/file", ctx.base_prefix).as_str()))
+        .with_context(|| "RemotePath conversion")?;
+
+    let data = "remote blob data here".as_bytes();
+    let data_len = data.len() as u64;
+
+    ctx.client
+        .upload(std::io::Cursor::new(data), data.len(), &path, None)
+        .await?;
+
+    async fn download_and_compare(mut dl: Download) -> anyhow::Result<Vec<u8>> {
+        let mut buf = Vec::new();
+        tokio::io::copy(&mut dl.download_stream, &mut buf).await?;
+        Ok(buf)
+    }
+    // Normal download request
+    let dl = ctx.client.download(&path).await?;
+    let buf = download_and_compare(dl).await?;
+    assert_eq!(buf, data);
+
+    // Full range (end specified)
+    let dl = ctx
+        .client
+        .download_byte_range(&path, 0, Some(data_len))
+        .await?;
+    let buf = download_and_compare(dl).await?;
+    assert_eq!(buf, data);
+
+    // partial range (end specified)
+    let dl = ctx.client.download_byte_range(&path, 4, Some(10)).await?;
+    let buf = download_and_compare(dl).await?;
+    assert_eq!(buf, data[4..10]);
+
+    // partial range (end beyond real end)
+    let dl = ctx
+        .client
+        .download_byte_range(&path, 8, Some(data_len * 100))
+        .await?;
+    let buf = download_and_compare(dl).await?;
+    assert_eq!(buf, data[8..]);
+
+    // Partial range (end unspecified)
+    let dl = ctx.client.download_byte_range(&path, 4, None).await?;
+    let buf = download_and_compare(dl).await?;
+    assert_eq!(buf, data[4..]);
+
+    // Full range (end unspecified)
+    let dl = ctx.client.download_byte_range(&path, 0, None).await?;
+    let buf = download_and_compare(dl).await?;
+    assert_eq!(buf, data);
+
+    debug!("Cleanup: deleting file at path {path:?}");
+    ctx.client
+        .delete(&path)
+        .await
+        .with_context(|| format!("{path:?} removal"))?;
+
+    Ok(())
+}
+
+fn ensure_logging_ready() {
+    LOGGING_DONE.get_or_init(|| {
+        utils::logging::init(
+            utils::logging::LogFormat::Test,
+            utils::logging::TracingErrorLayerEnablement::Disabled,
+        )
+        .expect("logging init failed");
+    });
+}
+
+struct EnabledAzure {
+    client: Arc<GenericRemoteStorage>,
+    base_prefix: &'static str,
+}
+
+impl EnabledAzure {
+    async fn setup(max_keys_in_list_response: Option<i32>) -> Self {
+        let client = create_azure_client(max_keys_in_list_response)
+            .context("Azure client creation")
+            .expect("Azure client creation failed");
+
+        EnabledAzure {
+            client,
+            base_prefix: BASE_PREFIX,
+        }
+    }
+}
+
+enum MaybeEnabledAzure {
+    Enabled(EnabledAzure),
+    Disabled,
+}
+
+#[async_trait::async_trait]
+impl AsyncTestContext for MaybeEnabledAzure {
+    async fn setup() -> Self {
+        ensure_logging_ready();
+
+        if env::var(ENABLE_REAL_AZURE_REMOTE_STORAGE_ENV_VAR_NAME).is_err() {
+            info!(
+                "`{}` env variable is not set, skipping the test",
+                ENABLE_REAL_AZURE_REMOTE_STORAGE_ENV_VAR_NAME
+            );
+            return Self::Disabled;
+        }
+
+        Self::Enabled(EnabledAzure::setup(None).await)
+    }
+}
+
+enum MaybeEnabledAzureWithTestBlobs {
+    Enabled(AzureWithTestBlobs),
+    Disabled,
+    UploadsFailed(anyhow::Error, AzureWithTestBlobs),
+}
+
+struct AzureWithTestBlobs {
+    enabled: EnabledAzure,
+    remote_prefixes: HashSet<RemotePath>,
+    remote_blobs: HashSet<RemotePath>,
+}
+
+#[async_trait::async_trait]
+impl AsyncTestContext for MaybeEnabledAzureWithTestBlobs {
+    async fn setup() -> Self {
+        ensure_logging_ready();
+        if env::var(ENABLE_REAL_AZURE_REMOTE_STORAGE_ENV_VAR_NAME).is_err() {
+            info!(
+                "`{}` env variable is not set, skipping the test",
+                ENABLE_REAL_AZURE_REMOTE_STORAGE_ENV_VAR_NAME
+            );
+            return Self::Disabled;
+        }
+
+        let max_keys_in_list_response = 10;
+        let upload_tasks_count = 1 + (2 * usize::try_from(max_keys_in_list_response).unwrap());
+
+        let enabled = EnabledAzure::setup(Some(max_keys_in_list_response)).await;
+
+        match upload_azure_data(&enabled.client, enabled.base_prefix, upload_tasks_count).await {
+            ControlFlow::Continue(uploads) => {
+                info!("Remote objects created successfully");
+
+                Self::Enabled(AzureWithTestBlobs {
+                    enabled,
+                    remote_prefixes: uploads.prefixes,
+                    remote_blobs: uploads.blobs,
+                })
+            }
+            ControlFlow::Break(uploads) => Self::UploadsFailed(
+                anyhow::anyhow!("One or multiple blobs failed to upload to Azure"),
+                AzureWithTestBlobs {
+                    enabled,
+                    remote_prefixes: uploads.prefixes,
+                    remote_blobs: uploads.blobs,
+                },
+            ),
+        }
+    }
+
+    async fn teardown(self) {
+        match self {
+            Self::Disabled => {}
+            Self::Enabled(ctx) | Self::UploadsFailed(_, ctx) => {
+                cleanup(&ctx.enabled.client, ctx.remote_blobs).await;
+            }
+        }
+    }
+}
+
+// NOTE: the setups for the list_prefixes test and the list_files test are very similar
+// However, they are not idential. The list_prefixes function is concerned with listing prefixes,
+// whereas the list_files function is concerned with listing files.
+// See `RemoteStorage::list_files` documentation for more details
+enum MaybeEnabledAzureWithSimpleTestBlobs {
+    Enabled(AzureWithSimpleTestBlobs),
+    Disabled,
+    UploadsFailed(anyhow::Error, AzureWithSimpleTestBlobs),
+}
+struct AzureWithSimpleTestBlobs {
+    enabled: EnabledAzure,
+    remote_blobs: HashSet<RemotePath>,
+}
+
+#[async_trait::async_trait]
+impl AsyncTestContext for MaybeEnabledAzureWithSimpleTestBlobs {
+    async fn setup() -> Self {
+        ensure_logging_ready();
+        if env::var(ENABLE_REAL_AZURE_REMOTE_STORAGE_ENV_VAR_NAME).is_err() {
+            info!(
+                "`{}` env variable is not set, skipping the test",
+                ENABLE_REAL_AZURE_REMOTE_STORAGE_ENV_VAR_NAME
+            );
+            return Self::Disabled;
+        }
+
+        let max_keys_in_list_response = 10;
+        let upload_tasks_count = 1 + (2 * usize::try_from(max_keys_in_list_response).unwrap());
+
+        let enabled = EnabledAzure::setup(Some(max_keys_in_list_response)).await;
+
+        match upload_simple_azure_data(&enabled.client, upload_tasks_count).await {
+            ControlFlow::Continue(uploads) => {
+                info!("Remote objects created successfully");
+
+                Self::Enabled(AzureWithSimpleTestBlobs {
+                    enabled,
+                    remote_blobs: uploads,
+                })
+            }
+            ControlFlow::Break(uploads) => Self::UploadsFailed(
+                anyhow::anyhow!("One or multiple blobs failed to upload to Azure"),
+                AzureWithSimpleTestBlobs {
+                    enabled,
+                    remote_blobs: uploads,
+                },
+            ),
+        }
+    }
+
+    async fn teardown(self) {
+        match self {
+            Self::Disabled => {}
+            Self::Enabled(ctx) | Self::UploadsFailed(_, ctx) => {
+                cleanup(&ctx.enabled.client, ctx.remote_blobs).await;
+            }
+        }
+    }
+}
+
+fn create_azure_client(
+    max_keys_per_list_response: Option<i32>,
+) -> anyhow::Result<Arc<GenericRemoteStorage>> {
+    use rand::Rng;
+
+    let remote_storage_azure_container = env::var("REMOTE_STORAGE_AZURE_CONTAINER").context(
+        "`REMOTE_STORAGE_AZURE_CONTAINER` env var is not set, but real Azure tests are enabled",
+    )?;
+    let remote_storage_azure_region = env::var("REMOTE_STORAGE_AZURE_REGION").context(
+        "`REMOTE_STORAGE_AZURE_REGION` env var is not set, but real Azure tests are enabled",
+    )?;
+
+    // due to how time works, we've had test runners use the same nanos as bucket prefixes.
+    // millis is just a debugging aid for easier finding the prefix later.
+    let millis = std::time::SystemTime::now()
+        .duration_since(UNIX_EPOCH)
+        .context("random Azure test prefix part calculation")?
+        .as_millis();
+
+    // because nanos can be the same for two threads so can millis, add randomness
+    let random = rand::thread_rng().gen::<u32>();
+
+    let remote_storage_config = RemoteStorageConfig {
+        max_concurrent_syncs: NonZeroUsize::new(100).unwrap(),
+        max_sync_errors: NonZeroU32::new(5).unwrap(),
+        storage: RemoteStorageKind::AzureContainer(AzureConfig {
+            container_name: remote_storage_azure_container,
+            container_region: remote_storage_azure_region,
+            prefix_in_container: Some(format!("test_{millis}_{random:08x}/")),
+            concurrency_limit: NonZeroUsize::new(100).unwrap(),
+            max_keys_per_list_response,
+        }),
+    };
+    Ok(Arc::new(
+        GenericRemoteStorage::from_config(&remote_storage_config).context("remote storage init")?,
+    ))
+}
+
+struct Uploads {
+    prefixes: HashSet<RemotePath>,
+    blobs: HashSet<RemotePath>,
+}
+
+async fn upload_azure_data(
+    client: &Arc<GenericRemoteStorage>,
+    base_prefix_str: &'static str,
+    upload_tasks_count: usize,
+) -> ControlFlow<Uploads, Uploads> {
+    info!("Creating {upload_tasks_count} Azure files");
+    let mut upload_tasks = JoinSet::new();
+    for i in 1..upload_tasks_count + 1 {
+        let task_client = Arc::clone(client);
+        upload_tasks.spawn(async move {
+            let prefix = format!("{base_prefix_str}/sub_prefix_{i}/");
+            let blob_prefix = RemotePath::new(Utf8Path::new(&prefix))
+                .with_context(|| format!("{prefix:?} to RemotePath conversion"))?;
+            let blob_path = blob_prefix.join(Utf8Path::new(&format!("blob_{i}")));
+            debug!("Creating remote item {i} at path {blob_path:?}");
+
+            let data = format!("remote blob data {i}").into_bytes();
+            let data_len = data.len();
+            task_client
+                .upload(std::io::Cursor::new(data), data_len, &blob_path, None)
+                .await?;
+
+            Ok::<_, anyhow::Error>((blob_prefix, blob_path))
+        });
+    }
+
+    let mut upload_tasks_failed = false;
+    let mut uploaded_prefixes = HashSet::with_capacity(upload_tasks_count);
+    let mut uploaded_blobs = HashSet::with_capacity(upload_tasks_count);
+    while let Some(task_run_result) = upload_tasks.join_next().await {
+        match task_run_result
+            .context("task join failed")
+            .and_then(|task_result| task_result.context("upload task failed"))
+        {
+            Ok((upload_prefix, upload_path)) => {
+                uploaded_prefixes.insert(upload_prefix);
+                uploaded_blobs.insert(upload_path);
+            }
+            Err(e) => {
+                error!("Upload task failed: {e:?}");
+                upload_tasks_failed = true;
+            }
+        }
+    }
+
+    let uploads = Uploads {
+        prefixes: uploaded_prefixes,
+        blobs: uploaded_blobs,
+    };
+    if upload_tasks_failed {
+        ControlFlow::Break(uploads)
+    } else {
+        ControlFlow::Continue(uploads)
+    }
+}
+
+async fn cleanup(client: &Arc<GenericRemoteStorage>, objects_to_delete: HashSet<RemotePath>) {
+    info!(
+        "Removing {} objects from the remote storage during cleanup",
+        objects_to_delete.len()
+    );
+    let mut delete_tasks = JoinSet::new();
+    for object_to_delete in objects_to_delete {
+        let task_client = Arc::clone(client);
+        delete_tasks.spawn(async move {
+            debug!("Deleting remote item at path {object_to_delete:?}");
+            task_client
+                .delete(&object_to_delete)
+                .await
+                .with_context(|| format!("{object_to_delete:?} removal"))
+        });
+    }
+
+    while let Some(task_run_result) = delete_tasks.join_next().await {
+        match task_run_result {
+            Ok(task_result) => match task_result {
+                Ok(()) => {}
+                Err(e) => error!("Delete task failed: {e:?}"),
+            },
+            Err(join_err) => error!("Delete task did not finish correctly: {join_err}"),
+        }
+    }
+}
+
+// Uploads files `folder{j}/blob{i}.txt`. See test description for more details.
+async fn upload_simple_azure_data(
+    client: &Arc<GenericRemoteStorage>,
+    upload_tasks_count: usize,
+) -> ControlFlow<HashSet<RemotePath>, HashSet<RemotePath>> {
+    info!("Creating {upload_tasks_count} Azure files");
+    let mut upload_tasks = JoinSet::new();
+    for i in 1..upload_tasks_count + 1 {
+        let task_client = Arc::clone(client);
+        upload_tasks.spawn(async move {
+            let blob_path = PathBuf::from(format!("folder{}/blob_{}.txt", i / 7, i));
+            let blob_path = RemotePath::new(
+                Utf8Path::from_path(blob_path.as_path()).expect("must be valid blob path"),
+            )
+            .with_context(|| format!("{blob_path:?} to RemotePath conversion"))?;
+            debug!("Creating remote item {i} at path {blob_path:?}");
+
+            let data = format!("remote blob data {i}").into_bytes();
+            let data_len = data.len();
+            task_client
+                .upload(std::io::Cursor::new(data), data_len, &blob_path, None)
+                .await?;
+
+            Ok::<_, anyhow::Error>(blob_path)
+        });
+    }
+
+    let mut upload_tasks_failed = false;
+    let mut uploaded_blobs = HashSet::with_capacity(upload_tasks_count);
+    while let Some(task_run_result) = upload_tasks.join_next().await {
+        match task_run_result
+            .context("task join failed")
+            .and_then(|task_result| task_result.context("upload task failed"))
+        {
+            Ok(upload_path) => {
+                uploaded_blobs.insert(upload_path);
+            }
+            Err(e) => {
+                error!("Upload task failed: {e:?}");
+                upload_tasks_failed = true;
+            }
+        }
+    }
+
+    if upload_tasks_failed {
+        ControlFlow::Break(uploaded_blobs)
+    } else {
+        ControlFlow::Continue(uploaded_blobs)
+    }
+}

libs/vm_monitor/README.md

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

libs/vm_monitor/src/cgroup.rs

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

libs/vm_monitor/src/dispatcher.rs

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

libs/vm_monitor/src/runner.rs

@@ -5,18 +5,16 @@
 //! all functionality.
 
 use std::fmt::Debug;
-use std::sync::Arc;
 use std::time::{Duration, Instant};
 
 use anyhow::{bail, Context};
 use axum::extract::ws::{Message, WebSocket};
 use futures::StreamExt;
-use tokio::sync::broadcast;
-use tokio::sync::mpsc;
+use tokio::sync::{broadcast, watch};
 use tokio_util::sync::CancellationToken;
 use tracing::{error, info, warn};
 
-use crate::cgroup::{CgroupWatcher, Sequenced};
+use crate::cgroup::{self, CgroupWatcher};
 use crate::dispatcher::Dispatcher;
 use crate::filecache::{FileCacheConfig, FileCacheState};
 use crate::protocol::{InboundMsg, InboundMsgKind, OutboundMsg, OutboundMsgKind, Resources};
@@ -28,7 +26,7 @@ use crate::{bytes_to_mebibytes, get_total_system_memory, spawn_with_cancel, Args
 pub struct Runner {
     config: Config,
     filecache: Option<FileCacheState>,
-    cgroup: Option<Arc<CgroupWatcher>>,
+    cgroup: Option<CgroupState>,
     dispatcher: Dispatcher,
 
     /// We "mint" new message ids by incrementing this counter and taking the value.
@@ -45,6 +43,14 @@ pub struct Runner {
     kill: broadcast::Receiver<()>,
 }
 
+#[derive(Debug)]
+struct CgroupState {
+    watcher: watch::Receiver<(Instant, cgroup::MemoryHistory)>,
+    /// If [`cgroup::MemoryHistory::avg_non_reclaimable`] exceeds `threshold`, we send upscale
+    /// requests.
+    threshold: u64,
+}
+
 /// Configuration for a `Runner`
 #[derive(Debug)]
 pub struct Config {
@@ -62,16 +68,56 @@ pub struct Config {
     /// upscale resource amounts (because we might not *actually* have been upscaled yet). This field
     /// should be removed once we have a better solution there.
     sys_buffer_bytes: u64,
+
+    /// Minimum fraction of total system memory reserved *before* the the cgroup threshold; in
+    /// other words, providing a ceiling for the highest value of the threshold by enforcing that
+    /// there's at least `cgroup_min_overhead_fraction` of the total memory remaining beyond the
+    /// threshold.
+    ///
+    /// For example, a value of `0.1` means that 10% of total memory must remain after exceeding
+    /// the threshold, so the value of the cgroup threshold would always be capped at 90% of total
+    /// memory.
+    ///
+    /// The default value of `0.15` means that we *guarantee* sending upscale requests if the
+    /// cgroup is using more than 85% of total memory (even if we're *not* separately reserving
+    /// memory for the file cache).
+    cgroup_min_overhead_fraction: f64,
+
+    cgroup_downscale_threshold_buffer_bytes: u64,
 }
 
 impl Default for Config {
     fn default() -> Self {
         Self {
             sys_buffer_bytes: 100 * MiB,
+            cgroup_min_overhead_fraction: 0.15,
+            cgroup_downscale_threshold_buffer_bytes: 100 * MiB,
         }
     }
 }
 
+impl Config {
+    fn cgroup_threshold(&self, total_mem: u64, file_cache_disk_size: u64) -> u64 {
+        // If the file cache is in tmpfs, then it will count towards shmem usage of the cgroup,
+        // and thus be non-reclaimable, so we should allow for additional memory usage.
+        //
+        // If the file cache sits on disk, our desired stable system state is for it to be fully
+        // page cached (its contents should only be paged to/from disk in situations where we can't
+        // upscale fast enough). Page-cached memory is reclaimable, so we need to lower the
+        // threshold for non-reclaimable memory so we scale up *before* the kernel starts paging
+        // out the file cache.
+        let memory_remaining_for_cgroup = total_mem.saturating_sub(file_cache_disk_size);
+
+        // Even if we're not separately making room for the file cache (if it's in tmpfs), we still
+        // want our threshold to be met gracefully instead of letting postgres get OOM-killed.
+        // So we guarantee that there's at least `cgroup_min_overhead_fraction` of total memory
+        // remaining above the threshold.
+        let max_threshold = (total_mem as f64 * (1.0 - self.cgroup_min_overhead_fraction)) as u64;
+
+        memory_remaining_for_cgroup.min(max_threshold)
+    }
+}
+
 impl Runner {
     /// Create a new monitor.
     #[tracing::instrument(skip_all, fields(?config, ?args))]
@@ -87,12 +133,7 @@ impl Runner {
             "invalid monitor Config: sys_buffer_bytes cannot be 0"
         );
 
-        // *NOTE*: the dispatcher and cgroup manager talk through these channels
-        // so make sure they each get the correct half, nothing is droppped, etc.
-        let (notified_send, notified_recv) = mpsc::channel(1);
-        let (requesting_send, requesting_recv) = mpsc::channel(1);
-
-        let dispatcher = Dispatcher::new(ws, notified_send, requesting_recv)
+        let dispatcher = Dispatcher::new(ws)
             .await
             .context("error creating new dispatcher")?;
 
@@ -106,46 +147,10 @@ impl Runner {
             kill,
         };
 
-        // If we have both the cgroup and file cache integrations enabled, it's possible for
-        // temporary failures to result in cgroup throttling (from memory.high), that in turn makes
-        // it near-impossible to connect to the file cache (because it times out). Unfortunately,
-        // we *do* still want to determine the file cache size before setting the cgroup's
-        // memory.high, so it's not as simple as just swapping the order.
-        //
-        // Instead, the resolution here is that on vm-monitor startup (note: happens on each
-        // connection from autoscaler-agent, possibly multiple times per compute_ctl lifecycle), we
-        // temporarily unset memory.high, to allow any existing throttling to dissipate. It's a bit
-        // of a hacky solution, but helps with reliability.
-        if let Some(name) = &args.cgroup {
-            // Best not to set up cgroup stuff more than once, so we'll initialize cgroup state
-            // now, and then set limits later.
-            info!("initializing cgroup");
-
-            let (cgroup, cgroup_event_stream) = CgroupWatcher::new(name.clone(), requesting_send)
-                .context("failed to create cgroup manager")?;
-
-            info!("temporarily unsetting memory.high");
-
-            // Temporarily un-set cgroup memory.high; see above.
-            cgroup
-                .unset_memory_high()
-                .context("failed to unset memory.high")?;
-
-            let cgroup = Arc::new(cgroup);
-
-            let cgroup_clone = Arc::clone(&cgroup);
-            spawn_with_cancel(
-                token.clone(),
-                |_| error!("cgroup watcher terminated"),
-                async move { cgroup_clone.watch(notified_recv, cgroup_event_stream).await },
-            );
-
-            state.cgroup = Some(cgroup);
-        }
-
-        let mut file_cache_reserved_bytes = 0;
         let mem = get_total_system_memory();
 
+        let mut file_cache_disk_size = 0;
+
         // We need to process file cache initialization before cgroup initialization, so that the memory
         // allocated to the file cache is appropriately taken into account when we decide the cgroup's
         // memory limits.
@@ -156,7 +161,7 @@ impl Runner {
                 false => FileCacheConfig::default_in_memory(),
             };
 
-            let mut file_cache = FileCacheState::new(connstr, config, token)
+            let mut file_cache = FileCacheState::new(connstr, config, token.clone())
                 .await
                 .context("failed to create file cache")?;
 
@@ -181,23 +186,40 @@ impl Runner {
             if actual_size != new_size {
                 info!("file cache size actually got set to {actual_size}")
             }
-            // Mark the resources given to the file cache as reserved, but only if it's in memory.
-            if !args.file_cache_on_disk {
-                file_cache_reserved_bytes = actual_size;
+
+            if args.file_cache_on_disk {
+                file_cache_disk_size = actual_size;
             }
 
             state.filecache = Some(file_cache);
         }
 
-        if let Some(cgroup) = &state.cgroup {
-            let available = mem - file_cache_reserved_bytes;
-            let value = cgroup.config.calculate_memory_high_value(available);
+        if let Some(name) = &args.cgroup {
+            // Best not to set up cgroup stuff more than once, so we'll initialize cgroup state
+            // now, and then set limits later.
+            info!("initializing cgroup");
 
-            info!(value, "setting memory.high");
+            let cgroup =
+                CgroupWatcher::new(name.clone()).context("failed to create cgroup manager")?;
 
-            cgroup
-                .set_memory_high_bytes(value)
-                .context("failed to set cgroup memory.high")?;
+            let init_value = cgroup::MemoryHistory {
+                avg_non_reclaimable: 0,
+                samples_count: 0,
+                samples_span: Duration::ZERO,
+            };
+            let (hist_tx, hist_rx) = watch::channel((Instant::now(), init_value));
+
+            spawn_with_cancel(token, |_| error!("cgroup watcher terminated"), async move {
+                cgroup.watch(hist_tx).await
+            });
+
+            let threshold = state.config.cgroup_threshold(mem, file_cache_disk_size);
+            info!(threshold, "set initial cgroup threshold",);
+
+            state.cgroup = Some(CgroupState {
+                watcher: hist_rx,
+                threshold,
+            });
         }
 
         Ok(state)
@@ -217,28 +239,51 @@ impl Runner {
 
         let requested_mem = target.mem;
         let usable_system_memory = requested_mem.saturating_sub(self.config.sys_buffer_bytes);
-        let expected_file_cache_mem_usage = self
+        let (expected_file_cache_size, expected_file_cache_disk_size) = self
             .filecache
             .as_ref()
-            .map(|file_cache| file_cache.config.calculate_cache_size(usable_system_memory))
-            .unwrap_or(0);
-        let mut new_cgroup_mem_high = 0;
+            .map(|file_cache| {
+                let size = file_cache.config.calculate_cache_size(usable_system_memory);
+                match file_cache.config.in_memory {
+                    true => (size, 0),
+                    false => (size, size),
+                }
+            })
+            .unwrap_or((0, 0));
         if let Some(cgroup) = &self.cgroup {
-            new_cgroup_mem_high = cgroup
+            let (last_time, last_history) = *cgroup.watcher.borrow();
+
+            // NB: The ordering of these conditions is intentional. During startup, we should deny
+            // downscaling until we have enough information to determine that it's safe to do so
+            // (i.e. enough samples have come in). But if it's been a while and we *still* haven't
+            // received any information, we should *fail* instead of just denying downscaling.
+            //
+            // `last_time` is set to `Instant::now()` on startup, so checking `last_time.elapsed()`
+            // serves double-duty: it trips if we haven't received *any* metrics for long enough,
+            // OR if we haven't received metrics *recently enough*.
+            //
+            // TODO: make the duration here configurable.
+            if last_time.elapsed() > Duration::from_secs(5) {
+                bail!("haven't gotten cgroup memory stats recently enough to determine downscaling information");
+            } else if last_history.samples_count <= 1 {
+                let status = "haven't received enough cgroup memory stats yet";
+                info!(status, "discontinuing downscale");
+                return Ok((false, status.to_owned()));
+            }
+
+            let new_threshold = self
                 .config
-                .calculate_memory_high_value(usable_system_memory - expected_file_cache_mem_usage);
+                .cgroup_threshold(usable_system_memory, expected_file_cache_disk_size);
 
-            let current = cgroup
-                .current_memory_usage()
-                .context("failed to fetch cgroup memory")?;
+            let current = last_history.avg_non_reclaimable;
 
-            if new_cgroup_mem_high < current + cgroup.config.memory_high_buffer_bytes {
+            if new_threshold < current + self.config.cgroup_downscale_threshold_buffer_bytes {
                 let status = format!(
-                    "{}: {} MiB (new high) < {} (current usage) + {} (buffer)",
-                    "calculated memory.high too low",
-                    bytes_to_mebibytes(new_cgroup_mem_high),
+                    "{}: {} MiB (new threshold) < {} (current usage) + {} (downscale buffer)",
+                    "calculated memory threshold too low",
+                    bytes_to_mebibytes(new_threshold),
                     bytes_to_mebibytes(current),
-                    bytes_to_mebibytes(cgroup.config.memory_high_buffer_bytes)
+                    bytes_to_mebibytes(self.config.cgroup_downscale_threshold_buffer_bytes)
                 );
 
                 info!(status, "discontinuing downscale");
@@ -249,14 +294,14 @@ impl Runner {
 
         // The downscaling has been approved. Downscale the file cache, then the cgroup.
         let mut status = vec![];
-        let mut file_cache_mem_usage = 0;
+        let mut file_cache_disk_size = 0;
         if let Some(file_cache) = &mut self.filecache {
             let actual_usage = file_cache
-                .set_file_cache_size(expected_file_cache_mem_usage)
+                .set_file_cache_size(expected_file_cache_size)
                 .await
                 .context("failed to set file cache size")?;
-            if file_cache.config.in_memory {
-                file_cache_mem_usage = actual_usage;
+            if !file_cache.config.in_memory {
+                file_cache_disk_size = actual_usage;
             }
             let message = format!(
                 "set file cache size to {} MiB (in memory = {})",
@@ -267,24 +312,18 @@ impl Runner {
             status.push(message);
         }
 
-        if let Some(cgroup) = &self.cgroup {
-            let available_memory = usable_system_memory - file_cache_mem_usage;
-
-            if file_cache_mem_usage != expected_file_cache_mem_usage {
-                new_cgroup_mem_high = cgroup.config.calculate_memory_high_value(available_memory);
-            }
-
-            // new_cgroup_mem_high is initialized to 0 but it is guaranteed to not be here
-            // since it is properly initialized in the previous cgroup if let block
-            cgroup
-                .set_memory_high_bytes(new_cgroup_mem_high)
-                .context("failed to set cgroup memory.high")?;
+        if let Some(cgroup) = &mut self.cgroup {
+            let new_threshold = self
+                .config
+                .cgroup_threshold(usable_system_memory, file_cache_disk_size);
 
             let message = format!(
-                "set cgroup memory.high to {} MiB, of new max {} MiB",
-                bytes_to_mebibytes(new_cgroup_mem_high),
-                bytes_to_mebibytes(available_memory)
+                "set cgroup memory threshold from {} MiB to {} MiB, of new total {} MiB",
+                bytes_to_mebibytes(cgroup.threshold),
+                bytes_to_mebibytes(new_threshold),
+                bytes_to_mebibytes(usable_system_memory)
             );
+            cgroup.threshold = new_threshold;
             info!("downscale: {message}");
             status.push(message);
         }
@@ -305,8 +344,7 @@ impl Runner {
         let new_mem = resources.mem;
         let usable_system_memory = new_mem.saturating_sub(self.config.sys_buffer_bytes);
 
-        // Get the file cache's expected contribution to the memory usage
-        let mut file_cache_mem_usage = 0;
+        let mut file_cache_disk_size = 0;
         if let Some(file_cache) = &mut self.filecache {
             let expected_usage = file_cache.config.calculate_cache_size(usable_system_memory);
             info!(
@@ -319,8 +357,8 @@ impl Runner {
                 .set_file_cache_size(expected_usage)
                 .await
                 .context("failed to set file cache size")?;
-            if file_cache.config.in_memory {
-                file_cache_mem_usage = actual_usage;
+            if !file_cache.config.in_memory {
+                file_cache_disk_size = actual_usage;
             }
 
             if actual_usage != expected_usage {
@@ -332,18 +370,18 @@ impl Runner {
             }
         }
 
-        if let Some(cgroup) = &self.cgroup {
-            let available_memory = usable_system_memory - file_cache_mem_usage;
-            let new_cgroup_mem_high = cgroup.config.calculate_memory_high_value(available_memory);
+        if let Some(cgroup) = &mut self.cgroup {
+            let new_threshold = self
+                .config
+                .cgroup_threshold(usable_system_memory, file_cache_disk_size);
+
             info!(
-                target = bytes_to_mebibytes(new_cgroup_mem_high),
-                total = bytes_to_mebibytes(new_mem),
-                name = cgroup.path(),
-                "updating cgroup memory.high",
+                "set cgroup memory threshold from {} MiB to {} MiB of new total {} MiB",
+                bytes_to_mebibytes(cgroup.threshold),
+                bytes_to_mebibytes(new_threshold),
+                bytes_to_mebibytes(usable_system_memory)
             );
-            cgroup
-                .set_memory_high_bytes(new_cgroup_mem_high)
-                .context("failed to set cgroup memory.high")?;
+            cgroup.threshold = new_threshold;
         }
 
         Ok(())
@@ -361,10 +399,6 @@ impl Runner {
                 self.handle_upscale(granted)
                     .await
                     .context("failed to handle upscale")?;
-                self.dispatcher
-                    .notify_upscale(Sequenced::new(granted))
-                    .await
-                    .context("failed to notify notify cgroup of upscale")?;
                 Ok(Some(OutboundMsg::new(
                     OutboundMsgKind::UpscaleConfirmation {},
                     id,
@@ -408,33 +442,53 @@ impl Runner {
                         Err(e) => bail!("failed to receive kill signal: {e}")
                     }
                 }
-                // we need to propagate an upscale request
-                request = self.dispatcher.request_upscale_events.recv(), if self.cgroup.is_some() => {
-                    if request.is_none() {
-                        bail!("failed to listen for upscale event from cgroup")
+
+                // New memory stats from the cgroup, *may* need to request upscaling, if we've
+                // exceeded the threshold
+                result = self.cgroup.as_mut().unwrap().watcher.changed(), if self.cgroup.is_some() => {
+                    result.context("failed to receive from cgroup memory stats watcher")?;
+
+                    let cgroup = self.cgroup.as_ref().unwrap();
+
+                    let (_time, cgroup_mem_stat) = *cgroup.watcher.borrow();
+
+                    // If we haven't exceeded the threshold, then we're all ok
+                    if cgroup_mem_stat.avg_non_reclaimable < cgroup.threshold {
+                        continue;
                     }
 
-                    // If it's been less than 1 second since the last time we requested upscaling,
-                    // ignore the event, to avoid spamming the agent (otherwise, this can happen
-                    // ~1k times per second).
+                    // Otherwise, we generally want upscaling. But, if it's been less than 1 second
+                    // since the last time we requested upscaling, ignore the event, to avoid
+                    // spamming the agent.
                     if let Some(t) = self.last_upscale_request_at {
                         let elapsed = t.elapsed();
                         if elapsed < Duration::from_secs(1) {
-                            info!(elapsed_millis = elapsed.as_millis(), "cgroup asked for upscale but too soon to forward the request, ignoring");
+                            info!(
+                                elapsed_millis = elapsed.as_millis(),
+                                avg_non_reclaimable = bytes_to_mebibytes(cgroup_mem_stat.avg_non_reclaimable),
+                                threshold = bytes_to_mebibytes(cgroup.threshold),
+                                "cgroup memory stats are high enough to upscale but too soon to forward the request, ignoring",
+                            );
                             continue;
                         }
                     }
 
                     self.last_upscale_request_at = Some(Instant::now());
 
-                    info!("cgroup asking for upscale; forwarding request");
+                    info!(
+                        avg_non_reclaimable = bytes_to_mebibytes(cgroup_mem_stat.avg_non_reclaimable),
+                        threshold = bytes_to_mebibytes(cgroup.threshold),
+                        "cgroup memory stats are high enough to upscale, requesting upscale",
+                    );
+
                     self.counter += 2; // Increment, preserving parity (i.e. keep the
                                        // counter odd). See the field comment for more.
                     self.dispatcher
                         .send(OutboundMsg::new(OutboundMsgKind::UpscaleRequest {}, self.counter))
                         .await
                         .context("failed to send message")?;
-                }
+                },
+
                 // there is a message from the agent
                 msg = self.dispatcher.source.next() => {
                     if let Some(msg) = msg {
@@ -462,11 +516,14 @@ impl Runner {
                                     Ok(Some(out)) => out,
                                     Ok(None) => continue,
                                     Err(e) => {
-                                        let error = e.to_string();
-                                        warn!(?error, "error handling message");
+                                        // use {:#} for our logging because the display impl only
+                                        // gives the outermost cause, and the debug impl
+                                        // pretty-prints the error, whereas {:#} contains all the
+                                        // causes, but is compact (no newlines).
+                                        warn!(error = format!("{e:#}"), "error handling message");
                                         OutboundMsg::new(
                                             OutboundMsgKind::InternalError {
-                                                error
+                                                error: e.to_string(),
                                             },
                                             message.id
                                         )

libs/walproposer/Cargo.toml

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

libs/walproposer/bindgen_deps.h

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

libs/walproposer/build.rs

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

libs/walproposer/src/api_bindings.rs

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

libs/walproposer/src/lib.rs

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

libs/walproposer/src/walproposer.rs

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

pageserver/benches/bench_walredo.rs

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

pageserver/src/basebackup.rs

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

pageserver/src/bin/pageserver.rs

@@ -2,6 +2,7 @@
 
 use std::env::{var, VarError};
 use std::sync::Arc;
+use std::time::Duration;
 use std::{env, ops::ControlFlow, str::FromStr};
 
 use anyhow::{anyhow, Context};
@@ -200,6 +201,51 @@ fn initialize_config(
     })
 }
 
+struct WaitForPhaseResult<F: std::future::Future + Unpin> {
+    timeout_remaining: Duration,
+    skipped: Option<F>,
+}
+
+/// During startup, we apply a timeout to our waits for readiness, to avoid
+/// stalling the whole service if one Tenant experiences some problem.  Each
+/// phase may consume some of the timeout: this function returns the updated
+/// timeout for use in the next call.
+async fn wait_for_phase<F>(phase: &str, mut fut: F, timeout: Duration) -> WaitForPhaseResult<F>
+where
+    F: std::future::Future + Unpin,
+{
+    let initial_t = Instant::now();
+    let skipped = match tokio::time::timeout(timeout, &mut fut).await {
+        Ok(_) => None,
+        Err(_) => {
+            tracing::info!(
+                timeout_millis = timeout.as_millis(),
+                %phase,
+                "Startup phase timed out, proceeding anyway"
+            );
+            Some(fut)
+        }
+    };
+
+    WaitForPhaseResult {
+        timeout_remaining: timeout
+            .checked_sub(Instant::now().duration_since(initial_t))
+            .unwrap_or(Duration::ZERO),
+        skipped,
+    }
+}
+
+fn startup_checkpoint(started_at: Instant, phase: &str, human_phase: &str) {
+    let elapsed = started_at.elapsed();
+    let secs = elapsed.as_secs_f64();
+    STARTUP_DURATION.with_label_values(&[phase]).set(secs);
+
+    info!(
+        elapsed_ms = elapsed.as_millis(),
+        "{human_phase} ({secs:.3}s since start)"
+    )
+}
+
 fn start_pageserver(
     launch_ts: &'static LaunchTimestamp,
     conf: &'static PageServerConf,
@@ -207,16 +253,6 @@ fn start_pageserver(
     // Monotonic time for later calculating startup duration
     let started_startup_at = Instant::now();
 
-    let startup_checkpoint = move |phase: &str, human_phase: &str| {
-        let elapsed = started_startup_at.elapsed();
-        let secs = elapsed.as_secs_f64();
-        STARTUP_DURATION.with_label_values(&[phase]).set(secs);
-        info!(
-            elapsed_ms = elapsed.as_millis(),
-            "{human_phase} ({secs:.3}s since start)"
-        )
-    };
-
     // Print version and launch timestamp to the log,
     // and expose them as prometheus metrics.
     // A changed version string indicates changed software.
@@ -341,7 +377,7 @@ fn start_pageserver(
 
     // Up to this point no significant I/O has been done: this should have been fast.  Record
     // duration prior to starting I/O intensive phase of startup.
-    startup_checkpoint("initial", "Starting loading tenants");
+    startup_checkpoint(started_startup_at, "initial", "Starting loading tenants");
     STARTUP_IS_LOADING.set(1);
 
     // Startup staging or optimizing:
@@ -355,6 +391,7 @@ fn start_pageserver(
     // consumer side) will be dropped once we can start the background jobs. Currently it is behind
     // completing all initial logical size calculations (init_logical_size_done_rx) and a timeout
     // (background_task_maximum_delay).
+    let (init_remote_done_tx, init_remote_done_rx) = utils::completion::channel();
     let (init_done_tx, init_done_rx) = utils::completion::channel();
 
     let (init_logical_size_done_tx, init_logical_size_done_rx) = utils::completion::channel();
@@ -362,7 +399,8 @@ fn start_pageserver(
     let (background_jobs_can_start, background_jobs_barrier) = utils::completion::channel();
 
     let order = pageserver::InitializationOrder {
-        initial_tenant_load: Some(init_done_tx),
+        initial_tenant_load_remote: Some(init_done_tx),
+        initial_tenant_load: Some(init_remote_done_tx),
         initial_logical_size_can_start: init_done_rx.clone(),
         initial_logical_size_attempt: Some(init_logical_size_done_tx),
         background_jobs_can_start: background_jobs_barrier.clone(),
@@ -386,55 +424,93 @@ fn start_pageserver(
         let shutdown_pageserver = shutdown_pageserver.clone();
         let drive_init = async move {
             // NOTE: unlike many futures in pageserver, this one is cancellation-safe
-            let guard = scopeguard::guard_on_success((), |_| tracing::info!("Cancelled before initial load completed"));
+            let guard = scopeguard::guard_on_success((), |_| {
+                tracing::info!("Cancelled before initial load completed")
+            });
 
-            init_done_rx.wait().await;
-            startup_checkpoint("initial_tenant_load", "Initial load completed");
-            STARTUP_IS_LOADING.set(0);
+            let timeout = conf.background_task_maximum_delay;
+
+            let init_remote_done = std::pin::pin!(async {
+                init_remote_done_rx.wait().await;
+                startup_checkpoint(
+                    started_startup_at,
+                    "initial_tenant_load_remote",
+                    "Remote part of initial load completed",
+                );
+            });
+
+            let WaitForPhaseResult {
+                timeout_remaining: timeout,
+                skipped: init_remote_skipped,
+            } = wait_for_phase("initial_tenant_load_remote", init_remote_done, timeout).await;
+
+            let init_load_done = std::pin::pin!(async {
+                init_done_rx.wait().await;
+                startup_checkpoint(
+                    started_startup_at,
+                    "initial_tenant_load",
+                    "Initial load completed",
+                );
+                STARTUP_IS_LOADING.set(0);
+            });
+
+            let WaitForPhaseResult {
+                timeout_remaining: timeout,
+                skipped: init_load_skipped,
+            } = wait_for_phase("initial_tenant_load", init_load_done, timeout).await;
 
             // initial logical sizes can now start, as they were waiting on init_done_rx.
 
             scopeguard::ScopeGuard::into_inner(guard);
 
-            let mut init_sizes_done = std::pin::pin!(init_logical_size_done_rx.wait());
+            let guard = scopeguard::guard_on_success((), |_| {
+                tracing::info!("Cancelled before initial logical sizes completed")
+            });
 
-            let timeout = conf.background_task_maximum_delay;
+            let logical_sizes_done = std::pin::pin!(async {
+                init_logical_size_done_rx.wait().await;
+                startup_checkpoint(
+                    started_startup_at,
+                    "initial_logical_sizes",
+                    "Initial logical sizes completed",
+                );
+            });
 
-            let guard = scopeguard::guard_on_success((), |_| tracing::info!("Cancelled before initial logical sizes completed"));
-
-            let init_sizes_done = match tokio::time::timeout(timeout, &mut init_sizes_done).await {
-                Ok(_) => {
-                    startup_checkpoint("initial_logical_sizes", "Initial logical sizes completed");
-                    None
-                }
-                Err(_) => {
-                    tracing::info!(
-                        timeout_millis = timeout.as_millis(),
-                        "Initial logical size timeout elapsed; starting background jobs"
-                    );
-                    Some(init_sizes_done)
-                }
-            };
+            let WaitForPhaseResult {
+                timeout_remaining: _,
+                skipped: logical_sizes_skipped,
+            } = wait_for_phase("initial_logical_sizes", logical_sizes_done, timeout).await;
 
             scopeguard::ScopeGuard::into_inner(guard);
 
-            // allow background jobs to start
+            // allow background jobs to start: we either completed prior stages, or they reached timeout
+            // and were skipped.  It is important that we do not let them block background jobs indefinitely,
+            // because things like consumption metrics for billing are blocked by this barrier.
             drop(background_jobs_can_start);
-            startup_checkpoint("background_jobs_can_start", "Starting background jobs");
-
-            if let Some(init_sizes_done) = init_sizes_done {
-                // ending up here is not a bug; at the latest logical sizes will be queried by
-                // consumption metrics.
-                let guard = scopeguard::guard_on_success((), |_| tracing::info!("Cancelled before initial logical sizes completed"));
-                init_sizes_done.await;
-
-                scopeguard::ScopeGuard::into_inner(guard);
-
-                startup_checkpoint("initial_logical_sizes", "Initial logical sizes completed after timeout (background jobs already started)");
+            startup_checkpoint(
+                started_startup_at,
+                "background_jobs_can_start",
+                "Starting background jobs",
+            );
 
+            // We are done. If we skipped any phases due to timeout, run them to completion here so that
+            // they will eventually update their startup_checkpoint, and so that we do not declare the
+            // 'complete' stage until all the other stages are really done.
+            let guard = scopeguard::guard_on_success((), |_| {
+                tracing::info!("Cancelled before waiting for skipped phases done")
+            });
+            if let Some(f) = init_remote_skipped {
+                f.await;
             }
+            if let Some(f) = init_load_skipped {
+                f.await;
+            }
+            if let Some(f) = logical_sizes_skipped {
+                f.await;
+            }
+            scopeguard::ScopeGuard::into_inner(guard);
 
-            startup_checkpoint("complete", "Startup complete");
+            startup_checkpoint(started_startup_at, "complete", "Startup complete");
         };
 
         async move {
@@ -574,6 +650,7 @@ fn start_pageserver(
                     pageserver_listener,
                     conf.pg_auth_type,
                     libpq_ctx,
+                    task_mgr::shutdown_token(),
                 )
                 .await
             },

pageserver/src/config.rs

@@ -1479,6 +1479,8 @@ threshold = "20m"
             Some(DiskUsageEvictionTaskConfig {
                 max_usage_pct: Percent::new(80).unwrap(),
                 min_avail_bytes: 0,
+                target_avail_bytes: None,
+                target_usage_pct: None,
                 period: Duration::from_secs(10),
                 #[cfg(feature = "testing")]
                 mock_statvfs: None,

pageserver/src/consumption_metrics.rs

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

pageserver/src/deletion_queue.rs

@@ -1298,10 +1298,6 @@ pub(crate) mod mock {
             }
         }
 
-        pub fn get_executed(&self) -> usize {
-            self.executed.load(Ordering::Relaxed)
-        }
-
         #[allow(clippy::await_holding_lock)]
         pub async fn pump(&self) {
             if let Some(remote_storage) = &self.remote_storage {

pageserver/src/disk_usage_eviction_task.rs

@@ -67,16 +67,40 @@ use crate::{
 pub struct DiskUsageEvictionTaskConfig {
     pub max_usage_pct: Percent,
     pub min_avail_bytes: u64,
+
+    // Control how far we will go when evicting: when usage exceeds max_usage_pct or min_avail_bytes,
+    // we will keep evicting layers until we reach the target.  The resulting disk usage should look
+    // like a sawtooth bouncing between the upper max/min line and the lower target line.
+    #[serde(default)]
+    pub target_usage_pct: Option<Percent>,
+    #[serde(default)]
+    pub target_avail_bytes: Option<u64>,
+
     #[serde(with = "humantime_serde")]
     pub period: Duration,
     #[cfg(feature = "testing")]
     pub mock_statvfs: Option<crate::statvfs::mock::Behavior>,
 }
 
+#[derive(Default)]
+enum Status {
+    /// We are within disk limits, and not currently doing any eviction
+    #[default]
+    Idle,
+    /// Disk limits have been exceeded: we will evict soon
+    UnderPressure,
+    /// We are currently doing an eviction pass.
+    Evicting,
+}
+
 #[derive(Default)]
 pub struct State {
     /// Exclude http requests and background task from running at the same time.
     mutex: tokio::sync::Mutex<()>,
+
+    /// Publish the current status of eviction work, for visibility to other subsystems
+    /// that modify their behavior if disk pressure is high or if eviction is going on.
+    status: std::sync::RwLock<Status>,
 }
 
 pub fn launch_disk_usage_global_eviction_task(
@@ -176,7 +200,9 @@ async fn disk_usage_eviction_task(
 }
 
 pub trait Usage: Clone + Copy + std::fmt::Debug {
-    fn has_pressure(&self) -> bool;
+    fn pressure(&self) -> f64;
+    fn over_pressure(&self) -> bool;
+    fn no_pressure(&self) -> bool;
     fn add_available_bytes(&mut self, bytes: u64);
 }
 
@@ -189,13 +215,19 @@ async fn disk_usage_eviction_task_iteration(
 ) -> anyhow::Result<()> {
     let usage_pre = filesystem_level_usage::get(tenants_dir, task_config)
         .context("get filesystem-level disk usage before evictions")?;
+
+    if usage_pre.over_pressure() {
+        *state.status.write().unwrap() = Status::Evicting;
+    }
+
     let res = disk_usage_eviction_task_iteration_impl(state, storage, usage_pre, cancel).await;
     match res {
         Ok(outcome) => {
             debug!(?outcome, "disk_usage_eviction_iteration finished");
-            match outcome {
+            let new_status = match outcome {
                 IterationOutcome::NoPressure | IterationOutcome::Cancelled => {
                     // nothing to do, select statement below will handle things
+                    Status::Idle
                 }
                 IterationOutcome::Finished(outcome) => {
                     // Verify with statvfs whether we made any real progress
@@ -205,21 +237,30 @@ async fn disk_usage_eviction_task_iteration(
 
                     debug!(?after, "disk usage");
 
-                    if after.has_pressure() {
+                    if after.over_pressure() {
                         // Don't bother doing an out-of-order iteration here now.
                         // In practice, the task period is set to a value in the tens-of-seconds range,
                         // which will cause another iteration to happen soon enough.
                         // TODO: deltas between the three different usages would be helpful,
                         // consider MiB, GiB, TiB
                         warn!(?outcome, ?after, "disk usage still high");
+                        Status::UnderPressure
                     } else {
                         info!(?outcome, ?after, "disk usage pressure relieved");
+                        Status::Idle
                     }
                 }
-            }
+            };
+
+            *state.status.write().unwrap() = new_status;
         }
         Err(e) => {
             error!("disk_usage_eviction_iteration failed: {:#}", e);
+            *state.status.write().unwrap() = if usage_pre.over_pressure() {
+                Status::UnderPressure
+            } else {
+                Status::Idle
+            };
         }
     }
 
@@ -285,8 +326,10 @@ pub async fn disk_usage_eviction_task_iteration_impl<U: Usage>(
 
     debug!(?usage_pre, "disk usage");
 
-    if !usage_pre.has_pressure() {
+    if !usage_pre.over_pressure() {
         return Ok(IterationOutcome::NoPressure);
+    } else {
+        *state.status.write().unwrap() = Status::Evicting;
     }
 
     warn!(
@@ -334,7 +377,7 @@ pub async fn disk_usage_eviction_task_iteration_impl<U: Usage>(
     let mut warned = None;
     let mut usage_planned = usage_pre;
     for (i, (partition, candidate)) in candidates.into_iter().enumerate() {
-        if !usage_planned.has_pressure() {
+        if usage_planned.no_pressure() {
             debug!(
                 no_candidates_evicted = i,
                 "took enough candidates for pressure to be relieved"
@@ -411,6 +454,11 @@ pub async fn disk_usage_eviction_task_iteration_impl<U: Usage>(
                                 evictions_failed.file_sizes += file_size;
                                 evictions_failed.count += 1;
                             }
+                            Some(Err(EvictionError::MetadataInconsistency(detail))) => {
+                                warn!(%layer, "failed to evict layer: {detail}");
+                                evictions_failed.file_sizes += file_size;
+                                evictions_failed.count += 1;
+                            }
                             None => {
                                 assert!(cancel.is_cancelled());
                                 return;
@@ -639,22 +687,57 @@ mod filesystem_level_usage {
     }
 
     impl super::Usage for Usage<'_> {
-        fn has_pressure(&self) -> bool {
-            let usage_pct =
-                (100.0 * (1.0 - ((self.avail_bytes as f64) / (self.total_bytes as f64)))) as u64;
+        /// Does the pressure exceed 1.0, i.e. has the disk usage exceeded upper bounds?
+        ///
+        /// This is the condition for starting eviction.
+        fn over_pressure(&self) -> bool {
+            self.pressure() >= 1.0
+        }
 
-            let pressures = [
-                (
-                    "min_avail_bytes",
-                    self.avail_bytes < self.config.min_avail_bytes,
-                ),
-                (
-                    "max_usage_pct",
-                    usage_pct >= self.config.max_usage_pct.get() as u64,
-                ),
-            ];
+        /// Is the pressure <0, ie.. has disk usage gone below the target bound?
+        ///
+        /// This is the condition for dropping out of eviction.
+        fn no_pressure(&self) -> bool {
+            self.pressure() <= 0.0
+        }
 
-            pressures.into_iter().any(|(_, has_pressure)| has_pressure)
+        fn pressure(&self) -> f64 {
+            let max_usage = std::cmp::min(
+                self.total_bytes - self.config.min_avail_bytes,
+                (self.total_bytes as f64 * (self.config.max_usage_pct.get() as f64 / 100.0)) as u64,
+            );
+
+            let mut target_usage = max_usage;
+            if let Some(target_avail_bytes) = self.config.target_avail_bytes {
+                target_usage = std::cmp::min(target_usage, self.total_bytes - target_avail_bytes);
+            }
+            if let Some(target_usage_pct) = self.config.target_usage_pct {
+                target_usage = std::cmp::min(
+                    target_usage,
+                    (self.total_bytes as f64 * (target_usage_pct.get() as f64 / 100.0)) as u64,
+                );
+            };
+
+            let usage = self.total_bytes - self.avail_bytes;
+            eprintln!(
+                "pressure: {} {}, current {}",
+                target_usage, max_usage, usage
+            );
+            if target_usage == max_usage {
+                // We are configured with a zero sized range: treat anything at+beyond limit as pressure 1.0, else 0.0
+                if usage >= max_usage {
+                    1.0
+                } else {
+                    0.0
+                }
+            } else if usage <= target_usage {
+                // No pressure.
+                0.0
+            } else {
+                // We are above target: pressure is the ratio of how much we exceed target to the size of the gap
+                let range_size = (max_usage - target_usage) as f64;
+                (usage - target_usage) as f64 / range_size
+            }
         }
 
         fn add_available_bytes(&mut self, bytes: u64) {
@@ -708,6 +791,8 @@ mod filesystem_level_usage {
             config: &DiskUsageEvictionTaskConfig {
                 max_usage_pct: Percent::new(85).unwrap(),
                 min_avail_bytes: 0,
+                target_avail_bytes: None,
+                target_usage_pct: None,
                 period: Duration::MAX,
                 #[cfg(feature = "testing")]
                 mock_statvfs: None,
@@ -716,24 +801,24 @@ mod filesystem_level_usage {
             avail_bytes: 0,
         };
 
-        assert!(usage.has_pressure(), "expected pressure at 100%");
+        assert!(usage.over_pressure(), "expected pressure at 100%");
 
         usage.add_available_bytes(14_000);
-        assert!(usage.has_pressure(), "expected pressure at 86%");
+        assert!(usage.over_pressure(), "expected pressure at 86%");
 
         usage.add_available_bytes(999);
-        assert!(usage.has_pressure(), "expected pressure at 85.001%");
+        assert!(usage.over_pressure(), "expected pressure at 85.001%");
 
         usage.add_available_bytes(1);
-        assert!(usage.has_pressure(), "expected pressure at precisely 85%");
+        assert!(usage.over_pressure(), "expected pressure at precisely 85%");
 
         usage.add_available_bytes(1);
-        assert!(!usage.has_pressure(), "no pressure at 84.999%");
+        assert!(!usage.over_pressure(), "no pressure at 84.999%");
 
         usage.add_available_bytes(999);
-        assert!(!usage.has_pressure(), "no pressure at 84%");
+        assert!(!usage.over_pressure(), "no pressure at 84%");
 
         usage.add_available_bytes(16_000);
-        assert!(!usage.has_pressure());
+        assert!(!usage.over_pressure());
     }
 }

pageserver/src/http/openapi_spec.yml

@@ -306,6 +306,67 @@ paths:
               schema:
                 $ref: "#/components/schemas/ServiceUnavailableError"
 
+  /v1/tenant/{tenant_id}/timeline/{timeline_id}/get_timestamp_of_lsn:
+    parameters:
+      - name: tenant_id
+        in: path
+        required: true
+        schema:
+          type: string
+          format: hex
+      - name: timeline_id
+        in: path
+        required: true
+        schema:
+          type: string
+          format: hex
+    get:
+      description: Get timestamp for a given LSN
+      parameters:
+        - name: lsn
+          in: query
+          required: true
+          schema:
+            type: integer
+          description: A LSN to get the timestamp
+      responses:
+        "200":
+          description: OK
+          content:
+            application/json:
+              schema:
+                type: string
+                format: date-time
+        "400":
+          description: Error when no tenant id found in path, no timeline id or invalid timestamp
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/Error"
+        "401":
+          description: Unauthorized Error
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/UnauthorizedError"
+        "403":
+          description: Forbidden Error
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/ForbiddenError"
+        "404":
+          description: Timeline not found, or there is no timestamp information for the given lsn
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/NotFoundError"
+        "500":
+          description: Generic operation error
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/Error"
 
   /v1/tenant/{tenant_id}/timeline/{timeline_id}/get_lsn_by_timestamp:
     parameters:

pageserver/src/http/routes.rs

@@ -2,10 +2,12 @@
 //! Management HTTP API
 //!
 use std::collections::HashMap;
+use std::str::FromStr;
 use std::sync::Arc;
 
 use anyhow::{anyhow, Context, Result};
 use futures::TryFutureExt;
+use humantime::format_rfc3339;
 use hyper::header::CONTENT_TYPE;
 use hyper::StatusCode;
 use hyper::{Body, Request, Response, Uri};
@@ -136,9 +138,7 @@ impl From<PageReconstructError> for ApiError {
             PageReconstructError::AncestorStopping(_) => {
                 ApiError::ResourceUnavailable(format!("{pre}").into())
             }
-            PageReconstructError::WalRedo(pre) => {
-                ApiError::InternalServerError(anyhow::Error::new(pre))
-            }
+            PageReconstructError::WalRedo(pre) => ApiError::InternalServerError(pre),
         }
     }
 }
@@ -504,6 +504,33 @@ async fn get_lsn_by_timestamp_handler(
     json_response(StatusCode::OK, result)
 }
 
+async fn get_timestamp_of_lsn_handler(
+    request: Request<Body>,
+    _cancel: CancellationToken,
+) -> Result<Response<Body>, ApiError> {
+    let tenant_id: TenantId = parse_request_param(&request, "tenant_id")?;
+    check_permission(&request, Some(tenant_id))?;
+
+    let timeline_id: TimelineId = parse_request_param(&request, "timeline_id")?;
+
+    let lsn_str = must_get_query_param(&request, "lsn")?;
+    let lsn = Lsn::from_str(&lsn_str)
+        .with_context(|| format!("Invalid LSN: {lsn_str:?}"))
+        .map_err(ApiError::BadRequest)?;
+
+    let ctx = RequestContext::new(TaskKind::MgmtRequest, DownloadBehavior::Download);
+    let timeline = active_timeline_of_active_tenant(tenant_id, timeline_id).await?;
+    let result = timeline.get_timestamp_for_lsn(lsn, &ctx).await?;
+
+    match result {
+        Some(time) => {
+            let time = format_rfc3339(postgres_ffi::from_pg_timestamp(time)).to_string();
+            json_response(StatusCode::OK, time)
+        }
+        None => json_response(StatusCode::NOT_FOUND, ()),
+    }
+}
+
 async fn tenant_attach_handler(
     mut request: Request<Body>,
     _cancel: CancellationToken,
@@ -1036,9 +1063,17 @@ async fn put_tenant_location_config_handler(
     // The `Detached` state is special, it doesn't upsert a tenant, it removes
     // its local disk content and drops it from memory.
     if let LocationConfigMode::Detached = request_data.config.mode {
-        mgr::detach_tenant(conf, tenant_id, true, &state.deletion_queue_client)
+        if let Err(e) = mgr::detach_tenant(conf, tenant_id, true, &state.deletion_queue_client)
             .instrument(info_span!("tenant_detach", %tenant_id))
-            .await?;
+            .await
+        {
+            match e {
+                TenantStateError::NotFound(_) => {
+                    // This API is idempotent: a NotFound on a detach is fine.
+                }
+                _ => return Err(e.into()),
+            }
+        }
         return json_response(StatusCode::OK, ());
     }
 
@@ -1417,10 +1452,22 @@ async fn disk_usage_eviction_run(
     }
 
     impl crate::disk_usage_eviction_task::Usage for Usage {
-        fn has_pressure(&self) -> bool {
+        fn over_pressure(&self) -> bool {
             self.config.evict_bytes > self.freed_bytes
         }
 
+        fn no_pressure(&self) -> bool {
+            !self.over_pressure()
+        }
+
+        fn pressure(&self) -> f64 {
+            if self.over_pressure() {
+                1.0
+            } else {
+                0.0
+            }
+        }
+
         fn add_available_bytes(&mut self, bytes: u64) {
             self.freed_bytes += bytes;
         }
@@ -1674,6 +1721,10 @@ pub fn make_router(
             "/v1/tenant/:tenant_id/timeline/:timeline_id/get_lsn_by_timestamp",
             |r| api_handler(r, get_lsn_by_timestamp_handler),
         )
+        .get(
+            "/v1/tenant/:tenant_id/timeline/:timeline_id/get_timestamp_of_lsn",
+            |r| api_handler(r, get_timestamp_of_lsn_handler),
+        )
         .put("/v1/tenant/:tenant_id/timeline/:timeline_id/do_gc", |r| {
             api_handler(r, timeline_gc_handler)
         })

pageserver/src/lib.rs

@@ -173,6 +173,9 @@ fn is_walkdir_io_not_found(e: &walkdir::Error) -> bool {
 /// delaying is needed.
 #[derive(Clone)]
 pub struct InitializationOrder {
+    /// Each initial tenant load task carries this until it is done loading timelines from remote storage
+    pub initial_tenant_load_remote: Option<utils::completion::Completion>,
+
     /// Each initial tenant load task carries this until completion.
     pub initial_tenant_load: Option<utils::completion::Completion>,
 

pageserver/src/metrics.rs

@@ -1067,6 +1067,26 @@ pub(crate) static TENANT_TASK_EVENTS: Lazy<IntCounterVec> = Lazy::new(|| {
     .expect("Failed to register tenant_task_events metric")
 });
 
+pub(crate) static BACKGROUND_LOOP_SEMAPHORE_WAIT_START_COUNT: Lazy<IntCounterVec> =
+    Lazy::new(|| {
+        register_int_counter_vec!(
+            "pageserver_background_loop_semaphore_wait_start_count",
+            "Counter for background loop concurrency-limiting semaphore acquire calls started",
+            &["task"],
+        )
+        .unwrap()
+    });
+
+pub(crate) static BACKGROUND_LOOP_SEMAPHORE_WAIT_FINISH_COUNT: Lazy<IntCounterVec> =
+    Lazy::new(|| {
+        register_int_counter_vec!(
+            "pageserver_background_loop_semaphore_wait_finish_count",
+            "Counter for background loop concurrency-limiting semaphore acquire calls finished",
+            &["task"],
+        )
+        .unwrap()
+    });
+
 pub(crate) static BACKGROUND_LOOP_PERIOD_OVERRUN_COUNT: Lazy<IntCounterVec> = Lazy::new(|| {
     register_int_counter_vec!(
         "pageserver_background_loop_period_overrun_count",

pageserver/src/page_cache.rs

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

pageserver/src/page_service.rs

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

pageserver/src/pgdatadir_mapping.rs

@@ -19,6 +19,7 @@ use postgres_ffi::BLCKSZ;
 use postgres_ffi::{Oid, TimestampTz, TransactionId};
 use serde::{Deserialize, Serialize};
 use std::collections::{hash_map, HashMap, HashSet};
+use std::ops::ControlFlow;
 use std::ops::Range;
 use tokio_util::sync::CancellationToken;
 use tracing::{debug, trace, warn};
@@ -370,7 +371,6 @@ impl Timeline {
         }
     }
 
-    ///
     /// Subroutine of find_lsn_for_timestamp(). Returns true, if there are any
     /// commits that committed after 'search_timestamp', at LSN 'probe_lsn'.
     ///
@@ -385,6 +385,50 @@ impl Timeline {
         found_larger: &mut bool,
         ctx: &RequestContext,
     ) -> Result<bool, PageReconstructError> {
+        self.map_all_timestamps(probe_lsn, ctx, |timestamp| {
+            if timestamp >= search_timestamp {
+                *found_larger = true;
+                return ControlFlow::Break(true);
+            } else {
+                *found_smaller = true;
+            }
+            ControlFlow::Continue(())
+        })
+        .await
+    }
+
+    /// Obtain the possible timestamp range for the given lsn.
+    ///
+    /// If the lsn has no timestamps, returns None. returns `(min, max, median)` if it has timestamps.
+    pub async fn get_timestamp_for_lsn(
+        &self,
+        probe_lsn: Lsn,
+        ctx: &RequestContext,
+    ) -> Result<Option<TimestampTz>, PageReconstructError> {
+        let mut max: Option<TimestampTz> = None;
+        self.map_all_timestamps(probe_lsn, ctx, |timestamp| {
+            if let Some(max_prev) = max {
+                max = Some(max_prev.max(timestamp));
+            } else {
+                max = Some(timestamp);
+            }
+            ControlFlow::Continue(())
+        })
+        .await?;
+
+        Ok(max)
+    }
+
+    /// Runs the given function on all the timestamps for a given lsn
+    ///
+    /// The return value is either given by the closure, or set to the `Default`
+    /// impl's output.
+    async fn map_all_timestamps<T: Default>(
+        &self,
+        probe_lsn: Lsn,
+        ctx: &RequestContext,
+        mut f: impl FnMut(TimestampTz) -> ControlFlow<T>,
+    ) -> Result<T, PageReconstructError> {
         for segno in self
             .list_slru_segments(SlruKind::Clog, probe_lsn, ctx)
             .await?
@@ -402,16 +446,14 @@ impl Timeline {
                     timestamp_bytes.copy_from_slice(&clog_page[BLCKSZ as usize..]);
                     let timestamp = TimestampTz::from_be_bytes(timestamp_bytes);
 
-                    if timestamp >= search_timestamp {
-                        *found_larger = true;
-                        return Ok(true);
-                    } else {
-                        *found_smaller = true;
+                    match f(timestamp) {
+                        ControlFlow::Break(b) => return Ok(b),
+                        ControlFlow::Continue(()) => (),
                     }
                 }
             }
         }
-        Ok(false)
+        Ok(Default::default())
     }
 
     /// Get a list of SLRU segments
@@ -499,6 +541,23 @@ impl Timeline {
         self.get(CHECKPOINT_KEY, lsn, ctx).await
     }
 
+    pub async fn list_aux_files(
+        &self,
+        lsn: Lsn,
+        ctx: &RequestContext,
+    ) -> Result<HashMap<String, Bytes>, PageReconstructError> {
+        match self.get(AUX_FILES_KEY, lsn, ctx).await {
+            Ok(buf) => match AuxFilesDirectory::des(&buf).context("deserialization failure") {
+                Ok(dir) => Ok(dir.files),
+                Err(e) => Err(PageReconstructError::from(e)),
+            },
+            Err(e) => {
+                warn!("Failed to get info about AUX files: {}", e);
+                Ok(HashMap::new())
+            }
+        }
+    }
+
     /// Does the same as get_current_logical_size but counted on demand.
     /// Used to initialize the logical size tracking on startup.
     ///
@@ -616,6 +675,7 @@ impl Timeline {
 
         result.add_key(CONTROLFILE_KEY);
         result.add_key(CHECKPOINT_KEY);
+        result.add_key(AUX_FILES_KEY);
 
         Ok(result.to_keyspace())
     }
@@ -692,6 +752,12 @@ impl<'a> DatadirModification<'a> {
         })?;
         self.put(DBDIR_KEY, Value::Image(buf.into()));
 
+        // Create AuxFilesDirectory
+        let buf = AuxFilesDirectory::ser(&AuxFilesDirectory {
+            files: HashMap::new(),
+        })?;
+        self.put(AUX_FILES_KEY, Value::Image(Bytes::from(buf)));
+
         let buf = TwoPhaseDirectory::ser(&TwoPhaseDirectory {
             xids: HashSet::new(),
         })?;
@@ -796,6 +862,12 @@ impl<'a> DatadirModification<'a> {
             // 'true', now write the updated 'dbdirs' map back.
             let buf = DbDirectory::ser(&dbdir)?;
             self.put(DBDIR_KEY, Value::Image(buf.into()));
+
+            // Create AuxFilesDirectory as well
+            let buf = AuxFilesDirectory::ser(&AuxFilesDirectory {
+                files: HashMap::new(),
+            })?;
+            self.put(AUX_FILES_KEY, Value::Image(Bytes::from(buf)));
         }
         if r.is_none() {
             // Create RelDirectory
@@ -1120,6 +1192,36 @@ impl<'a> DatadirModification<'a> {
         Ok(())
     }
 
+    pub async fn put_file(
+        &mut self,
+        path: &str,
+        content: &[u8],
+        ctx: &RequestContext,
+    ) -> anyhow::Result<()> {
+        let mut dir = match self.get(AUX_FILES_KEY, ctx).await {
+            Ok(buf) => AuxFilesDirectory::des(&buf)?,
+            Err(e) => {
+                warn!("Failed to get info about AUX files: {}", e);
+                AuxFilesDirectory {
+                    files: HashMap::new(),
+                }
+            }
+        };
+        let path = path.to_string();
+        if content.is_empty() {
+            dir.files.remove(&path);
+        } else {
+            dir.files.insert(path, Bytes::copy_from_slice(content));
+        }
+        self.put(
+            AUX_FILES_KEY,
+            Value::Image(Bytes::from(
+                AuxFilesDirectory::ser(&dir).context("serialize")?,
+            )),
+        );
+        Ok(())
+    }
+
     ///
     /// Flush changes accumulated so far to the underlying repository.
     ///
@@ -1255,6 +1357,11 @@ struct RelDirectory {
     rels: HashSet<(Oid, u8)>,
 }
 
+#[derive(Debug, Serialize, Deserialize, Default)]
+struct AuxFilesDirectory {
+    files: HashMap<String, Bytes>,
+}
+
 #[derive(Debug, Serialize, Deserialize)]
 struct RelSizeEntry {
     nblocks: u32,
@@ -1303,10 +1410,12 @@ static ZERO_PAGE: Bytes = Bytes::from_static(&[0u8; BLCKSZ as usize]);
 // 02 pg_twophase
 //
 // 03 misc
-//    controlfile
+//    Controlfile
 //    checkpoint
 //    pg_version
 //
+// 04 aux files
+//
 // Below is a full list of the keyspace allocation:
 //
 // DbDir:
@@ -1344,6 +1453,11 @@ static ZERO_PAGE: Bytes = Bytes::from_static(&[0u8; BLCKSZ as usize]);
 //
 // Checkpoint:
 // 03 00000000 00000000 00000000 00   00000001
+//
+// AuxFiles:
+// 03 00000000 00000000 00000000 00   00000002
+//
+
 //-- Section 01: relation data and metadata
 
 const DBDIR_KEY: Key = Key {
@@ -1567,6 +1681,15 @@ const CHECKPOINT_KEY: Key = Key {
     field6: 1,
 };
 
+const AUX_FILES_KEY: Key = Key {
+    field1: 0x03,
+    field2: 0,
+    field3: 0,
+    field4: 0,
+    field5: 0,
+    field6: 2,
+};
+
 // Reverse mappings for a few Keys.
 // These are needed by WAL redo manager.
 

pageserver/src/tenant.rs

@@ -18,17 +18,20 @@ use pageserver_api::models::TimelineState;
 use remote_storage::DownloadError;
 use remote_storage::GenericRemoteStorage;
 use storage_broker::BrokerClientChannel;
+use tokio::runtime::Handle;
 use tokio::sync::watch;
 use tokio::task::JoinSet;
 use tokio_util::sync::CancellationToken;
 use tracing::*;
 use utils::completion;
+use utils::completion::Completion;
 use utils::crashsafe::path_with_suffix_extension;
 
 use std::cmp::min;
 use std::collections::hash_map::Entry;
 use std::collections::BTreeSet;
 use std::collections::HashMap;
+use std::collections::HashSet;
 use std::fmt::Debug;
 use std::fmt::Display;
 use std::fs;
@@ -75,12 +78,12 @@ use crate::tenant::remote_timeline_client::MaybeDeletedIndexPart;
 use crate::tenant::storage_layer::DeltaLayer;
 use crate::tenant::storage_layer::ImageLayer;
 use crate::InitializationOrder;
+use crate::METADATA_FILE_NAME;
 
 use crate::tenant::timeline::delete::DeleteTimelineFlow;
 use crate::tenant::timeline::uninit::cleanup_timeline_directory;
 use crate::virtual_file::VirtualFile;
 use crate::walredo::PostgresRedoManager;
-use crate::walredo::WalRedoManager;
 use crate::TEMP_FILE_SUFFIX;
 pub use pageserver_api::models::TenantState;
 
@@ -185,6 +188,11 @@ impl AttachedTenantConf {
         }
     }
 }
+struct TimelinePreload {
+    timeline_id: TimelineId,
+    client: RemoteTimelineClient,
+    index_part: Result<MaybeDeletedIndexPart, DownloadError>,
+}
 
 ///
 /// Tenant consists of multiple timelines. Keep them in a hash table.
@@ -222,7 +230,7 @@ pub struct Tenant {
     // with timelines, which in turn may cause dropping replication connection, expiration of wait_for_lsn
     // timeout...
     gc_cs: tokio::sync::Mutex<()>,
-    walredo_mgr: Arc<dyn WalRedoManager + Send + Sync>,
+    walredo_mgr: Arc<WalRedoManager>,
 
     // provides access to timeline data sitting in the remote storage
     pub(crate) remote_storage: Option<GenericRemoteStorage>,
@@ -239,67 +247,43 @@ pub struct Tenant {
     pub(crate) delete_progress: Arc<tokio::sync::Mutex<DeleteTenantFlow>>,
 }
 
-// We should not blindly overwrite local metadata with remote one.
-// For example, consider the following case:
-//     Image layer is flushed to disk as a new delta layer, we update local metadata and start upload task but after that
-//     pageserver crashes. During startup we'll load new metadata, and then reset it
-//     to the state of remote one. But current layermap will have layers from the old
-//     metadata which is inconsistent.
-//     And with current logic it wont disgard them during load because during layermap
-//     load it sees local disk consistent lsn which is ahead of layer lsns.
-//     If we treat remote as source of truth we need to completely sync with it,
-//     i e delete local files which are missing on the remote. This will add extra work,
-//     wal for these layers needs to be reingested for example
-//
-// So the solution is to take remote metadata only when we're attaching.
-pub fn merge_local_remote_metadata<'a>(
-    local: Option<&'a TimelineMetadata>,
-    remote: Option<&'a TimelineMetadata>,
-) -> anyhow::Result<(&'a TimelineMetadata, bool)> {
-    match (local, remote) {
-        (None, None) => anyhow::bail!("we should have either local metadata or remote"),
-        (Some(local), None) => Ok((local, true)),
-        // happens if we crash during attach, before writing out the metadata file
-        (None, Some(remote)) => Ok((remote, false)),
-        // This is the regular case where we crash/exit before finishing queued uploads.
-        // Also, it happens if we crash during attach after writing the metadata file
-        // but before removing the attaching marker file.
-        (Some(local), Some(remote)) => {
-            let consistent_lsn_cmp = local
-                .disk_consistent_lsn()
-                .cmp(&remote.disk_consistent_lsn());
-            let gc_cutoff_lsn_cmp = local
-                .latest_gc_cutoff_lsn()
-                .cmp(&remote.latest_gc_cutoff_lsn());
-            use std::cmp::Ordering::*;
-            match (consistent_lsn_cmp, gc_cutoff_lsn_cmp) {
-                // It wouldn't matter, but pick the local one so that we don't rewrite the metadata file.
-                (Equal, Equal) => Ok((local, true)),
-                // Local state is clearly ahead of the remote.
-                (Greater, Greater) => Ok((local, true)),
-                // We have local layer files that aren't on the remote, but GC horizon is on par.
-                (Greater, Equal) => Ok((local, true)),
-                // Local GC started running but we couldn't sync it to the remote.
-                (Equal, Greater) => Ok((local, true)),
+pub(crate) enum WalRedoManager {
+    Prod(PostgresRedoManager),
+    #[cfg(test)]
+    Test(harness::TestRedoManager),
+}
 
-                // We always update the local value first, so something else must have
-                // updated the remote value, probably a different pageserver.
-                // The control plane is supposed to prevent this from happening.
-                // Bail out.
-                (Less, Less)
-                | (Less, Equal)
-                | (Equal, Less)
-                | (Less, Greater)
-                | (Greater, Less) => {
-                    anyhow::bail!(
-                        r#"remote metadata appears to be ahead of local metadata:
-local:
-  {local:#?}
-remote:
-  {remote:#?}
-"#
-                    );
-                }
+impl From<PostgresRedoManager> for WalRedoManager {
+    fn from(mgr: PostgresRedoManager) -> Self {
+        Self::Prod(mgr)
+    }
+}
+
+#[cfg(test)]
+impl From<harness::TestRedoManager> for WalRedoManager {
+    fn from(mgr: harness::TestRedoManager) -> Self {
+        Self::Test(mgr)
+    }
+}
+
+impl WalRedoManager {
+    pub async fn request_redo(
+        &self,
+        key: crate::repository::Key,
+        lsn: Lsn,
+        base_img: Option<(Lsn, bytes::Bytes)>,
+        records: Vec<(Lsn, crate::walrecord::NeonWalRecord)>,
+        pg_version: u32,
+    ) -> anyhow::Result<bytes::Bytes> {
+        match self {
+            Self::Prod(mgr) => {
+                mgr.request_redo(key, lsn, base_img, records, pg_version)
+                    .await
+            }
+            #[cfg(test)]
+            Self::Test(mgr) => {
+                mgr.request_redo(key, lsn, base_img, records, pg_version)
+                    .await
             }
         }
     }
@@ -368,11 +352,6 @@ impl Debug for SetStoppingError {
     }
 }
 
-struct RemoteStartupData {
-    index_part: IndexPart,
-    remote_metadata: TimelineMetadata,
-}
-
 #[derive(Debug, thiserror::Error)]
 pub(crate) enum WaitToBecomeActiveError {
     WillNotBecomeActive {
@@ -413,6 +392,12 @@ pub enum CreateTimelineError {
     Other(#[from] anyhow::Error),
 }
 
+/// spawn_attach argument for whether the caller is using attachment markers
+pub(super) enum AttachMarkerMode {
+    Expect,
+    Ignore,
+}
+
 struct TenantDirectoryScan {
     sorted_timelines_to_load: Vec<(TimelineId, TimelineMetadata)>,
     timelines_to_resume_deletion: Vec<(TimelineId, Option<TimelineMetadata>)>,
@@ -439,24 +424,17 @@ impl Tenant {
         &self,
         timeline_id: TimelineId,
         resources: TimelineResources,
-        remote_startup_data: Option<RemoteStartupData>,
-        local_metadata: Option<TimelineMetadata>,
+        index_part: Option<IndexPart>,
+        metadata: TimelineMetadata,
         ancestor: Option<Arc<Timeline>>,
         init_order: Option<&InitializationOrder>,
         _ctx: &RequestContext,
     ) -> anyhow::Result<()> {
         let tenant_id = self.tenant_id;
 
-        let (up_to_date_metadata, picked_local) = merge_local_remote_metadata(
-            local_metadata.as_ref(),
-            remote_startup_data.as_ref().map(|r| &r.remote_metadata),
-        )
-        .context("merge_local_remote_metadata")?
-        .to_owned();
-
         let timeline = self.create_timeline_struct(
             timeline_id,
-            up_to_date_metadata,
+            &metadata,
             ancestor.clone(),
             resources,
             init_order,
@@ -469,20 +447,11 @@ impl Tenant {
         );
         assert_eq!(
             disk_consistent_lsn,
-            up_to_date_metadata.disk_consistent_lsn(),
+            metadata.disk_consistent_lsn(),
             "these are used interchangeably"
         );
 
-        // Save the metadata file to local disk.
-        if !picked_local {
-            save_metadata(self.conf, &tenant_id, &timeline_id, up_to_date_metadata)
-                .await
-                .context("save_metadata")?;
-        }
-
-        let index_part = remote_startup_data.as_ref().map(|x| &x.index_part);
-
-        if let Some(index_part) = index_part {
+        if let Some(index_part) = index_part.as_ref() {
             timeline
                 .remote_client
                 .as_ref()
@@ -495,15 +464,12 @@ impl Tenant {
             // If control plane retries timeline creation in the meantime, the mgmt API handler
             // for timeline creation will coalesce on the upload we queue here.
             let rtc = timeline.remote_client.as_ref().unwrap();
-            rtc.init_upload_queue_for_empty_remote(up_to_date_metadata)?;
-            rtc.schedule_index_upload_for_metadata_update(up_to_date_metadata)?;
+            rtc.init_upload_queue_for_empty_remote(&metadata)?;
+            rtc.schedule_index_upload_for_metadata_update(&metadata)?;
         }
 
         timeline
-            .load_layer_map(
-                disk_consistent_lsn,
-                remote_startup_data.map(|x| x.index_part),
-            )
+            .load_layer_map(disk_consistent_lsn, index_part)
             .await
             .with_context(|| {
                 format!("Failed to load layermap for timeline {tenant_id}/{timeline_id}")
@@ -560,10 +526,13 @@ impl Tenant {
         resources: TenantSharedResources,
         attached_conf: AttachedTenantConf,
         tenants: &'static tokio::sync::RwLock<TenantsMap>,
+        expect_marker: AttachMarkerMode,
         ctx: &RequestContext,
     ) -> anyhow::Result<Arc<Tenant>> {
         // TODO dedup with spawn_load
-        let wal_redo_manager = Arc::new(PostgresRedoManager::new(conf, tenant_id));
+        let wal_redo_manager = Arc::new(WalRedoManager::from(PostgresRedoManager::new(
+            conf, tenant_id,
+        )));
 
         let TenantSharedResources {
             broker_client,
@@ -641,7 +610,7 @@ impl Tenant {
                     }
                 }
 
-                match tenant_clone.attach(&ctx).await {
+                match tenant_clone.attach(&ctx, expect_marker).await {
                     Ok(()) => {
                         info!("attach finished, activating");
                         tenant_clone.activate(broker_client, None, &ctx);
@@ -666,17 +635,23 @@ impl Tenant {
     ///
     /// No background tasks are started as part of this routine.
     ///
-    async fn attach(self: &Arc<Tenant>, ctx: &RequestContext) -> anyhow::Result<()> {
+    async fn attach(
+        self: &Arc<Tenant>,
+        ctx: &RequestContext,
+        expect_marker: AttachMarkerMode,
+    ) -> anyhow::Result<()> {
         span::debug_assert_current_span_has_tenant_id();
 
         let marker_file = self.conf.tenant_attaching_mark_file_path(&self.tenant_id);
-        if !tokio::fs::try_exists(&marker_file)
-            .await
-            .context("check for existence of marker file")?
-        {
-            anyhow::bail!(
-                "implementation error: marker file should exist at beginning of this function"
-            );
+        if let AttachMarkerMode::Expect = expect_marker {
+            if !tokio::fs::try_exists(&marker_file)
+                .await
+                .context("check for existence of marker file")?
+            {
+                anyhow::bail!(
+                    "implementation error: marker file should exist at beginning of this function"
+                );
+            }
         }
 
         // Get list of remote timelines
@@ -798,10 +773,12 @@ impl Tenant {
             .map_err(LoadLocalTimelineError::ResumeDeletion)?;
         }
 
-        std::fs::remove_file(&marker_file)
-            .with_context(|| format!("unlink attach marker file {marker_file}"))?;
-        crashsafe::fsync(marker_file.parent().expect("marker file has parent dir"))
-            .context("fsync tenant directory after unlinking attach marker file")?;
+        if let AttachMarkerMode::Expect = expect_marker {
+            std::fs::remove_file(&marker_file)
+                .with_context(|| format!("unlink attach marker file {marker_file}"))?;
+            crashsafe::fsync(marker_file.parent().expect("marker file has parent dir"))
+                .context("fsync tenant directory after unlinking attach marker file")?;
+        }
 
         crate::failpoint_support::sleep_millis_async!("attach-before-activate");
 
@@ -854,21 +831,23 @@ impl Tenant {
             None
         };
 
-        // Even if there is local metadata it cannot be ahead of the remote one
-        // since we're attaching. Even if we resume interrupted attach remote one
-        // cannot be older than the local one
-        let local_metadata = None;
+        // we can load remote timelines during init, but they are assumed to be so rare that
+        // initialization order is not passed to here.
+        let init_order = None;
+
+        // timeline loading after attach expects to find metadata file for each metadata
+        save_metadata(self.conf, &self.tenant_id, &timeline_id, &remote_metadata)
+            .await
+            .context("save_metadata")
+            .map_err(LoadLocalTimelineError::Load)?;
 
         self.timeline_init_and_sync(
             timeline_id,
             resources,
-            Some(RemoteStartupData {
-                index_part,
-                remote_metadata,
-            }),
-            local_metadata,
+            Some(index_part),
+            remote_metadata,
             ancestor,
-            None,
+            init_order,
             ctx,
         )
         .await
@@ -880,7 +859,9 @@ impl Tenant {
         tenant_id: TenantId,
         reason: String,
     ) -> Arc<Tenant> {
-        let wal_redo_manager = Arc::new(PostgresRedoManager::new(conf, tenant_id));
+        let wal_redo_manager = Arc::new(WalRedoManager::from(PostgresRedoManager::new(
+            conf, tenant_id,
+        )));
         Arc::new(Tenant::new(
             TenantState::Broken {
                 reason,
@@ -919,7 +900,9 @@ impl Tenant {
         let broker_client = resources.broker_client;
         let remote_storage = resources.remote_storage;
 
-        let wal_redo_manager = Arc::new(PostgresRedoManager::new(conf, tenant_id));
+        let wal_redo_manager = Arc::new(WalRedoManager::from(PostgresRedoManager::new(
+            conf, tenant_id,
+        )));
         let tenant = Tenant::new(
             TenantState::Loading,
             conf,
@@ -962,6 +945,9 @@ impl Tenant {
                 let _completion = init_order
                     .as_mut()
                     .and_then(|x| x.initial_tenant_load.take());
+                let remote_load_completion = init_order
+                    .as_mut()
+                    .and_then(|x| x.initial_tenant_load_remote.take());
 
                 // Dont block pageserver startup on figuring out deletion status
                 let pending_deletion = {
@@ -986,6 +972,7 @@ impl Tenant {
                     // as we are no longer loading, signal completion by dropping
                     // the completion while we resume deletion
                     drop(_completion);
+                    drop(remote_load_completion);
                     // do not hold to initial_logical_size_attempt as it will prevent loading from proceeding without timeout
                     let _ = init_order
                         .as_mut()
@@ -1011,7 +998,10 @@ impl Tenant {
                 let background_jobs_can_start =
                     init_order.as_ref().map(|x| &x.background_jobs_can_start);
 
-                match tenant_clone.load(init_order.as_ref(), &ctx).await {
+                match tenant_clone
+                    .load(init_order.as_ref(), remote_load_completion, &ctx)
+                    .await
+                {
                     Ok(()) => {
                         debug!("load finished");
 
@@ -1175,6 +1165,52 @@ impl Tenant {
         })
     }
 
+    async fn load_timeline_metadata(
+        self: &Arc<Tenant>,
+        timeline_ids: HashSet<TimelineId>,
+        remote_storage: &GenericRemoteStorage,
+    ) -> anyhow::Result<HashMap<TimelineId, TimelinePreload>> {
+        let mut part_downloads = JoinSet::new();
+        for timeline_id in timeline_ids {
+            let client = RemoteTimelineClient::new(
+                remote_storage.clone(),
+                self.deletion_queue_client.clone(),
+                self.conf,
+                self.tenant_id,
+                timeline_id,
+                self.generation,
+            );
+            part_downloads.spawn(
+                async move {
+                    debug!("starting index part download");
+
+                    let index_part = client.download_index_file().await;
+
+                    debug!("finished index part download");
+
+                    Result::<_, anyhow::Error>::Ok(TimelinePreload {
+                        client,
+                        timeline_id,
+                        index_part,
+                    })
+                }
+                .map(move |res| {
+                    res.with_context(|| format!("download index part for timeline {timeline_id}"))
+                })
+                .instrument(info_span!("download_index_part", %timeline_id)),
+            );
+        }
+
+        let mut timeline_preloads: HashMap<TimelineId, TimelinePreload> = HashMap::new();
+        while let Some(result) = part_downloads.join_next().await {
+            let preload_result = result.context("join preload task")?;
+            let preload = preload_result?;
+            timeline_preloads.insert(preload.timeline_id, preload);
+        }
+
+        Ok(timeline_preloads)
+    }
+
     ///
     /// Background task to load in-memory data structures for this tenant, from
     /// files on disk. Used at pageserver startup.
@@ -1183,14 +1219,13 @@ impl Tenant {
     async fn load(
         self: &Arc<Tenant>,
         init_order: Option<&InitializationOrder>,
+        remote_completion: Option<Completion>,
         ctx: &RequestContext,
     ) -> anyhow::Result<()> {
         span::debug_assert_current_span_has_tenant_id();
 
         debug!("loading tenant task");
 
-        crate::failpoint_support::sleep_millis_async!("before-loading-tenant");
-
         // Load in-memory state to reflect the local files on disk
         //
         // Scan the directory, peek into the metadata file of each timeline, and
@@ -1209,10 +1244,38 @@ impl Tenant {
         // FIXME original collect_timeline_files contained one more check:
         //    1. "Timeline has no ancestor and no layer files"
 
+        // Load remote content for timelines in this tenant
+        let all_timeline_ids = scan
+            .sorted_timelines_to_load
+            .iter()
+            .map(|i| i.0)
+            .chain(scan.timelines_to_resume_deletion.iter().map(|i| i.0))
+            .collect();
+        let mut preload = if let Some(remote_storage) = &self.remote_storage {
+            Some(
+                self.load_timeline_metadata(all_timeline_ids, remote_storage)
+                    .await?,
+            )
+        } else {
+            None
+        };
+
+        drop(remote_completion);
+
+        crate::failpoint_support::sleep_millis_async!("before-loading-tenant");
+
         // Process loadable timelines first
         for (timeline_id, local_metadata) in scan.sorted_timelines_to_load {
+            let timeline_preload = preload.as_mut().map(|p| p.remove(&timeline_id).unwrap());
             if let Err(e) = self
-                .load_local_timeline(timeline_id, local_metadata, init_order, ctx, false)
+                .load_local_timeline(
+                    timeline_id,
+                    local_metadata,
+                    timeline_preload,
+                    init_order,
+                    ctx,
+                    false,
+                )
                 .await
             {
                 match e {
@@ -1245,16 +1308,25 @@ impl Tenant {
                     }
                 }
                 Some(local_metadata) => {
+                    let timeline_preload =
+                        preload.as_mut().map(|p| p.remove(&timeline_id).unwrap());
                     if let Err(e) = self
-                        .load_local_timeline(timeline_id, local_metadata, init_order, ctx, true)
+                        .load_local_timeline(
+                            timeline_id,
+                            local_metadata,
+                            timeline_preload,
+                            init_order,
+                            ctx,
+                            true,
+                        )
                         .await
                     {
                         match e {
                             LoadLocalTimelineError::Load(source) => {
                                 // We tried to load deleted timeline, this is a bug.
                                 return Err(anyhow::anyhow!(source).context(
-                                "This is a bug. We tried to load deleted timeline which is wrong and loading failed. Timeline: {timeline_id}"
-                            ));
+                                    format!("This is a bug. We tried to load deleted timeline which is wrong and loading failed. Timeline: {timeline_id}")
+                                ));
                             }
                             LoadLocalTimelineError::ResumeDeletion(source) => {
                                 // Make sure resumed deletion wont fail loading for entire tenant.
@@ -1274,11 +1346,12 @@ impl Tenant {
     /// Subroutine of `load_tenant`, to load an individual timeline
     ///
     /// NB: The parent is assumed to be already loaded!
-    #[instrument(skip(self, local_metadata, init_order, ctx))]
+    #[instrument(skip(self, local_metadata, init_order, preload, ctx))]
     async fn load_local_timeline(
         self: &Arc<Self>,
         timeline_id: TimelineId,
         local_metadata: TimelineMetadata,
+        preload: Option<TimelinePreload>,
         init_order: Option<&InitializationOrder>,
         ctx: &RequestContext,
         found_delete_mark: bool,
@@ -1287,76 +1360,147 @@ impl Tenant {
 
         let mut resources = self.build_timeline_resources(timeline_id);
 
-        let (remote_startup_data, remote_client) = match resources.remote_client {
-            Some(remote_client) => match remote_client.download_index_file().await {
-                Ok(index_part) => {
-                    let index_part = match index_part {
-                        MaybeDeletedIndexPart::IndexPart(index_part) => index_part,
-                        MaybeDeletedIndexPart::Deleted(index_part) => {
-                            // TODO: we won't reach here if remote storage gets de-configured after start of the deletion operation.
-                            // Example:
-                            //  start deletion operation
-                            //  finishes upload of index part
-                            //  pageserver crashes
-                            //  remote storage gets de-configured
-                            //  pageserver starts
-                            //
-                            // We don't really anticipate remote storage to be de-configured, so, for now, this is fine.
-                            // Also, maybe we'll remove that option entirely in the future, see https://github.com/neondatabase/neon/issues/4099.
-                            info!("is_deleted is set on remote, resuming removal of timeline data originally done by timeline deletion handler");
+        struct RemoteStartupData {
+            index_part: IndexPart,
+            remote_metadata: TimelineMetadata,
+        }
 
-                            remote_client
-                                .init_upload_queue_stopped_to_continue_deletion(&index_part)
-                                .context("init queue stopped")
+        let (remote_startup_data, remote_client) = match preload {
+            Some(preload) => {
+                let TimelinePreload {
+                    index_part,
+                    client: remote_client,
+                    timeline_id: _timeline_id,
+                } = preload;
+                match index_part {
+                    Ok(index_part) => {
+                        let index_part = match index_part {
+                            MaybeDeletedIndexPart::IndexPart(index_part) => index_part,
+                            MaybeDeletedIndexPart::Deleted(index_part) => {
+                                // TODO: we won't reach here if remote storage gets de-configured after start of the deletion operation.
+                                // Example:
+                                //  start deletion operation
+                                //  finishes upload of index part
+                                //  pageserver crashes
+                                //  remote storage gets de-configured
+                                //  pageserver starts
+                                //
+                                // We don't really anticipate remote storage to be de-configured, so, for now, this is fine.
+                                // Also, maybe we'll remove that option entirely in the future, see https://github.com/neondatabase/neon/issues/4099.
+                                info!("is_deleted is set on remote, resuming removal of timeline data originally done by timeline deletion handler");
+
+                                remote_client
+                                    .init_upload_queue_stopped_to_continue_deletion(&index_part)
+                                    .context("init queue stopped")
+                                    .map_err(LoadLocalTimelineError::ResumeDeletion)?;
+
+                                DeleteTimelineFlow::resume_deletion(
+                                    Arc::clone(self),
+                                    timeline_id,
+                                    &local_metadata,
+                                    Some(remote_client),
+                                    self.deletion_queue_client.clone(),
+                                    init_order,
+                                )
+                                .await
+                                .context("resume deletion")
                                 .map_err(LoadLocalTimelineError::ResumeDeletion)?;
 
-                            DeleteTimelineFlow::resume_deletion(
-                                Arc::clone(self),
+                                return Ok(());
+                            }
+                        };
+
+                        let remote_metadata = index_part.metadata.clone();
+                        (
+                            Some(RemoteStartupData {
+                                index_part,
+                                remote_metadata,
+                            }),
+                            Some(remote_client),
+                        )
+                    }
+                    Err(DownloadError::NotFound) => {
+                        info!(found_delete_mark, "no index file was found on the remote, resuming deletion or cleaning unuploaded up");
+
+                        if found_delete_mark {
+                            // We could've resumed at a point where remote index was deleted, but metadata file wasnt.
+                            // Cleanup:
+                            return DeleteTimelineFlow::cleanup_remaining_timeline_fs_traces(
+                                self,
                                 timeline_id,
-                                &local_metadata,
-                                Some(remote_client),
-                                self.deletion_queue_client.clone(),
-                                init_order,
                             )
                             .await
-                            .context("resume deletion")
-                            .map_err(LoadLocalTimelineError::ResumeDeletion)?;
-
-                            return Ok(());
+                            .context("cleanup_remaining_timeline_fs_traces")
+                            .map_err(LoadLocalTimelineError::ResumeDeletion);
                         }
-                    };
 
-                    let remote_metadata = index_part.metadata.clone();
-                    (
-                        Some(RemoteStartupData {
-                            index_part,
-                            remote_metadata,
-                        }),
-                        Some(remote_client),
-                    )
-                }
-                Err(DownloadError::NotFound) => {
-                    info!("no index file was found on the remote, found_delete_mark: {found_delete_mark}");
+                        // as the remote index_part.json did not exist, this timeline is a
+                        // not-yet-uploaded one. it should be deleted now, because the branching might
+                        // not have been valid as it's ancestor may have been restored to earlier state
+                        // as well. in practice, control plane will keep retrying.
+                        //
+                        // first ensure that the un-uploaded timeline looks like it should, as in we
+                        // are not accidentially deleting a timeline which was ever active:
+                        // - root timelines have metadata and one possibly partial layer
+                        // - branched timelines have metadata
+                        //
+                        // if the timeline does not look like expected, fail loading of the tenant.
+                        // cleaning the timeline up manually and reloading the tenant is possible via
+                        // the above log message.
+                        let path = self.conf.timeline_path(&self.tenant_id, &timeline_id);
 
-                    if found_delete_mark {
-                        // We could've resumed at a point where remote index was deleted, but metadata file wasnt.
-                        // Cleanup:
-                        return DeleteTimelineFlow::cleanup_remaining_timeline_fs_traces(
-                            self,
-                            timeline_id,
-                        )
-                        .await
-                        .context("cleanup_remaining_timeline_fs_traces")
-                        .map_err(LoadLocalTimelineError::ResumeDeletion);
+                        let span = tracing::Span::current();
+
+                        return tokio::task::spawn_blocking({
+                        move || {
+                            use std::str::FromStr;
+                            use crate::tenant::storage_layer::LayerFileName;
+
+                            let _e = span.entered();
+                            let mut metadata = false;
+                            let mut layers = 0;
+                            let mut others = 0;
+                            for dentry in path.read_dir_utf8()? {
+                                let dentry = dentry?;
+                                let file_name = dentry.file_name();
+
+                                if file_name == METADATA_FILE_NAME {
+                                    metadata = true;
+                                    continue;
+                                }
+
+                                if LayerFileName::from_str(file_name).is_ok()
+                                {
+                                    layers += 1;
+                                    continue;
+                                }
+
+                                others += 1;
+                            }
+
+                            // bootstrapped have the one image layer file, or one partial temp
+                            // file, branched have just the metadata
+                            if !(metadata && layers + others <= 1) {
+                                anyhow::bail!("unexpected assumed unuploaded, never been active timeline: found metadata={}, layers={}, others={}", metadata, layers, others);
+                            }
+
+                            let tmp_path =
+                                path.with_file_name(format!("{timeline_id}{}", TEMP_FILE_SUFFIX));
+                            std::fs::rename(path, &tmp_path)?;
+                            std::fs::remove_dir_all(&tmp_path)?;
+                            Ok(())
+                        }
+                    })
+                    .await
+                    .map_err(anyhow::Error::new)
+                    .and_then(|x| x)
+                    .context("delete assumed unuploaded fresh timeline")
+                    .map_err(LoadLocalTimelineError::Load);
                     }
-
-                    // We're loading fresh timeline that didnt yet make it into remote.
-                    (None, Some(remote_client))
+                    Err(e) => return Err(LoadLocalTimelineError::Load(anyhow::Error::new(e))),
                 }
-                Err(e) => return Err(LoadLocalTimelineError::Load(anyhow::Error::new(e))),
-            },
+            }
             None => {
-                // No remote client
                 if found_delete_mark {
                     // There is no remote client, we found local metadata.
                     // Continue cleaning up local disk.
@@ -1388,11 +1532,27 @@ impl Tenant {
             None
         };
 
+        let (index_part, metadata) = match remote_startup_data {
+            Some(RemoteStartupData {
+                index_part,
+                remote_metadata,
+            }) => {
+                // always choose the remote metadata to be crash consistent (see RFC 27)
+                save_metadata(self.conf, &self.tenant_id, &timeline_id, &remote_metadata)
+                    .await
+                    .context("save_metadata")
+                    .map_err(LoadLocalTimelineError::Load)?;
+
+                (Some(index_part), remote_metadata)
+            }
+            None => (None, local_metadata),
+        };
+
         self.timeline_init_and_sync(
             timeline_id,
             resources,
-            remote_startup_data,
-            Some(local_metadata),
+            index_part,
+            metadata,
             ancestor,
             init_order,
             ctx,
@@ -2307,7 +2467,7 @@ impl Tenant {
         state: TenantState,
         conf: &'static PageServerConf,
         attached_conf: AttachedTenantConf,
-        walredo_mgr: Arc<dyn WalRedoManager + Send + Sync>,
+        walredo_mgr: Arc<WalRedoManager>,
         tenant_id: TenantId,
         remote_storage: Option<GenericRemoteStorage>,
         deletion_queue_client: DeletionQueueClient,
@@ -2455,6 +2615,7 @@ impl Tenant {
     ) -> anyhow::Result<()> {
         let legacy_config_path = conf.tenant_config_path(tenant_id);
         let config_path = conf.tenant_location_config_path(tenant_id);
+
         Self::persist_tenant_config_at(tenant_id, &config_path, &legacy_config_path, location_conf)
             .await
     }
@@ -2493,12 +2654,20 @@ impl Tenant {
         // Convert the config to a toml file.
         conf_content += &toml_edit::ser::to_string_pretty(&location_conf)?;
 
-        let conf_content = conf_content.as_bytes();
-
         let temp_path = path_with_suffix_extension(config_path, TEMP_FILE_SUFFIX);
-        VirtualFile::crashsafe_overwrite(config_path, &temp_path, conf_content)
-            .await
-            .with_context(|| format!("write tenant {tenant_id} config to {config_path}"))?;
+
+        let tenant_id = *tenant_id;
+        let config_path = config_path.to_owned();
+        tokio::task::spawn_blocking(move || {
+            Handle::current().block_on(async move {
+                let conf_content = conf_content.as_bytes();
+                VirtualFile::crashsafe_overwrite(&config_path, &temp_path, conf_content)
+                    .await
+                    .with_context(|| format!("write tenant {tenant_id} config to {config_path}"))
+            })
+        })
+        .await??;
+
         Ok(())
     }
 
@@ -2520,12 +2689,21 @@ impl Tenant {
         // Convert the config to a toml file.
         conf_content += &toml_edit::ser::to_string(&tenant_conf)?;
 
-        let conf_content = conf_content.as_bytes();
-
         let temp_path = path_with_suffix_extension(target_config_path, TEMP_FILE_SUFFIX);
-        VirtualFile::crashsafe_overwrite(target_config_path, &temp_path, conf_content)
-            .await
-            .with_context(|| format!("write tenant {tenant_id} config to {target_config_path}"))?;
+
+        let tenant_id = *tenant_id;
+        let target_config_path = target_config_path.to_owned();
+        tokio::task::spawn_blocking(move || {
+            Handle::current().block_on(async move {
+                let conf_content = conf_content.as_bytes();
+                VirtualFile::crashsafe_overwrite(&target_config_path, &temp_path, conf_content)
+                    .await
+                    .with_context(|| {
+                        format!("write tenant {tenant_id} config to {target_config_path}")
+                    })
+            })
+        })
+        .await??;
         Ok(())
     }
 
@@ -3438,7 +3616,7 @@ pub async fn dump_layerfile_from_path(
 }
 
 #[cfg(test)]
-pub mod harness {
+pub(crate) mod harness {
     use bytes::{Bytes, BytesMut};
     use once_cell::sync::OnceCell;
     use std::fs;
@@ -3448,11 +3626,7 @@ pub mod harness {
 
     use crate::deletion_queue::mock::MockDeletionQueue;
     use crate::{
-        config::PageServerConf,
-        repository::Key,
-        tenant::Tenant,
-        walrecord::NeonWalRecord,
-        walredo::{WalRedoError, WalRedoManager},
+        config::PageServerConf, repository::Key, tenant::Tenant, walrecord::NeonWalRecord,
     };
 
     use super::*;
@@ -3513,17 +3687,21 @@ pub mod harness {
 
     static LOG_HANDLE: OnceCell<()> = OnceCell::new();
 
+    pub(crate) fn setup_logging() {
+        LOG_HANDLE.get_or_init(|| {
+            logging::init(
+                logging::LogFormat::Test,
+                // enable it in case the tests exercise code paths that use
+                // debug_assert_current_span_has_tenant_and_timeline_id
+                logging::TracingErrorLayerEnablement::EnableWithRustLogFilter,
+            )
+            .expect("Failed to init test logging")
+        });
+    }
+
     impl TenantHarness {
         pub fn create(test_name: &'static str) -> anyhow::Result<Self> {
-            LOG_HANDLE.get_or_init(|| {
-                logging::init(
-                    logging::LogFormat::Test,
-                    // enable it in case in case the tests exercise code paths that use
-                    // debug_assert_current_span_has_tenant_and_timeline_id
-                    logging::TracingErrorLayerEnablement::EnableWithRustLogFilter,
-                )
-                .expect("Failed to init test logging")
-            });
+            setup_logging();
 
             let repo_dir = PageServerConf::test_repo_dir(test_name);
             let _ = fs::remove_dir_all(&repo_dir);
@@ -3581,7 +3759,7 @@ pub mod harness {
         }
 
         pub async fn try_load(&self, ctx: &RequestContext) -> anyhow::Result<Arc<Tenant>> {
-            let walredo_mgr = Arc::new(TestRedoManager);
+            let walredo_mgr = Arc::new(WalRedoManager::from(TestRedoManager));
 
             let tenant = Arc::new(Tenant::new(
                 TenantState::Loading,
@@ -3597,7 +3775,7 @@ pub mod harness {
                 self.deletion_queue.new_client(),
             ));
             tenant
-                .load(None, ctx)
+                .load(None, None, ctx)
                 .instrument(info_span!("try_load", tenant_id=%self.tenant_id))
                 .await?;
 
@@ -3615,17 +3793,17 @@ pub mod harness {
     }
 
     // Mock WAL redo manager that doesn't do much
-    pub struct TestRedoManager;
+    pub(crate) struct TestRedoManager;
 
-    impl WalRedoManager for TestRedoManager {
-        fn request_redo(
+    impl TestRedoManager {
+        pub async fn request_redo(
             &self,
             key: Key,
             lsn: Lsn,
             base_img: Option<(Lsn, Bytes)>,
             records: Vec<(Lsn, NeonWalRecord)>,
             _pg_version: u32,
-        ) -> Result<Bytes, WalRedoError> {
+        ) -> anyhow::Result<Bytes> {
             let s = format!(
                 "redo for {} to get to {}, with {} and {} records",
                 key,

pageserver/src/tenant/delete.rs

@@ -432,7 +432,7 @@ impl DeleteTenantFlow {
         // Tenant may not be loadable if we fail late in cleanup_remaining_fs_traces (e g remove timelines dir)
         let timelines_path = tenant.conf.timelines_path(&tenant.tenant_id);
         if timelines_path.exists() {
-            tenant.load(init_order, ctx).await.context("load")?;
+            tenant.load(init_order, None, ctx).await.context("load")?;
         }
 
         Self::background(
@@ -458,7 +458,10 @@ impl DeleteTenantFlow {
             .await
             .expect("cant be stopping or broken");
 
-        tenant.attach(ctx).await.context("attach")?;
+        tenant
+            .attach(ctx, super::AttachMarkerMode::Expect)
+            .await
+            .context("attach")?;
 
         Self::background(
             guard,

pageserver/src/tenant/ephemeral_file.rs

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

pageserver/src/tenant/mgr.rs

@@ -1,7 +1,7 @@
 //! This module acts as a switchboard to access different repositories managed by this
 //! page server.
 
-use camino::{Utf8Path, Utf8PathBuf};
+use camino::{Utf8DirEntry, Utf8Path, Utf8PathBuf};
 use rand::{distributions::Alphanumeric, Rng};
 use std::collections::{hash_map, HashMap};
 use std::sync::Arc;
@@ -27,7 +27,8 @@ use crate::task_mgr::{self, TaskKind};
 use crate::tenant::config::{AttachmentMode, LocationConf, LocationMode, TenantConfOpt};
 use crate::tenant::delete::DeleteTenantFlow;
 use crate::tenant::{
-    create_tenant_files, AttachedTenantConf, CreateTenantFilesMode, Tenant, TenantState,
+    create_tenant_files, AttachMarkerMode, AttachedTenantConf, CreateTenantFilesMode, Tenant,
+    TenantState,
 };
 use crate::{InitializationOrder, IGNORED_TENANT_FILE_NAME, TEMP_FILE_SUFFIX};
 
@@ -151,6 +152,49 @@ async fn safe_rename_tenant_dir(path: impl AsRef<Utf8Path>) -> std::io::Result<U
 
 static TENANTS: Lazy<RwLock<TenantsMap>> = Lazy::new(|| RwLock::new(TenantsMap::Initializing));
 
+/// Create a directory, including parents.  This does no fsyncs and makes
+/// no guarantees about the persistence of the resulting metadata: for
+/// use when creating dirs for use as cache.
+async fn unsafe_create_dir_all(path: &Utf8PathBuf) -> std::io::Result<()> {
+    let mut dirs_to_create = Vec::new();
+    let mut path: &Utf8Path = path.as_ref();
+
+    // Figure out which directories we need to create.
+    loop {
+        let meta = tokio::fs::metadata(path).await;
+        match meta {
+            Ok(metadata) if metadata.is_dir() => break,
+            Ok(_) => {
+                return Err(std::io::Error::new(
+                    std::io::ErrorKind::AlreadyExists,
+                    format!("non-directory found in path: {path}"),
+                ));
+            }
+            Err(ref e) if e.kind() == std::io::ErrorKind::NotFound => {}
+            Err(e) => return Err(e),
+        }
+
+        dirs_to_create.push(path);
+
+        match path.parent() {
+            Some(parent) => path = parent,
+            None => {
+                return Err(std::io::Error::new(
+                    std::io::ErrorKind::InvalidInput,
+                    format!("can't find parent of path '{path}'"),
+                ));
+            }
+        }
+    }
+
+    // Create directories from parent to child.
+    for &path in dirs_to_create.iter().rev() {
+        tokio::fs::create_dir(path).await?;
+    }
+
+    Ok(())
+}
+
 fn emergency_generations(
     tenant_confs: &HashMap<TenantId, anyhow::Result<LocationConf>>,
 ) -> HashMap<TenantId, Generation> {
@@ -212,83 +256,99 @@ async fn init_load_generations(
     Ok(Some(generations))
 }
 
+/// Given a directory discovered in the pageserver's tenants/ directory, attempt
+/// to load a tenant config from it.
+///
+/// If file is missing, return Ok(None)
+fn load_tenant_config(
+    conf: &'static PageServerConf,
+    dentry: Utf8DirEntry,
+) -> anyhow::Result<Option<(TenantId, anyhow::Result<LocationConf>)>> {
+    let tenant_dir_path = dentry.path().to_path_buf();
+    if crate::is_temporary(&tenant_dir_path) {
+        info!("Found temporary tenant directory, removing: {tenant_dir_path}");
+        // No need to use safe_remove_tenant_dir_all because this is already
+        // a temporary path
+        if let Err(e) = std::fs::remove_dir_all(&tenant_dir_path) {
+            error!(
+                "Failed to remove temporary directory '{}': {:?}",
+                tenant_dir_path, e
+            );
+        }
+        return Ok(None);
+    }
+
+    // This case happens if we crash during attachment before writing a config into the dir
+    let is_empty = tenant_dir_path
+        .is_empty_dir()
+        .with_context(|| format!("Failed to check whether {tenant_dir_path:?} is an empty dir"))?;
+    if is_empty {
+        info!("removing empty tenant directory {tenant_dir_path:?}");
+        if let Err(e) = std::fs::remove_dir(&tenant_dir_path) {
+            error!(
+                "Failed to remove empty tenant directory '{}': {e:#}",
+                tenant_dir_path
+            )
+        }
+        return Ok(None);
+    }
+
+    let tenant_ignore_mark_file = tenant_dir_path.join(IGNORED_TENANT_FILE_NAME);
+    if tenant_ignore_mark_file.exists() {
+        info!("Found an ignore mark file {tenant_ignore_mark_file:?}, skipping the tenant");
+        return Ok(None);
+    }
+
+    let tenant_id = match tenant_dir_path
+        .file_name()
+        .unwrap_or_default()
+        .parse::<TenantId>()
+    {
+        Ok(id) => id,
+        Err(_) => {
+            warn!("Invalid tenant path (garbage in our repo directory?): {tenant_dir_path}",);
+            return Ok(None);
+        }
+    };
+
+    Ok(Some((
+        tenant_id,
+        Tenant::load_tenant_config(conf, &tenant_id),
+    )))
+}
+
 /// Initial stage of load: walk the local tenants directory, clean up any temp files,
 /// and load configurations for the tenants we found.
+///
+/// Do this in parallel, because we expect 10k+ tenants, so serial execution can take
+/// seconds even on reasonably fast drives.
 async fn init_load_tenant_configs(
     conf: &'static PageServerConf,
 ) -> anyhow::Result<HashMap<TenantId, anyhow::Result<LocationConf>>> {
     let tenants_dir = conf.tenants_path();
 
-    let mut dir_entries = tenants_dir
-        .read_dir_utf8()
-        .with_context(|| format!("Failed to list tenants dir {tenants_dir:?}"))?;
+    let dentries = tokio::task::spawn_blocking(move || -> anyhow::Result<Vec<Utf8DirEntry>> {
+        let dir_entries = tenants_dir
+            .read_dir_utf8()
+            .with_context(|| format!("Failed to list tenants dir {tenants_dir:?}"))?;
+
+        Ok(dir_entries.collect::<Result<Vec<_>, std::io::Error>>()?)
+    })
+    .await??;
 
     let mut configs = HashMap::new();
 
-    loop {
-        match dir_entries.next() {
-            None => break,
-            Some(Ok(dentry)) => {
-                let tenant_dir_path = dentry.path().to_path_buf();
-                if crate::is_temporary(&tenant_dir_path) {
-                    info!("Found temporary tenant directory, removing: {tenant_dir_path}");
-                    // No need to use safe_remove_tenant_dir_all because this is already
-                    // a temporary path
-                    if let Err(e) = fs::remove_dir_all(&tenant_dir_path).await {
-                        error!(
-                            "Failed to remove temporary directory '{}': {:?}",
-                            tenant_dir_path, e
-                        );
-                    }
-                    continue;
-                }
+    let mut join_set = JoinSet::new();
+    for dentry in dentries {
+        join_set.spawn_blocking(move || load_tenant_config(conf, dentry));
+    }
 
-                // This case happens if we:
-                // * crash during attach before creating the attach marker file
-                // * crash during tenant delete before removing tenant directory
-                let is_empty = tenant_dir_path.is_empty_dir().with_context(|| {
-                    format!("Failed to check whether {tenant_dir_path:?} is an empty dir")
-                })?;
-                if is_empty {
-                    info!("removing empty tenant directory {tenant_dir_path:?}");
-                    if let Err(e) = fs::remove_dir(&tenant_dir_path).await {
-                        error!(
-                            "Failed to remove empty tenant directory '{}': {e:#}",
-                            tenant_dir_path
-                        )
-                    }
-                    continue;
-                }
-
-                let tenant_ignore_mark_file = tenant_dir_path.join(IGNORED_TENANT_FILE_NAME);
-                if tenant_ignore_mark_file.exists() {
-                    info!("Found an ignore mark file {tenant_ignore_mark_file:?}, skipping the tenant");
-                    continue;
-                }
-
-                let tenant_id = match tenant_dir_path
-                    .file_name()
-                    .unwrap_or_default()
-                    .parse::<TenantId>()
-                {
-                    Ok(id) => id,
-                    Err(_) => {
-                        warn!(
-                            "Invalid tenant path (garbage in our repo directory?): {tenant_dir_path}",
-                        );
-                        continue;
-                    }
-                };
-
-                configs.insert(tenant_id, Tenant::load_tenant_config(conf, &tenant_id));
-            }
-            Some(Err(e)) => {
-                // An error listing the top level directory indicates serious problem
-                // with local filesystem: we will fail to load, and fail to start.
-                anyhow::bail!(e);
-            }
+    while let Some(r) = join_set.join_next().await {
+        if let Some((tenant_id, tenant_config)) = r?? {
+            configs.insert(tenant_id, tenant_config);
         }
     }
+
     Ok(configs)
 }
 
@@ -446,7 +506,15 @@ pub(crate) fn schedule_local_tenant_processing(
                 "attaching mark file present but no remote storage configured".to_string(),
             )
         } else {
-            match Tenant::spawn_attach(conf, tenant_id, resources, location_conf, tenants, ctx) {
+            match Tenant::spawn_attach(
+                conf,
+                tenant_id,
+                resources,
+                location_conf,
+                tenants,
+                AttachMarkerMode::Expect,
+                ctx,
+            ) {
                 Ok(tenant) => tenant,
                 Err(e) => {
                     error!("Failed to spawn_attach tenant {tenant_id}, reason: {e:#}");
@@ -655,7 +723,7 @@ pub(crate) async fn set_new_tenant_config(
     Ok(())
 }
 
-#[instrument(skip_all, fields(tenant_id, new_location_config))]
+#[instrument(skip_all, fields(%tenant_id))]
 pub(crate) async fn upsert_location(
     conf: &'static PageServerConf,
     tenant_id: TenantId,
@@ -734,36 +802,61 @@ pub(crate) async fn upsert_location(
             }
 
             let new_slot = match &new_location_config.mode {
-                LocationMode::Secondary(_) => TenantSlot::Secondary,
-                LocationMode::Attached(_attach_config) => {
-                    // Do a schedule_local_tenant_processing
-                    // FIXME: should avoid doing this disk I/O inside the TenantsMap lock,
-                    // we have the same problem in load_tenant/attach_tenant.  Probably
-                    // need a lock in TenantSlot to fix this.
+                LocationMode::Secondary(_) => {
+                    let tenant_path = conf.tenant_path(&tenant_id);
+                    // Directory doesn't need to be fsync'd because if we crash it can
+                    // safely be recreated next time this tenant location is configured.
+                    unsafe_create_dir_all(&tenant_path)
+                        .await
+                        .with_context(|| format!("Creating {tenant_path}"))?;
+
                     Tenant::persist_tenant_config(conf, &tenant_id, &new_location_config)
                         .await
                         .map_err(SetNewTenantConfigError::Persist)?;
-                    let tenant_path = conf.tenant_path(&tenant_id);
-                    let resources = TenantSharedResources {
-                        broker_client,
-                        remote_storage,
-                        deletion_queue_client,
-                    };
-                    let new_tenant = schedule_local_tenant_processing(
+
+                    TenantSlot::Secondary
+                }
+                LocationMode::Attached(_attach_config) => {
+                    // FIXME: should avoid doing this disk I/O inside the TenantsMap lock,
+                    // we have the same problem in load_tenant/attach_tenant.  Probably
+                    // need a lock in TenantSlot to fix this.
+                    let timelines_path = conf.timelines_path(&tenant_id);
+
+                    // Directory doesn't need to be fsync'd because we do not depend on
+                    // it to exist after crashes: it may be recreated when tenant is
+                    // re-attached, see https://github.com/neondatabase/neon/issues/5550
+                    unsafe_create_dir_all(&timelines_path)
+                        .await
+                        .with_context(|| format!("Creating {timelines_path}"))?;
+
+                    Tenant::persist_tenant_config(conf, &tenant_id, &new_location_config)
+                        .await
+                        .map_err(SetNewTenantConfigError::Persist)?;
+
+                    let tenant = match Tenant::spawn_attach(
                         conf,
                         tenant_id,
-                        &tenant_path,
+                        TenantSharedResources {
+                            broker_client,
+                            remote_storage,
+                            deletion_queue_client,
+                        },
                         AttachedTenantConf::try_from(new_location_config)?,
-                        resources,
-                        None,
                         &TENANTS,
+                        // The LocationConf API does not use marker files, because we have Secondary
+                        // locations where the directory's existence is not a signal that it contains
+                        // all timelines.  See https://github.com/neondatabase/neon/issues/5550
+                        AttachMarkerMode::Ignore,
                         ctx,
-                    )
-                    .with_context(|| {
-                        format!("Failed to schedule tenant processing in path {tenant_path:?}")
-                    })?;
+                    ) {
+                        Ok(tenant) => tenant,
+                        Err(e) => {
+                            error!("Failed to spawn_attach tenant {tenant_id}, reason: {e:#}");
+                            Tenant::create_broken_tenant(conf, tenant_id, format!("{e:#}"))
+                        }
+                    };
 
-                    TenantSlot::Attached(new_tenant)
+                    TenantSlot::Attached(tenant)
                 }
             };
 
@@ -771,7 +864,6 @@ pub(crate) async fn upsert_location(
         })
         .await?;
     }
-
     Ok(())
 }
 

pageserver/src/tenant/remote_timeline_client.rs

@@ -1419,6 +1419,13 @@ impl RemoteTimelineClient {
             }
         }
     }
+
+    pub(crate) fn get_layer_metadata(
+        &self,
+        name: &LayerFileName,
+    ) -> anyhow::Result<Option<LayerFileMetadata>> {
+        self.upload_queue.lock().unwrap().get_layer_metadata(name)
+    }
 }
 
 pub fn remote_timelines_path(tenant_id: &TenantId) -> RemotePath {

pageserver/src/tenant/remote_timeline_client/download.rs

@@ -18,7 +18,7 @@ use crate::config::PageServerConf;
 use crate::tenant::remote_timeline_client::{remote_layer_path, remote_timelines_path};
 use crate::tenant::storage_layer::LayerFileName;
 use crate::tenant::timeline::span::debug_assert_current_span_has_tenant_and_timeline_id;
-use crate::tenant::Generation;
+use crate::tenant::{Generation, TENANT_DELETED_MARKER_FILE_NAME};
 use remote_storage::{DownloadError, GenericRemoteStorage};
 use utils::crashsafe::path_with_suffix_extension;
 use utils::id::{TenantId, TimelineId};
@@ -190,6 +190,12 @@ pub async fn list_remote_timelines(
     let mut timeline_ids = HashSet::new();
 
     for timeline_remote_storage_key in timelines {
+        if timeline_remote_storage_key.object_name() == Some(TENANT_DELETED_MARKER_FILE_NAME) {
+            // A `deleted` key within `timelines/` is a marker file, not a timeline.  Ignore it.
+            // This code will be removed in https://github.com/neondatabase/neon/pull/5580
+            continue;
+        }
+
         let object_name = timeline_remote_storage_key.object_name().ok_or_else(|| {
             anyhow::anyhow!("failed to get timeline id for remote tenant {tenant_id}")
         })?;

pageserver/src/tenant/remote_timeline_client/upload.rs

@@ -60,6 +60,8 @@ pub(super) async fn upload_timeline_layer<'a>(
         bail!("failpoint before-upload-layer")
     });
 
+    pausable_failpoint!("before-upload-layer-pausable");
+
     let storage_path = remote_path(conf, source_path, generation)?;
     let source_file_res = fs::File::open(&source_path).await;
     let source_file = match source_file_res {

pageserver/src/tenant/storage_layer/filename.rs

@@ -226,6 +226,14 @@ impl LayerFileName {
             _ => false,
         }
     }
+
+    pub(crate) fn kind(&self) -> &'static str {
+        use LayerFileName::*;
+        match self {
+            Delta(_) => "delta",
+            Image(_) => "image",
+        }
+    }
 }
 
 impl fmt::Display for LayerFileName {

pageserver/src/tenant/storage_layer/remote_layer.rs

@@ -25,7 +25,7 @@ use super::{
 };
 
 /// RemoteLayer is a not yet downloaded [`ImageLayer`] or
-/// [`DeltaLayer`](super::DeltaLayer).
+/// [`DeltaLayer`].
 ///
 /// RemoteLayer might be downloaded on-demand during operations which are
 /// allowed download remote layers and during which, it gets replaced with a

pageserver/src/tenant/tasks.rs

@@ -14,6 +14,73 @@ use tokio_util::sync::CancellationToken;
 use tracing::*;
 use utils::completion;
 
+static CONCURRENT_BACKGROUND_TASKS: once_cell::sync::Lazy<tokio::sync::Semaphore> =
+    once_cell::sync::Lazy::new(|| {
+        let total_threads = *task_mgr::BACKGROUND_RUNTIME_WORKER_THREADS;
+        let permits = usize::max(
+            1,
+            // while a lot of the work is done on spawn_blocking, we still do
+            // repartitioning in the async context. this should give leave us some workers
+            // unblocked to be blocked on other work, hopefully easing any outside visible
+            // effects of restarts.
+            //
+            // 6/8 is a guess; previously we ran with unlimited 8 and more from
+            // spawn_blocking.
+            (total_threads * 3).checked_div(4).unwrap_or(0),
+        );
+        assert_ne!(permits, 0, "we will not be adding in permits later");
+        assert!(
+            permits < total_threads,
+            "need threads avail for shorter work"
+        );
+        tokio::sync::Semaphore::new(permits)
+    });
+
+#[derive(Debug, PartialEq, Eq, Clone, Copy, strum_macros::IntoStaticStr)]
+#[strum(serialize_all = "snake_case")]
+pub(crate) enum BackgroundLoopKind {
+    Compaction,
+    Gc,
+    Eviction,
+    ConsumptionMetricsCollectMetrics,
+    ConsumptionMetricsSyntheticSizeWorker,
+}
+
+impl BackgroundLoopKind {
+    fn as_static_str(&self) -> &'static str {
+        let s: &'static str = self.into();
+        s
+    }
+}
+
+pub(crate) enum RateLimitError {
+    Cancelled,
+}
+
+pub(crate) async fn concurrent_background_tasks_rate_limit(
+    loop_kind: BackgroundLoopKind,
+    _ctx: &RequestContext,
+    cancel: &CancellationToken,
+) -> Result<impl Drop, RateLimitError> {
+    crate::metrics::BACKGROUND_LOOP_SEMAPHORE_WAIT_START_COUNT
+        .with_label_values(&[loop_kind.as_static_str()])
+        .inc();
+    scopeguard::defer!(
+        crate::metrics::BACKGROUND_LOOP_SEMAPHORE_WAIT_FINISH_COUNT.with_label_values(&[loop_kind.as_static_str()]).inc();
+    );
+    tokio::select! {
+        permit = CONCURRENT_BACKGROUND_TASKS.acquire() => {
+            match permit {
+                Ok(permit) => Ok(permit),
+                Err(_closed) => unreachable!("we never close the semaphore"),
+            }
+        },
+        _ = cancel.cancelled() => {
+            Err(RateLimitError::Cancelled)
+        }
+    }
+}
+
 /// Start per tenant background loops: compaction and gc.
 pub fn start_background_loops(
     tenant: &Arc<Tenant>,
@@ -116,7 +183,7 @@ async fn compaction_loop(tenant: Arc<Tenant>, cancel: CancellationToken) {
                 }
             };
 
-            warn_when_period_overrun(started_at.elapsed(), period, "compaction");
+            warn_when_period_overrun(started_at.elapsed(), period, BackgroundLoopKind::Compaction);
 
             // Sleep
             if tokio::time::timeout(sleep_duration, cancel.cancelled())
@@ -184,7 +251,7 @@ async fn gc_loop(tenant: Arc<Tenant>, cancel: CancellationToken) {
                 }
             };
 
-            warn_when_period_overrun(started_at.elapsed(), period, "gc");
+            warn_when_period_overrun(started_at.elapsed(), period, BackgroundLoopKind::Gc);
 
             // Sleep
             if tokio::time::timeout(sleep_duration, cancel.cancelled())
@@ -258,7 +325,11 @@ pub(crate) async fn random_init_delay(
 }
 
 /// Attention: the `task` and `period` beocme labels of a pageserver-wide prometheus metric.
-pub(crate) fn warn_when_period_overrun(elapsed: Duration, period: Duration, task: &str) {
+pub(crate) fn warn_when_period_overrun(
+    elapsed: Duration,
+    period: Duration,
+    task: BackgroundLoopKind,
+) {
     // Duration::ZERO will happen because it's the "disable [bgtask]" value.
     if elapsed >= period && period != Duration::ZERO {
         // humantime does no significant digits clamping whereas Duration's debug is a bit more
@@ -267,11 +338,11 @@ pub(crate) fn warn_when_period_overrun(elapsed: Duration, period: Duration, task
         warn!(
             ?elapsed,
             period = %humantime::format_duration(period),
-            task,
+            ?task,
             "task iteration took longer than the configured period"
         );
         crate::metrics::BACKGROUND_LOOP_PERIOD_OVERRUN_COUNT
-            .with_label_values(&[task, &format!("{}", period.as_secs())])
+            .with_label_values(&[task.as_static_str(), &format!("{}", period.as_secs())])
             .inc();
     }
 }

pageserver/src/tenant/timeline.rs

@@ -44,6 +44,7 @@ use crate::tenant::storage_layer::delta_layer::DeltaEntry;
 use crate::tenant::storage_layer::{
     DeltaLayerWriter, ImageLayerWriter, InMemoryLayer, LayerAccessStats, LayerFileName, RemoteLayer,
 };
+use crate::tenant::tasks::{BackgroundLoopKind, RateLimitError};
 use crate::tenant::timeline::logical_size::CurrentLogicalSize;
 use crate::tenant::{
     layer_map::{LayerMap, SearchResult},
@@ -80,7 +81,6 @@ use crate::repository::GcResult;
 use crate::repository::{Key, Value};
 use crate::task_mgr;
 use crate::task_mgr::TaskKind;
-use crate::walredo::WalRedoManager;
 use crate::ZERO_PAGE;
 
 use self::delete::DeleteTimelineFlow;
@@ -200,7 +200,7 @@ pub struct Timeline {
     last_freeze_ts: RwLock<Instant>,
 
     // WAL redo manager
-    walredo_mgr: Arc<dyn WalRedoManager + Sync + Send>,
+    walredo_mgr: Arc<super::WalRedoManager>,
 
     /// Remote storage client.
     /// See [`remote_timeline_client`](super::remote_timeline_client) module comment for details.
@@ -370,7 +370,7 @@ pub enum PageReconstructError {
 
     /// An error happened replaying WAL records
     #[error(transparent)]
-    WalRedo(#[from] crate::walredo::WalRedoError),
+    WalRedo(anyhow::Error),
 }
 
 impl std::fmt::Debug for PageReconstructError {
@@ -684,37 +684,17 @@ impl Timeline {
     ) -> anyhow::Result<()> {
         const ROUNDS: usize = 2;
 
-        static CONCURRENT_COMPACTIONS: once_cell::sync::Lazy<tokio::sync::Semaphore> =
-            once_cell::sync::Lazy::new(|| {
-                let total_threads = *task_mgr::BACKGROUND_RUNTIME_WORKER_THREADS;
-                let permits = usize::max(
-                    1,
-                    // while a lot of the work is done on spawn_blocking, we still do
-                    // repartitioning in the async context. this should give leave us some workers
-                    // unblocked to be blocked on other work, hopefully easing any outside visible
-                    // effects of restarts.
-                    //
-                    // 6/8 is a guess; previously we ran with unlimited 8 and more from
-                    // spawn_blocking.
-                    (total_threads * 3).checked_div(4).unwrap_or(0),
-                );
-                assert_ne!(permits, 0, "we will not be adding in permits later");
-                assert!(
-                    permits < total_threads,
-                    "need threads avail for shorter work"
-                );
-                tokio::sync::Semaphore::new(permits)
-            });
-
         // this wait probably never needs any "long time spent" logging, because we already nag if
         // compaction task goes over it's period (20s) which is quite often in production.
-        let _permit = tokio::select! {
-            permit = CONCURRENT_COMPACTIONS.acquire() => {
-                permit
-            },
-            _ = cancel.cancelled() => {
-                return Ok(());
-            }
+        let _permit = match super::tasks::concurrent_background_tasks_rate_limit(
+            BackgroundLoopKind::Compaction,
+            ctx,
+            cancel,
+        )
+        .await
+        {
+            Ok(permit) => permit,
+            Err(RateLimitError::Cancelled) => return Ok(()),
         };
 
         let last_record_lsn = self.get_last_record_lsn();
@@ -1294,7 +1274,23 @@ impl Timeline {
                 Ok(delta) => Some(delta),
             };
 
-        let layer_metadata = LayerFileMetadata::new(layer_file_size, self.generation);
+        // RemoteTimelineClient holds the metadata on layers' remote generations, so
+        // query it to construct a RemoteLayer.
+        let layer_metadata = self
+            .remote_client
+            .as_ref()
+            .expect("Eviction is not called without remote storage")
+            .get_layer_metadata(&local_layer.filename())
+            .map_err(EvictionError::LayerNotFound)?
+            .ok_or_else(|| {
+                EvictionError::LayerNotFound(anyhow::anyhow!("Layer not in remote metadata"))
+            })?;
+        if layer_metadata.file_size() != layer_file_size {
+            return Err(EvictionError::MetadataInconsistency(format!(
+                "Layer size {layer_file_size} doesn't match remote metadata file size {}",
+                layer_metadata.file_size()
+            )));
+        }
 
         let new_remote_layer = Arc::new(match local_layer.filename() {
             LayerFileName::Image(image_name) => RemoteLayer::new_img(
@@ -1373,6 +1369,10 @@ pub(crate) enum EvictionError {
     /// different objects in memory.
     #[error("layer was no longer part of LayerMap")]
     LayerNotFound(#[source] anyhow::Error),
+
+    /// This should never happen
+    #[error("Metadata inconsistency")]
+    MetadataInconsistency(String),
 }
 
 /// Number of times we will compute partition within a checkpoint distance.
@@ -1470,7 +1470,7 @@ impl Timeline {
         timeline_id: TimelineId,
         tenant_id: TenantId,
         generation: Generation,
-        walredo_mgr: Arc<dyn WalRedoManager + Send + Sync>,
+        walredo_mgr: Arc<super::WalRedoManager>,
         resources: TimelineResources,
         pg_version: u32,
         initial_logical_size_can_start: Option<completion::Barrier>,
@@ -1699,7 +1699,7 @@ impl Timeline {
         disk_consistent_lsn: Lsn,
         index_part: Option<IndexPart>,
     ) -> anyhow::Result<()> {
-        use init::{Decision::*, Discovered, FutureLayer};
+        use init::{Decision::*, Discovered, DismissedLayer};
         use LayerFileName::*;
 
         let mut guard = self.layers.write().await;
@@ -1715,7 +1715,7 @@ impl Timeline {
         // Copy to move into the task we're about to spawn
         let generation = self.generation;
 
-        let (loaded_layers, to_sync, total_physical_size) = tokio::task::spawn_blocking({
+        let (loaded_layers, needs_cleanup, total_physical_size) = tokio::task::spawn_blocking({
             move || {
                 let _g = span.entered();
                 let discovered = init::scan_timeline_dir(&timeline_path)?;
@@ -1764,7 +1764,6 @@ impl Timeline {
                 );
 
                 let mut loaded_layers = Vec::new();
-                let mut needs_upload = Vec::new();
                 let mut needs_cleanup = Vec::new();
                 let mut total_physical_size = 0;
 
@@ -1785,7 +1784,7 @@ impl Timeline {
                             }
                         }
                         Ok(decision) => decision,
-                        Err(FutureLayer { local }) => {
+                        Err(DismissedLayer::Future { local }) => {
                             if local.is_some() {
                                 path.push(name.file_name());
                                 init::cleanup_future_layer(&path, &name, disk_consistent_lsn)?;
@@ -1794,6 +1793,13 @@ impl Timeline {
                             needs_cleanup.push(name);
                             continue;
                         }
+                        Err(DismissedLayer::LocalOnly(local)) => {
+                            path.push(name.file_name());
+                            init::cleanup_local_only_file(&path, &name, &local)?;
+                            path.pop();
+                            // this file never existed remotely, we will have to do rework
+                            continue;
+                        }
                     };
 
                     match &name {
@@ -1802,14 +1808,16 @@ impl Timeline {
                     }
 
                     let status = match &decision {
-                        UseLocal(_) | NeedsUpload(_) => LayerResidenceStatus::Resident,
+                        UseLocal(_) => LayerResidenceStatus::Resident,
                         Evicted(_) | UseRemote { .. } => LayerResidenceStatus::Evicted,
                     };
 
+                    tracing::debug!(layer=%name, ?decision, ?status, "applied");
+
                     let stats = LayerAccessStats::for_loading_layer(status);
 
                     let layer: Arc<dyn PersistentLayer> = match (name, &decision) {
-                        (Delta(d), UseLocal(m) | NeedsUpload(m)) => {
+                        (Delta(d), UseLocal(m)) => {
                             total_physical_size += m.file_size();
                             Arc::new(DeltaLayer::new(
                                 conf,
@@ -1820,7 +1828,7 @@ impl Timeline {
                                 stats,
                             ))
                         }
-                        (Image(i), UseLocal(m) | NeedsUpload(m)) => {
+                        (Image(i), UseLocal(m)) => {
                             total_physical_size += m.file_size();
                             Arc::new(ImageLayer::new(
                                 conf,
@@ -1839,17 +1847,9 @@ impl Timeline {
                         ),
                     };
 
-                    if let NeedsUpload(m) = decision {
-                        needs_upload.push((layer.clone(), m));
-                    }
-
                     loaded_layers.push(layer);
                 }
-                Ok((
-                    loaded_layers,
-                    (needs_upload, needs_cleanup),
-                    total_physical_size,
-                ))
+                Ok((loaded_layers, needs_cleanup, total_physical_size))
             }
         })
         .await
@@ -1861,10 +1861,6 @@ impl Timeline {
         guard.initialize_local_layers(loaded_layers, disk_consistent_lsn + 1);
 
         if let Some(rtc) = self.remote_client.as_ref() {
-            let (needs_upload, needs_cleanup) = to_sync;
-            for (layer, m) in needs_upload {
-                rtc.schedule_layer_file_upload(&layer.layer_desc().filename(), &m)?;
-            }
             rtc.schedule_layer_file_deletion(needs_cleanup)?;
             rtc.schedule_index_upload_for_file_changes()?;
             // Tenant::create_timeline will wait for these uploads to happen before returning, or
@@ -2797,10 +2793,13 @@ impl Timeline {
                 )
             };
 
+        let disk_consistent_lsn = Lsn(lsn_range.end.0 - 1);
+        let old_disk_consistent_lsn = self.disk_consistent_lsn.load();
+
         // The new on-disk layers are now in the layer map. We can remove the
         // in-memory layer from the map now. The flushed layer is stored in
         // the mapping in `create_delta_layer`.
-        {
+        let metadata = {
             let mut guard = self.layers.write().await;
 
             if let Some(ref l) = delta_layer_to_add {
@@ -2816,8 +2815,17 @@ impl Timeline {
             }
 
             guard.finish_flush_l0_layer(delta_layer_to_add, &frozen_layer);
+            if disk_consistent_lsn != old_disk_consistent_lsn {
+                assert!(disk_consistent_lsn > old_disk_consistent_lsn);
+                self.disk_consistent_lsn.store(disk_consistent_lsn);
+
+                // Schedule remote uploads that will reflect our new disk_consistent_lsn
+                Some(self.schedule_uploads(disk_consistent_lsn, layer_paths_to_upload)?)
+            } else {
+                None
+            }
             // release lock on 'layers'
-        }
+        };
 
         // FIXME: between create_delta_layer and the scheduling of the upload in `update_metadata_file`,
         // a compaction can delete the file and then it won't be available for uploads any more.
@@ -2833,28 +2841,22 @@ impl Timeline {
         //
         // TODO: This perhaps should be done in 'flush_frozen_layers', after flushing
         // *all* the layers, to avoid fsyncing the file multiple times.
-        let disk_consistent_lsn = Lsn(lsn_range.end.0 - 1);
-        let old_disk_consistent_lsn = self.disk_consistent_lsn.load();
 
-        // If we were able to advance 'disk_consistent_lsn', save it the metadata file.
-        // After crash, we will restart WAL streaming and processing from that point.
-        if disk_consistent_lsn != old_disk_consistent_lsn {
-            assert!(disk_consistent_lsn > old_disk_consistent_lsn);
-            self.update_metadata_file(disk_consistent_lsn, layer_paths_to_upload)
+        // If we updated our disk_consistent_lsn, persist the updated metadata to local disk.
+        if let Some(metadata) = metadata {
+            save_metadata(self.conf, &self.tenant_id, &self.timeline_id, &metadata)
                 .await
-                .context("update_metadata_file")?;
-            // Also update the in-memory copy
-            self.disk_consistent_lsn.store(disk_consistent_lsn);
+                .context("save_metadata")?;
         }
         Ok(())
     }
 
     /// Update metadata file
-    async fn update_metadata_file(
+    fn schedule_uploads(
         &self,
         disk_consistent_lsn: Lsn,
         layer_paths_to_upload: HashMap<LayerFileName, LayerFileMetadata>,
-    ) -> anyhow::Result<()> {
+    ) -> anyhow::Result<TimelineMetadata> {
         // We can only save a valid 'prev_record_lsn' value on disk if we
         // flushed *all* in-memory changes to disk. We only track
         // 'prev_record_lsn' in memory for the latest processed record, so we
@@ -2891,10 +2893,6 @@ impl Timeline {
             x.unwrap()
         ));
 
-        save_metadata(self.conf, &self.tenant_id, &self.timeline_id, &metadata)
-            .await
-            .context("save_metadata")?;
-
         if let Some(remote_client) = &self.remote_client {
             for (path, layer_metadata) in layer_paths_to_upload {
                 remote_client.schedule_layer_file_upload(&path, &layer_metadata)?;
@@ -2902,6 +2900,20 @@ impl Timeline {
             remote_client.schedule_index_upload_for_metadata_update(&metadata)?;
         }
 
+        Ok(metadata)
+    }
+
+    async fn update_metadata_file(
+        &self,
+        disk_consistent_lsn: Lsn,
+        layer_paths_to_upload: HashMap<LayerFileName, LayerFileMetadata>,
+    ) -> anyhow::Result<()> {
+        let metadata = self.schedule_uploads(disk_consistent_lsn, layer_paths_to_upload)?;
+
+        save_metadata(self.conf, &self.tenant_id, &self.timeline_id, &metadata)
+            .await
+            .context("save_metadata")?;
+
         Ok(())
     }
 
@@ -4327,6 +4339,7 @@ impl Timeline {
                 let img = match self
                     .walredo_mgr
                     .request_redo(key, request_lsn, data.img, data.records, self.pg_version)
+                    .await
                     .context("Failed to reconstruct a page image:")
                 {
                     Ok(img) => img,

pageserver/src/tenant/timeline/eviction_task.rs

@@ -30,6 +30,7 @@ use crate::{
     tenant::{
         config::{EvictionPolicy, EvictionPolicyLayerAccessThreshold},
         storage_layer::PersistentLayer,
+        tasks::{BackgroundLoopKind, RateLimitError},
         timeline::EvictionError,
         LogicalSizeCalculationCause, Tenant,
     },
@@ -129,7 +130,11 @@ impl Timeline {
                     ControlFlow::Continue(()) => (),
                 }
                 let elapsed = start.elapsed();
-                crate::tenant::tasks::warn_when_period_overrun(elapsed, p.period, "eviction");
+                crate::tenant::tasks::warn_when_period_overrun(
+                    elapsed,
+                    p.period,
+                    BackgroundLoopKind::Eviction,
+                );
                 crate::metrics::EVICTION_ITERATION_DURATION
                     .get_metric_with_label_values(&[
                         &format!("{}", p.period.as_secs()),
@@ -150,6 +155,17 @@ impl Timeline {
     ) -> ControlFlow<()> {
         let now = SystemTime::now();
 
+        let _permit = match crate::tenant::tasks::concurrent_background_tasks_rate_limit(
+            BackgroundLoopKind::Eviction,
+            ctx,
+            cancel,
+        )
+        .await
+        {
+            Ok(permit) => permit,
+            Err(RateLimitError::Cancelled) => return ControlFlow::Break(()),
+        };
+
         // If we evict layers but keep cached values derived from those layers, then
         // we face a storm of on-demand downloads after pageserver restart.
         // The reason is that the restart empties the caches, and so, the values
@@ -285,6 +301,10 @@ impl Timeline {
                     warn!(layer = %l, "failed to evict layer: {e}");
                     stats.not_evictable += 1;
                 }
+                Some(Err(EvictionError::MetadataInconsistency(detail))) => {
+                    warn!(layer = %l, "failed to evict layer: {detail}");
+                    stats.not_evictable += 1;
+                }
             }
         }
         if stats.candidates == stats.not_evictable {

pageserver/src/tenant/timeline/init.rs

@@ -72,7 +72,7 @@ pub(super) fn scan_timeline_dir(path: &Utf8Path) -> anyhow::Result<Vec<Discovere
 }
 
 /// Decision on what to do with a layer file after considering its local and remote metadata.
-#[derive(Clone)]
+#[derive(Clone, Debug)]
 pub(super) enum Decision {
     /// The layer is not present locally.
     Evicted(LayerFileMetadata),
@@ -84,27 +84,30 @@ pub(super) enum Decision {
     },
     /// The layer is present locally, and metadata matches.
     UseLocal(LayerFileMetadata),
-    /// The layer is only known locally, it needs to be uploaded.
-    NeedsUpload(LayerFileMetadata),
 }
 
-/// The related layer is is in future compared to disk_consistent_lsn, it must not be loaded.
+/// A layer needs to be left out of the layer map.
 #[derive(Debug)]
-pub(super) struct FutureLayer {
-    /// The local metadata. `None` if the layer is only known through [`IndexPart`].
-    pub(super) local: Option<LayerFileMetadata>,
+pub(super) enum DismissedLayer {
+    /// The related layer is is in future compared to disk_consistent_lsn, it must not be loaded.
+    Future {
+        /// The local metadata. `None` if the layer is only known through [`IndexPart`].
+        local: Option<LayerFileMetadata>,
+    },
+    /// The layer only exists locally.
+    ///
+    /// In order to make crash safe updates to layer map, we must dismiss layers which are only
+    /// found locally or not yet included in the remote `index_part.json`.
+    LocalOnly(LayerFileMetadata),
 }
 
 /// Merges local discoveries and remote [`IndexPart`] to a collection of decisions.
-///
-/// This function should not gain additional reasons to fail than [`FutureLayer`], consider adding
-/// the checks earlier to [`scan_timeline_dir`].
 pub(super) fn reconcile(
     discovered: Vec<(LayerFileName, u64)>,
     index_part: Option<&IndexPart>,
     disk_consistent_lsn: Lsn,
     generation: Generation,
-) -> Vec<(LayerFileName, Result<Decision, FutureLayer>)> {
+) -> Vec<(LayerFileName, Result<Decision, DismissedLayer>)> {
     use Decision::*;
 
     // name => (local, remote)
@@ -142,17 +145,19 @@ pub(super) fn reconcile(
         .into_iter()
         .map(|(name, (local, remote))| {
             let decision = if name.is_in_future(disk_consistent_lsn) {
-                Err(FutureLayer { local })
+                Err(DismissedLayer::Future { local })
             } else {
-                Ok(match (local, remote) {
-                    (Some(local), Some(remote)) if local != remote => UseRemote { local, remote },
-                    (Some(x), Some(_)) => UseLocal(x),
-                    (None, Some(x)) => Evicted(x),
-                    (Some(x), None) => NeedsUpload(x),
+                match (local, remote) {
+                    (Some(local), Some(remote)) if local != remote => {
+                        Ok(UseRemote { local, remote })
+                    }
+                    (Some(x), Some(_)) => Ok(UseLocal(x)),
+                    (None, Some(x)) => Ok(Evicted(x)),
+                    (Some(x), None) => Err(DismissedLayer::LocalOnly(x)),
                     (None, None) => {
                         unreachable!("there must not be any non-local non-remote files")
                     }
-                })
+                }
             };
 
             (name, decision)
@@ -192,14 +197,21 @@ pub(super) fn cleanup_future_layer(
     name: &LayerFileName,
     disk_consistent_lsn: Lsn,
 ) -> anyhow::Result<()> {
-    use LayerFileName::*;
-    let kind = match name {
-        Delta(_) => "delta",
-        Image(_) => "image",
-    };
     // future image layers are allowed to be produced always for not yet flushed to disk
     // lsns stored in InMemoryLayer.
+    let kind = name.kind();
     tracing::info!("found future {kind} layer {name} disk_consistent_lsn is {disk_consistent_lsn}");
-    crate::tenant::timeline::rename_to_backup(path)?;
+    std::fs::remove_file(path)?;
+    Ok(())
+}
+
+pub(super) fn cleanup_local_only_file(
+    path: &Utf8Path,
+    name: &LayerFileName,
+    local: &LayerFileMetadata,
+) -> anyhow::Result<()> {
+    let kind = name.kind();
+    tracing::info!("found local-only {kind} layer {name}, metadata {local:?}");
+    std::fs::remove_file(path)?;
     Ok(())
 }

pageserver/src/tenant/upload_queue.rs

@@ -203,6 +203,18 @@ impl UploadQueue {
             UploadQueue::Stopped(stopped) => Ok(stopped),
         }
     }
+
+    pub(crate) fn get_layer_metadata(
+        &self,
+        name: &LayerFileName,
+    ) -> anyhow::Result<Option<LayerFileMetadata>> {
+        match self {
+            UploadQueue::Stopped(_) | UploadQueue::Uninitialized => {
+                anyhow::bail!("queue is in state {}", self.as_str())
+            }
+            UploadQueue::Initialized(inner) => Ok(inner.latest_files.get(name).cloned()),
+        }
+    }
 }
 
 /// An in-progress upload or delete task.

pageserver/src/walingest.rs

@@ -338,11 +338,20 @@ impl<'a> WalIngest<'a> {
         } else if decoded.xl_rmid == pg_constants::RM_LOGICALMSG_ID {
             let info = decoded.xl_info & pg_constants::XLR_RMGR_INFO_MASK;
             if info == pg_constants::XLOG_LOGICAL_MESSAGE {
-                // This is a convenient way to make the WAL ingestion pause at
-                // particular point in the WAL. For more fine-grained control,
-                // we could peek into the message and only pause if it contains
-                // a particular string, for example, but this is enough for now.
-                crate::failpoint_support::sleep_millis_async!("wal-ingest-logical-message-sleep");
+                let xlrec = XlLogicalMessage::decode(&mut buf);
+                let prefix = std::str::from_utf8(&buf[0..xlrec.prefix_size - 1])?;
+                let message = &buf[xlrec.prefix_size..xlrec.prefix_size + xlrec.message_size];
+                if prefix == "neon-test" {
+                    // This is a convenient way to make the WAL ingestion pause at
+                    // particular point in the WAL. For more fine-grained control,
+                    // we could peek into the message and only pause if it contains
+                    // a particular string, for example, but this is enough for now.
+                    crate::failpoint_support::sleep_millis_async!(
+                        "wal-ingest-logical-message-sleep"
+                    );
+                } else if let Some(path) = prefix.strip_prefix("neon-file:") {
+                    modification.put_file(path, message, ctx).await?;
+                }
             }
         }
 
@@ -459,7 +468,6 @@ impl<'a> WalIngest<'a> {
                         }
                     } else if info == pg_constants::XLOG_HEAP_DELETE {
                         let xlrec = v14::XlHeapDelete::decode(buf);
-                        assert_eq!(0, buf.remaining());
                         if (xlrec.flags & pg_constants::XLH_DELETE_ALL_VISIBLE_CLEARED) != 0 {
                             new_heap_blkno = Some(decoded.blocks[0].blkno);
                         }
@@ -527,7 +535,6 @@ impl<'a> WalIngest<'a> {
                         }
                     } else if info == pg_constants::XLOG_HEAP_DELETE {
                         let xlrec = v15::XlHeapDelete::decode(buf);
-                        assert_eq!(0, buf.remaining());
                         if (xlrec.flags & pg_constants::XLH_DELETE_ALL_VISIBLE_CLEARED) != 0 {
                             new_heap_blkno = Some(decoded.blocks[0].blkno);
                         }
@@ -595,7 +602,6 @@ impl<'a> WalIngest<'a> {
                         }
                     } else if info == pg_constants::XLOG_HEAP_DELETE {
                         let xlrec = v16::XlHeapDelete::decode(buf);
-                        assert_eq!(0, buf.remaining());
                         if (xlrec.flags & pg_constants::XLH_DELETE_ALL_VISIBLE_CLEARED) != 0 {
                             new_heap_blkno = Some(decoded.blocks[0].blkno);
                         }
@@ -771,7 +777,6 @@ impl<'a> WalIngest<'a> {
                     }
                     pg_constants::XLOG_NEON_HEAP_DELETE => {
                         let xlrec = v16::rm_neon::XlNeonHeapDelete::decode(buf);
-                        assert_eq!(0, buf.remaining());
                         if (xlrec.flags & pg_constants::XLH_DELETE_ALL_VISIBLE_CLEARED) != 0 {
                             new_heap_blkno = Some(decoded.blocks[0].blkno);
                         }

pageserver/src/walrecord.rs

@@ -748,6 +748,26 @@ impl XlMultiXactTruncate {
     }
 }
 
+#[repr(C)]
+#[derive(Debug)]
+pub struct XlLogicalMessage {
+    pub db_id: Oid,
+    pub transactional: bool,
+    pub prefix_size: usize,
+    pub message_size: usize,
+}
+
+impl XlLogicalMessage {
+    pub fn decode(buf: &mut Bytes) -> XlLogicalMessage {
+        XlLogicalMessage {
+            db_id: buf.get_u32_le(),
+            transactional: buf.get_u32_le() != 0, // 4-bytes alignment
+            prefix_size: buf.get_u64_le() as usize,
+            message_size: buf.get_u64_le() as usize,
+        }
+    }
+}
+
 /// Main routine to decode a WAL record and figure out which blocks are modified
 //
 // See xlogrecord.h for details

pageserver/src/walredo.rs

@@ -18,38 +18,37 @@
 //! any WAL records, so that even if an attacker hijacks the Postgres
 //! process, he cannot escape out of it.
 //!
+use anyhow::Context;
 use byteorder::{ByteOrder, LittleEndian};
 use bytes::{BufMut, Bytes, BytesMut};
 use nix::poll::*;
 use serde::Serialize;
 use std::collections::VecDeque;
+use std::io;
 use std::io::prelude::*;
-use std::io::{Error, ErrorKind};
 use std::ops::{Deref, DerefMut};
-use std::os::unix::io::{AsRawFd, RawFd};
+use std::os::unix::io::AsRawFd;
 use std::os::unix::prelude::CommandExt;
 use std::process::Stdio;
-use std::process::{Child, ChildStderr, ChildStdin, ChildStdout, Command};
-use std::sync::{Mutex, MutexGuard};
+use std::process::{Child, ChildStdin, ChildStdout, Command};
+use std::sync::{Arc, Mutex, MutexGuard, RwLock};
 use std::time::Duration;
 use std::time::Instant;
-use std::{fs, io};
+use tokio_util::sync::CancellationToken;
 use tracing::*;
-use utils::crashsafe::path_with_suffix_extension;
 use utils::{bin_ser::BeSer, id::TenantId, lsn::Lsn, nonblock::set_nonblock};
 
 #[cfg(feature = "testing")]
 use std::sync::atomic::{AtomicUsize, Ordering};
 
+use crate::config::PageServerConf;
 use crate::metrics::{
     WAL_REDO_BYTES_HISTOGRAM, WAL_REDO_RECORDS_HISTOGRAM, WAL_REDO_RECORD_COUNTER, WAL_REDO_TIME,
     WAL_REDO_WAIT_TIME,
 };
 use crate::pgdatadir_mapping::{key_to_rel_block, key_to_slru_block};
 use crate::repository::Key;
-use crate::task_mgr::BACKGROUND_RUNTIME;
 use crate::walrecord::NeonWalRecord;
-use crate::{config::PageServerConf, TEMP_FILE_SUFFIX};
 use pageserver_api::reltag::{RelTag, SlruKind};
 use postgres_ffi::pg_constants;
 use postgres_ffi::relfile_utils::VISIBILITYMAP_FORKNUM;
@@ -66,37 +65,13 @@ use postgres_ffi::BLCKSZ;
 /// [See more related comments here](https://github.com/postgres/postgres/blob/99c5852e20a0987eca1c38ba0c09329d4076b6a0/src/include/storage/buf_internals.h#L91).
 ///
 #[derive(Debug, PartialEq, Eq, PartialOrd, Ord, Clone, Copy, Serialize)]
-pub struct BufferTag {
+pub(crate) struct BufferTag {
     pub rel: RelTag,
     pub blknum: u32,
 }
 
-///
-/// WAL Redo Manager is responsible for replaying WAL records.
-///
-/// Callers use the WAL redo manager through this abstract interface,
-/// which makes it easy to mock it in tests.
-pub trait WalRedoManager: Send + Sync {
-    /// Apply some WAL records.
-    ///
-    /// The caller passes an old page image, and WAL records that should be
-    /// applied over it. The return value is a new page image, after applying
-    /// the reords.
-    fn request_redo(
-        &self,
-        key: Key,
-        lsn: Lsn,
-        base_img: Option<(Lsn, Bytes)>,
-        records: Vec<(Lsn, NeonWalRecord)>,
-        pg_version: u32,
-    ) -> Result<Bytes, WalRedoError>;
-}
-
 struct ProcessInput {
-    child: NoLeakChild,
     stdin: ChildStdin,
-    stderr_fd: RawFd,
-    stdout_fd: RawFd,
     n_requests: usize,
 }
 
@@ -116,13 +91,7 @@ struct ProcessOutput {
 pub struct PostgresRedoManager {
     tenant_id: TenantId,
     conf: &'static PageServerConf,
-    /// Counter to separate same sized walredo inputs failing at the same millisecond.
-    #[cfg(feature = "testing")]
-    dump_sequence: AtomicUsize,
-
-    stdout: Mutex<Option<ProcessOutput>>,
-    stdin: Mutex<Option<ProcessInput>>,
-    stderr: Mutex<Option<ChildStderr>>,
+    redo_process: RwLock<Option<Arc<WalRedoProcess>>>,
 }
 
 /// Can this request be served by neon redo functions
@@ -140,41 +109,27 @@ fn can_apply_in_neon(rec: &NeonWalRecord) -> bool {
     }
 }
 
-/// An error happened in WAL redo
-#[derive(Debug, thiserror::Error)]
-pub enum WalRedoError {
-    #[error(transparent)]
-    IoError(#[from] std::io::Error),
-
-    #[error("cannot perform WAL redo now")]
-    InvalidState,
-    #[error("cannot perform WAL redo for this request")]
-    InvalidRequest,
-    #[error("cannot perform WAL redo for this record")]
-    InvalidRecord,
-}
-
 ///
 /// Public interface of WAL redo manager
 ///
-impl WalRedoManager for PostgresRedoManager {
+impl PostgresRedoManager {
     ///
     /// Request the WAL redo manager to apply some WAL records
     ///
     /// The WAL redo is handled by a separate thread, so this just sends a request
     /// to the thread and waits for response.
     ///
-    fn request_redo(
+    /// CANCEL SAFETY: NOT CANCEL SAFE.
+    pub async fn request_redo(
         &self,
         key: Key,
         lsn: Lsn,
         base_img: Option<(Lsn, Bytes)>,
         records: Vec<(Lsn, NeonWalRecord)>,
         pg_version: u32,
-    ) -> Result<Bytes, WalRedoError> {
+    ) -> anyhow::Result<Bytes> {
         if records.is_empty() {
-            error!("invalid WAL redo request with no records");
-            return Err(WalRedoError::InvalidRequest);
+            anyhow::bail!("invalid WAL redo request with no records");
         }
 
         let base_img_lsn = base_img.as_ref().map(|p| p.0).unwrap_or(Lsn::INVALID);
@@ -197,6 +152,7 @@ impl WalRedoManager for PostgresRedoManager {
                         self.conf.wal_redo_timeout,
                         pg_version,
                     )
+                    .await
                 };
                 img = Some(result?);
 
@@ -217,6 +173,7 @@ impl WalRedoManager for PostgresRedoManager {
                 self.conf.wal_redo_timeout,
                 pg_version,
             )
+            .await
         }
     }
 }
@@ -230,28 +187,15 @@ impl PostgresRedoManager {
         PostgresRedoManager {
             tenant_id,
             conf,
-            #[cfg(feature = "testing")]
-            dump_sequence: AtomicUsize::default(),
-            stdin: Mutex::new(None),
-            stdout: Mutex::new(None),
-            stderr: Mutex::new(None),
+            redo_process: RwLock::new(None),
         }
     }
 
-    /// Launch process pre-emptively. Should not be needed except for benchmarking.
-    pub fn launch_process(&self, pg_version: u32) -> anyhow::Result<()> {
-        let mut proc = self.stdin.lock().unwrap();
-        if proc.is_none() {
-            self.launch(&mut proc, pg_version)?;
-        }
-        Ok(())
-    }
-
     ///
     /// Process one request for WAL redo using wal-redo postgres
     ///
     #[allow(clippy::too_many_arguments)]
-    fn apply_batch_postgres(
+    async fn apply_batch_postgres(
         &self,
         key: Key,
         lsn: Lsn,
@@ -260,26 +204,45 @@ impl PostgresRedoManager {
         records: &[(Lsn, NeonWalRecord)],
         wal_redo_timeout: Duration,
         pg_version: u32,
-    ) -> Result<Bytes, WalRedoError> {
-        let (rel, blknum) = key_to_rel_block(key).or(Err(WalRedoError::InvalidRecord))?;
+    ) -> anyhow::Result<Bytes> {
+        let (rel, blknum) = key_to_rel_block(key).context("invalid record")?;
         const MAX_RETRY_ATTEMPTS: u32 = 1;
         let start_time = Instant::now();
         let mut n_attempts = 0u32;
         loop {
-            let mut proc = self.stdin.lock().unwrap();
             let lock_time = Instant::now();
 
             // launch the WAL redo process on first use
-            if proc.is_none() {
-                self.launch(&mut proc, pg_version)?;
-            }
+            let proc: Arc<WalRedoProcess> = {
+                let proc_guard = self.redo_process.read().unwrap();
+                match &*proc_guard {
+                    None => {
+                        // "upgrade" to write lock to launch the process
+                        drop(proc_guard);
+                        let mut proc_guard = self.redo_process.write().unwrap();
+                        match &*proc_guard {
+                            None => {
+                                let proc = Arc::new(
+                                    WalRedoProcess::launch(self.conf, self.tenant_id, pg_version)
+                                        .context("launch walredo process")?,
+                                );
+                                *proc_guard = Some(Arc::clone(&proc));
+                                proc
+                            }
+                            Some(proc) => Arc::clone(proc),
+                        }
+                    }
+                    Some(proc) => Arc::clone(proc),
+                }
+            };
+
             WAL_REDO_WAIT_TIME.observe(lock_time.duration_since(start_time).as_secs_f64());
 
             // Relational WAL records are applied using wal-redo-postgres
             let buf_tag = BufferTag { rel, blknum };
-            let result = self
-                .apply_wal_records(proc, buf_tag, &base_img, records, wal_redo_timeout)
-                .map_err(WalRedoError::IoError);
+            let result = proc
+                .apply_wal_records(buf_tag, &base_img, records, wal_redo_timeout)
+                .context("apply_wal_records");
 
             let end_time = Instant::now();
             let duration = end_time.duration_since(lock_time);
@@ -309,32 +272,50 @@ impl PostgresRedoManager {
             // next request will launch a new one.
             if let Err(e) = result.as_ref() {
                 error!(
-                    n_attempts,
-                    "error applying {} WAL records {}..{} ({} bytes) to base image with LSN {} to reconstruct page image at LSN {}: {}",
+                    "error applying {} WAL records {}..{} ({} bytes) to base image with LSN {} to reconstruct page image at LSN {} n_attempts={}: {:?}",
                     records.len(),
                     records.first().map(|p| p.0).unwrap_or(Lsn(0)),
                     records.last().map(|p| p.0).unwrap_or(Lsn(0)),
                     nbytes,
                     base_img_lsn,
                     lsn,
-                    utils::error::report_compact_sources(e),
+                    n_attempts,
+                    e,
                 );
-                // self.stdin only holds stdin & stderr as_raw_fd().
-                // Dropping it as part of take() doesn't close them.
-                // The owning objects (ChildStdout and ChildStderr) are stored in
-                // self.stdout and self.stderr, respsectively.
-                // We intentionally keep them open here to avoid a race between
-                // currently running `apply_wal_records()` and a `launch()` call
-                // after we return here.
-                // The currently running `apply_wal_records()` must not read from
-                // the newly launched process.
-                // By keeping self.stdout and self.stderr open here, `launch()` will
-                // get other file descriptors for the new child's stdout and stderr,
-                // and hence the current `apply_wal_records()` calls will observe
-                //  `output.stdout.as_raw_fd() != stdout_fd` .
-                if let Some(proc) = self.stdin.lock().unwrap().take() {
-                    proc.child.kill_and_wait();
+                // Avoid concurrent callers hitting the same issue.
+                // We can't prevent it from happening because we want to enable parallelism.
+                {
+                    let mut guard = self.redo_process.write().unwrap();
+                    match &*guard {
+                        Some(current_field_value) => {
+                            if Arc::ptr_eq(current_field_value, &proc) {
+                                // We're the first to observe an error from `proc`, it's our job to take it out of rotation.
+                                *guard = None;
+                            }
+                        }
+                        None => {
+                            // Another thread was faster to observe the error, and already took the process out of rotation.
+                        }
+                    }
                 }
+                // NB: there may still be other concurrent threads using `proc`.
+                // The last one will send SIGKILL when the underlying Arc reaches refcount 0.
+                // NB: it's important to drop(proc) after drop(guard). Otherwise we'd keep
+                // holding the lock while waiting for the process to exit.
+                // NB: the drop impl blocks the current threads with a wait() system call for
+                // the child process. We dropped the `guard` above so that other threads aren't
+                // affected. But, it's good that the current thread _does_ block to wait.
+                // If we instead deferred the waiting into the background / to tokio, it could
+                // happen that if walredo always fails immediately, we spawn processes faster
+                // than we can SIGKILL & `wait` for them to exit. By doing it the way we do here,
+                // we limit this risk of run-away to at most $num_runtimes * $num_executor_threads.
+                // This probably needs revisiting at some later point.
+                let mut wait_done = proc.stderr_logger_task_done.clone();
+                drop(proc);
+                wait_done
+                    .wait_for(|v| *v)
+                    .await
+                    .expect("we use scopeguard to ensure we always send `true` to the channel before dropping the sender");
             } else if n_attempts != 0 {
                 info!(n_attempts, "retried walredo succeeded");
             }
@@ -354,7 +335,7 @@ impl PostgresRedoManager {
         lsn: Lsn,
         base_img: Option<Bytes>,
         records: &[(Lsn, NeonWalRecord)],
-    ) -> Result<Bytes, WalRedoError> {
+    ) -> anyhow::Result<Bytes> {
         let start_time = Instant::now();
 
         let mut page = BytesMut::new();
@@ -363,8 +344,7 @@ impl PostgresRedoManager {
             page.extend_from_slice(&fpi[..]);
         } else {
             // All the current WAL record types that we can handle require a base image.
-            error!("invalid neon WAL redo request with no base image");
-            return Err(WalRedoError::InvalidRequest);
+            anyhow::bail!("invalid neon WAL redo request with no base image");
         }
 
         // Apply all the WAL records in the batch
@@ -392,14 +372,13 @@ impl PostgresRedoManager {
         page: &mut BytesMut,
         _record_lsn: Lsn,
         record: &NeonWalRecord,
-    ) -> Result<(), WalRedoError> {
+    ) -> anyhow::Result<()> {
         match record {
             NeonWalRecord::Postgres {
                 will_init: _,
                 rec: _,
             } => {
-                error!("tried to pass postgres wal record to neon WAL redo");
-                return Err(WalRedoError::InvalidRequest);
+                anyhow::bail!("tried to pass postgres wal record to neon WAL redo");
             }
             NeonWalRecord::ClearVisibilityMapFlags {
                 new_heap_blkno,
@@ -407,7 +386,7 @@ impl PostgresRedoManager {
                 flags,
             } => {
                 // sanity check that this is modifying the correct relation
-                let (rel, blknum) = key_to_rel_block(key).or(Err(WalRedoError::InvalidRecord))?;
+                let (rel, blknum) = key_to_rel_block(key).context("invalid record")?;
                 assert!(
                     rel.forknum == VISIBILITYMAP_FORKNUM,
                     "ClearVisibilityMapFlags record on unexpected rel {}",
@@ -445,7 +424,7 @@ impl PostgresRedoManager {
             // same effects as the corresponding Postgres WAL redo function.
             NeonWalRecord::ClogSetCommitted { xids, timestamp } => {
                 let (slru_kind, segno, blknum) =
-                    key_to_slru_block(key).or(Err(WalRedoError::InvalidRecord))?;
+                    key_to_slru_block(key).context("invalid record")?;
                 assert_eq!(
                     slru_kind,
                     SlruKind::Clog,
@@ -495,7 +474,7 @@ impl PostgresRedoManager {
             }
             NeonWalRecord::ClogSetAborted { xids } => {
                 let (slru_kind, segno, blknum) =
-                    key_to_slru_block(key).or(Err(WalRedoError::InvalidRecord))?;
+                    key_to_slru_block(key).context("invalid record")?;
                 assert_eq!(
                     slru_kind,
                     SlruKind::Clog,
@@ -526,7 +505,7 @@ impl PostgresRedoManager {
             }
             NeonWalRecord::MultixactOffsetCreate { mid, moff } => {
                 let (slru_kind, segno, blknum) =
-                    key_to_slru_block(key).or(Err(WalRedoError::InvalidRecord))?;
+                    key_to_slru_block(key).context("invalid record")?;
                 assert_eq!(
                     slru_kind,
                     SlruKind::MultiXactOffsets,
@@ -559,7 +538,7 @@ impl PostgresRedoManager {
             }
             NeonWalRecord::MultixactMembersCreate { moff, members } => {
                 let (slru_kind, segno, blknum) =
-                    key_to_slru_block(key).or(Err(WalRedoError::InvalidRecord))?;
+                    key_to_slru_block(key).context("invalid record")?;
                 assert_eq!(
                     slru_kind,
                     SlruKind::MultiXactMembers,
@@ -639,44 +618,33 @@ impl<C: CommandExt> CloseFileDescriptors for C {
     }
 }
 
-impl PostgresRedoManager {
+struct WalRedoProcess {
+    #[allow(dead_code)]
+    conf: &'static PageServerConf,
+    tenant_id: TenantId,
+    // Some() on construction, only becomes None on Drop.
+    child: Option<NoLeakChild>,
+    stdout: Mutex<ProcessOutput>,
+    stdin: Mutex<ProcessInput>,
+    stderr_logger_cancel: CancellationToken,
+    stderr_logger_task_done: tokio::sync::watch::Receiver<bool>,
+    /// Counter to separate same sized walredo inputs failing at the same millisecond.
+    #[cfg(feature = "testing")]
+    dump_sequence: AtomicUsize,
+}
+
+impl WalRedoProcess {
     //
     // Start postgres binary in special WAL redo mode.
     //
-    #[instrument(skip_all,fields(tenant_id=%self.tenant_id, pg_version=pg_version))]
+    #[instrument(skip_all,fields(tenant_id=%tenant_id, pg_version=pg_version))]
     fn launch(
-        &self,
-        input: &mut MutexGuard<Option<ProcessInput>>,
+        conf: &'static PageServerConf,
+        tenant_id: TenantId,
         pg_version: u32,
-    ) -> Result<(), Error> {
-        // Previous versions of wal-redo required data directory and that directories
-        // occupied some space on disk. Remove it if we face it.
-        //
-        // This code could be dropped after one release cycle.
-        let legacy_datadir = path_with_suffix_extension(
-            self.conf
-                .tenant_path(&self.tenant_id)
-                .join("wal-redo-datadir"),
-            TEMP_FILE_SUFFIX,
-        );
-        if legacy_datadir.exists() {
-            info!("legacy wal-redo datadir {legacy_datadir:?} exists, removing");
-            fs::remove_dir_all(&legacy_datadir).map_err(|e| {
-                Error::new(
-                    e.kind(),
-                    format!("legacy wal-redo datadir {legacy_datadir:?} removal failure: {e}"),
-                )
-            })?;
-        }
-
-        let pg_bin_dir_path = self
-            .conf
-            .pg_bin_dir(pg_version)
-            .map_err(|e| Error::new(ErrorKind::Other, format!("incorrect pg_bin_dir path: {e}")))?;
-        let pg_lib_dir_path = self
-            .conf
-            .pg_lib_dir(pg_version)
-            .map_err(|e| Error::new(ErrorKind::Other, format!("incorrect pg_lib_dir path: {e}")))?;
+    ) -> anyhow::Result<Self> {
+        let pg_bin_dir_path = conf.pg_bin_dir(pg_version).context("pg_bin_dir")?; // TODO these should be infallible.
+        let pg_lib_dir_path = conf.pg_lib_dir(pg_version).context("pg_lib_dir")?;
 
         // Start postgres itself
         let child = Command::new(pg_bin_dir_path.join("postgres"))
@@ -697,13 +665,8 @@ impl PostgresRedoManager {
             // as close-on-exec by default, but that's not enough, since we use
             // libraries that directly call libc open without setting that flag.
             .close_fds()
-            .spawn_no_leak_child(self.tenant_id)
-            .map_err(|e| {
-                Error::new(
-                    e.kind(),
-                    format!("postgres --wal-redo command failed to start: {}", e),
-                )
-            })?;
+            .spawn_no_leak_child(tenant_id)
+            .context("spawn process")?;
 
         let mut child = scopeguard::guard(child, |child| {
             error!("killing wal-redo-postgres process due to a problem during launch");
@@ -713,7 +676,6 @@ impl PostgresRedoManager {
         let stdin = child.stdin.take().unwrap();
         let stdout = child.stdout.take().unwrap();
         let stderr = child.stderr.take().unwrap();
-
         macro_rules! set_nonblock_or_log_err {
             ($file:ident) => {{
                 let res = set_nonblock($file.as_raw_fd());
@@ -727,39 +689,108 @@ impl PostgresRedoManager {
         set_nonblock_or_log_err!(stdout)?;
         set_nonblock_or_log_err!(stderr)?;
 
+        let mut stderr = tokio::io::unix::AsyncFd::new(stderr).context("AsyncFd::with_interest")?;
+
         // all fallible operations post-spawn are complete, so get rid of the guard
         let child = scopeguard::ScopeGuard::into_inner(child);
 
-        **input = Some(ProcessInput {
-            child,
-            stdout_fd: stdout.as_raw_fd(),
-            stderr_fd: stderr.as_raw_fd(),
-            stdin,
-            n_requests: 0,
+        let stderr_logger_cancel = CancellationToken::new();
+        let (stderr_logger_task_done_tx, stderr_logger_task_done_rx) =
+            tokio::sync::watch::channel(false);
+        tokio::spawn({
+            let stderr_logger_cancel = stderr_logger_cancel.clone();
+            async move {
+                scopeguard::defer! {
+                    debug!("wal-redo-postgres stderr_logger_task finished");
+                    let _ = stderr_logger_task_done_tx.send(true);
+                }
+                debug!("wal-redo-postgres stderr_logger_task started");
+                loop {
+                    // NB: we purposefully don't do a select! for the cancellation here.
+                    // The cancellation would likely cause us to miss stderr messages.
+                    // We can rely on this to return from .await because when we SIGKILL
+                    // the child, the writing end of the stderr pipe gets closed.
+                    match stderr.readable_mut().await {
+                        Ok(mut guard) => {
+                            let mut errbuf = [0; 16384];
+                            let res = guard.try_io(|fd| {
+                                use std::io::Read;
+                                fd.get_mut().read(&mut errbuf)
+                            });
+                            match res {
+                                Ok(Ok(0)) => {
+                                    // it closed the stderr pipe
+                                    break;
+                                }
+                                Ok(Ok(n)) => {
+                                    // The message might not be split correctly into lines here. But this is
+                                    // good enough, the important thing is to get the message to the log.
+                                    let output = String::from_utf8_lossy(&errbuf[0..n]).to_string();
+                                    error!(output, "received output");
+                                },
+                                Ok(Err(e)) => {
+                                    error!(error = ?e, "read() error, waiting for cancellation");
+                                    stderr_logger_cancel.cancelled().await;
+                                    error!(error = ?e, "read() error, cancellation complete");
+                                    break;
+                                }
+                                Err(e) => {
+                                    let _e: tokio::io::unix::TryIoError = e;
+                                    // the read() returned WouldBlock, that's expected
+                                }
+                            }
+                        }
+                        Err(e) => {
+                            error!(error = ?e, "read() error, waiting for cancellation");
+                            stderr_logger_cancel.cancelled().await;
+                            error!(error = ?e, "read() error, cancellation complete");
+                            break;
+                        }
+                    }
+                }
+            }.instrument(tracing::info_span!(parent: None, "wal-redo-postgres-stderr", pid = child.id(), tenant_id = %tenant_id, %pg_version))
         });
 
-        *self.stdout.lock().unwrap() = Some(ProcessOutput {
-            stdout,
-            pending_responses: VecDeque::new(),
-            n_processed_responses: 0,
-        });
-        *self.stderr.lock().unwrap() = Some(stderr);
+        Ok(Self {
+            conf,
+            tenant_id,
+            child: Some(child),
+            stdin: Mutex::new(ProcessInput {
+                stdin,
+                n_requests: 0,
+            }),
+            stdout: Mutex::new(ProcessOutput {
+                stdout,
+                pending_responses: VecDeque::new(),
+                n_processed_responses: 0,
+            }),
+            stderr_logger_cancel,
+            stderr_logger_task_done: stderr_logger_task_done_rx,
+            #[cfg(feature = "testing")]
+            dump_sequence: AtomicUsize::default(),
+        })
+    }
 
-        Ok(())
+    fn id(&self) -> u32 {
+        self.child
+            .as_ref()
+            .expect("must not call this during Drop")
+            .id()
     }
 
     // Apply given WAL records ('records') over an old page image. Returns
     // new page image.
     //
-    #[instrument(skip_all, fields(tenant_id=%self.tenant_id, pid=%input.as_ref().unwrap().child.id()))]
+    #[instrument(skip_all, fields(tenant_id=%self.tenant_id, pid=%self.id()))]
     fn apply_wal_records(
         &self,
-        input: MutexGuard<Option<ProcessInput>>,
         tag: BufferTag,
         base_img: &Option<Bytes>,
         records: &[(Lsn, NeonWalRecord)],
         wal_redo_timeout: Duration,
-    ) -> Result<Bytes, std::io::Error> {
+    ) -> anyhow::Result<Bytes> {
+        let input = self.stdin.lock().unwrap();
+
         // Serialize all the messages to send the WAL redo process first.
         //
         // This could be problematic if there are millions of records to replay,
@@ -782,10 +813,7 @@ impl PostgresRedoManager {
             {
                 build_apply_record_msg(*lsn, postgres_rec, &mut writebuf);
             } else {
-                return Err(Error::new(
-                    ErrorKind::Other,
-                    "tried to pass neon wal record to postgres WAL redo",
-                ));
+                anyhow::bail!("tried to pass neon wal record to postgres WAL redo");
             }
         }
         build_get_page_msg(tag, &mut writebuf);
@@ -805,77 +833,38 @@ impl PostgresRedoManager {
     fn apply_wal_records0(
         &self,
         writebuf: &[u8],
-        mut input: MutexGuard<Option<ProcessInput>>,
+        input: MutexGuard<ProcessInput>,
         wal_redo_timeout: Duration,
-    ) -> Result<Bytes, std::io::Error> {
-        let proc = input.as_mut().unwrap();
+    ) -> anyhow::Result<Bytes> {
+        let mut proc = { input }; // TODO: remove this legacy rename, but this keep the patch small.
         let mut nwrite = 0usize;
-        let stdout_fd = proc.stdout_fd;
 
-        // Prepare for calling poll()
-        let mut pollfds = [
-            PollFd::new(proc.stdin.as_raw_fd(), PollFlags::POLLOUT),
-            PollFd::new(proc.stderr_fd, PollFlags::POLLIN),
-            PollFd::new(stdout_fd, PollFlags::POLLIN),
-        ];
+        let mut stdin_pollfds = [PollFd::new(proc.stdin.as_raw_fd(), PollFlags::POLLOUT)];
 
-        // We do two things simultaneously: send the old base image and WAL records to
-        // the child process's stdin and forward any logging
-        // information that the child writes to its stderr to the page server's log.
         while nwrite < writebuf.len() {
             let n = loop {
-                match nix::poll::poll(&mut pollfds[0..2], wal_redo_timeout.as_millis() as i32) {
+                match nix::poll::poll(&mut stdin_pollfds[..], wal_redo_timeout.as_millis() as i32) {
                     Err(nix::errno::Errno::EINTR) => continue,
                     res => break res,
                 }
             }?;
 
             if n == 0 {
-                return Err(Error::new(ErrorKind::Other, "WAL redo timed out"));
-            }
-
-            // If we have some messages in stderr, forward them to the log.
-            let err_revents = pollfds[1].revents().unwrap();
-            if err_revents & (PollFlags::POLLERR | PollFlags::POLLIN) != PollFlags::empty() {
-                let mut errbuf: [u8; 16384] = [0; 16384];
-                let mut stderr_guard = self.stderr.lock().unwrap();
-                let stderr = stderr_guard.as_mut().unwrap();
-                let len = stderr.read(&mut errbuf)?;
-
-                // The message might not be split correctly into lines here. But this is
-                // good enough, the important thing is to get the message to the log.
-                if len > 0 {
-                    error!(
-                        "wal-redo-postgres: {}",
-                        String::from_utf8_lossy(&errbuf[0..len])
-                    );
-
-                    // To make sure we capture all log from the process if it fails, keep
-                    // reading from the stderr, before checking the stdout.
-                    continue;
-                }
-            } else if err_revents.contains(PollFlags::POLLHUP) {
-                return Err(Error::new(
-                    ErrorKind::BrokenPipe,
-                    "WAL redo process closed its stderr unexpectedly",
-                ));
+                anyhow::bail!("WAL redo timed out");
             }
 
             // If 'stdin' is writeable, do write.
-            let in_revents = pollfds[0].revents().unwrap();
+            let in_revents = stdin_pollfds[0].revents().unwrap();
             if in_revents & (PollFlags::POLLERR | PollFlags::POLLOUT) != PollFlags::empty() {
                 nwrite += proc.stdin.write(&writebuf[nwrite..])?;
             } else if in_revents.contains(PollFlags::POLLHUP) {
                 // We still have more data to write, but the process closed the pipe.
-                return Err(Error::new(
-                    ErrorKind::BrokenPipe,
-                    "WAL redo process closed its stdin unexpectedly",
-                ));
+                anyhow::bail!("WAL redo process closed its stdin unexpectedly");
             }
         }
         let request_no = proc.n_requests;
         proc.n_requests += 1;
-        drop(input);
+        drop(proc);
 
         // To improve walredo performance we separate sending requests and receiving
         // responses. Them are protected by different mutexes (output and input).
@@ -889,23 +878,8 @@ impl PostgresRedoManager {
         // pending responses ring buffer and truncate all empty elements from the front,
         // advancing processed responses number.
 
-        let mut output_guard = self.stdout.lock().unwrap();
-        let output = output_guard.as_mut().unwrap();
-        if output.stdout.as_raw_fd() != stdout_fd {
-            // If stdout file descriptor is changed then it means that walredo process is crashed and restarted.
-            // As far as ProcessInput and ProcessOutout are protected by different mutexes,
-            // it can happen that we send request to one process and waiting response from another.
-            // To prevent such situation we compare stdout file descriptors.
-            // As far as old stdout pipe is destroyed only after new one is created,
-            // it can not reuse the same file descriptor, so this check is safe.
-            //
-            // Cross-read this with the comment in apply_batch_postgres if result.is_err().
-            // That's where we kill the child process.
-            return Err(Error::new(
-                ErrorKind::BrokenPipe,
-                "WAL redo process closed its stdout unexpectedly",
-            ));
-        }
+        let mut output = self.stdout.lock().unwrap();
+        let mut stdout_pollfds = [PollFd::new(output.stdout.as_raw_fd(), PollFlags::POLLIN)];
         let n_processed_responses = output.n_processed_responses;
         while n_processed_responses + output.pending_responses.len() <= request_no {
             // We expect the WAL redo process to respond with an 8k page image. We read it
@@ -916,52 +890,25 @@ impl PostgresRedoManager {
                 // We do two things simultaneously: reading response from stdout
                 // and forward any logging information that the child writes to its stderr to the page server's log.
                 let n = loop {
-                    match nix::poll::poll(&mut pollfds[1..3], wal_redo_timeout.as_millis() as i32) {
+                    match nix::poll::poll(
+                        &mut stdout_pollfds[..],
+                        wal_redo_timeout.as_millis() as i32,
+                    ) {
                         Err(nix::errno::Errno::EINTR) => continue,
                         res => break res,
                     }
                 }?;
 
                 if n == 0 {
-                    return Err(Error::new(ErrorKind::Other, "WAL redo timed out"));
-                }
-
-                // If we have some messages in stderr, forward them to the log.
-                let err_revents = pollfds[1].revents().unwrap();
-                if err_revents & (PollFlags::POLLERR | PollFlags::POLLIN) != PollFlags::empty() {
-                    let mut errbuf: [u8; 16384] = [0; 16384];
-                    let mut stderr_guard = self.stderr.lock().unwrap();
-                    let stderr = stderr_guard.as_mut().unwrap();
-                    let len = stderr.read(&mut errbuf)?;
-
-                    // The message might not be split correctly into lines here. But this is
-                    // good enough, the important thing is to get the message to the log.
-                    if len > 0 {
-                        error!(
-                            "wal-redo-postgres: {}",
-                            String::from_utf8_lossy(&errbuf[0..len])
-                        );
-
-                        // To make sure we capture all log from the process if it fails, keep
-                        // reading from the stderr, before checking the stdout.
-                        continue;
-                    }
-                } else if err_revents.contains(PollFlags::POLLHUP) {
-                    return Err(Error::new(
-                        ErrorKind::BrokenPipe,
-                        "WAL redo process closed its stderr unexpectedly",
-                    ));
+                    anyhow::bail!("WAL redo timed out");
                 }
 
                 // If we have some data in stdout, read it to the result buffer.
-                let out_revents = pollfds[2].revents().unwrap();
+                let out_revents = stdout_pollfds[0].revents().unwrap();
                 if out_revents & (PollFlags::POLLERR | PollFlags::POLLIN) != PollFlags::empty() {
                     nresult += output.stdout.read(&mut resultbuf[nresult..])?;
                 } else if out_revents.contains(PollFlags::POLLHUP) {
-                    return Err(Error::new(
-                        ErrorKind::BrokenPipe,
-                        "WAL redo process closed its stdout unexpectedly",
-                    ));
+                    anyhow::bail!("WAL redo process closed its stdout unexpectedly");
                 }
             }
             output
@@ -1047,6 +994,17 @@ impl PostgresRedoManager {
     fn record_and_log(&self, _: &[u8]) {}
 }
 
+impl Drop for WalRedoProcess {
+    fn drop(&mut self) {
+        self.child
+            .take()
+            .expect("we only do this once")
+            .kill_and_wait();
+        self.stderr_logger_cancel.cancel();
+        // no way to wait for stderr_logger_task from Drop because that is async only
+    }
+}
+
 /// Wrapper type around `std::process::Child` which guarantees that the child
 /// will be killed and waited-for by this process before being dropped.
 struct NoLeakChild {
@@ -1125,7 +1083,7 @@ impl Drop for NoLeakChild {
         // Offload the kill+wait of the child process into the background.
         // If someone stops the runtime, we'll leak the child process.
         // We can ignore that case because we only stop the runtime on pageserver exit.
-        BACKGROUND_RUNTIME.spawn(async move {
+        tokio::runtime::Handle::current().spawn(async move {
             tokio::task::spawn_blocking(move || {
                 // Intentionally don't inherit the tracing context from whoever is dropping us.
                 // This thread here is going to outlive of our dropper.
@@ -1194,15 +1152,15 @@ fn build_get_page_msg(tag: BufferTag, buf: &mut Vec<u8>) {
 
 #[cfg(test)]
 mod tests {
-    use super::{PostgresRedoManager, WalRedoManager};
+    use super::PostgresRedoManager;
     use crate::repository::Key;
     use crate::{config::PageServerConf, walrecord::NeonWalRecord};
     use bytes::Bytes;
     use std::str::FromStr;
     use utils::{id::TenantId, lsn::Lsn};
 
-    #[test]
-    fn short_v14_redo() {
+    #[tokio::test]
+    async fn short_v14_redo() {
         let expected = std::fs::read("fixtures/short_v14_redo.page").unwrap();
 
         let h = RedoHarness::new().unwrap();
@@ -1223,13 +1181,14 @@ mod tests {
                 short_records(),
                 14,
             )
+            .await
             .unwrap();
 
         assert_eq!(&expected, &*page);
     }
 
-    #[test]
-    fn short_v14_fails_for_wrong_key_but_returns_zero_page() {
+    #[tokio::test]
+    async fn short_v14_fails_for_wrong_key_but_returns_zero_page() {
         let h = RedoHarness::new().unwrap();
 
         let page = h
@@ -1249,6 +1208,7 @@ mod tests {
                 short_records(),
                 14,
             )
+            .await
             .unwrap();
 
         // TODO: there will be some stderr printout, which is forwarded to tracing that could
@@ -1256,6 +1216,22 @@ mod tests {
         assert_eq!(page, crate::ZERO_PAGE);
     }
 
+    #[tokio::test]
+    async fn test_stderr() {
+        let h = RedoHarness::new().unwrap();
+        h
+            .manager
+            .request_redo(
+                Key::from_i128(0),
+                Lsn::INVALID,
+                None,
+                short_records(),
+                16, /* 16 currently produces stderr output on startup, which adds a nice extra edge */
+            )
+            .await
+            .unwrap_err();
+    }
+
     #[allow(clippy::octal_escapes)]
     fn short_records() -> Vec<(Lsn, NeonWalRecord)> {
         vec![
@@ -1284,6 +1260,8 @@ mod tests {
 
     impl RedoHarness {
         fn new() -> anyhow::Result<Self> {
+            crate::tenant::harness::setup_logging();
+
             let repo_dir = camino_tempfile::tempdir()?;
             let conf = PageServerConf::dummy_conf(repo_dir.path().to_path_buf());
             let conf = Box::leak(Box::new(conf));

pgxn/neon/Makefile

@@ -23,6 +23,23 @@ EXTENSION = neon
 DATA = neon--1.0.sql
 PGFILEDESC = "neon - cloud storage for PostgreSQL"
 
+EXTRA_CLEAN = \
+	libwalproposer.a
+
+WALPROP_OBJS = \
+	$(WIN32RES) \
+	walproposer.o \
+	neon_utils.o \
+	walproposer_compat.o
+
+.PHONY: walproposer-lib
+walproposer-lib: CPPFLAGS += -DWALPROPOSER_LIB
+walproposer-lib: libwalproposer.a;
+
+.PHONY: libwalproposer.a
+libwalproposer.a: $(WALPROP_OBJS)
+	rm -f $@
+	$(AR) $(AROPT) $@ $^
 
 PG_CONFIG = pg_config
 PGXS := $(shell $(PG_CONFIG) --pgxs)

pgxn/neon/control_plane_connector.c

@@ -741,13 +741,6 @@ NeonProcessUtility(
 			break;
 		case T_DropdbStmt:
 			HandleDropDb(castNode(DropdbStmt, parseTree));
-			/*
-			 * We do this here to hack around the fact that Postgres performs the drop
-			 * INSIDE of standard_ProcessUtility, which means that if we try to
-			 * abort the drop normally it'll be too late. DROP DATABASE can't be inside
-			 * of a transaction block anyway, so this should be fine to do.
-			 */
-			NeonXactCallback(XACT_EVENT_PRE_COMMIT, NULL);
 			break;
 		case T_CreateRoleStmt:
 			HandleCreateRole(castNode(CreateRoleStmt, parseTree));

pgxn/neon/pagestore_smgr.c

@@ -63,7 +63,6 @@
 #include "storage/md.h"
 #include "pgstat.h"
 
-
 #if PG_VERSION_NUM >= 150000
 #include "access/xlogutils.h"
 #include "access/xlogrecovery.h"
@@ -1395,12 +1394,6 @@ neon_get_request_lsn(bool *latest, NRelFileInfo rinfo, ForkNumber forknum, Block
 		elog(DEBUG1, "neon_get_request_lsn GetXLogReplayRecPtr %X/%X request lsn 0 ",
 			 (uint32) ((lsn) >> 32), (uint32) (lsn));
 	}
-	else if (am_walsender)
-	{
-		*latest = true;
-		lsn = InvalidXLogRecPtr;
-		elog(DEBUG1, "am walsender neon_get_request_lsn lsn 0 ");
-	}
 	else
 	{
 		XLogRecPtr	flushlsn;

pgxn/neon/walproposer.c

@@ -79,7 +79,7 @@ static int	CompareLsn(const void *a, const void *b);
 static char *FormatSafekeeperState(SafekeeperState state);
 static void AssertEventsOkForState(uint32 events, Safekeeper *sk);
 static uint32 SafekeeperStateDesiredEvents(SafekeeperState state);
-static char *FormatEvents(uint32 events);
+static char *FormatEvents(WalProposer *wp, uint32 events);
 
 WalProposer *
 WalProposerCreate(WalProposerConfig *config, walproposer_api api)
@@ -98,7 +98,7 @@ WalProposerCreate(WalProposerConfig *config, walproposer_api api)
 		port = strchr(host, ':');
 		if (port == NULL)
 		{
-			elog(FATAL, "port is not specified");
+			walprop_log(FATAL, "port is not specified");
 		}
 		*port++ = '\0';
 		sep = strchr(port, ',');
@@ -106,12 +106,11 @@ WalProposerCreate(WalProposerConfig *config, walproposer_api api)
 			*sep++ = '\0';
 		if (wp->n_safekeepers + 1 >= MAX_SAFEKEEPERS)
 		{
-			elog(FATAL, "Too many safekeepers");
+			walprop_log(FATAL, "Too many safekeepers");
 		}
 		wp->safekeeper[wp->n_safekeepers].host = host;
 		wp->safekeeper[wp->n_safekeepers].port = port;
 		wp->safekeeper[wp->n_safekeepers].state = SS_OFFLINE;
-		wp->safekeeper[wp->n_safekeepers].conn = NULL;
 		wp->safekeeper[wp->n_safekeepers].wp = wp;
 
 		{
@@ -122,13 +121,11 @@ WalProposerCreate(WalProposerConfig *config, walproposer_api api)
 							   "host=%s port=%s dbname=replication options='-c timeline_id=%s tenant_id=%s'",
 							   sk->host, sk->port, wp->config->neon_timeline, wp->config->neon_tenant);
 			if (written > MAXCONNINFO || written < 0)
-				elog(FATAL, "could not create connection string for safekeeper %s:%s", sk->host, sk->port);
+				walprop_log(FATAL, "could not create connection string for safekeeper %s:%s", sk->host, sk->port);
 		}
 
 		initStringInfo(&wp->safekeeper[wp->n_safekeepers].outbuf);
-		wp->safekeeper[wp->n_safekeepers].xlogreader = wp->api.wal_reader_allocate();
-		if (wp->safekeeper[wp->n_safekeepers].xlogreader == NULL)
-			elog(FATAL, "Failed to allocate xlog reader");
+		wp->api.wal_reader_allocate(&wp->safekeeper[wp->n_safekeepers]);
 		wp->safekeeper[wp->n_safekeepers].flushWrite = false;
 		wp->safekeeper[wp->n_safekeepers].startStreamingAt = InvalidXLogRecPtr;
 		wp->safekeeper[wp->n_safekeepers].streamingAt = InvalidXLogRecPtr;
@@ -136,7 +133,7 @@ WalProposerCreate(WalProposerConfig *config, walproposer_api api)
 	}
 	if (wp->n_safekeepers < 1)
 	{
-		elog(FATAL, "Safekeepers addresses are not specified");
+		walprop_log(FATAL, "Safekeepers addresses are not specified");
 	}
 	wp->quorum = wp->n_safekeepers / 2 + 1;
 
@@ -144,27 +141,47 @@ WalProposerCreate(WalProposerConfig *config, walproposer_api api)
 	wp->greetRequest.tag = 'g';
 	wp->greetRequest.protocolVersion = SK_PROTOCOL_VERSION;
 	wp->greetRequest.pgVersion = PG_VERSION_NUM;
-	wp->api.strong_random(&wp->greetRequest.proposerId, sizeof(wp->greetRequest.proposerId));
+	wp->api.strong_random(wp, &wp->greetRequest.proposerId, sizeof(wp->greetRequest.proposerId));
 	wp->greetRequest.systemId = wp->config->systemId;
 	if (!wp->config->neon_timeline)
-		elog(FATAL, "neon.timeline_id is not provided");
+		walprop_log(FATAL, "neon.timeline_id is not provided");
 	if (*wp->config->neon_timeline != '\0' &&
 		!HexDecodeString(wp->greetRequest.timeline_id, wp->config->neon_timeline, 16))
-		elog(FATAL, "Could not parse neon.timeline_id, %s", wp->config->neon_timeline);
+		walprop_log(FATAL, "Could not parse neon.timeline_id, %s", wp->config->neon_timeline);
 	if (!wp->config->neon_tenant)
-		elog(FATAL, "neon.tenant_id is not provided");
+		walprop_log(FATAL, "neon.tenant_id is not provided");
 	if (*wp->config->neon_tenant != '\0' &&
 		!HexDecodeString(wp->greetRequest.tenant_id, wp->config->neon_tenant, 16))
-		elog(FATAL, "Could not parse neon.tenant_id, %s", wp->config->neon_tenant);
+		walprop_log(FATAL, "Could not parse neon.tenant_id, %s", wp->config->neon_tenant);
 
-	wp->greetRequest.timeline = wp->api.get_timeline_id();
+	wp->greetRequest.timeline = wp->config->pgTimeline;
 	wp->greetRequest.walSegSize = wp->config->wal_segment_size;
 
-	wp->api.init_event_set(wp->n_safekeepers);
+	wp->api.init_event_set(wp);
 
 	return wp;
 }
 
+void
+WalProposerFree(WalProposer *wp)
+{
+	for (int i = 0; i < wp->n_safekeepers; i++)
+	{
+		Safekeeper *sk = &wp->safekeeper[i];
+
+		Assert(sk->outbuf.data != NULL);
+		pfree(sk->outbuf.data);
+		if (sk->voteResponse.termHistory.entries)
+			pfree(sk->voteResponse.termHistory.entries);
+		sk->voteResponse.termHistory.entries = NULL;
+	}
+	if (wp->propTermHistory.entries != NULL)
+		pfree(wp->propTermHistory.entries);
+	wp->propTermHistory.entries = NULL;
+	
+	pfree(wp);
+}
+
 /*
  * Create new AppendRequest message and start sending it. This function is
  * called from walsender every time the new WAL is available.
@@ -190,10 +207,10 @@ WalProposerPoll(WalProposer *wp)
 		Safekeeper *sk = NULL;
 		int			rc = 0;
 		uint32		events = 0;
-		TimestampTz now = wp->api.get_current_timestamp();
+		TimestampTz now = wp->api.get_current_timestamp(wp);
 		long		timeout = TimeToReconnect(wp, now);
 
-		rc = wp->api.wait_event_set(timeout, &sk, &events);
+		rc = wp->api.wait_event_set(wp, timeout, &sk, &events);
 
 		/* Exit loop if latch is set (we got new WAL) */
 		if ((rc == 1 && events & WL_LATCH_SET))
@@ -224,14 +241,14 @@ WalProposerPoll(WalProposer *wp)
 			 */
 			if (!wp->config->syncSafekeepers)
 			{
-				XLogRecPtr	flushed = wp->api.get_flush_rec_ptr();
+				XLogRecPtr	flushed = wp->api.get_flush_rec_ptr(wp);
 
 				if (flushed > wp->availableLsn)
 					break;
 			}
 		}
 
-		now = wp->api.get_current_timestamp();
+		now = wp->api.get_current_timestamp(wp);
 		/* timeout expired: poll state */
 		if (rc == 0 || TimeToReconnect(wp, now) <= 0)
 		{
@@ -249,7 +266,7 @@ WalProposerPoll(WalProposer *wp)
 			/*
 			 * Abandon connection attempts which take too long.
 			 */
-			now = wp->api.get_current_timestamp();
+			now = wp->api.get_current_timestamp(wp);
 			for (int i = 0; i < wp->n_safekeepers; i++)
 			{
 				Safekeeper *sk = &wp->safekeeper[i];
@@ -257,7 +274,7 @@ WalProposerPoll(WalProposer *wp)
 				if (TimestampDifferenceExceeds(sk->latestMsgReceivedAt, now,
 											   wp->config->safekeeper_connection_timeout))
 				{
-					elog(WARNING, "terminating connection to safekeeper '%s:%s' in '%s' state: no messages received during the last %dms or connection attempt took longer than that",
+					walprop_log(WARNING, "terminating connection to safekeeper '%s:%s' in '%s' state: no messages received during the last %dms or connection attempt took longer than that",
 						 sk->host, sk->port, FormatSafekeeperState(sk->state), wp->config->safekeeper_connection_timeout);
 					ShutdownConnection(sk);
 				}
@@ -296,10 +313,10 @@ HackyRemoveWalProposerEvent(Safekeeper *to_remove)
 {
 	WalProposer *wp = to_remove->wp;
 
-	/* Remove the existing event set */
-	wp->api.free_event_set();
+	/* Remove the existing event set, assign sk->eventPos = -1 */
+	wp->api.free_event_set(wp);
 	/* Re-initialize it without adding any safekeeper events */
-	wp->api.init_event_set(wp->n_safekeepers);
+	wp->api.init_event_set(wp);
 
 	/*
 	 * loop through the existing safekeepers. If they aren't the one we're
@@ -311,13 +328,11 @@ HackyRemoveWalProposerEvent(Safekeeper *to_remove)
 		uint32		desired_events = WL_NO_EVENTS;
 		Safekeeper *sk = &wp->safekeeper[i];
 
-		sk->eventPos = -1;
-
 		if (sk == to_remove)
 			continue;
 
 		/* If this safekeeper isn't offline, add an event for it! */
-		if (sk->conn != NULL)
+		if (sk->state != SS_OFFLINE)
 		{
 			desired_events = SafekeeperStateDesiredEvents(sk->state);
 			/* will set sk->eventPos */
@@ -330,9 +345,7 @@ HackyRemoveWalProposerEvent(Safekeeper *to_remove)
 static void
 ShutdownConnection(Safekeeper *sk)
 {
-	if (sk->conn)
-		sk->wp->api.conn_finish(sk->conn);
-	sk->conn = NULL;
+	sk->wp->api.conn_finish(sk);
 	sk->state = SS_OFFLINE;
 	sk->flushWrite = false;
 	sk->streamingAt = InvalidXLogRecPtr;
@@ -361,23 +374,16 @@ ResetConnection(Safekeeper *sk)
 	}
 
 	/*
-	 * Try to establish new connection
+	 * Try to establish new connection, it will update sk->conn.
 	 */
-	sk->conn = wp->api.conn_connect_start((char *) &sk->conninfo);
-
-	/*
-	 * "If the result is null, then libpq has been unable to allocate a new
-	 * PGconn structure"
-	 */
-	if (!sk->conn)
-		elog(FATAL, "failed to allocate new PGconn object");
+	wp->api.conn_connect_start(sk);
 
 	/*
 	 * PQconnectStart won't actually start connecting until we run
 	 * PQconnectPoll. Before we do that though, we need to check that it
 	 * didn't immediately fail.
 	 */
-	if (wp->api.conn_status(sk->conn) == WP_CONNECTION_BAD)
+	if (wp->api.conn_status(sk) == WP_CONNECTION_BAD)
 	{
 		/*---
 		 * According to libpq docs:
@@ -388,15 +394,14 @@ ResetConnection(Safekeeper *sk)
 		 *
 		 * https://www.postgresql.org/docs/devel/libpq-connect.html#LIBPQ-PQCONNECTSTARTPARAMS
 		 */
-		elog(WARNING, "Immediate failure to connect with node '%s:%s':\n\terror: %s",
-			 sk->host, sk->port, wp->api.conn_error_message(sk->conn));
+		walprop_log(WARNING, "Immediate failure to connect with node '%s:%s':\n\terror: %s",
+			 sk->host, sk->port, wp->api.conn_error_message(sk));
 
 		/*
 		 * Even though the connection failed, we still need to clean up the
 		 * object
 		 */
-		wp->api.conn_finish(sk->conn);
-		sk->conn = NULL;
+		wp->api.conn_finish(sk);
 		return;
 	}
 
@@ -413,10 +418,10 @@ ResetConnection(Safekeeper *sk)
 	 * (see libpqrcv_connect, defined in
 	 * src/backend/replication/libpqwalreceiver/libpqwalreceiver.c)
 	 */
-	elog(LOG, "connecting with node %s:%s", sk->host, sk->port);
+	walprop_log(LOG, "connecting with node %s:%s", sk->host, sk->port);
 
 	sk->state = SS_CONNECTING_WRITE;
-	sk->latestMsgReceivedAt = wp->api.get_current_timestamp();
+	sk->latestMsgReceivedAt = wp->api.get_current_timestamp(wp);
 
 	wp->api.add_safekeeper_event_set(sk, WL_SOCKET_WRITEABLE);
 	return;
@@ -447,7 +452,7 @@ TimeToReconnect(WalProposer *wp, TimestampTz now)
 static void
 ReconnectSafekeepers(WalProposer *wp)
 {
-	TimestampTz now = wp->api.get_current_timestamp();
+	TimestampTz now = wp->api.get_current_timestamp(wp);
 
 	if (TimeToReconnect(wp, now) == 0)
 	{
@@ -467,6 +472,8 @@ ReconnectSafekeepers(WalProposer *wp)
 static void
 AdvancePollState(Safekeeper *sk, uint32 events)
 {
+	WalProposer *wp = sk->wp;
+
 	/*
 	 * Sanity check. We assume further down that the operations don't block
 	 * because the socket is ready.
@@ -481,7 +488,7 @@ AdvancePollState(Safekeeper *sk, uint32 events)
 			 * ResetConnection
 			 */
 		case SS_OFFLINE:
-			elog(FATAL, "Unexpected safekeeper %s:%s state advancement: is offline",
+			walprop_log(FATAL, "Unexpected safekeeper %s:%s state advancement: is offline",
 				 sk->host, sk->port);
 			break;				/* actually unreachable, but prevents
 								 * -Wimplicit-fallthrough */
@@ -517,7 +524,7 @@ AdvancePollState(Safekeeper *sk, uint32 events)
 			 * requests.
 			 */
 		case SS_VOTING:
-			elog(WARNING, "EOF from node %s:%s in %s state", sk->host,
+			walprop_log(WARNING, "EOF from node %s:%s in %s state", sk->host,
 				 sk->port, FormatSafekeeperState(sk->state));
 			ResetConnection(sk);
 			return;
@@ -546,7 +553,7 @@ AdvancePollState(Safekeeper *sk, uint32 events)
 			 * Idle state for waiting votes from quorum.
 			 */
 		case SS_IDLE:
-			elog(WARNING, "EOF from node %s:%s in %s state", sk->host,
+			walprop_log(WARNING, "EOF from node %s:%s in %s state", sk->host,
 				 sk->port, FormatSafekeeperState(sk->state));
 			ResetConnection(sk);
 			return;
@@ -564,7 +571,7 @@ static void
 HandleConnectionEvent(Safekeeper *sk)
 {
 	WalProposer *wp = sk->wp;
-	WalProposerConnectPollStatusType result = wp->api.conn_connect_poll(sk->conn);
+	WalProposerConnectPollStatusType result = wp->api.conn_connect_poll(sk);
 
 	/* The new set of events we'll wait on, after updating */
 	uint32		new_events = WL_NO_EVENTS;
@@ -572,9 +579,9 @@ HandleConnectionEvent(Safekeeper *sk)
 	switch (result)
 	{
 		case WP_CONN_POLLING_OK:
-			elog(LOG, "connected with node %s:%s", sk->host,
+			walprop_log(LOG, "connected with node %s:%s", sk->host,
 				 sk->port);
-			sk->latestMsgReceivedAt = wp->api.get_current_timestamp();
+			sk->latestMsgReceivedAt = wp->api.get_current_timestamp(wp);
 
 			/*
 			 * We have to pick some event to update event set. We'll
@@ -596,8 +603,8 @@ HandleConnectionEvent(Safekeeper *sk)
 			break;
 
 		case WP_CONN_POLLING_FAILED:
-			elog(WARNING, "failed to connect to node '%s:%s': %s",
-				 sk->host, sk->port, wp->api.conn_error_message(sk->conn));
+			walprop_log(WARNING, "failed to connect to node '%s:%s': %s",
+				 sk->host, sk->port, wp->api.conn_error_message(sk));
 
 			/*
 			 * If connecting failed, we don't want to restart the connection
@@ -631,10 +638,10 @@ SendStartWALPush(Safekeeper *sk)
 {
 	WalProposer *wp = sk->wp;
 
-	if (!wp->api.conn_send_query(sk->conn, "START_WAL_PUSH"))
+	if (!wp->api.conn_send_query(sk, "START_WAL_PUSH"))
 	{
-		elog(WARNING, "Failed to send 'START_WAL_PUSH' query to safekeeper %s:%s: %s",
-			 sk->host, sk->port, wp->api.conn_error_message(sk->conn));
+		walprop_log(WARNING, "Failed to send 'START_WAL_PUSH' query to safekeeper %s:%s: %s",
+			 sk->host, sk->port, wp->api.conn_error_message(sk));
 		ShutdownConnection(sk);
 		return;
 	}
@@ -647,7 +654,7 @@ RecvStartWALPushResult(Safekeeper *sk)
 {
 	WalProposer *wp = sk->wp;
 
-	switch (wp->api.conn_get_query_result(sk->conn))
+	switch (wp->api.conn_get_query_result(sk))
 	{
 			/*
 			 * Successful result, move on to starting the handshake
@@ -670,8 +677,8 @@ RecvStartWALPushResult(Safekeeper *sk)
 			break;
 
 		case WP_EXEC_FAILED:
-			elog(WARNING, "Failed to send query to safekeeper %s:%s: %s",
-				 sk->host, sk->port, wp->api.conn_error_message(sk->conn));
+			walprop_log(WARNING, "Failed to send query to safekeeper %s:%s: %s",
+				 sk->host, sk->port, wp->api.conn_error_message(sk));
 			ShutdownConnection(sk);
 			return;
 
@@ -681,7 +688,7 @@ RecvStartWALPushResult(Safekeeper *sk)
 			 * wrong"
 			 */
 		case WP_EXEC_UNEXPECTED_SUCCESS:
-			elog(WARNING, "Received bad response from safekeeper %s:%s query execution",
+			walprop_log(WARNING, "Received bad response from safekeeper %s:%s query execution",
 				 sk->host, sk->port);
 			ShutdownConnection(sk);
 			return;
@@ -717,7 +724,7 @@ RecvAcceptorGreeting(Safekeeper *sk)
 	if (!AsyncReadMessage(sk, (AcceptorProposerMessage *) &sk->greetResponse))
 		return;
 
-	elog(LOG, "received AcceptorGreeting from safekeeper %s:%s", sk->host, sk->port);
+	walprop_log(LOG, "received AcceptorGreeting from safekeeper %s:%s", sk->host, sk->port);
 
 	/* Protocol is all good, move to voting. */
 	sk->state = SS_VOTING;
@@ -737,7 +744,7 @@ RecvAcceptorGreeting(Safekeeper *sk)
 		if (wp->n_connected == wp->quorum)
 		{
 			wp->propTerm++;
-			elog(LOG, "proposer connected to quorum (%d) safekeepers, propTerm=" INT64_FORMAT, wp->quorum, wp->propTerm);
+			walprop_log(LOG, "proposer connected to quorum (%d) safekeepers, propTerm=" INT64_FORMAT, wp->quorum, wp->propTerm);
 
 			wp->voteRequest = (VoteRequest)
 			{
@@ -750,7 +757,7 @@ RecvAcceptorGreeting(Safekeeper *sk)
 	else if (sk->greetResponse.term > wp->propTerm)
 	{
 		/* Another compute with higher term is running. */
-		elog(FATAL, "WAL acceptor %s:%s with term " INT64_FORMAT " rejects our connection request with term " INT64_FORMAT "",
+		walprop_log(FATAL, "WAL acceptor %s:%s with term " INT64_FORMAT " rejects our connection request with term " INT64_FORMAT "",
 			 sk->host, sk->port,
 			 sk->greetResponse.term, wp->propTerm);
 	}
@@ -792,7 +799,7 @@ SendVoteRequest(Safekeeper *sk)
 	WalProposer *wp = sk->wp;
 
 	/* We have quorum for voting, send our vote request */
-	elog(LOG, "requesting vote from %s:%s for term " UINT64_FORMAT, sk->host, sk->port, wp->voteRequest.term);
+	walprop_log(LOG, "requesting vote from %s:%s for term " UINT64_FORMAT, sk->host, sk->port, wp->voteRequest.term);
 	/* On failure, logging & resetting is handled */
 	if (!BlockingWrite(sk, &wp->voteRequest, sizeof(wp->voteRequest), SS_WAIT_VERDICT))
 		return;
@@ -809,7 +816,7 @@ RecvVoteResponse(Safekeeper *sk)
 	if (!AsyncReadMessage(sk, (AcceptorProposerMessage *) &sk->voteResponse))
 		return;
 
-	elog(LOG,
+	walprop_log(LOG,
 		 "got VoteResponse from acceptor %s:%s, voteGiven=" UINT64_FORMAT ", epoch=" UINT64_FORMAT ", flushLsn=%X/%X, truncateLsn=%X/%X, timelineStartLsn=%X/%X",
 		 sk->host, sk->port, sk->voteResponse.voteGiven, GetHighestTerm(&sk->voteResponse.termHistory),
 		 LSN_FORMAT_ARGS(sk->voteResponse.flushLsn),
@@ -824,7 +831,7 @@ RecvVoteResponse(Safekeeper *sk)
 	if ((!sk->voteResponse.voteGiven) &&
 		(sk->voteResponse.term > wp->propTerm || wp->n_votes < wp->quorum))
 	{
-		elog(FATAL, "WAL acceptor %s:%s with term " INT64_FORMAT " rejects our connection request with term " INT64_FORMAT "",
+		walprop_log(FATAL, "WAL acceptor %s:%s with term " INT64_FORMAT " rejects our connection request with term " INT64_FORMAT "",
 			 sk->host, sk->port,
 			 sk->voteResponse.term, wp->propTerm);
 	}
@@ -869,19 +876,19 @@ HandleElectedProposer(WalProposer *wp)
 	 */
 	if (wp->truncateLsn < wp->propEpochStartLsn)
 	{
-		elog(LOG,
+		walprop_log(LOG,
 			 "start recovery because truncateLsn=%X/%X is not "
 			 "equal to epochStartLsn=%X/%X",
 			 LSN_FORMAT_ARGS(wp->truncateLsn),
 			 LSN_FORMAT_ARGS(wp->propEpochStartLsn));
 		/* Perform recovery */
 		if (!wp->api.recovery_download(&wp->safekeeper[wp->donor], wp->greetRequest.timeline, wp->truncateLsn, wp->propEpochStartLsn))
-			elog(FATAL, "Failed to recover state");
+			walprop_log(FATAL, "Failed to recover state");
 	}
 	else if (wp->config->syncSafekeepers)
 	{
 		/* Sync is not needed: just exit */
-		wp->api.finish_sync_safekeepers(wp->propEpochStartLsn);
+		wp->api.finish_sync_safekeepers(wp, wp->propEpochStartLsn);
 		/* unreachable */
 	}
 
@@ -982,7 +989,7 @@ DetermineEpochStartLsn(WalProposer *wp)
 				if (wp->timelineStartLsn != InvalidXLogRecPtr &&
 					wp->timelineStartLsn != wp->safekeeper[i].voteResponse.timelineStartLsn)
 				{
-					elog(WARNING,
+					walprop_log(WARNING,
 						 "inconsistent timelineStartLsn: current %X/%X, received %X/%X",
 						 LSN_FORMAT_ARGS(wp->timelineStartLsn),
 						 LSN_FORMAT_ARGS(wp->safekeeper[i].voteResponse.timelineStartLsn));
@@ -998,12 +1005,12 @@ DetermineEpochStartLsn(WalProposer *wp)
 	 */
 	if (wp->propEpochStartLsn == InvalidXLogRecPtr && !wp->config->syncSafekeepers)
 	{
-		wp->propEpochStartLsn = wp->truncateLsn = wp->api.get_redo_start_lsn();
+		wp->propEpochStartLsn = wp->truncateLsn = wp->api.get_redo_start_lsn(wp);
 		if (wp->timelineStartLsn == InvalidXLogRecPtr)
 		{
-			wp->timelineStartLsn = wp->api.get_redo_start_lsn();
+			wp->timelineStartLsn = wp->api.get_redo_start_lsn(wp);
 		}
-		elog(LOG, "bumped epochStartLsn to the first record %X/%X", LSN_FORMAT_ARGS(wp->propEpochStartLsn));
+		walprop_log(LOG, "bumped epochStartLsn to the first record %X/%X", LSN_FORMAT_ARGS(wp->propEpochStartLsn));
 	}
 
 	/*
@@ -1030,7 +1037,7 @@ DetermineEpochStartLsn(WalProposer *wp)
 	wp->propTermHistory.entries[wp->propTermHistory.n_entries - 1].term = wp->propTerm;
 	wp->propTermHistory.entries[wp->propTermHistory.n_entries - 1].lsn = wp->propEpochStartLsn;
 
-	elog(LOG, "got votes from majority (%d) of nodes, term " UINT64_FORMAT ", epochStartLsn %X/%X, donor %s:%s, truncate_lsn %X/%X",
+	walprop_log(LOG, "got votes from majority (%d) of nodes, term " UINT64_FORMAT ", epochStartLsn %X/%X, donor %s:%s, truncate_lsn %X/%X",
 		 wp->quorum,
 		 wp->propTerm,
 		 LSN_FORMAT_ARGS(wp->propEpochStartLsn),
@@ -1044,7 +1051,7 @@ DetermineEpochStartLsn(WalProposer *wp)
 	 */
 	if (!wp->config->syncSafekeepers)
 	{
-		WalproposerShmemState *walprop_shared = wp->api.get_shmem_state();
+		WalproposerShmemState *walprop_shared = wp->api.get_shmem_state(wp);
 
 		/*
 		 * Basebackup LSN always points to the beginning of the record (not
@@ -1052,7 +1059,7 @@ DetermineEpochStartLsn(WalProposer *wp)
 		 * Safekeepers don't skip header as they need continious stream of
 		 * data, so correct LSN for comparison.
 		 */
-		if (SkipXLogPageHeader(wp, wp->propEpochStartLsn) != wp->api.get_redo_start_lsn())
+		if (SkipXLogPageHeader(wp, wp->propEpochStartLsn) != wp->api.get_redo_start_lsn(wp))
 		{
 			/*
 			 * However, allow to proceed if previously elected leader was me;
@@ -1062,14 +1069,21 @@ DetermineEpochStartLsn(WalProposer *wp)
 			if (!((dth->n_entries >= 1) && (dth->entries[dth->n_entries - 1].term ==
 											walprop_shared->mineLastElectedTerm)))
 			{
-				elog(PANIC,
+				walprop_log(PANIC,
 					 "collected propEpochStartLsn %X/%X, but basebackup LSN %X/%X",
 					 LSN_FORMAT_ARGS(wp->propEpochStartLsn),
-					 LSN_FORMAT_ARGS(wp->api.get_redo_start_lsn()));
+					 LSN_FORMAT_ARGS(wp->api.get_redo_start_lsn(wp)));
 			}
 		}
 		walprop_shared->mineLastElectedTerm = wp->propTerm;
 	}
+
+	/*
+	 * WalProposer has just elected itself and initialized history, so
+	 * we can call election callback. Usually it updates truncateLsn to
+	 * fetch WAL for logical replication.
+	 */
+	wp->api.after_election(wp);
 }
 
 /*
@@ -1140,7 +1154,7 @@ SendProposerElected(Safekeeper *sk)
 			 */
 			sk->startStreamingAt = wp->truncateLsn;
 
-			elog(WARNING, "empty safekeeper joined cluster as %s:%s, historyStart=%X/%X, sk->startStreamingAt=%X/%X",
+			walprop_log(WARNING, "empty safekeeper joined cluster as %s:%s, historyStart=%X/%X, sk->startStreamingAt=%X/%X",
 				 sk->host, sk->port, LSN_FORMAT_ARGS(wp->propTermHistory.entries[0].lsn),
 				 LSN_FORMAT_ARGS(sk->startStreamingAt));
 		}
@@ -1175,7 +1189,7 @@ SendProposerElected(Safekeeper *sk)
 	msg.timelineStartLsn = wp->timelineStartLsn;
 
 	lastCommonTerm = i >= 0 ? wp->propTermHistory.entries[i].term : 0;
-	elog(LOG,
+	walprop_log(LOG,
 		 "sending elected msg to node " UINT64_FORMAT " term=" UINT64_FORMAT ", startStreamingAt=%X/%X (lastCommonTerm=" UINT64_FORMAT "), termHistory.n_entries=%u to %s:%s, timelineStartLsn=%X/%X",
 		 sk->greetResponse.nodeId, msg.term, LSN_FORMAT_ARGS(msg.startStreamingAt), lastCommonTerm, msg.termHistory->n_entries, sk->host, sk->port, LSN_FORMAT_ARGS(msg.timelineStartLsn));
 
@@ -1340,13 +1354,12 @@ SendAppendRequests(Safekeeper *sk)
 		req = &sk->appendRequest;
 		PrepareAppendRequest(sk->wp, &sk->appendRequest, sk->streamingAt, endLsn);
 
-		ereport(DEBUG2,
-				(errmsg("sending message len %ld beginLsn=%X/%X endLsn=%X/%X commitLsn=%X/%X truncateLsn=%X/%X to %s:%s",
+		walprop_log(DEBUG2, "sending message len %ld beginLsn=%X/%X endLsn=%X/%X commitLsn=%X/%X truncateLsn=%X/%X to %s:%s",
 						req->endLsn - req->beginLsn,
 						LSN_FORMAT_ARGS(req->beginLsn),
 						LSN_FORMAT_ARGS(req->endLsn),
 						LSN_FORMAT_ARGS(req->commitLsn),
-						LSN_FORMAT_ARGS(wp->truncateLsn), sk->host, sk->port)));
+						LSN_FORMAT_ARGS(wp->truncateLsn), sk->host, sk->port);
 
 		resetStringInfo(&sk->outbuf);
 
@@ -1356,13 +1369,13 @@ SendAppendRequests(Safekeeper *sk)
 		/* write the WAL itself */
 		enlargeStringInfo(&sk->outbuf, req->endLsn - req->beginLsn);
 		/* wal_read will raise error on failure */
-		wp->api.wal_read(sk->xlogreader,
+		wp->api.wal_read(sk,
 						 &sk->outbuf.data[sk->outbuf.len],
 						 req->beginLsn,
 						 req->endLsn - req->beginLsn);
 		sk->outbuf.len += req->endLsn - req->beginLsn;
 
-		writeResult = wp->api.conn_async_write(sk->conn, sk->outbuf.data, sk->outbuf.len);
+		writeResult = wp->api.conn_async_write(sk, sk->outbuf.data, sk->outbuf.len);
 
 		/* Mark current message as sent, whatever the result is */
 		sk->streamingAt = endLsn;
@@ -1384,9 +1397,9 @@ SendAppendRequests(Safekeeper *sk)
 				return true;
 
 			case PG_ASYNC_WRITE_FAIL:
-				elog(WARNING, "Failed to send to node %s:%s in %s state: %s",
+				walprop_log(WARNING, "Failed to send to node %s:%s in %s state: %s",
 					 sk->host, sk->port, FormatSafekeeperState(sk->state),
-					 wp->api.conn_error_message(sk->conn));
+					 wp->api.conn_error_message(sk));
 				ShutdownConnection(sk);
 				return false;
 			default:
@@ -1424,17 +1437,16 @@ RecvAppendResponses(Safekeeper *sk)
 		if (!AsyncReadMessage(sk, (AcceptorProposerMessage *) &sk->appendResponse))
 			break;
 
-		ereport(DEBUG2,
-				(errmsg("received message term=" INT64_FORMAT " flushLsn=%X/%X commitLsn=%X/%X from %s:%s",
+		walprop_log(DEBUG2, "received message term=" INT64_FORMAT " flushLsn=%X/%X commitLsn=%X/%X from %s:%s",
 						sk->appendResponse.term,
 						LSN_FORMAT_ARGS(sk->appendResponse.flushLsn),
 						LSN_FORMAT_ARGS(sk->appendResponse.commitLsn),
-						sk->host, sk->port)));
+						sk->host, sk->port);
 
 		if (sk->appendResponse.term > wp->propTerm)
 		{
 			/* Another compute with higher term is running. */
-			elog(PANIC, "WAL acceptor %s:%s with term " INT64_FORMAT " rejected our request, our term " INT64_FORMAT "",
+			walprop_log(PANIC, "WAL acceptor %s:%s with term " INT64_FORMAT " rejected our request, our term " INT64_FORMAT "",
 				 sk->host, sk->port,
 				 sk->appendResponse.term, wp->propTerm);
 		}
@@ -1462,7 +1474,7 @@ RecvAppendResponses(Safekeeper *sk)
 
 /* Parse a PageserverFeedback message, or the PageserverFeedback part of an AppendResponse */
 void
-ParsePageserverFeedbackMessage(StringInfo reply_message, PageserverFeedback *rf)
+ParsePageserverFeedbackMessage(WalProposer *wp, StringInfo reply_message, PageserverFeedback *rf)
 {
 	uint8		nkeys;
 	int			i;
@@ -1480,7 +1492,7 @@ ParsePageserverFeedbackMessage(StringInfo reply_message, PageserverFeedback *rf)
 			pq_getmsgint(reply_message, sizeof(int32));
 			/* read value length */
 			rf->currentClusterSize = pq_getmsgint64(reply_message);
-			elog(DEBUG2, "ParsePageserverFeedbackMessage: current_timeline_size %lu",
+			walprop_log(DEBUG2, "ParsePageserverFeedbackMessage: current_timeline_size %lu",
 				 rf->currentClusterSize);
 		}
 		else if ((strcmp(key, "ps_writelsn") == 0) || (strcmp(key, "last_received_lsn") == 0))
@@ -1488,7 +1500,7 @@ ParsePageserverFeedbackMessage(StringInfo reply_message, PageserverFeedback *rf)
 			pq_getmsgint(reply_message, sizeof(int32));
 			/* read value length */
 			rf->last_received_lsn = pq_getmsgint64(reply_message);
-			elog(DEBUG2, "ParsePageserverFeedbackMessage: last_received_lsn %X/%X",
+			walprop_log(DEBUG2, "ParsePageserverFeedbackMessage: last_received_lsn %X/%X",
 				 LSN_FORMAT_ARGS(rf->last_received_lsn));
 		}
 		else if ((strcmp(key, "ps_flushlsn") == 0) || (strcmp(key, "disk_consistent_lsn") == 0))
@@ -1496,7 +1508,7 @@ ParsePageserverFeedbackMessage(StringInfo reply_message, PageserverFeedback *rf)
 			pq_getmsgint(reply_message, sizeof(int32));
 			/* read value length */
 			rf->disk_consistent_lsn = pq_getmsgint64(reply_message);
-			elog(DEBUG2, "ParsePageserverFeedbackMessage: disk_consistent_lsn %X/%X",
+			walprop_log(DEBUG2, "ParsePageserverFeedbackMessage: disk_consistent_lsn %X/%X",
 				 LSN_FORMAT_ARGS(rf->disk_consistent_lsn));
 		}
 		else if ((strcmp(key, "ps_applylsn") == 0) || (strcmp(key, "remote_consistent_lsn") == 0))
@@ -1504,7 +1516,7 @@ ParsePageserverFeedbackMessage(StringInfo reply_message, PageserverFeedback *rf)
 			pq_getmsgint(reply_message, sizeof(int32));
 			/* read value length */
 			rf->remote_consistent_lsn = pq_getmsgint64(reply_message);
-			elog(DEBUG2, "ParsePageserverFeedbackMessage: remote_consistent_lsn %X/%X",
+			walprop_log(DEBUG2, "ParsePageserverFeedbackMessage: remote_consistent_lsn %X/%X",
 				 LSN_FORMAT_ARGS(rf->remote_consistent_lsn));
 		}
 		else if ((strcmp(key, "ps_replytime") == 0) || (strcmp(key, "replytime") == 0))
@@ -1517,7 +1529,7 @@ ParsePageserverFeedbackMessage(StringInfo reply_message, PageserverFeedback *rf)
 
 				/* Copy because timestamptz_to_str returns a static buffer */
 				replyTimeStr = pstrdup(timestamptz_to_str(rf->replytime));
-				elog(DEBUG2, "ParsePageserverFeedbackMessage: replytime %lu reply_time: %s",
+				walprop_log(DEBUG2, "ParsePageserverFeedbackMessage: replytime %lu reply_time: %s",
 					 rf->replytime, replyTimeStr);
 
 				pfree(replyTimeStr);
@@ -1532,7 +1544,7 @@ ParsePageserverFeedbackMessage(StringInfo reply_message, PageserverFeedback *rf)
 			 * Skip unknown keys to support backward compatibile protocol
 			 * changes
 			 */
-			elog(LOG, "ParsePageserverFeedbackMessage: unknown key: %s len %d", key, len);
+			walprop_log(LOG, "ParsePageserverFeedbackMessage: unknown key: %s len %d", key, len);
 			pq_getmsgbytes(reply_message, len);
 		};
 	}
@@ -1615,7 +1627,7 @@ HandleSafekeeperResponse(WalProposer *wp)
 		 * Advance the replication slot to free up old WAL files. Note that
 		 * slot doesn't exist if we are in syncSafekeepers mode.
 		 */
-		wp->api.confirm_wal_streamed(wp->truncateLsn);
+		wp->api.confirm_wal_streamed(wp, wp->truncateLsn);
 	}
 
 	/*
@@ -1662,7 +1674,7 @@ HandleSafekeeperResponse(WalProposer *wp)
 			 */
 			BroadcastAppendRequest(wp);
 
-			wp->api.finish_sync_safekeepers(wp->propEpochStartLsn);
+			wp->api.finish_sync_safekeepers(wp, wp->propEpochStartLsn);
 			/* unreachable */
 		}
 	}
@@ -1677,7 +1689,7 @@ AsyncRead(Safekeeper *sk, char **buf, int *buf_size)
 {
 	WalProposer *wp = sk->wp;
 
-	switch (wp->api.conn_async_read(sk->conn, buf, buf_size))
+	switch (wp->api.conn_async_read(sk, buf, buf_size))
 	{
 		case PG_ASYNC_READ_SUCCESS:
 			return true;
@@ -1687,9 +1699,9 @@ AsyncRead(Safekeeper *sk, char **buf, int *buf_size)
 			return false;
 
 		case PG_ASYNC_READ_FAIL:
-			elog(WARNING, "Failed to read from node %s:%s in %s state: %s", sk->host,
+			walprop_log(WARNING, "Failed to read from node %s:%s in %s state: %s", sk->host,
 				 sk->port, FormatSafekeeperState(sk->state),
-				 wp->api.conn_error_message(sk->conn));
+				 wp->api.conn_error_message(sk));
 			ShutdownConnection(sk);
 			return false;
 	}
@@ -1727,12 +1739,12 @@ AsyncReadMessage(Safekeeper *sk, AcceptorProposerMessage *anymsg)
 	tag = pq_getmsgint64_le(&s);
 	if (tag != anymsg->tag)
 	{
-		elog(WARNING, "unexpected message tag %c from node %s:%s in state %s", (char) tag, sk->host,
+		walprop_log(WARNING, "unexpected message tag %c from node %s:%s in state %s", (char) tag, sk->host,
 			 sk->port, FormatSafekeeperState(sk->state));
 		ResetConnection(sk);
 		return false;
 	}
-	sk->latestMsgReceivedAt = wp->api.get_current_timestamp();
+	sk->latestMsgReceivedAt = wp->api.get_current_timestamp(wp);
 	switch (tag)
 	{
 		case 'g':
@@ -1776,7 +1788,7 @@ AsyncReadMessage(Safekeeper *sk, AcceptorProposerMessage *anymsg)
 				msg->hs.xmin.value = pq_getmsgint64_le(&s);
 				msg->hs.catalog_xmin.value = pq_getmsgint64_le(&s);
 				if (buf_size > APPENDRESPONSE_FIXEDPART_SIZE)
-					ParsePageserverFeedbackMessage(&s, &msg->rf);
+					ParsePageserverFeedbackMessage(wp, &s, &msg->rf);
 				pq_getmsgend(&s);
 				return true;
 			}
@@ -1801,11 +1813,11 @@ BlockingWrite(Safekeeper *sk, void *msg, size_t msg_size, SafekeeperState succes
 	WalProposer *wp = sk->wp;
 	uint32		events;
 
-	if (!wp->api.conn_blocking_write(sk->conn, msg, msg_size))
+	if (!wp->api.conn_blocking_write(sk, msg, msg_size))
 	{
-		elog(WARNING, "Failed to send to node %s:%s in %s state: %s",
+		walprop_log(WARNING, "Failed to send to node %s:%s in %s state: %s",
 			 sk->host, sk->port, FormatSafekeeperState(sk->state),
-			 wp->api.conn_error_message(sk->conn));
+			 wp->api.conn_error_message(sk));
 		ShutdownConnection(sk);
 		return false;
 	}
@@ -1835,7 +1847,7 @@ AsyncWrite(Safekeeper *sk, void *msg, size_t msg_size, SafekeeperState flush_sta
 {
 	WalProposer *wp = sk->wp;
 
-	switch (wp->api.conn_async_write(sk->conn, msg, msg_size))
+	switch (wp->api.conn_async_write(sk, msg, msg_size))
 	{
 		case PG_ASYNC_WRITE_SUCCESS:
 			return true;
@@ -1850,9 +1862,9 @@ AsyncWrite(Safekeeper *sk, void *msg, size_t msg_size, SafekeeperState flush_sta
 			wp->api.update_event_set(sk, WL_SOCKET_READABLE | WL_SOCKET_WRITEABLE);
 			return false;
 		case PG_ASYNC_WRITE_FAIL:
-			elog(WARNING, "Failed to send to node %s:%s in %s state: %s",
+			walprop_log(WARNING, "Failed to send to node %s:%s in %s state: %s",
 				 sk->host, sk->port, FormatSafekeeperState(sk->state),
-				 wp->api.conn_error_message(sk->conn));
+				 wp->api.conn_error_message(sk));
 			ShutdownConnection(sk);
 			return false;
 		default:
@@ -1880,7 +1892,7 @@ AsyncFlush(Safekeeper *sk)
 	 *   1 if unable to send everything yet [call PQflush again]
 	 *  -1 if it failed                     [emit an error]
 	 */
-	switch (wp->api.conn_flush(sk->conn))
+	switch (wp->api.conn_flush(sk))
 	{
 		case 0:
 			/* flush is done */
@@ -1889,9 +1901,9 @@ AsyncFlush(Safekeeper *sk)
 			/* Nothing to do; try again when the socket's ready */
 			return false;
 		case -1:
-			elog(WARNING, "Failed to flush write to node %s:%s in %s state: %s",
+			walprop_log(WARNING, "Failed to flush write to node %s:%s in %s state: %s",
 				 sk->host, sk->port, FormatSafekeeperState(sk->state),
-				 wp->api.conn_error_message(sk->conn));
+				 wp->api.conn_error_message(sk));
 			ResetConnection(sk);
 			return false;
 		default:
@@ -1920,11 +1932,11 @@ CompareLsn(const void *a, const void *b)
  *
  * The strings are intended to be used as a prefix to "state", e.g.:
  *
- *   elog(LOG, "currently in %s state", FormatSafekeeperState(sk->state));
+ *   walprop_log(LOG, "currently in %s state", FormatSafekeeperState(sk->state));
  *
  * If this sort of phrasing doesn't fit the message, instead use something like:
  *
- *   elog(LOG, "currently in state [%s]", FormatSafekeeperState(sk->state));
+ *   walprop_log(LOG, "currently in state [%s]", FormatSafekeeperState(sk->state));
  */
 static char *
 FormatSafekeeperState(SafekeeperState state)
@@ -1972,6 +1984,7 @@ FormatSafekeeperState(SafekeeperState state)
 static void
 AssertEventsOkForState(uint32 events, Safekeeper *sk)
 {
+	WalProposer *wp = sk->wp;
 	uint32		expected = SafekeeperStateDesiredEvents(sk->state);
 
 	/*
@@ -1994,8 +2007,8 @@ AssertEventsOkForState(uint32 events, Safekeeper *sk)
 		 * To give a descriptive message in the case of failure, we use elog
 		 * and then an assertion that's guaranteed to fail.
 		 */
-		elog(WARNING, "events %s mismatched for safekeeper %s:%s in state [%s]",
-			 FormatEvents(events), sk->host, sk->port, FormatSafekeeperState(sk->state));
+		walprop_log(WARNING, "events %s mismatched for safekeeper %s:%s in state [%s]",
+			 FormatEvents(wp, events), sk->host, sk->port, FormatSafekeeperState(sk->state));
 		Assert(events_ok_for_state);
 	}
 }
@@ -2068,7 +2081,7 @@ SafekeeperStateDesiredEvents(SafekeeperState state)
  * The string should not be freed. It should also not be expected to remain the same between
  * function calls. */
 static char *
-FormatEvents(uint32 events)
+FormatEvents(WalProposer *wp, uint32 events)
 {
 	static char return_str[8];
 
@@ -2097,7 +2110,7 @@ FormatEvents(uint32 events)
 
 	if (events & (~all_flags))
 	{
-		elog(WARNING, "Event formatting found unexpected component %d",
+		walprop_log(WARNING, "Event formatting found unexpected component %d",
 			 events & (~all_flags));
 		return_str[6] = '*';
 		return_str[7] = '\0';

pgxn/neon/walproposer.h

@@ -333,24 +333,11 @@ typedef struct Safekeeper
 	 */
 	char		conninfo[MAXCONNINFO];
 
-	/*
-	 * postgres protocol connection to the WAL acceptor
-	 *
-	 * Equals NULL only when state = SS_OFFLINE. Nonblocking is set once we
-	 * reach SS_ACTIVE; not before.
-	 */
-	WalProposerConn *conn;
-
 	/*
 	 * Temporary buffer for the message being sent to the safekeeper.
 	 */
 	StringInfoData outbuf;
 
-	/*
-	 * WAL reader, allocated for each safekeeper.
-	 */
-	XLogReaderState *xlogreader;
-
 	/*
 	 * Streaming will start here; must be record boundary.
 	 */
@@ -361,13 +348,43 @@ typedef struct Safekeeper
 	XLogRecPtr	streamingAt;	/* current streaming position */
 	AppendRequestHeader appendRequest;	/* request for sending to safekeeper */
 
-	int			eventPos;		/* position in wait event set. Equal to -1 if*
-								 * no event */
 	SafekeeperState state;		/* safekeeper state machine state */
 	TimestampTz latestMsgReceivedAt;	/* when latest msg is received */
 	AcceptorGreeting greetResponse; /* acceptor greeting */
 	VoteResponse voteResponse;	/* the vote */
 	AppendResponse appendResponse;	/* feedback for master */
+
+
+	/* postgres-specific fields */
+	#ifndef WALPROPOSER_LIB
+	/*
+	 * postgres protocol connection to the WAL acceptor
+	 *
+	 * Equals NULL only when state = SS_OFFLINE. Nonblocking is set once we
+	 * reach SS_ACTIVE; not before.
+	 */
+	WalProposerConn *conn;
+
+	/*
+	 * WAL reader, allocated for each safekeeper.
+	 */
+	XLogReaderState *xlogreader;
+
+	/*
+	 * Position in wait event set. Equal to -1 if no event
+	 */
+	int			eventPos;
+	#endif
+
+
+	/* WalProposer library specifics */
+	#ifdef WALPROPOSER_LIB
+	/*
+	 * Buffer for incoming messages. Usually Rust vector is stored here.
+	 * Caller is responsible for freeing the buffer.
+	 */
+	StringInfoData inbuf;
+	#endif
 } Safekeeper;
 
 /* Re-exported PostgresPollingStatusType */
@@ -433,7 +450,7 @@ typedef struct walproposer_api
 	 * Get WalproposerShmemState. This is used to store information about last
 	 * elected term.
 	 */
-	WalproposerShmemState *(*get_shmem_state) (void);
+	WalproposerShmemState *(*get_shmem_state) (WalProposer *wp);
 
 	/*
 	 * Start receiving notifications about new WAL. This is an infinite loop
@@ -443,61 +460,63 @@ typedef struct walproposer_api
 	void		(*start_streaming) (WalProposer *wp, XLogRecPtr startpos);
 
 	/* Get pointer to the latest available WAL. */
-	XLogRecPtr	(*get_flush_rec_ptr) (void);
+	XLogRecPtr	(*get_flush_rec_ptr) (WalProposer *wp);
 
 	/* Get current time. */
-	TimestampTz (*get_current_timestamp) (void);
-
-	/* Get postgres timeline. */
-	TimeLineID	(*get_timeline_id) (void);
+	TimestampTz (*get_current_timestamp) (WalProposer *wp);
 
 	/* Current error message, aka PQerrorMessage. */
-	char	   *(*conn_error_message) (WalProposerConn *conn);
+	char	   *(*conn_error_message) (Safekeeper *sk);
 
 	/* Connection status, aka PQstatus. */
-	WalProposerConnStatusType (*conn_status) (WalProposerConn *conn);
+	WalProposerConnStatusType (*conn_status) (Safekeeper *sk);
 
 	/* Start the connection, aka PQconnectStart. */
-	WalProposerConn *(*conn_connect_start) (char *conninfo);
+	void (*conn_connect_start) (Safekeeper *sk);
 
 	/* Poll an asynchronous connection, aka PQconnectPoll. */
-	WalProposerConnectPollStatusType (*conn_connect_poll) (WalProposerConn *conn);
+	WalProposerConnectPollStatusType (*conn_connect_poll) (Safekeeper *sk);
 
 	/* Send a blocking SQL query, aka PQsendQuery. */
-	bool		(*conn_send_query) (WalProposerConn *conn, char *query);
+	bool		(*conn_send_query) (Safekeeper *sk, char *query);
 
 	/* Read the query result, aka PQgetResult. */
-	WalProposerExecStatusType (*conn_get_query_result) (WalProposerConn *conn);
+	WalProposerExecStatusType (*conn_get_query_result) (Safekeeper *sk);
 
 	/* Flush buffer to the network, aka PQflush. */
-	int			(*conn_flush) (WalProposerConn *conn);
+	int			(*conn_flush) (Safekeeper *sk);
 
 	/* Close the connection, aka PQfinish. */
-	void		(*conn_finish) (WalProposerConn *conn);
+	void		(*conn_finish) (Safekeeper *sk);
 
-	/* Try to read CopyData message, aka PQgetCopyData. */
-	PGAsyncReadResult (*conn_async_read) (WalProposerConn *conn, char **buf, int *amount);
+	/*
+	 * Try to read CopyData message from the safekeeper, aka PQgetCopyData. 
+	 *
+	 * On success, the data is placed in *buf. It is valid until the next call
+	 * to this function.
+	 */
+	PGAsyncReadResult (*conn_async_read) (Safekeeper *sk, char **buf, int *amount);
 
 	/* Try to write CopyData message, aka PQputCopyData. */
-	PGAsyncWriteResult (*conn_async_write) (WalProposerConn *conn, void const *buf, size_t size);
+	PGAsyncWriteResult (*conn_async_write) (Safekeeper *sk, void const *buf, size_t size);
 
 	/* Blocking CopyData write, aka PQputCopyData + PQflush. */
-	bool		(*conn_blocking_write) (WalProposerConn *conn, void const *buf, size_t size);
+	bool		(*conn_blocking_write) (Safekeeper *sk, void const *buf, size_t size);
 
 	/* Download WAL from startpos to endpos and make it available locally. */
 	bool		(*recovery_download) (Safekeeper *sk, TimeLineID timeline, XLogRecPtr startpos, XLogRecPtr endpos);
 
 	/* Read WAL from disk to buf. */
-	void		(*wal_read) (XLogReaderState *state, char *buf, XLogRecPtr startptr, Size count);
+	void		(*wal_read) (Safekeeper *sk, char *buf, XLogRecPtr startptr, Size count);
 
 	/* Allocate WAL reader. */
-	XLogReaderState *(*wal_reader_allocate) (void);
+	void (*wal_reader_allocate) (Safekeeper *sk);
 
 	/* Deallocate event set. */
-	void		(*free_event_set) (void);
+	void		(*free_event_set) (WalProposer *wp);
 
 	/* Initialize event set. */
-	void		(*init_event_set) (int n_safekeepers);
+	void		(*init_event_set) (WalProposer *wp);
 
 	/* Update events for an existing safekeeper connection. */
 	void		(*update_event_set) (Safekeeper *sk, uint32 events);
@@ -513,22 +532,22 @@ typedef struct walproposer_api
 	 * events mask to indicate events and sets sk to the safekeeper which has
 	 * an event.
 	 */
-	int			(*wait_event_set) (long timeout, Safekeeper **sk, uint32 *events);
+	int			(*wait_event_set) (WalProposer *wp, long timeout, Safekeeper **sk, uint32 *events);
 
 	/* Read random bytes. */
-	bool		(*strong_random) (void *buf, size_t len);
+	bool		(*strong_random) (WalProposer *wp, void *buf, size_t len);
 
 	/*
 	 * Get a basebackup LSN. Used to cross-validate with the latest available
 	 * LSN on the safekeepers.
 	 */
-	XLogRecPtr	(*get_redo_start_lsn) (void);
+	XLogRecPtr	(*get_redo_start_lsn) (WalProposer *wp);
 
 	/*
 	 * Finish sync safekeepers with the given LSN. This function should not
 	 * return and should exit the program.
 	 */
-	void		(*finish_sync_safekeepers) (XLogRecPtr lsn);
+	void		(*finish_sync_safekeepers) (WalProposer *wp, XLogRecPtr lsn);
 
 	/*
 	 * Called after every new message from the safekeeper. Used to propagate
@@ -541,7 +560,22 @@ typedef struct walproposer_api
 	 * Called on peer_horizon_lsn updates. Used to advance replication slot
 	 * and to free up disk space by deleting unnecessary WAL.
 	 */
-	void		(*confirm_wal_streamed) (XLogRecPtr lsn);
+	void		(*confirm_wal_streamed) (WalProposer *wp, XLogRecPtr lsn);
+
+	/*
+	 * Write a log message to the internal log processor. This is used only
+	 * when walproposer is compiled as a library. Otherwise, all logging is
+	 * handled by elog().
+	 */
+	void		(*log_internal) (WalProposer *wp, int level, const char *line);
+
+	/*
+	 * Called right after the proposer was elected, but before it started
+	 * recovery and sent ProposerElected message to the safekeepers.
+	 * 
+	 * Used by logical replication to update truncateLsn.
+	 */
+	void		(*after_election) (WalProposer *wp);
 } walproposer_api;
 
 /*
@@ -590,6 +624,13 @@ typedef struct WalProposerConfig
 
 	/* Will be passed to safekeepers in greet request. */
 	uint64		systemId;
+
+	/* Will be passed to safekeepers in greet request. */
+	TimeLineID  pgTimeline;
+
+#ifdef WALPROPOSER_LIB
+	void *callback_data;
+#endif
 } WalProposerConfig;
 
 
@@ -666,7 +707,16 @@ extern WalProposer *WalProposerCreate(WalProposerConfig *config, walproposer_api
 extern void WalProposerStart(WalProposer *wp);
 extern void WalProposerBroadcast(WalProposer *wp, XLogRecPtr startpos, XLogRecPtr endpos);
 extern void WalProposerPoll(WalProposer *wp);
-extern void ParsePageserverFeedbackMessage(StringInfo reply_message,
-										   PageserverFeedback *rf);
+extern void WalProposerFree(WalProposer *wp);
+
+
+#define WPEVENT		1337	/* special log level for walproposer internal events */
+
+#ifdef WALPROPOSER_LIB
+void WalProposerLibLog(WalProposer *wp, int elevel, char *fmt, ...);
+#define walprop_log(elevel, ...) WalProposerLibLog(wp, elevel, __VA_ARGS__)
+#else
+#define walprop_log(elevel, ...) elog(elevel, __VA_ARGS__)
+#endif
 
 #endif							/* __NEON_WALPROPOSER_H__ */

pgxn/neon/walproposer_compat.c

@@ -0,0 +1,192 @@
+/*
+ * Contains copied/adapted functions from libpq and some internal postgres functions.
+ * This is needed to avoid linking to full postgres server installation. This file
+ * is compiled as a part of libwalproposer static library.
+ */
+
+#include <stdio.h>
+#include "walproposer.h"
+#include "utils/datetime.h"
+#include "miscadmin.h"
+
+void ExceptionalCondition(const char *conditionName,
+						  const char *fileName, int lineNumber)
+{
+	fprintf(stderr, "ExceptionalCondition: %s:%d: %s\n",
+			fileName, lineNumber, conditionName);
+	fprintf(stderr, "aborting...\n");
+	exit(1);
+}
+
+void
+pq_copymsgbytes(StringInfo msg, char *buf, int datalen)
+{
+	if (datalen < 0 || datalen > (msg->len - msg->cursor))
+		ExceptionalCondition("insufficient data left in message", __FILE__, __LINE__);
+	memcpy(buf, &msg->data[msg->cursor], datalen);
+	msg->cursor += datalen;
+}
+
+/* --------------------------------
+ *		pq_getmsgint	- get a binary integer from a message buffer
+ *
+ *		Values are treated as unsigned.
+ * --------------------------------
+ */
+unsigned int
+pq_getmsgint(StringInfo msg, int b)
+{
+	unsigned int result;
+	unsigned char n8;
+	uint16		n16;
+	uint32		n32;
+
+	switch (b)
+	{
+		case 1:
+			pq_copymsgbytes(msg, (char *) &n8, 1);
+			result = n8;
+			break;
+		case 2:
+			pq_copymsgbytes(msg, (char *) &n16, 2);
+			result = pg_ntoh16(n16);
+			break;
+		case 4:
+			pq_copymsgbytes(msg, (char *) &n32, 4);
+			result = pg_ntoh32(n32);
+			break;
+		default:
+			fprintf(stderr, "unsupported integer size %d\n", b);
+			ExceptionalCondition("unsupported integer size", __FILE__, __LINE__);
+			result = 0;			/* keep compiler quiet */
+			break;
+	}
+	return result;
+}
+
+/* --------------------------------
+ *		pq_getmsgint64	- get a binary 8-byte int from a message buffer
+ *
+ * It is tempting to merge this with pq_getmsgint, but we'd have to make the
+ * result int64 for all data widths --- that could be a big performance
+ * hit on machines where int64 isn't efficient.
+ * --------------------------------
+ */
+int64
+pq_getmsgint64(StringInfo msg)
+{
+	uint64		n64;
+
+	pq_copymsgbytes(msg, (char *) &n64, sizeof(n64));
+
+	return pg_ntoh64(n64);
+}
+
+/* --------------------------------
+ *		pq_getmsgbyte	- get a raw byte from a message buffer
+ * --------------------------------
+ */
+int
+pq_getmsgbyte(StringInfo msg)
+{
+	if (msg->cursor >= msg->len)
+		ExceptionalCondition("no data left in message", __FILE__, __LINE__);
+	return (unsigned char) msg->data[msg->cursor++];
+}
+
+/* --------------------------------
+ *		pq_getmsgbytes	- get raw data from a message buffer
+ *
+ *		Returns a pointer directly into the message buffer; note this
+ *		may not have any particular alignment.
+ * --------------------------------
+ */
+const char *
+pq_getmsgbytes(StringInfo msg, int datalen)
+{
+	const char *result;
+
+	if (datalen < 0 || datalen > (msg->len - msg->cursor))
+		ExceptionalCondition("insufficient data left in message", __FILE__, __LINE__);
+	result = &msg->data[msg->cursor];
+	msg->cursor += datalen;
+	return result;
+}
+
+/* --------------------------------
+ *		pq_getmsgstring - get a null-terminated text string (with conversion)
+ *
+ *		May return a pointer directly into the message buffer, or a pointer
+ *		to a palloc'd conversion result.
+ * --------------------------------
+ */
+const char *
+pq_getmsgstring(StringInfo msg)
+{
+	char	   *str;
+	int			slen;
+
+	str = &msg->data[msg->cursor];
+
+	/*
+	 * It's safe to use strlen() here because a StringInfo is guaranteed to
+	 * have a trailing null byte.  But check we found a null inside the
+	 * message.
+	 */
+	slen = strlen(str);
+	if (msg->cursor + slen >= msg->len)
+		ExceptionalCondition("invalid string in message", __FILE__, __LINE__);
+	msg->cursor += slen + 1;
+
+	return str;
+}
+
+/* --------------------------------
+ *		pq_getmsgend	- verify message fully consumed
+ * --------------------------------
+ */
+void
+pq_getmsgend(StringInfo msg)
+{
+	if (msg->cursor != msg->len)
+		ExceptionalCondition("invalid msg format", __FILE__, __LINE__);
+}
+
+
+/*
+ * Produce a C-string representation of a TimestampTz.
+ *
+ * This is mostly for use in emitting messages.
+ */
+const char *
+timestamptz_to_str(TimestampTz t)
+{
+	static char buf[MAXDATELEN + 1];
+
+	snprintf(buf, sizeof(buf), "TimestampTz(%ld)", t);
+	return buf;
+}
+
+bool
+TimestampDifferenceExceeds(TimestampTz start_time,
+								TimestampTz stop_time,
+								int msec)
+{
+	TimestampTz diff = stop_time - start_time;
+	return (diff >= msec * INT64CONST(1000));
+}
+
+void
+WalProposerLibLog(WalProposer *wp, int elevel, char *fmt, ...)
+{
+	char buf[1024];
+	va_list		args;
+
+	fmt = _(fmt);
+
+	va_start(args, fmt);
+	vsnprintf(buf, sizeof(buf), fmt, args);
+	va_end(args);
+
+	wp->api.log_internal(wp, elevel, buf);
+}

pgxn/neon/walproposer_pg.c

@@ -73,7 +73,8 @@ static void walprop_register_bgworker(void);
 static void walprop_pg_init_standalone_sync_safekeepers(void);
 static void walprop_pg_init_walsender(void);
 static void walprop_pg_init_bgworker(void);
-static TimestampTz walprop_pg_get_current_timestamp(void);
+static TimestampTz walprop_pg_get_current_timestamp(WalProposer *wp);
+static TimeLineID walprop_pg_get_timeline_id(void);
 static void walprop_pg_load_libpqwalreceiver(void);
 
 static process_interrupts_callback_t PrevProcessInterruptsCallback;
@@ -104,6 +105,7 @@ init_walprop_config(bool syncSafekeepers)
 		walprop_config.systemId = GetSystemIdentifier();
 	else
 		walprop_config.systemId = 0;
+	walprop_config.pgTimeline = walprop_pg_get_timeline_id();
 }
 
 /*
@@ -136,7 +138,7 @@ WalProposerMain(Datum main_arg)
 	walprop_pg_load_libpqwalreceiver();
 
 	wp = WalProposerCreate(&walprop_config, walprop_pg);
-	wp->last_reconnect_attempt = walprop_pg_get_current_timestamp();
+	wp->last_reconnect_attempt = walprop_pg_get_current_timestamp(wp);
 
 	walprop_pg_init_walsender();
 	WalProposerStart(wp);
@@ -379,7 +381,7 @@ nwp_shmem_startup_hook(void)
 }
 
 static WalproposerShmemState *
-walprop_pg_get_shmem_state(void)
+walprop_pg_get_shmem_state(WalProposer *wp)
 {
 	Assert(walprop_shared != NULL);
 	return walprop_shared;
@@ -505,7 +507,7 @@ walprop_pg_init_bgworker(void)
 }
 
 static XLogRecPtr
-walprop_pg_get_flush_rec_ptr(void)
+walprop_pg_get_flush_rec_ptr(WalProposer *wp)
 {
 #if PG_MAJORVERSION_NUM < 15
 	return GetFlushRecPtr();
@@ -515,7 +517,7 @@ walprop_pg_get_flush_rec_ptr(void)
 }
 
 static TimestampTz
-walprop_pg_get_current_timestamp(void)
+walprop_pg_get_current_timestamp(WalProposer *wp)
 {
 	return GetCurrentTimestamp();
 }
@@ -565,15 +567,15 @@ ensure_nonblocking_status(WalProposerConn *conn, bool is_nonblocking)
 
 /* Exported function definitions */
 static char *
-walprop_error_message(WalProposerConn *conn)
+walprop_error_message(Safekeeper *sk)
 {
-	return PQerrorMessage(conn->pg_conn);
+	return PQerrorMessage(sk->conn->pg_conn);
 }
 
 static WalProposerConnStatusType
-walprop_status(WalProposerConn *conn)
+walprop_status(Safekeeper *sk)
 {
-	switch (PQstatus(conn->pg_conn))
+	switch (PQstatus(sk->conn->pg_conn))
 	{
 		case CONNECTION_OK:
 			return WP_CONNECTION_OK;
@@ -584,16 +586,17 @@ walprop_status(WalProposerConn *conn)
 	}
 }
 
-static WalProposerConn *
-walprop_connect_start(char *conninfo)
+static void
+walprop_connect_start(Safekeeper *sk)
 {
-	WalProposerConn *conn;
 	PGconn	   *pg_conn;
 	const char *keywords[3];
 	const char *values[3];
 	int			n;
 	char	   *password = neon_auth_token;
 
+	Assert(sk->conn == NULL);
+
 	/*
 	 * Connect using the given connection string. If the NEON_AUTH_TOKEN
 	 * environment variable was set, use that as the password.
@@ -611,7 +614,7 @@ walprop_connect_start(char *conninfo)
 		n++;
 	}
 	keywords[n] = "dbname";
-	values[n] = conninfo;
+	values[n] = sk->conninfo;
 	n++;
 	keywords[n] = NULL;
 	values[n] = NULL;
@@ -619,11 +622,11 @@ walprop_connect_start(char *conninfo)
 	pg_conn = PQconnectStartParams(keywords, values, 1);
 
 	/*
-	 * Allocation of a PQconn can fail, and will return NULL. We want to fully
-	 * replicate the behavior of PQconnectStart here.
+	 * "If the result is null, then libpq has been unable to allocate a new
+	 * PGconn structure"
 	 */
 	if (!pg_conn)
-		return NULL;
+		elog(FATAL, "failed to allocate new PGconn object");
 
 	/*
 	 * And in theory this allocation can fail as well, but it's incredibly
@@ -632,20 +635,19 @@ walprop_connect_start(char *conninfo)
 	 * palloc will exit on failure though, so there's not much we could do if
 	 * it *did* fail.
 	 */
-	conn = palloc(sizeof(WalProposerConn));
-	conn->pg_conn = pg_conn;
-	conn->is_nonblocking = false;	/* connections always start in blocking
+	sk->conn = palloc(sizeof(WalProposerConn));
+	sk->conn->pg_conn = pg_conn;
+	sk->conn->is_nonblocking = false;	/* connections always start in blocking
 									 * mode */
-	conn->recvbuf = NULL;
-	return conn;
+	sk->conn->recvbuf = NULL;
 }
 
 static WalProposerConnectPollStatusType
-walprop_connect_poll(WalProposerConn *conn)
+walprop_connect_poll(Safekeeper *sk)
 {
 	WalProposerConnectPollStatusType return_val;
 
-	switch (PQconnectPoll(conn->pg_conn))
+	switch (PQconnectPoll(sk->conn->pg_conn))
 	{
 		case PGRES_POLLING_FAILED:
 			return_val = WP_CONN_POLLING_FAILED;
@@ -682,24 +684,24 @@ walprop_connect_poll(WalProposerConn *conn)
 }
 
 static bool
-walprop_send_query(WalProposerConn *conn, char *query)
+walprop_send_query(Safekeeper *sk, char *query)
 {
 	/*
 	 * We need to be in blocking mode for sending the query to run without
 	 * requiring a call to PQflush
 	 */
-	if (!ensure_nonblocking_status(conn, false))
+	if (!ensure_nonblocking_status(sk->conn, false))
 		return false;
 
 	/* PQsendQuery returns 1 on success, 0 on failure */
-	if (!PQsendQuery(conn->pg_conn, query))
+	if (!PQsendQuery(sk->conn->pg_conn, query))
 		return false;
 
 	return true;
 }
 
 static WalProposerExecStatusType
-walprop_get_query_result(WalProposerConn *conn)
+walprop_get_query_result(Safekeeper *sk)
 {
 	PGresult   *result;
 	WalProposerExecStatusType return_val;
@@ -708,14 +710,14 @@ walprop_get_query_result(WalProposerConn *conn)
 	char	   *unexpected_success = NULL;
 
 	/* Consume any input that we might be missing */
-	if (!PQconsumeInput(conn->pg_conn))
+	if (!PQconsumeInput(sk->conn->pg_conn))
 		return WP_EXEC_FAILED;
 
-	if (PQisBusy(conn->pg_conn))
+	if (PQisBusy(sk->conn->pg_conn))
 		return WP_EXEC_NEEDS_INPUT;
 
 
-	result = PQgetResult(conn->pg_conn);
+	result = PQgetResult(sk->conn->pg_conn);
 
 	/*
 	 * PQgetResult returns NULL only if getting the result was successful &
@@ -777,24 +779,28 @@ walprop_get_query_result(WalProposerConn *conn)
 }
 
 static pgsocket
-walprop_socket(WalProposerConn *conn)
+walprop_socket(Safekeeper *sk)
 {
-	return PQsocket(conn->pg_conn);
+	return PQsocket(sk->conn->pg_conn);
 }
 
 static int
-walprop_flush(WalProposerConn *conn)
+walprop_flush(Safekeeper *sk)
 {
-	return (PQflush(conn->pg_conn));
+	return (PQflush(sk->conn->pg_conn));
 }
 
 static void
-walprop_finish(WalProposerConn *conn)
+walprop_finish(Safekeeper *sk)
 {
-	if (conn->recvbuf != NULL)
-		PQfreemem(conn->recvbuf);
-	PQfinish(conn->pg_conn);
-	pfree(conn);
+	if (!sk->conn)
+		return;
+
+	if (sk->conn->recvbuf != NULL)
+		PQfreemem(sk->conn->recvbuf);
+	PQfinish(sk->conn->pg_conn);
+	pfree(sk->conn);
+	sk->conn = NULL;
 }
 
 /*
@@ -804,18 +810,18 @@ walprop_finish(WalProposerConn *conn)
  * to this function.
  */
 static PGAsyncReadResult
-walprop_async_read(WalProposerConn *conn, char **buf, int *amount)
+walprop_async_read(Safekeeper *sk, char **buf, int *amount)
 {
 	int			result;
 
-	if (conn->recvbuf != NULL)
+	if (sk->conn->recvbuf != NULL)
 	{
-		PQfreemem(conn->recvbuf);
-		conn->recvbuf = NULL;
+		PQfreemem(sk->conn->recvbuf);
+		sk->conn->recvbuf = NULL;
 	}
 
 	/* Call PQconsumeInput so that we have the data we need */
-	if (!PQconsumeInput(conn->pg_conn))
+	if (!PQconsumeInput(sk->conn->pg_conn))
 	{
 		*amount = 0;
 		*buf = NULL;
@@ -833,7 +839,7 @@ walprop_async_read(WalProposerConn *conn, char **buf, int *amount)
 	 * sometimes be triggered by the server returning an ErrorResponse (which
 	 * also happens to have the effect that the copy is done).
 	 */
-	switch (result = PQgetCopyData(conn->pg_conn, &conn->recvbuf, true))
+	switch (result = PQgetCopyData(sk->conn->pg_conn, &sk->conn->recvbuf, true))
 	{
 		case 0:
 			*amount = 0;
@@ -848,7 +854,7 @@ walprop_async_read(WalProposerConn *conn, char **buf, int *amount)
 				 * We can check PQgetResult to make sure that the server
 				 * failed; it'll always result in PGRES_FATAL_ERROR
 				 */
-				ExecStatusType status = PQresultStatus(PQgetResult(conn->pg_conn));
+				ExecStatusType status = PQresultStatus(PQgetResult(sk->conn->pg_conn));
 
 				if (status != PGRES_FATAL_ERROR)
 					elog(FATAL, "unexpected result status %d after failed PQgetCopyData", status);
@@ -869,18 +875,18 @@ walprop_async_read(WalProposerConn *conn, char **buf, int *amount)
 		default:
 			/* Positive values indicate the size of the returned result */
 			*amount = result;
-			*buf = conn->recvbuf;
+			*buf = sk->conn->recvbuf;
 			return PG_ASYNC_READ_SUCCESS;
 	}
 }
 
 static PGAsyncWriteResult
-walprop_async_write(WalProposerConn *conn, void const *buf, size_t size)
+walprop_async_write(Safekeeper *sk, void const *buf, size_t size)
 {
 	int			result;
 
 	/* If we aren't in non-blocking mode, switch to it. */
-	if (!ensure_nonblocking_status(conn, true))
+	if (!ensure_nonblocking_status(sk->conn, true))
 		return PG_ASYNC_WRITE_FAIL;
 
 	/*
@@ -888,7 +894,7 @@ walprop_async_write(WalProposerConn *conn, void const *buf, size_t size)
 	 * queued, 0 if it was not queued because of full buffers, or -1 if an
 	 * error occurred
 	 */
-	result = PQputCopyData(conn->pg_conn, buf, size);
+	result = PQputCopyData(sk->conn->pg_conn, buf, size);
 
 	/*
 	 * We won't get a result of zero because walproposer always empties the
@@ -916,7 +922,7 @@ walprop_async_write(WalProposerConn *conn, void const *buf, size_t size)
 	 * sucessful, 1 if it was unable to send all the data in the send queue
 	 * yet -1 if it failed for some reason
 	 */
-	switch (result = PQflush(conn->pg_conn))
+	switch (result = PQflush(sk->conn->pg_conn))
 	{
 		case 0:
 			return PG_ASYNC_WRITE_SUCCESS;
@@ -934,22 +940,22 @@ walprop_async_write(WalProposerConn *conn, void const *buf, size_t size)
  * information, refer to the comments there.
  */
 static bool
-walprop_blocking_write(WalProposerConn *conn, void const *buf, size_t size)
+walprop_blocking_write(Safekeeper *sk, void const *buf, size_t size)
 {
 	int			result;
 
 	/* If we are in non-blocking mode, switch out of it. */
-	if (!ensure_nonblocking_status(conn, false))
+	if (!ensure_nonblocking_status(sk->conn, false))
 		return false;
 
-	if ((result = PQputCopyData(conn->pg_conn, buf, size)) == -1)
+	if ((result = PQputCopyData(sk->conn->pg_conn, buf, size)) == -1)
 		return false;
 
 	Assert(result == 1);
 
 	/* Because the connection is non-blocking, flushing returns 0 or -1 */
 
-	if ((result = PQflush(conn->pg_conn)) == -1)
+	if ((result = PQflush(sk->conn->pg_conn)) == -1)
 		return false;
 
 	Assert(result == 0);
@@ -1381,11 +1387,11 @@ XLogWalPropClose(XLogRecPtr recptr)
 }
 
 static void
-walprop_pg_wal_read(XLogReaderState *state, char *buf, XLogRecPtr startptr, Size count)
+walprop_pg_wal_read(Safekeeper *sk, char *buf, XLogRecPtr startptr, Size count)
 {
 	WALReadError errinfo;
 
-	if (!WALRead(state,
+	if (!WALRead(sk->xlogreader,
 				 buf,
 				 startptr,
 				 count,
@@ -1396,31 +1402,38 @@ walprop_pg_wal_read(XLogReaderState *state, char *buf, XLogRecPtr startptr, Size
 	}
 }
 
-static XLogReaderState *
-walprop_pg_wal_reader_allocate(void)
+static void
+walprop_pg_wal_reader_allocate(Safekeeper *sk)
 {
-	return XLogReaderAllocate(wal_segment_size, NULL, XL_ROUTINE(.segment_open = wal_segment_open,.segment_close = wal_segment_close), NULL);
+	sk->xlogreader = XLogReaderAllocate(wal_segment_size, NULL, XL_ROUTINE(.segment_open = wal_segment_open,.segment_close = wal_segment_close), NULL);
+	if (sk->xlogreader == NULL)
+		elog(FATAL, "Failed to allocate xlog reader");
 }
 
 static WaitEventSet *waitEvents;
 
 static void
-walprop_pg_free_event_set(void)
+walprop_pg_free_event_set(WalProposer *wp)
 {
 	if (waitEvents)
 	{
 		FreeWaitEventSet(waitEvents);
 		waitEvents = NULL;
 	}
+
+	for (int i = 0; i < wp->n_safekeepers; i++)
+	{
+		wp->safekeeper[i].eventPos = -1;
+	}
 }
 
 static void
-walprop_pg_init_event_set(int n_safekeepers)
+walprop_pg_init_event_set(WalProposer *wp)
 {
 	if (waitEvents)
 		elog(FATAL, "double-initialization of event set");
 
-	waitEvents = CreateWaitEventSet(TopMemoryContext, 2 + n_safekeepers);
+	waitEvents = CreateWaitEventSet(TopMemoryContext, 2 + wp->n_safekeepers);
 	AddWaitEventToSet(waitEvents, WL_LATCH_SET, PGINVALID_SOCKET,
 					  MyLatch, NULL);
 	AddWaitEventToSet(waitEvents, WL_EXIT_ON_PM_DEATH, PGINVALID_SOCKET,
@@ -1439,11 +1452,11 @@ walprop_pg_update_event_set(Safekeeper *sk, uint32 events)
 static void
 walprop_pg_add_safekeeper_event_set(Safekeeper *sk, uint32 events)
 {
-	sk->eventPos = AddWaitEventToSet(waitEvents, events, walprop_socket(sk->conn), NULL, sk);
+	sk->eventPos = AddWaitEventToSet(waitEvents, events, walprop_socket(sk), NULL, sk);
 }
 
 static int
-walprop_pg_wait_event_set(long timeout, Safekeeper **sk, uint32 *events)
+walprop_pg_wait_event_set(WalProposer *wp, long timeout, Safekeeper **sk, uint32 *events)
 {
 	WaitEvent	event = {0};
 	int			rc = 0;
@@ -1499,7 +1512,7 @@ walprop_pg_wait_event_set(long timeout, Safekeeper **sk, uint32 *events)
 }
 
 static void
-walprop_pg_finish_sync_safekeepers(XLogRecPtr lsn)
+walprop_pg_finish_sync_safekeepers(WalProposer *wp, XLogRecPtr lsn)
 {
 	fprintf(stdout, "%X/%X\n", LSN_FORMAT_ARGS(lsn));
 	exit(0);
@@ -1611,7 +1624,7 @@ walprop_pg_process_safekeeper_feedback(WalProposer *wp, XLogRecPtr commitLsn)
 			 * pageserver.
 			 */
 								quorumFeedback.rf.disk_consistent_lsn,
-								walprop_pg_get_current_timestamp(), false);
+								walprop_pg_get_current_timestamp(wp), false);
 	}
 
 	CombineHotStanbyFeedbacks(&hsFeedback, wp);
@@ -1628,18 +1641,65 @@ walprop_pg_process_safekeeper_feedback(WalProposer *wp, XLogRecPtr commitLsn)
 }
 
 static void
-walprop_pg_confirm_wal_streamed(XLogRecPtr lsn)
+walprop_pg_confirm_wal_streamed(WalProposer *wp, XLogRecPtr lsn)
 {
 	if (MyReplicationSlot)
 		PhysicalConfirmReceivedLocation(lsn);
 }
 
+static XLogRecPtr
+walprop_pg_get_redo_start_lsn(WalProposer *wp)
+{
+	return GetRedoStartLsn();
+}
+
+static bool
+walprop_pg_strong_random(WalProposer *wp, void *buf, size_t len)
+{
+	return pg_strong_random(buf, len);
+}
+
+static void
+walprop_pg_log_internal(WalProposer *wp, int level, const char *line)
+{
+	elog(FATAL, "unexpected log_internal message at level %d: %s", level, line);
+}
+
+static void
+walprop_pg_after_election(WalProposer *wp)
+{
+	FILE* f;
+	XLogRecPtr lrRestartLsn;
+
+	/* We don't need to do anything in syncSafekeepers mode.*/
+	if (wp->config->syncSafekeepers)
+		return;
+
+	/*
+	 * If there are active logical replication subscription we need
+	 * to provide enough WAL for their WAL senders based on th position
+	 * of their replication slots.
+	 */
+	f = fopen("restart.lsn", "rb");
+	if (f != NULL && !wp->config->syncSafekeepers)
+	{
+		fread(&lrRestartLsn, sizeof(lrRestartLsn), 1, f);
+		fclose(f);
+		if (lrRestartLsn != InvalidXLogRecPtr)
+		{
+			elog(LOG, "Logical replication restart LSN %X/%X",  LSN_FORMAT_ARGS(lrRestartLsn));
+			/* start from the beginning of the segment to fetch page headers verifed by XLogReader */
+			lrRestartLsn = lrRestartLsn - XLogSegmentOffset(lrRestartLsn, wal_segment_size);
+			wp->truncateLsn = Min(wp->truncateLsn, lrRestartLsn);
+		}
+	}
+}
+
 static const walproposer_api walprop_pg = {
 	.get_shmem_state = walprop_pg_get_shmem_state,
 	.start_streaming = walprop_pg_start_streaming,
 	.get_flush_rec_ptr = walprop_pg_get_flush_rec_ptr,
 	.get_current_timestamp = walprop_pg_get_current_timestamp,
-	.get_timeline_id = walprop_pg_get_timeline_id,
 	.conn_error_message = walprop_error_message,
 	.conn_status = walprop_status,
 	.conn_connect_start = walprop_connect_start,
@@ -1659,9 +1719,11 @@ static const walproposer_api walprop_pg = {
 	.update_event_set = walprop_pg_update_event_set,
 	.add_safekeeper_event_set = walprop_pg_add_safekeeper_event_set,
 	.wait_event_set = walprop_pg_wait_event_set,
-	.strong_random = pg_strong_random,
-	.get_redo_start_lsn = GetRedoStartLsn,
+	.strong_random = walprop_pg_strong_random,
+	.get_redo_start_lsn = walprop_pg_get_redo_start_lsn,
 	.finish_sync_safekeepers = walprop_pg_finish_sync_safekeepers,
 	.process_safekeeper_feedback = walprop_pg_process_safekeeper_feedback,
 	.confirm_wal_streamed = walprop_pg_confirm_wal_streamed,
+	.log_internal = walprop_pg_log_internal,
+	.after_election = walprop_pg_after_election,
 };

poetry.lock

@@ -2415,13 +2415,13 @@ files = [
 
 [[package]]
 name = "urllib3"
-version = "1.26.17"
+version = "1.26.18"
 description = "HTTP library with thread-safe connection pooling, file post, and more."
 optional = false
 python-versions = ">=2.7, !=3.0.*, !=3.1.*, !=3.2.*, !=3.3.*, !=3.4.*, !=3.5.*"
 files = [
-    {file = "urllib3-1.26.17-py2.py3-none-any.whl", hash = "sha256:94a757d178c9be92ef5539b8840d48dc9cf1b2709c9d6b588232a055c524458b"},
-    {file = "urllib3-1.26.17.tar.gz", hash = "sha256:24d6a242c28d29af46c3fae832c36db3bbebcc533dd1bb549172cd739c82df21"},
+    {file = "urllib3-1.26.18-py2.py3-none-any.whl", hash = "sha256:34b97092d7e0a3a8cf7cd10e386f401b3737364026c45e622aa02903dffe0f07"},
+    {file = "urllib3-1.26.18.tar.gz", hash = "sha256:f8ecc1bba5667413457c529ab955bf8c67b45db799d159066261719e328580a0"},
 ]
 
 [package.extras]
@@ -2488,6 +2488,16 @@ files = [
     {file = "wrapt-1.14.1-cp310-cp310-musllinux_1_1_x86_64.whl", hash = "sha256:8ad85f7f4e20964db4daadcab70b47ab05c7c1cf2a7c1e51087bfaa83831854c"},
     {file = "wrapt-1.14.1-cp310-cp310-win32.whl", hash = "sha256:a9a52172be0b5aae932bef82a79ec0a0ce87288c7d132946d645eba03f0ad8a8"},
     {file = "wrapt-1.14.1-cp310-cp310-win_amd64.whl", hash = "sha256:6d323e1554b3d22cfc03cd3243b5bb815a51f5249fdcbb86fda4bf62bab9e164"},
+    {file = "wrapt-1.14.1-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:ecee4132c6cd2ce5308e21672015ddfed1ff975ad0ac8d27168ea82e71413f55"},
+    {file = "wrapt-1.14.1-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:2020f391008ef874c6d9e208b24f28e31bcb85ccff4f335f15a3251d222b92d9"},
+    {file = "wrapt-1.14.1-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:2feecf86e1f7a86517cab34ae6c2f081fd2d0dac860cb0c0ded96d799d20b335"},
+    {file = "wrapt-1.14.1-cp311-cp311-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:240b1686f38ae665d1b15475966fe0472f78e71b1b4903c143a842659c8e4cb9"},
+    {file = "wrapt-1.14.1-cp311-cp311-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:a9008dad07d71f68487c91e96579c8567c98ca4c3881b9b113bc7b33e9fd78b8"},
+    {file = "wrapt-1.14.1-cp311-cp311-musllinux_1_1_aarch64.whl", hash = "sha256:6447e9f3ba72f8e2b985a1da758767698efa72723d5b59accefd716e9e8272bf"},
+    {file = "wrapt-1.14.1-cp311-cp311-musllinux_1_1_i686.whl", hash = "sha256:acae32e13a4153809db37405f5eba5bac5fbe2e2ba61ab227926a22901051c0a"},
+    {file = "wrapt-1.14.1-cp311-cp311-musllinux_1_1_x86_64.whl", hash = "sha256:49ef582b7a1152ae2766557f0550a9fcbf7bbd76f43fbdc94dd3bf07cc7168be"},
+    {file = "wrapt-1.14.1-cp311-cp311-win32.whl", hash = "sha256:358fe87cc899c6bb0ddc185bf3dbfa4ba646f05b1b0b9b5a27c2cb92c2cea204"},
+    {file = "wrapt-1.14.1-cp311-cp311-win_amd64.whl", hash = "sha256:26046cd03936ae745a502abf44dac702a5e6880b2b01c29aea8ddf3353b68224"},
     {file = "wrapt-1.14.1-cp35-cp35m-manylinux1_i686.whl", hash = "sha256:43ca3bbbe97af00f49efb06e352eae40434ca9d915906f77def219b88e85d907"},
     {file = "wrapt-1.14.1-cp35-cp35m-manylinux1_x86_64.whl", hash = "sha256:6b1a564e6cb69922c7fe3a678b9f9a3c54e72b469875aa8018f18b4d1dd1adf3"},
     {file = "wrapt-1.14.1-cp35-cp35m-manylinux2010_i686.whl", hash = "sha256:00b6d4ea20a906c0ca56d84f93065b398ab74b927a7a3dbd470f6fc503f95dc3"},

proxy/src/auth/backend.rs

@@ -3,9 +3,12 @@ mod hacks;
 mod link;
 
 pub use link::LinkAuthError;
+use tokio_postgres::config::AuthKeys;
 
+use crate::proxy::{handle_try_wake, retry_after};
 use crate::{
     auth::{self, ClientCredentials},
+    config::AuthenticationConfig,
     console::{
         self,
         provider::{CachedNodeInfo, ConsoleReqExtra},
@@ -15,8 +18,9 @@ use crate::{
 };
 use futures::TryFutureExt;
 use std::borrow::Cow;
+use std::ops::ControlFlow;
 use tokio::io::{AsyncRead, AsyncWrite};
-use tracing::info;
+use tracing::{error, info, warn};
 
 /// A product of successful authentication.
 pub struct AuthSuccess<T> {
@@ -116,21 +120,27 @@ impl<'a, T, E> BackendType<'a, Result<T, E>> {
     }
 }
 
+pub enum ComputeCredentials {
+    Password(Vec<u8>),
+    AuthKeys(AuthKeys),
+}
+
 /// True to its name, this function encapsulates our current auth trade-offs.
 /// Here, we choose the appropriate auth flow based on circumstances.
-async fn auth_quirks(
+async fn auth_quirks_creds(
     api: &impl console::Api,
     extra: &ConsoleReqExtra<'_>,
     creds: &mut ClientCredentials<'_>,
     client: &mut stream::PqStream<impl AsyncRead + AsyncWrite + Unpin>,
     allow_cleartext: bool,
-) -> auth::Result<AuthSuccess<CachedNodeInfo>> {
+    config: &'static AuthenticationConfig,
+) -> auth::Result<AuthSuccess<ComputeCredentials>> {
     // If there's no project so far, that entails that client doesn't
     // support SNI or other means of passing the endpoint (project) name.
     // We now expect to see a very specific payload in the place of password.
     if creds.project.is_none() {
         // Password will be checked by the compute node later.
-        return hacks::password_hack(api, extra, creds, client).await;
+        return hacks::password_hack(creds, client).await;
     }
 
     // Password hack should set the project name.
@@ -141,11 +151,53 @@ async fn auth_quirks(
     // Currently, we use it for websocket connections (latency).
     if allow_cleartext {
         // Password will be checked by the compute node later.
-        return hacks::cleartext_hack(api, extra, creds, client).await;
+        return hacks::cleartext_hack(client).await;
     }
 
     // Finally, proceed with the main auth flow (SCRAM-based).
-    classic::authenticate(api, extra, creds, client).await
+    classic::authenticate(api, extra, creds, client, config).await
+}
+
+/// True to its name, this function encapsulates our current auth trade-offs.
+/// Here, we choose the appropriate auth flow based on circumstances.
+async fn auth_quirks(
+    api: &impl console::Api,
+    extra: &ConsoleReqExtra<'_>,
+    creds: &mut ClientCredentials<'_>,
+    client: &mut stream::PqStream<impl AsyncRead + AsyncWrite + Unpin>,
+    allow_cleartext: bool,
+    config: &'static AuthenticationConfig,
+) -> auth::Result<AuthSuccess<CachedNodeInfo>> {
+    let auth_stuff = auth_quirks_creds(api, extra, creds, client, allow_cleartext, config).await?;
+
+    let mut num_retries = 0;
+    let mut node = loop {
+        let wake_res = api.wake_compute(extra, creds).await;
+        match handle_try_wake(wake_res, num_retries) {
+            Err(e) => {
+                error!(error = ?e, num_retries, retriable = false, "couldn't wake compute node");
+                return Err(e.into());
+            }
+            Ok(ControlFlow::Continue(e)) => {
+                warn!(error = ?e, num_retries, retriable = true, "couldn't wake compute node");
+            }
+            Ok(ControlFlow::Break(n)) => break n,
+        }
+
+        let wait_duration = retry_after(num_retries);
+        num_retries += 1;
+        tokio::time::sleep(wait_duration).await;
+    };
+
+    match auth_stuff.value {
+        ComputeCredentials::Password(password) => node.config.password(password),
+        ComputeCredentials::AuthKeys(auth_keys) => node.config.auth_keys(auth_keys),
+    };
+
+    Ok(AuthSuccess {
+        reported_auth_ok: auth_stuff.reported_auth_ok,
+        value: node,
+    })
 }
 
 impl BackendType<'_, ClientCredentials<'_>> {
@@ -180,6 +232,7 @@ impl BackendType<'_, ClientCredentials<'_>> {
         extra: &ConsoleReqExtra<'_>,
         client: &mut stream::PqStream<impl AsyncRead + AsyncWrite + Unpin>,
         allow_cleartext: bool,
+        config: &'static AuthenticationConfig,
     ) -> auth::Result<AuthSuccess<CachedNodeInfo>> {
         use BackendType::*;
 
@@ -192,7 +245,7 @@ impl BackendType<'_, ClientCredentials<'_>> {
                 );
 
                 let api = api.as_ref();
-                auth_quirks(api, extra, creds, client, allow_cleartext).await?
+                auth_quirks(api, extra, creds, client, allow_cleartext, config).await?
             }
             Postgres(api, creds) => {
                 info!(
@@ -202,7 +255,7 @@ impl BackendType<'_, ClientCredentials<'_>> {
                 );
 
                 let api = api.as_ref();
-                auth_quirks(api, extra, creds, client, allow_cleartext).await?
+                auth_quirks(api, extra, creds, client, allow_cleartext, config).await?
             }
             // NOTE: this auth backend doesn't use client credentials.
             Link(url) => {

proxy/src/auth/backend/classic.rs

@@ -1,23 +1,22 @@
-use std::ops::ControlFlow;
-
-use super::AuthSuccess;
+use super::{AuthSuccess, ComputeCredentials};
 use crate::{
     auth::{self, AuthFlow, ClientCredentials},
     compute,
-    console::{self, AuthInfo, CachedNodeInfo, ConsoleReqExtra},
-    proxy::{handle_try_wake, retry_after},
+    config::AuthenticationConfig,
+    console::{self, AuthInfo, ConsoleReqExtra},
     sasl, scram,
     stream::PqStream,
 };
 use tokio::io::{AsyncRead, AsyncWrite};
-use tracing::{error, info, warn};
+use tracing::{info, warn};
 
 pub(super) async fn authenticate(
     api: &impl console::Api,
     extra: &ConsoleReqExtra<'_>,
     creds: &ClientCredentials<'_>,
     client: &mut PqStream<impl AsyncRead + AsyncWrite + Unpin>,
-) -> auth::Result<AuthSuccess<CachedNodeInfo>> {
+    config: &'static AuthenticationConfig,
+) -> auth::Result<AuthSuccess<ComputeCredentials>> {
     info!("fetching user's authentication info");
     let info = api.get_auth_info(extra, creds).await?.unwrap_or_else(|| {
         // If we don't have an authentication secret, we mock one to
@@ -42,7 +41,16 @@ pub(super) async fn authenticate(
                 error
             })?;
 
-            let auth_outcome = auth_flow.authenticate().await.map_err(|error| {
+            let auth_outcome = tokio::time::timeout(
+                config.scram_protocol_timeout,
+                auth_flow.authenticate(),
+            )
+            .await
+            .map_err(|error| {
+                warn!("error processing scram messages error = authentication timed out, execution time exeeded {} seconds", config.scram_protocol_timeout.as_secs());
+                auth::io::Error::new(auth::io::ErrorKind::TimedOut, error)
+            })?
+            .map_err(|error| {
                 warn!(?error, "error processing scram messages");
                 error
             })?;
@@ -55,38 +63,17 @@ pub(super) async fn authenticate(
                 }
             };
 
-            Some(compute::ScramKeys {
+            compute::ScramKeys {
                 client_key: client_key.as_bytes(),
                 server_key: secret.server_key.as_bytes(),
-            })
+            }
         }
     };
 
-    let mut num_retries = 0;
-    let mut node = loop {
-        let wake_res = api.wake_compute(extra, creds).await;
-        match handle_try_wake(wake_res, num_retries) {
-            Err(e) => {
-                error!(error = ?e, num_retries, retriable = false, "couldn't wake compute node");
-                return Err(e.into());
-            }
-            Ok(ControlFlow::Continue(e)) => {
-                warn!(error = ?e, num_retries, retriable = true, "couldn't wake compute node");
-            }
-            Ok(ControlFlow::Break(n)) => break n,
-        }
-
-        let wait_duration = retry_after(num_retries);
-        num_retries += 1;
-        tokio::time::sleep(wait_duration).await;
-    };
-    if let Some(keys) = scram_keys {
-        use tokio_postgres::config::AuthKeys;
-        node.config.auth_keys(AuthKeys::ScramSha256(keys));
-    }
-
     Ok(AuthSuccess {
         reported_auth_ok: false,
-        value: node,
+        value: ComputeCredentials::AuthKeys(tokio_postgres::config::AuthKeys::ScramSha256(
+            scram_keys,
+        )),
     })
 }

proxy/src/auth/backend/hacks.rs

@@ -1,10 +1,6 @@
-use super::AuthSuccess;
+use super::{AuthSuccess, ComputeCredentials};
 use crate::{
     auth::{self, AuthFlow, ClientCredentials},
-    console::{
-        self,
-        provider::{CachedNodeInfo, ConsoleReqExtra},
-    },
     stream,
 };
 use tokio::io::{AsyncRead, AsyncWrite};
@@ -15,11 +11,8 @@ use tracing::{info, warn};
 /// These properties are benefical for serverless JS workers, so we
 /// use this mechanism for websocket connections.
 pub async fn cleartext_hack(
-    api: &impl console::Api,
-    extra: &ConsoleReqExtra<'_>,
-    creds: &mut ClientCredentials<'_>,
     client: &mut stream::PqStream<impl AsyncRead + AsyncWrite + Unpin>,
-) -> auth::Result<AuthSuccess<CachedNodeInfo>> {
+) -> auth::Result<AuthSuccess<ComputeCredentials>> {
     warn!("cleartext auth flow override is enabled, proceeding");
     let password = AuthFlow::new(client)
         .begin(auth::CleartextPassword)
@@ -27,24 +20,19 @@ pub async fn cleartext_hack(
         .authenticate()
         .await?;
 
-    let mut node = api.wake_compute(extra, creds).await?;
-    node.config.password(password);
-
     // Report tentative success; compute node will check the password anyway.
     Ok(AuthSuccess {
         reported_auth_ok: false,
-        value: node,
+        value: ComputeCredentials::Password(password),
     })
 }
 
 /// Workaround for clients which don't provide an endpoint (project) name.
 /// Very similar to [`cleartext_hack`], but there's a specific password format.
 pub async fn password_hack(
-    api: &impl console::Api,
-    extra: &ConsoleReqExtra<'_>,
     creds: &mut ClientCredentials<'_>,
     client: &mut stream::PqStream<impl AsyncRead + AsyncWrite + Unpin>,
-) -> auth::Result<AuthSuccess<CachedNodeInfo>> {
+) -> auth::Result<AuthSuccess<ComputeCredentials>> {
     warn!("project not specified, resorting to the password hack auth flow");
     let payload = AuthFlow::new(client)
         .begin(auth::PasswordHack)
@@ -55,12 +43,9 @@ pub async fn password_hack(
     info!(project = &payload.endpoint, "received missing parameter");
     creds.project = Some(payload.endpoint);
 
-    let mut node = api.wake_compute(extra, creds).await?;
-    node.config.password(payload.password);
-
     // Report tentative success; compute node will check the password anyway.
     Ok(AuthSuccess {
         reported_auth_ok: false,
-        value: node,
+        value: ComputeCredentials::Password(payload.password),
     })
 }

proxy/src/bin/proxy.rs

@@ -1,5 +1,6 @@
 use futures::future::Either;
 use proxy::auth;
+use proxy::config::AuthenticationConfig;
 use proxy::config::HttpConfig;
 use proxy::console;
 use proxy::http;
@@ -83,6 +84,12 @@ struct ProxyCliArgs {
     /// timeout for http connections
     #[clap(long, default_value = "15s", value_parser = humantime::parse_duration)]
     sql_over_http_timeout: tokio::time::Duration,
+    /// timeout for scram authentication protocol
+    #[clap(long, default_value = "15s", value_parser = humantime::parse_duration)]
+    scram_protocol_timeout: tokio::time::Duration,
+    /// Require that all incoming requests have a Proxy Protocol V2 packet **and** have an IP address associated.
+    #[clap(long, default_value_t = false, value_parser = clap::builder::BoolishValueParser::new(), action = clap::ArgAction::Set)]
+    require_client_ip: bool,
 }
 
 #[tokio::main]
@@ -227,12 +234,17 @@ fn build_config(args: &ProxyCliArgs) -> anyhow::Result<&'static ProxyConfig> {
     let http_config = HttpConfig {
         sql_over_http_timeout: args.sql_over_http_timeout,
     };
+    let authentication_config = AuthenticationConfig {
+        scram_protocol_timeout: args.scram_protocol_timeout,
+    };
     let config = Box::leak(Box::new(ProxyConfig {
         tls_config,
         auth_backend,
         metric_collection,
         allow_self_signed_compute: args.allow_self_signed_compute,
         http_config,
+        authentication_config,
+        require_client_ip: args.require_client_ip,
     }));
 
     Ok(config)

proxy/src/cancellation.rs

@@ -1,5 +1,5 @@
-use anyhow::{anyhow, Context};
-use hashbrown::HashMap;
+use anyhow::{bail, Context};
+use dashmap::DashMap;
 use pq_proto::CancelKeyData;
 use std::net::SocketAddr;
 use tokio::net::TcpStream;
@@ -8,7 +8,7 @@ use tracing::info;
 
 /// Enables serving `CancelRequest`s.
 #[derive(Default)]
-pub struct CancelMap(parking_lot::RwLock<HashMap<CancelKeyData, Option<CancelClosure>>>);
+pub struct CancelMap(DashMap<CancelKeyData, Option<CancelClosure>>);
 
 impl CancelMap {
     /// Cancel a running query for the corresponding connection.
@@ -16,7 +16,6 @@ impl CancelMap {
         // NB: we should immediately release the lock after cloning the token.
         let cancel_closure = self
             .0
-            .read()
             .get(&key)
             .and_then(|x| x.clone())
             .with_context(|| format!("query cancellation key not found: {key}"))?;
@@ -40,15 +39,19 @@ impl CancelMap {
 
         // Random key collisions are unlikely to happen here, but they're still possible,
         // which is why we have to take care not to rewrite an existing key.
-        self.0
-            .write()
-            .try_insert(key, None)
-            .map_err(|_| anyhow!("query cancellation key already exists: {key}"))?;
+        match self.0.entry(key) {
+            dashmap::mapref::entry::Entry::Occupied(_) => {
+                bail!("query cancellation key already exists: {key}")
+            }
+            dashmap::mapref::entry::Entry::Vacant(e) => {
+                e.insert(None);
+            }
+        }
 
         // This will guarantee that the session gets dropped
         // as soon as the future is finished.
         scopeguard::defer! {
-            self.0.write().remove(&key);
+            self.0.remove(&key);
             info!("dropped query cancellation key {key}");
         }
 
@@ -59,12 +62,12 @@ impl CancelMap {
 
     #[cfg(test)]
     fn contains(&self, session: &Session) -> bool {
-        self.0.read().contains_key(&session.key)
+        self.0.contains_key(&session.key)
     }
 
     #[cfg(test)]
     fn is_empty(&self) -> bool {
-        self.0.read().is_empty()
+        self.0.is_empty()
     }
 }
 
@@ -113,10 +116,7 @@ impl Session<'_> {
     /// This enables query cancellation in `crate::proxy::prepare_client_connection`.
     pub fn enable_query_cancellation(self, cancel_closure: CancelClosure) -> CancelKeyData {
         info!("enabling query cancellation for this session");
-        self.cancel_map
-            .0
-            .write()
-            .insert(self.key, Some(cancel_closure));
+        self.cancel_map.0.insert(self.key, Some(cancel_closure));
 
         self.key
     }

proxy/src/config.rs

@@ -14,6 +14,8 @@ pub struct ProxyConfig {
     pub metric_collection: Option<MetricCollectionConfig>,
     pub allow_self_signed_compute: bool,
     pub http_config: HttpConfig,
+    pub authentication_config: AuthenticationConfig,
+    pub require_client_ip: bool,
 }
 
 #[derive(Debug)]
@@ -31,6 +33,10 @@ pub struct HttpConfig {
     pub sql_over_http_timeout: tokio::time::Duration,
 }
 
+pub struct AuthenticationConfig {
+    pub scram_protocol_timeout: tokio::time::Duration,
+}
+
 impl TlsConfig {
     pub fn to_server_config(&self) -> Arc<rustls::ServerConfig> {
         self.config.clone()

proxy/src/console/provider.rs

@@ -90,7 +90,11 @@ pub mod errors {
                     status: http::StatusCode::LOCKED,
                     ref text,
                 } => {
-                    !text.contains("written data quota exceeded")
+                    // written data quota exceeded
+                    // data transfer quota exceeded
+                    // compute time quota exceeded
+                    // logical size quota exceeded
+                    !text.contains("quota exceeded")
                         && !text.contains("the limit for current plan reached")
                 }
                 // retry server errors

proxy/src/http/conn_pool.rs

@@ -8,31 +8,34 @@ use pbkdf2::{
     Params, Pbkdf2,
 };
 use pq_proto::StartupMessageParams;
-use std::sync::atomic::{self, AtomicUsize};
 use std::{collections::HashMap, sync::Arc};
 use std::{
     fmt,
     task::{ready, Poll},
 };
+use std::{
+    ops::Deref,
+    sync::atomic::{self, AtomicUsize},
+};
 use tokio::time;
-use tokio_postgres::AsyncMessage;
+use tokio_postgres::{AsyncMessage, ReadyForQueryStatus};
 
 use crate::{
     auth, console,
     metrics::{Ids, MetricCounter, USAGE_METRICS},
-    proxy::{NUM_DB_CONNECTIONS_CLOSED_COUNTER, NUM_DB_CONNECTIONS_OPENED_COUNTER},
+    proxy::{LatencyTimer, NUM_DB_CONNECTIONS_CLOSED_COUNTER, NUM_DB_CONNECTIONS_OPENED_COUNTER},
 };
 use crate::{compute, config};
 
 use crate::proxy::ConnectMechanism;
 
-use tracing::{error, warn};
+use tracing::{error, warn, Span};
 use tracing::{info, info_span, Instrument};
 
 pub const APP_NAME: &str = "sql_over_http";
 const MAX_CONNS_PER_ENDPOINT: usize = 20;
 
-#[derive(Debug)]
+#[derive(Debug, Clone)]
 pub struct ConnInfo {
     pub username: String,
     pub dbname: String,
@@ -55,7 +58,7 @@ impl fmt::Display for ConnInfo {
 }
 
 struct ConnPoolEntry {
-    conn: Client,
+    conn: ClientInner,
     _last_access: std::time::Instant,
 }
 
@@ -133,12 +136,19 @@ impl GlobalConnPool {
     }
 
     pub async fn get(
-        &self,
+        self: &Arc<Self>,
         conn_info: &ConnInfo,
         force_new: bool,
         session_id: uuid::Uuid,
     ) -> anyhow::Result<Client> {
-        let mut client: Option<Client> = None;
+        let mut client: Option<ClientInner> = None;
+        let mut latency_timer = LatencyTimer::new("http");
+
+        let pool = if force_new {
+            None
+        } else {
+            Some((conn_info.clone(), self.clone()))
+        };
 
         let mut hash_valid = false;
         if !force_new {
@@ -182,15 +192,21 @@ impl GlobalConnPool {
         let new_client = if let Some(client) = client {
             if client.inner.is_closed() {
                 info!("pool: cached connection '{conn_info}' is closed, opening a new one");
-                connect_to_compute(self.proxy_config, conn_info, session_id).await
+                connect_to_compute(self.proxy_config, conn_info, session_id, latency_timer).await
             } else {
                 info!("pool: reusing connection '{conn_info}'");
                 client.session.send(session_id)?;
-                return Ok(client);
+                latency_timer.pool_hit();
+                latency_timer.success();
+                return Ok(Client {
+                    inner: Some(client),
+                    span: Span::current(),
+                    pool,
+                });
             }
         } else {
             info!("pool: opening a new connection '{conn_info}'");
-            connect_to_compute(self.proxy_config, conn_info, session_id).await
+            connect_to_compute(self.proxy_config, conn_info, session_id, latency_timer).await
         };
 
         match &new_client {
@@ -226,10 +242,14 @@ impl GlobalConnPool {
             _ => {}
         }
 
-        new_client
+        new_client.map(|inner| Client {
+            inner: Some(inner),
+            span: Span::current(),
+            pool,
+        })
     }
 
-    pub fn put(&self, conn_info: &ConnInfo, client: Client) -> anyhow::Result<()> {
+    fn put(&self, conn_info: &ConnInfo, client: ClientInner) -> anyhow::Result<()> {
         // We want to hold this open while we return. This ensures that the pool can't close
         // while we are in the middle of returning the connection.
         let closed = self.closed.read();
@@ -324,7 +344,7 @@ struct TokioMechanism<'a> {
 
 #[async_trait]
 impl ConnectMechanism for TokioMechanism<'_> {
-    type Connection = Client;
+    type Connection = ClientInner;
     type ConnectError = tokio_postgres::Error;
     type Error = anyhow::Error;
 
@@ -347,7 +367,8 @@ async fn connect_to_compute(
     config: &config::ProxyConfig,
     conn_info: &ConnInfo,
     session_id: uuid::Uuid,
-) -> anyhow::Result<Client> {
+    latency_timer: LatencyTimer,
+) -> anyhow::Result<ClientInner> {
     let tls = config.tls_config.as_ref();
     let common_names = tls.and_then(|tls| tls.common_names.clone());
 
@@ -386,6 +407,7 @@ async fn connect_to_compute(
         node_info,
         &extra,
         &creds,
+        latency_timer,
     )
     .await
 }
@@ -395,7 +417,7 @@ async fn connect_to_compute_once(
     conn_info: &ConnInfo,
     timeout: time::Duration,
     mut session: uuid::Uuid,
-) -> Result<Client, tokio_postgres::Error> {
+) -> Result<ClientInner, tokio_postgres::Error> {
     let mut config = (*node_info.config).clone();
 
     let (client, mut connection) = config
@@ -458,21 +480,99 @@ async fn connect_to_compute_once(
         .instrument(span)
     );
 
-    Ok(Client {
+    Ok(ClientInner {
         inner: client,
         session: tx,
         ids,
     })
 }
 
-pub struct Client {
-    pub inner: tokio_postgres::Client,
+struct ClientInner {
+    inner: tokio_postgres::Client,
     session: tokio::sync::watch::Sender<uuid::Uuid>,
     ids: Ids,
 }
 
 impl Client {
     pub fn metrics(&self) -> Arc<MetricCounter> {
-        USAGE_METRICS.register(self.ids.clone())
+        USAGE_METRICS.register(self.inner.as_ref().unwrap().ids.clone())
+    }
+}
+
+pub struct Client {
+    span: Span,
+    inner: Option<ClientInner>,
+    pool: Option<(ConnInfo, Arc<GlobalConnPool>)>,
+}
+
+pub struct Discard<'a> {
+    pool: &'a mut Option<(ConnInfo, Arc<GlobalConnPool>)>,
+}
+
+impl Client {
+    pub fn inner(&mut self) -> (&mut tokio_postgres::Client, Discard<'_>) {
+        let Self {
+            inner,
+            pool,
+            span: _,
+        } = self;
+        (
+            &mut inner
+                .as_mut()
+                .expect("client inner should not be removed")
+                .inner,
+            Discard { pool },
+        )
+    }
+
+    pub fn check_idle(&mut self, status: ReadyForQueryStatus) {
+        self.inner().1.check_idle(status)
+    }
+    pub fn discard(&mut self) {
+        self.inner().1.discard()
+    }
+}
+
+impl Discard<'_> {
+    pub fn check_idle(&mut self, status: ReadyForQueryStatus) {
+        if status != ReadyForQueryStatus::Idle {
+            if let Some((conn_info, _)) = self.pool.take() {
+                info!("pool: throwing away connection '{conn_info}' because connection is not idle")
+            }
+        }
+    }
+    pub fn discard(&mut self) {
+        if let Some((conn_info, _)) = self.pool.take() {
+            info!("pool: throwing away connection '{conn_info}' because connection is potentially in a broken state")
+        }
+    }
+}
+
+impl Deref for Client {
+    type Target = tokio_postgres::Client;
+
+    fn deref(&self) -> &Self::Target {
+        &self
+            .inner
+            .as_ref()
+            .expect("client inner should not be removed")
+            .inner
+    }
+}
+
+impl Drop for Client {
+    fn drop(&mut self) {
+        let client = self
+            .inner
+            .take()
+            .expect("client inner should not be removed");
+        if let Some((conn_info, conn_pool)) = self.pool.take() {
+            let current_span = self.span.clone();
+            // return connection to the pool
+            tokio::task::spawn_blocking(move || {
+                let _span = current_span.enter();
+                let _ = conn_pool.put(&conn_info, client);
+            });
+        }
     }
 }

proxy/src/http/sql_over_http.rs

@@ -17,7 +17,9 @@ use tokio_postgres::types::Kind;
 use tokio_postgres::types::Type;
 use tokio_postgres::GenericClient;
 use tokio_postgres::IsolationLevel;
+use tokio_postgres::ReadyForQueryStatus;
 use tokio_postgres::Row;
+use tokio_postgres::Transaction;
 use tracing::error;
 use tracing::instrument;
 use url::Url;
@@ -64,20 +66,18 @@ static HEADER_VALUE_TRUE: HeaderValue = HeaderValue::from_static("true");
 // Convert json non-string types to strings, so that they can be passed to Postgres
 // as parameters.
 //
-fn json_to_pg_text(json: Vec<Value>) -> Result<Vec<Option<String>>, serde_json::Error> {
+fn json_to_pg_text(json: Vec<Value>) -> Vec<Option<String>> {
     json.iter()
         .map(|value| {
             match value {
                 // special care for nulls
-                Value::Null => Ok(None),
+                Value::Null => None,
 
                 // convert to text with escaping
-                Value::Bool(_) => serde_json::to_string(value).map(Some),
-                Value::Number(_) => serde_json::to_string(value).map(Some),
-                Value::Object(_) => serde_json::to_string(value).map(Some),
+                v @ (Value::Bool(_) | Value::Number(_) | Value::Object(_)) => Some(v.to_string()),
 
                 // avoid escaping here, as we pass this as a parameter
-                Value::String(s) => Ok(Some(s.to_string())),
+                Value::String(s) => Some(s.to_string()),
 
                 // special care for arrays
                 Value::Array(_) => json_array_to_pg_array(value),
@@ -94,29 +94,26 @@ fn json_to_pg_text(json: Vec<Value>) -> Result<Vec<Option<String>>, serde_json::
 //
 // Example of the same escaping in node-postgres: packages/pg/lib/utils.js
 //
-fn json_array_to_pg_array(value: &Value) -> Result<Option<String>, serde_json::Error> {
+fn json_array_to_pg_array(value: &Value) -> Option<String> {
     match value {
         // special care for nulls
-        Value::Null => Ok(None),
+        Value::Null => None,
 
         // convert to text with escaping
-        Value::Bool(_) => serde_json::to_string(value).map(Some),
-        Value::Number(_) => serde_json::to_string(value).map(Some),
-
         // here string needs to be escaped, as it is part of the array
-        Value::Object(_) => json_array_to_pg_array(&Value::String(serde_json::to_string(value)?)),
-        Value::String(_) => serde_json::to_string(value).map(Some),
+        v @ (Value::Bool(_) | Value::Number(_) | Value::String(_)) => Some(v.to_string()),
+        v @ Value::Object(_) => json_array_to_pg_array(&Value::String(v.to_string())),
 
         // recurse into array
         Value::Array(arr) => {
             let vals = arr
                 .iter()
                 .map(json_array_to_pg_array)
-                .map(|r| r.map(|v| v.unwrap_or_else(|| "NULL".to_string())))
-                .collect::<Result<Vec<_>, _>>()?
+                .map(|v| v.unwrap_or_else(|| "NULL".to_string()))
+                .collect::<Vec<_>>()
                 .join(",");
 
-            Ok(Some(format!("{{{}}}", vals)))
+            Some(format!("{{{}}}", vals))
         }
     }
 }
@@ -315,83 +312,119 @@ async fn handle_inner(
     // Now execute the query and return the result
     //
     let mut size = 0;
-    let result = match payload {
-        Payload::Single(query) => {
-            query_to_json(&client.inner, query, &mut size, raw_output, array_mode).await
-        }
-        Payload::Batch(batch_query) => {
-            let mut results = Vec::new();
-            let mut builder = client.inner.build_transaction();
-            if let Some(isolation_level) = txn_isolation_level {
-                builder = builder.isolation_level(isolation_level);
+    let result =
+        match payload {
+            Payload::Single(stmt) => {
+                let (status, results) =
+                    query_to_json(&*client, stmt, &mut 0, raw_output, array_mode)
+                        .await
+                        .map_err(|e| {
+                            client.discard();
+                            e
+                        })?;
+                client.check_idle(status);
+                results
             }
-            if txn_read_only {
-                builder = builder.read_only(true);
-            }
-            if txn_deferrable {
-                builder = builder.deferrable(true);
-            }
-            let transaction = builder.start().await?;
-            for query in batch_query.queries {
-                let result =
-                    query_to_json(&transaction, query, &mut size, raw_output, array_mode).await;
-                match result {
-                    Ok(r) => results.push(r),
-                    Err(e) => {
-                        transaction.rollback().await?;
-                        return Err(e);
-                    }
+            Payload::Batch(statements) => {
+                let (inner, mut discard) = client.inner();
+                let mut builder = inner.build_transaction();
+                if let Some(isolation_level) = txn_isolation_level {
+                    builder = builder.isolation_level(isolation_level);
                 }
+                if txn_read_only {
+                    builder = builder.read_only(true);
+                }
+                if txn_deferrable {
+                    builder = builder.deferrable(true);
+                }
+
+                let transaction = builder.start().await.map_err(|e| {
+                    // if we cannot start a transaction, we should return immediately
+                    // and not return to the pool. connection is clearly broken
+                    discard.discard();
+                    e
+                })?;
+
+                let results =
+                    match query_batch(&transaction, statements, &mut size, raw_output, array_mode)
+                        .await
+                    {
+                        Ok(results) => {
+                            let status = transaction.commit().await.map_err(|e| {
+                                // if we cannot commit - for now don't return connection to pool
+                                // TODO: get a query status from the error
+                                discard.discard();
+                                e
+                            })?;
+                            discard.check_idle(status);
+                            results
+                        }
+                        Err(err) => {
+                            let status = transaction.rollback().await.map_err(|e| {
+                                // if we cannot rollback - for now don't return connection to pool
+                                // TODO: get a query status from the error
+                                discard.discard();
+                                e
+                            })?;
+                            discard.check_idle(status);
+                            return Err(err);
+                        }
+                    };
+
+                if txn_read_only {
+                    response = response.header(
+                        TXN_READ_ONLY.clone(),
+                        HeaderValue::try_from(txn_read_only.to_string())?,
+                    );
+                }
+                if txn_deferrable {
+                    response = response.header(
+                        TXN_DEFERRABLE.clone(),
+                        HeaderValue::try_from(txn_deferrable.to_string())?,
+                    );
+                }
+                if let Some(txn_isolation_level) = txn_isolation_level_raw {
+                    response = response.header(TXN_ISOLATION_LEVEL.clone(), txn_isolation_level);
+                }
+                json!({ "results": results })
             }
-            transaction.commit().await?;
-            if txn_read_only {
-                response = response.header(
-                    TXN_READ_ONLY.clone(),
-                    HeaderValue::try_from(txn_read_only.to_string())?,
-                );
-            }
-            if txn_deferrable {
-                response = response.header(
-                    TXN_DEFERRABLE.clone(),
-                    HeaderValue::try_from(txn_deferrable.to_string())?,
-                );
-            }
-            if let Some(txn_isolation_level) = txn_isolation_level_raw {
-                response = response.header(TXN_ISOLATION_LEVEL.clone(), txn_isolation_level);
-            }
-            Ok(json!({ "results": results }))
-        }
-    };
+        };
 
     let metrics = client.metrics();
 
-    if allow_pool {
-        let current_span = tracing::Span::current();
-        // return connection to the pool
-        tokio::task::spawn_blocking(move || {
-            let _span = current_span.enter();
-            let _ = conn_pool.put(&conn_info, client);
-        });
-    }
+    // how could this possibly fail
+    let body = serde_json::to_string(&result).expect("json serialization should not fail");
+    let len = body.len();
+    let response = response
+        .body(Body::from(body))
+        // only fails if invalid status code or invalid header/values are given.
+        // these are not user configurable so it cannot fail dynamically
+        .expect("building response payload should not fail");
 
-    match result {
-        Ok(value) => {
-            // how could this possibly fail
-            let body = serde_json::to_string(&value).expect("json serialization should not fail");
-            let len = body.len();
-            let response = response
-                .body(Body::from(body))
-                // only fails if invalid status code or invalid header/values are given.
-                // these are not user configurable so it cannot fail dynamically
-                .expect("building response payload should not fail");
+    // count the egress bytes - we miss the TLS and header overhead but oh well...
+    // moving this later in the stack is going to be a lot of effort and ehhhh
+    metrics.record_egress(len as u64);
 
-            // count the egress bytes - we miss the TLS and header overhead but oh well...
-            // moving this later in the stack is going to be a lot of effort and ehhhh
-            metrics.record_egress(len as u64);
-            Ok(response)
-        }
-        Err(e) => Err(e),
+    Ok(response)
+}
+
+async fn query_batch(
+    transaction: &Transaction<'_>,
+    queries: BatchQueryData,
+    total_size: &mut usize,
+    raw_output: bool,
+    array_mode: bool,
+) -> anyhow::Result<Vec<Value>> {
+    let mut results = Vec::with_capacity(queries.queries.len());
+    let mut current_size = 0;
+    for stmt in queries.queries {
+        // TODO: maybe we should check that the transaction bit is set here
+        let (_, values) =
+            query_to_json(transaction, stmt, &mut current_size, raw_output, array_mode).await?;
+        results.push(values);
     }
+    *total_size += current_size;
+    Ok(results)
 }
 
 async fn query_to_json<T: GenericClient>(
@@ -400,11 +433,9 @@ async fn query_to_json<T: GenericClient>(
     current_size: &mut usize,
     raw_output: bool,
     array_mode: bool,
-) -> anyhow::Result<Value> {
-    let query_params = json_to_pg_text(data.params)?;
-    let row_stream = client
-        .query_raw_txt::<String, _>(data.query, query_params)
-        .await?;
+) -> anyhow::Result<(ReadyForQueryStatus, Value)> {
+    let query_params = json_to_pg_text(data.params);
+    let row_stream = client.query_raw_txt(&data.query, query_params).await?;
 
     // Manually drain the stream into a vector to leave row_stream hanging
     // around to get a command tag. Also check that the response is not too
@@ -424,6 +455,8 @@ async fn query_to_json<T: GenericClient>(
         }
     }
 
+    let ready = row_stream.ready_status();
+
     // grab the command tag and number of rows affected
     let command_tag = row_stream.command_tag().unwrap_or_default();
     let mut command_tag_split = command_tag.split(' ');
@@ -464,13 +497,16 @@ async fn query_to_json<T: GenericClient>(
         .collect::<Result<Vec<_>, _>>()?;
 
     // resulting JSON format is based on the format of node-postgres result
-    Ok(json!({
-        "command": command_tag_name,
-        "rowCount": command_tag_count,
-        "rows": rows,
-        "fields": fields,
-        "rowAsArray": array_mode,
-    }))
+    Ok((
+        ready,
+        json!({
+            "command": command_tag_name,
+            "rowCount": command_tag_count,
+            "rows": rows,
+            "fields": fields,
+            "rowAsArray": array_mode,
+        }),
+    ))
 }
 
 //
@@ -655,22 +691,22 @@ mod tests {
     #[test]
     fn test_atomic_types_to_pg_params() {
         let json = vec![Value::Bool(true), Value::Bool(false)];
-        let pg_params = json_to_pg_text(json).unwrap();
+        let pg_params = json_to_pg_text(json);
         assert_eq!(
             pg_params,
             vec![Some("true".to_owned()), Some("false".to_owned())]
         );
 
         let json = vec![Value::Number(serde_json::Number::from(42))];
-        let pg_params = json_to_pg_text(json).unwrap();
+        let pg_params = json_to_pg_text(json);
         assert_eq!(pg_params, vec![Some("42".to_owned())]);
 
         let json = vec![Value::String("foo\"".to_string())];
-        let pg_params = json_to_pg_text(json).unwrap();
+        let pg_params = json_to_pg_text(json);
         assert_eq!(pg_params, vec![Some("foo\"".to_owned())]);
 
         let json = vec![Value::Null];
-        let pg_params = json_to_pg_text(json).unwrap();
+        let pg_params = json_to_pg_text(json);
         assert_eq!(pg_params, vec![None]);
     }
 
@@ -679,7 +715,7 @@ mod tests {
         // atoms and escaping
         let json = "[true, false, null, \"NULL\", 42, \"foo\", \"bar\\\"-\\\\\"]";
         let json: Value = serde_json::from_str(json).unwrap();
-        let pg_params = json_to_pg_text(vec![json]).unwrap();
+        let pg_params = json_to_pg_text(vec![json]);
         assert_eq!(
             pg_params,
             vec![Some(
@@ -690,7 +726,7 @@ mod tests {
         // nested arrays
         let json = "[[true, false], [null, 42], [\"foo\", \"bar\\\"-\\\\\"]]";
         let json: Value = serde_json::from_str(json).unwrap();
-        let pg_params = json_to_pg_text(vec![json]).unwrap();
+        let pg_params = json_to_pg_text(vec![json]);
         assert_eq!(
             pg_params,
             vec![Some(
@@ -700,7 +736,7 @@ mod tests {
         // array of objects
         let json = r#"[{"foo": 1},{"bar": 2}]"#;
         let json: Value = serde_json::from_str(json).unwrap();
-        let pg_params = json_to_pg_text(vec![json]).unwrap();
+        let pg_params = json_to_pg_text(vec![json]);
         assert_eq!(
             pg_params,
             vec![Some(r#"{"{\"foo\":1}","{\"bar\":2}"}"#.to_owned())]

proxy/src/http/websocket.rs

@@ -8,6 +8,7 @@ use crate::{
         NUM_CLIENT_CONNECTION_OPENED_COUNTER,
     },
 };
+use anyhow::bail;
 use bytes::{Buf, Bytes};
 use futures::{Sink, Stream, StreamExt};
 use hyper::{
@@ -22,7 +23,6 @@ use hyper_tungstenite::{tungstenite::Message, HyperWebsocket, WebSocketStream};
 use pin_project_lite::pin_project;
 
 use std::{
-    convert::Infallible,
     future::ready,
     pin::Pin,
     sync::Arc,
@@ -280,12 +280,18 @@ pub async fn task_main(
     let make_svc = hyper::service::make_service_fn(
         |stream: &tokio_rustls::server::TlsStream<WithClientIp<AddrStream>>| {
             let (io, tls) = stream.get_ref();
-            let peer_addr = io.client_addr().unwrap_or(io.inner.remote_addr());
+            let client_addr = io.client_addr();
+            let remote_addr = io.inner.remote_addr();
             let sni_name = tls.server_name().map(|s| s.to_string());
             let conn_pool = conn_pool.clone();
 
             async move {
-                Ok::<_, Infallible>(MetricService::new(hyper::service::service_fn(
+                let peer_addr = match client_addr {
+                    Some(addr) => addr,
+                    None if config.require_client_ip => bail!("missing required client ip"),
+                    None => remote_addr,
+                };
+                Ok(MetricService::new(hyper::service::service_fn(
                     move |req: Request<Body>| {
                         let sni_name = sni_name.clone();
                         let conn_pool = conn_pool.clone();

proxy/src/proxy.rs

@@ -5,7 +5,7 @@ use crate::{
     auth::{self, backend::AuthSuccess},
     cancellation::{self, CancelMap},
     compute::{self, PostgresConnection},
-    config::{ProxyConfig, TlsConfig},
+    config::{AuthenticationConfig, ProxyConfig, TlsConfig},
     console::{self, errors::WakeComputeError, messages::MetricsAuxInfo, Api},
     http::StatusCode,
     metrics::{Ids, USAGE_METRICS},
@@ -15,12 +15,11 @@ use crate::{
 use anyhow::{bail, Context};
 use async_trait::async_trait;
 use futures::TryFutureExt;
-use metrics::{
-    exponential_buckets, register_histogram, register_int_counter_vec, Histogram, IntCounterVec,
-};
+use metrics::{exponential_buckets, register_int_counter_vec, IntCounterVec};
 use once_cell::sync::Lazy;
 use pq_proto::{BeMessage as Be, FeStartupPacket, StartupMessageParams};
-use std::{error::Error, io, ops::ControlFlow, sync::Arc};
+use prometheus::{register_histogram_vec, HistogramVec};
+use std::{error::Error, io, ops::ControlFlow, sync::Arc, time::Instant};
 use tokio::{
     io::{AsyncRead, AsyncWrite, AsyncWriteExt},
     time,
@@ -93,16 +92,67 @@ pub static NUM_CONNECTIONS_CLOSED_COUNTER: Lazy<IntCounterVec> = Lazy::new(|| {
     .unwrap()
 });
 
-static COMPUTE_CONNECTION_LATENCY: Lazy<Histogram> = Lazy::new(|| {
-    register_histogram!(
+static COMPUTE_CONNECTION_LATENCY: Lazy<HistogramVec> = Lazy::new(|| {
+    register_histogram_vec!(
         "proxy_compute_connection_latency_seconds",
         "Time it took for proxy to establish a connection to the compute endpoint",
+        // http/ws/tcp, true/false, true/false, success/failure
+        // 3 * 2 * 2 * 2 = 24 counters
+        &["protocol", "cache_miss", "pool_miss", "outcome"],
         // largest bucket = 2^16 * 0.5ms = 32s
         exponential_buckets(0.0005, 2.0, 16).unwrap(),
     )
     .unwrap()
 });
 
+pub struct LatencyTimer {
+    start: Instant,
+    protocol: &'static str,
+    cache_miss: bool,
+    pool_miss: bool,
+    outcome: &'static str,
+}
+
+impl LatencyTimer {
+    pub fn new(protocol: &'static str) -> Self {
+        Self {
+            start: Instant::now(),
+            protocol,
+            cache_miss: false,
+            // by default we don't do pooling
+            pool_miss: true,
+            // assume failed unless otherwise specified
+            outcome: "failed",
+        }
+    }
+
+    pub fn cache_miss(&mut self) {
+        self.cache_miss = true;
+    }
+
+    pub fn pool_hit(&mut self) {
+        self.pool_miss = false;
+    }
+
+    pub fn success(mut self) {
+        self.outcome = "success";
+    }
+}
+
+impl Drop for LatencyTimer {
+    fn drop(&mut self) {
+        let duration = self.start.elapsed().as_secs_f64();
+        COMPUTE_CONNECTION_LATENCY
+            .with_label_values(&[
+                self.protocol,
+                bool_to_str(self.cache_miss),
+                bool_to_str(self.pool_miss),
+                self.outcome,
+            ])
+            .observe(duration)
+    }
+}
+
 static NUM_CONNECTION_FAILURES: Lazy<IntCounterVec> = Lazy::new(|| {
     register_int_counter_vec!(
         "proxy_connection_failures_total",
@@ -160,6 +210,8 @@ pub async fn task_main(
                         let mut socket = WithClientIp::new(socket);
                         if let Some(ip) = socket.wait_for_addr().await? {
                             tracing::Span::current().record("peer_addr", &tracing::field::display(ip));
+                        } else if config.require_client_ip {
+                            bail!("missing required client IP");
                         }
 
                         socket
@@ -298,7 +350,7 @@ pub async fn handle_client<S: AsyncRead + AsyncWrite + Unpin>(
         mode.allow_self_signed_compute(config),
     );
     cancel_map
-        .with_session(|session| client.connect_to_db(session, mode))
+        .with_session(|session| client.connect_to_db(session, mode, &config.authentication_config))
         .await
 }
 
@@ -495,24 +547,28 @@ pub async fn connect_to_compute<M: ConnectMechanism>(
     mut node_info: console::CachedNodeInfo,
     extra: &console::ConsoleReqExtra<'_>,
     creds: &auth::BackendType<'_, auth::ClientCredentials<'_>>,
+    mut latency_timer: LatencyTimer,
 ) -> Result<M::Connection, M::Error>
 where
     M::ConnectError: ShouldRetry + std::fmt::Debug,
     M::Error: From<WakeComputeError>,
 {
-    let _timer = COMPUTE_CONNECTION_LATENCY.start_timer();
-
     mechanism.update_connect_config(&mut node_info.config);
 
     // try once
     let (config, err) = match mechanism.connect_once(&node_info, CONNECT_TIMEOUT).await {
-        Ok(res) => return Ok(res),
+        Ok(res) => {
+            latency_timer.success();
+            return Ok(res);
+        }
         Err(e) => {
             error!(error = ?e, "could not connect to compute node");
             (invalidate_cache(node_info), e)
         }
     };
 
+    latency_timer.cache_miss();
+
     let mut num_retries = 1;
 
     // if we failed to connect, it's likely that the compute node was suspended, wake a new compute node
@@ -558,7 +614,10 @@ where
     info!("wake_compute success. attempting to connect");
     loop {
         match mechanism.connect_once(&node_info, CONNECT_TIMEOUT).await {
-            Ok(res) => return Ok(res),
+            Ok(res) => {
+                latency_timer.success();
+                return Ok(res);
+            }
             Err(e) => {
                 let retriable = e.should_retry(num_retries);
                 if !retriable {
@@ -775,6 +834,7 @@ impl<S: AsyncRead + AsyncWrite + Unpin> Client<'_, S> {
         self,
         session: cancellation::Session<'_>,
         mode: ClientMode,
+        config: &'static AuthenticationConfig,
     ) -> anyhow::Result<()> {
         let Self {
             mut stream,
@@ -789,8 +849,10 @@ impl<S: AsyncRead + AsyncWrite + Unpin> Client<'_, S> {
             application_name: params.get("application_name"),
         };
 
+        let latency_timer = LatencyTimer::new(mode.protocol_label());
+
         let auth_result = match creds
-            .authenticate(&extra, &mut stream, mode.allow_cleartext())
+            .authenticate(&extra, &mut stream, mode.allow_cleartext(), config)
             .await
         {
             Ok(auth_result) => auth_result,
@@ -812,9 +874,15 @@ impl<S: AsyncRead + AsyncWrite + Unpin> Client<'_, S> {
         node_info.allow_self_signed_compute = allow_self_signed_compute;
 
         let aux = node_info.aux.clone();
-        let mut node = connect_to_compute(&TcpMechanism { params }, node_info, &extra, &creds)
-            .or_else(|e| stream.throw_error(e))
-            .await?;
+        let mut node = connect_to_compute(
+            &TcpMechanism { params },
+            node_info,
+            &extra,
+            &creds,
+            latency_timer,
+        )
+        .or_else(|e| stream.throw_error(e))
+        .await?;
 
         let proto = mode.protocol_label();
         NUM_DB_CONNECTIONS_OPENED_COUNTER

proxy/src/proxy/tests.rs

@@ -450,7 +450,7 @@ async fn connect_to_compute_success() {
     use ConnectAction::*;
     let mechanism = TestConnectMechanism::new(vec![Connect]);
     let (cache, extra, creds) = helper_create_connect_info(&mechanism);
-    connect_to_compute(&mechanism, cache, &extra, &creds)
+    connect_to_compute(&mechanism, cache, &extra, &creds, LatencyTimer::new("test"))
         .await
         .unwrap();
     mechanism.verify();
@@ -461,7 +461,7 @@ async fn connect_to_compute_retry() {
     use ConnectAction::*;
     let mechanism = TestConnectMechanism::new(vec![Retry, Wake, Retry, Connect]);
     let (cache, extra, creds) = helper_create_connect_info(&mechanism);
-    connect_to_compute(&mechanism, cache, &extra, &creds)
+    connect_to_compute(&mechanism, cache, &extra, &creds, LatencyTimer::new("test"))
         .await
         .unwrap();
     mechanism.verify();
@@ -473,7 +473,7 @@ async fn connect_to_compute_non_retry_1() {
     use ConnectAction::*;
     let mechanism = TestConnectMechanism::new(vec![Retry, Wake, Retry, Fail]);
     let (cache, extra, creds) = helper_create_connect_info(&mechanism);
-    connect_to_compute(&mechanism, cache, &extra, &creds)
+    connect_to_compute(&mechanism, cache, &extra, &creds, LatencyTimer::new("test"))
         .await
         .unwrap_err();
     mechanism.verify();
@@ -485,7 +485,7 @@ async fn connect_to_compute_non_retry_2() {
     use ConnectAction::*;
     let mechanism = TestConnectMechanism::new(vec![Fail, Wake, Retry, Connect]);
     let (cache, extra, creds) = helper_create_connect_info(&mechanism);
-    connect_to_compute(&mechanism, cache, &extra, &creds)
+    connect_to_compute(&mechanism, cache, &extra, &creds, LatencyTimer::new("test"))
         .await
         .unwrap();
     mechanism.verify();
@@ -501,7 +501,7 @@ async fn connect_to_compute_non_retry_3() {
         Retry, Retry, Retry, Retry, /* the 17th time */ Retry,
     ]);
     let (cache, extra, creds) = helper_create_connect_info(&mechanism);
-    connect_to_compute(&mechanism, cache, &extra, &creds)
+    connect_to_compute(&mechanism, cache, &extra, &creds, LatencyTimer::new("test"))
         .await
         .unwrap_err();
     mechanism.verify();
@@ -513,7 +513,7 @@ async fn wake_retry() {
     use ConnectAction::*;
     let mechanism = TestConnectMechanism::new(vec![Retry, WakeRetry, Wake, Connect]);
     let (cache, extra, creds) = helper_create_connect_info(&mechanism);
-    connect_to_compute(&mechanism, cache, &extra, &creds)
+    connect_to_compute(&mechanism, cache, &extra, &creds, LatencyTimer::new("test"))
         .await
         .unwrap();
     mechanism.verify();
@@ -525,7 +525,7 @@ async fn wake_non_retry() {
     use ConnectAction::*;
     let mechanism = TestConnectMechanism::new(vec![Retry, WakeFail]);
     let (cache, extra, creds) = helper_create_connect_info(&mechanism);
-    connect_to_compute(&mechanism, cache, &extra, &creds)
+    connect_to_compute(&mechanism, cache, &extra, &creds, LatencyTimer::new("test"))
         .await
         .unwrap_err();
     mechanism.verify();

proxy/src/sasl/messages.rs

@@ -31,7 +31,7 @@ impl<'a> FirstMessage<'a> {
 
 /// A single SASL message.
 /// This struct is deliberately decoupled from lower-level
-/// [`BeAuthenticationSaslMessage`](pq_proto::BeAuthenticationSaslMessage).
+/// [`BeAuthenticationSaslMessage`].
 #[derive(Debug)]
 pub(super) enum ServerMessage<T> {
     /// We expect to see more steps.

rust-toolchain.toml

@@ -1,5 +1,5 @@
 [toolchain]
-channel = "1.72.1"
+channel = "1.73.0"
 profile = "default"
 # The default profile includes rustc, rust-std, cargo, rust-docs, rustfmt and clippy.
 # https://rust-lang.github.io/rustup/concepts/profiles.html

safekeeper/src/bin/safekeeper.rs

@@ -3,7 +3,7 @@
 //
 use anyhow::{bail, Context, Result};
 use camino::{Utf8Path, Utf8PathBuf};
-use clap::Parser;
+use clap::{ArgAction, Parser};
 use futures::future::BoxFuture;
 use futures::stream::FuturesUnordered;
 use futures::{FutureExt, StreamExt};
@@ -105,6 +105,9 @@ struct Args {
     /// it during this period passed as a human readable duration.
     #[arg(long, value_parser= humantime::parse_duration, default_value = DEFAULT_HEARTBEAT_TIMEOUT, verbatim_doc_comment)]
     heartbeat_timeout: Duration,
+    /// Enable/disable peer recovery.
+    #[arg(long, default_value = "false", action=ArgAction::Set)]
+    peer_recovery: bool,
     /// 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}
@@ -265,6 +268,7 @@ async fn main() -> anyhow::Result<()> {
         broker_endpoint: args.broker_endpoint,
         broker_keepalive_interval: args.broker_keepalive_interval,
         heartbeat_timeout: args.heartbeat_timeout,
+        peer_recovery_enabled: args.peer_recovery,
         remote_storage: args.remote_storage,
         max_offloader_lag_bytes: args.max_offloader_lag,
         wal_backup_enabled: !args.disable_wal_backup,

safekeeper/src/handler.rs

@@ -372,6 +372,13 @@ impl SafekeeperPostgresHandler {
     /// from a walproposer recovery function. This connection gets a special handling:
     /// safekeeper must stream all local WAL till the flush_lsn, whether committed or not.
     pub fn is_walproposer_recovery(&self) -> bool {
-        self.appname == Some("wal_proposer_recovery".to_string())
+        match &self.appname {
+            None => false,
+            Some(appname) => {
+                appname == "wal_proposer_recovery" ||
+                // set by safekeeper peer recovery
+                appname.starts_with("safekeeper")
+            }
+        }
     }
 }

safekeeper/src/http/routes.rs

@@ -16,8 +16,8 @@ use tokio::io::AsyncReadExt;
 use utils::http::endpoint::request_span;
 
 use crate::receive_wal::WalReceiverState;
-use crate::safekeeper::ServerInfo;
 use crate::safekeeper::Term;
+use crate::safekeeper::{ServerInfo, TermLsn};
 use crate::send_wal::WalSenderState;
 use crate::timeline::PeerInfo;
 use crate::{debug_dump, pull_timeline};
@@ -60,16 +60,25 @@ fn get_conf(request: &Request<Body>) -> &SafeKeeperConf {
         .as_ref()
 }
 
-/// Same as TermSwitchEntry, but serializes LSN using display serializer
+/// Same as TermLsn, but serializes LSN using display serializer
 /// in Postgres format, i.e. 0/FFFFFFFF. Used only for the API response.
 #[serde_as]
-#[derive(Debug, Serialize, Deserialize)]
+#[derive(Debug, Clone, Copy, Serialize, Deserialize)]
 pub struct TermSwitchApiEntry {
     pub term: Term,
     #[serde_as(as = "DisplayFromStr")]
     pub lsn: Lsn,
 }
 
+impl From<TermSwitchApiEntry> for TermLsn {
+    fn from(api_val: TermSwitchApiEntry) -> Self {
+        TermLsn {
+            term: api_val.term,
+            lsn: api_val.lsn,
+        }
+    }
+}
+
 /// Augment AcceptorState with epoch for convenience
 #[derive(Debug, Serialize, Deserialize)]
 pub struct AcceptorStateStatus {

safekeeper/src/lib.rs

@@ -62,6 +62,7 @@ pub struct SafeKeeperConf {
     pub broker_endpoint: Uri,
     pub broker_keepalive_interval: Duration,
     pub heartbeat_timeout: Duration,
+    pub peer_recovery_enabled: bool,
     pub remote_storage: Option<RemoteStorageConfig>,
     pub max_offloader_lag_bytes: u64,
     pub backup_parallel_jobs: usize,
@@ -100,6 +101,7 @@ impl SafeKeeperConf {
                 .parse()
                 .expect("failed to parse default broker endpoint"),
             broker_keepalive_interval: Duration::from_secs(5),
+            peer_recovery_enabled: true,
             wal_backup_enabled: true,
             backup_parallel_jobs: 1,
             pg_auth: None,

safekeeper/src/receive_wal.rs

@@ -55,9 +55,12 @@ impl WalReceivers {
 
     /// Register new walreceiver. Returned guard provides access to the slot and
     /// automatically deregisters in Drop.
-    pub fn register(self: &Arc<WalReceivers>) -> WalReceiverGuard {
+    pub fn register(self: &Arc<WalReceivers>, conn_id: Option<ConnectionId>) -> WalReceiverGuard {
         let slots = &mut self.mutex.lock().slots;
-        let walreceiver = WalReceiverState::Voting;
+        let walreceiver = WalReceiverState {
+            conn_id,
+            status: WalReceiverStatus::Voting,
+        };
         // find empty slot or create new one
         let pos = if let Some(pos) = slots.iter().position(|s| s.is_none()) {
             slots[pos] = Some(walreceiver);
@@ -96,6 +99,18 @@ impl WalReceivers {
         self.mutex.lock().slots.iter().flatten().cloned().collect()
     }
 
+    /// Get number of streaming walreceivers (normally 0 or 1) from compute.
+    pub fn get_num_streaming(self: &Arc<WalReceivers>) -> usize {
+        self.mutex
+            .lock()
+            .slots
+            .iter()
+            .flatten()
+            // conn_id.is_none skips recovery which also registers here
+            .filter(|s| s.conn_id.is_some() && matches!(s.status, WalReceiverStatus::Streaming))
+            .count()
+    }
+
     /// Unregister walsender.
     fn unregister(self: &Arc<WalReceivers>, id: WalReceiverId) {
         let mut shared = self.mutex.lock();
@@ -108,10 +123,17 @@ struct WalReceiversShared {
     slots: Vec<Option<WalReceiverState>>,
 }
 
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct WalReceiverState {
+    /// None means it is recovery initiated by us (this safekeeper).
+    pub conn_id: Option<ConnectionId>,
+    pub status: WalReceiverStatus,
+}
+
 /// Walreceiver status. Currently only whether it passed voting stage and
 /// started receiving the stream, but it is easy to add more if needed.
 #[derive(Debug, Clone, Serialize, Deserialize)]
-pub enum WalReceiverState {
+pub enum WalReceiverStatus {
     Voting,
     Streaming,
 }
@@ -136,8 +158,8 @@ impl Drop for WalReceiverGuard {
     }
 }
 
-const MSG_QUEUE_SIZE: usize = 256;
-const REPLY_QUEUE_SIZE: usize = 16;
+pub const MSG_QUEUE_SIZE: usize = 256;
+pub const REPLY_QUEUE_SIZE: usize = 16;
 
 impl SafekeeperPostgresHandler {
     /// Wrapper around handle_start_wal_push_guts handling result. Error is
@@ -261,7 +283,7 @@ impl<'a, IO: AsyncRead + AsyncWrite + Unpin> NetworkReader<'a, IO> {
             tli.clone(),
             msg_rx,
             reply_tx,
-            self.conn_id,
+            Some(self.conn_id),
         ));
 
         // Forward all messages to WalAcceptor
@@ -317,31 +339,41 @@ async fn network_write<IO: AsyncRead + AsyncWrite + Unpin>(
 // even when it writes a steady stream of messages.
 const KEEPALIVE_INTERVAL: Duration = Duration::from_secs(1);
 
-/// Takes messages from msg_rx, processes and pushes replies to reply_tx.
-struct WalAcceptor {
+/// Encapsulates a task which takes messages from msg_rx, processes and pushes
+/// replies to reply_tx; reading from socket and writing to disk in parallel is
+/// beneficial for performance, this struct provides writing to disk part.
+pub struct WalAcceptor {
     tli: Arc<Timeline>,
     msg_rx: Receiver<ProposerAcceptorMessage>,
     reply_tx: Sender<AcceptorProposerMessage>,
+    conn_id: Option<ConnectionId>,
 }
 
 impl WalAcceptor {
-    /// Spawn thread with WalAcceptor running, return handle to it.
-    fn spawn(
+    /// Spawn task with WalAcceptor running, return handle to it. Task returns
+    /// Ok(()) if either of channels has closed, and Err if any error during
+    /// message processing is encountered.
+    ///
+    /// conn_id None means WalAcceptor is used by recovery initiated at this safekeeper.
+    pub fn spawn(
         tli: Arc<Timeline>,
         msg_rx: Receiver<ProposerAcceptorMessage>,
         reply_tx: Sender<AcceptorProposerMessage>,
-        conn_id: ConnectionId,
+        conn_id: Option<ConnectionId>,
     ) -> JoinHandle<anyhow::Result<()>> {
         task::spawn(async move {
             let mut wa = WalAcceptor {
                 tli,
                 msg_rx,
                 reply_tx,
+                conn_id,
             };
 
             let span_ttid = wa.tli.ttid; // satisfy borrow checker
             wa.run()
-                .instrument(info_span!("WAL acceptor", cid = %conn_id, ttid = %span_ttid))
+                .instrument(
+                    info_span!("WAL acceptor", cid = %conn_id.unwrap_or(0), ttid = %span_ttid),
+                )
                 .await
         })
     }
@@ -355,7 +387,7 @@ impl WalAcceptor {
         let _compute_conn_guard = ComputeConnectionGuard {
             timeline: Arc::clone(&self.tli),
         };
-        let walreceiver_guard = self.tli.get_walreceivers().register();
+        let walreceiver_guard = self.tli.get_walreceivers().register(self.conn_id);
         self.tli.update_status_notify().await?;
 
         // After this timestamp we will stop processing AppendRequests and send a response
@@ -372,7 +404,7 @@ impl WalAcceptor {
 
             // Update walreceiver state in shmem for reporting.
             if let ProposerAcceptorMessage::Elected(_) = &next_msg {
-                *walreceiver_guard.get() = WalReceiverState::Streaming;
+                walreceiver_guard.get().status = WalReceiverStatus::Streaming;
             }
 
             let reply_msg = if matches!(next_msg, ProposerAcceptorMessage::AppendRequest(_)) {

safekeeper/src/recovery.rs

@@ -1,17 +1,41 @@
 //! This module implements pulling WAL from peer safekeepers if compute can't
 //! provide it, i.e. safekeeper lags too much.
 
-use std::sync::Arc;
+use std::time::SystemTime;
+use std::{fmt, pin::pin, sync::Arc};
 
-use tokio::{select, time::sleep, time::Duration};
-use tracing::{info, instrument};
+use anyhow::{bail, Context};
+use futures::StreamExt;
+use postgres_protocol::message::backend::ReplicationMessage;
+use tokio::sync::mpsc::{channel, Receiver, Sender};
+use tokio::time::timeout;
+use tokio::{
+    select,
+    time::sleep,
+    time::{self, Duration},
+};
+use tokio_postgres::replication::ReplicationStream;
+use tokio_postgres::types::PgLsn;
+use tracing::*;
+use utils::{id::NodeId, lsn::Lsn, postgres_client::wal_stream_connection_config};
 
-use crate::{timeline::Timeline, SafeKeeperConf};
+use crate::receive_wal::{WalAcceptor, REPLY_QUEUE_SIZE};
+use crate::safekeeper::{AppendRequest, AppendRequestHeader};
+use crate::{
+    http::routes::TimelineStatus,
+    receive_wal::MSG_QUEUE_SIZE,
+    safekeeper::{
+        AcceptorProposerMessage, ProposerAcceptorMessage, ProposerElected, Term, TermHistory,
+        TermLsn, VoteRequest,
+    },
+    timeline::{PeerInfo, Timeline},
+    SafeKeeperConf,
+};
 
 /// Entrypoint for per timeline task which always runs, checking whether
 /// recovery for this safekeeper is needed and starting it if so.
 #[instrument(name = "recovery task", skip_all, fields(ttid = %tli.ttid))]
-pub async fn recovery_main(tli: Arc<Timeline>, _conf: SafeKeeperConf) {
+pub async fn recovery_main(tli: Arc<Timeline>, conf: SafeKeeperConf) {
     info!("started");
     let mut cancellation_rx = match tli.get_cancellation_rx() {
         Ok(rx) => rx,
@@ -22,19 +46,387 @@ pub async fn recovery_main(tli: Arc<Timeline>, _conf: SafeKeeperConf) {
     };
 
     select! {
-        _ = recovery_main_loop(tli) => { unreachable!() }
+        _ = recovery_main_loop(tli, conf) => { unreachable!() }
         _ = cancellation_rx.changed() => {
             info!("stopped");
         }
     }
 }
 
+/// Result of Timeline::recovery_needed, contains donor(s) if recovery needed and
+/// fields to explain the choice.
+#[derive(Debug)]
+pub struct RecoveryNeededInfo {
+    /// my term
+    pub term: Term,
+    /// my last_log_term
+    pub last_log_term: Term,
+    /// my flush_lsn
+    pub flush_lsn: Lsn,
+    /// peers from which we can fetch WAL, for observability.
+    pub peers: Vec<PeerInfo>,
+    /// for observability
+    pub num_streaming_computes: usize,
+    pub donors: Vec<Donor>,
+}
+
+// Custom to omit not important fields from PeerInfo.
+impl fmt::Display for RecoveryNeededInfo {
+    fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
+        write!(f, "{{")?;
+        write!(
+            f,
+            "term: {}, last_log_term: {}, flush_lsn: {}, peers: {{",
+            self.term, self.last_log_term, self.flush_lsn
+        )?;
+        for p in self.peers.iter() {
+            write!(
+                f,
+                "PeerInfo {{ sk_id: {}, term: {}, last_log_term: {}, flush_lsn: {} }}, ",
+                p.sk_id, p.term, p.last_log_term, p.flush_lsn
+            )?;
+        }
+        write!(
+            f,
+            "}} num_streaming_computes: {}, donors: {:?}",
+            self.num_streaming_computes, self.donors
+        )
+    }
+}
+
+#[derive(Clone, Debug)]
+pub struct Donor {
+    pub sk_id: NodeId,
+    /// equals to last_log_term
+    pub term: Term,
+    pub flush_lsn: Lsn,
+    pub pg_connstr: String,
+    pub http_connstr: String,
+}
+
+impl From<&PeerInfo> for Donor {
+    fn from(p: &PeerInfo) -> Self {
+        Donor {
+            sk_id: p.sk_id,
+            term: p.term,
+            flush_lsn: p.flush_lsn,
+            pg_connstr: p.pg_connstr.clone(),
+            http_connstr: p.http_connstr.clone(),
+        }
+    }
+}
+
 const CHECK_INTERVAL_MS: u64 = 2000;
 
 /// Check regularly whether we need to start recovery.
-async fn recovery_main_loop(_tli: Arc<Timeline>) {
+async fn recovery_main_loop(tli: Arc<Timeline>, conf: SafeKeeperConf) {
     let check_duration = Duration::from_millis(CHECK_INTERVAL_MS);
     loop {
+        let recovery_needed_info = tli.recovery_needed(conf.heartbeat_timeout).await;
+        match recovery_needed_info.donors.first() {
+            Some(donor) => {
+                info!(
+                    "starting recovery from donor {}: {}",
+                    donor.sk_id, recovery_needed_info
+                );
+                match recover(tli.clone(), donor, &conf).await {
+                    // Note: 'write_wal rewrites WAL written before' error is
+                    // expected here and might happen if compute and recovery
+                    // concurrently write the same data. Eventually compute
+                    // should win.
+                    Err(e) => warn!("recovery failed: {:#}", e),
+                    Ok(msg) => info!("recovery finished: {}", msg),
+                }
+            }
+            None => {
+                trace!(
+                    "recovery not needed or not possible: {}",
+                    recovery_needed_info
+                );
+            }
+        }
         sleep(check_duration).await;
     }
 }
+
+/// Recover from the specified donor. Returns message explaining normal finish
+/// reason or error.
+async fn recover(
+    tli: Arc<Timeline>,
+    donor: &Donor,
+    conf: &SafeKeeperConf,
+) -> anyhow::Result<String> {
+    // Learn donor term switch history to figure out starting point.
+    let client = reqwest::Client::new();
+    let timeline_info: TimelineStatus = client
+        .get(format!(
+            "http://{}/v1/tenant/{}/timeline/{}",
+            donor.http_connstr, tli.ttid.tenant_id, tli.ttid.timeline_id
+        ))
+        .send()
+        .await?
+        .json()
+        .await?;
+    if timeline_info.acceptor_state.term != donor.term {
+        bail!(
+            "donor term changed from {} to {}",
+            donor.term,
+            timeline_info.acceptor_state.term
+        );
+    }
+    // convert from API TermSwitchApiEntry into TermLsn.
+    let donor_th = TermHistory(
+        timeline_info
+            .acceptor_state
+            .term_history
+            .iter()
+            .map(|tl| Into::<TermLsn>::into(*tl))
+            .collect(),
+    );
+
+    // Now understand our term history.
+    let vote_request = ProposerAcceptorMessage::VoteRequest(VoteRequest { term: donor.term });
+    let vote_response = match tli
+        .process_msg(&vote_request)
+        .await
+        .context("VoteRequest handling")?
+    {
+        Some(AcceptorProposerMessage::VoteResponse(vr)) => vr,
+        _ => {
+            bail!("unexpected VoteRequest response"); // unreachable
+        }
+    };
+    if vote_response.term != donor.term {
+        bail!(
+            "our term changed from {} to {}",
+            donor.term,
+            vote_response.term
+        );
+    }
+
+    let last_common_point = match TermHistory::find_highest_common_point(
+        &donor_th,
+        &vote_response.term_history,
+        vote_response.flush_lsn,
+    ) {
+        None => bail!(
+            "couldn't find common point in histories, donor {:?}, sk {:?}",
+            donor_th,
+            vote_response.term_history,
+        ),
+        Some(lcp) => lcp,
+    };
+    info!("found last common point at {:?}", last_common_point);
+
+    // truncate WAL locally
+    let pe = ProposerAcceptorMessage::Elected(ProposerElected {
+        term: donor.term,
+        start_streaming_at: last_common_point.lsn,
+        term_history: donor_th,
+        timeline_start_lsn: Lsn::INVALID,
+    });
+    // Successful ProposerElected handling always returns None. If term changed,
+    // we'll find out that during the streaming. Note: it is expected to get
+    // 'refusing to overwrite correct WAL' here if walproposer reconnected
+    // concurrently, restart helps here.
+    tli.process_msg(&pe)
+        .await
+        .context("ProposerElected handling")?;
+
+    recovery_stream(tli, donor, last_common_point.lsn, conf).await
+}
+
+// Pull WAL from donor, assuming handshake is already done.
+async fn recovery_stream(
+    tli: Arc<Timeline>,
+    donor: &Donor,
+    start_streaming_at: Lsn,
+    conf: &SafeKeeperConf,
+) -> anyhow::Result<String> {
+    // TODO: pass auth token
+    let cfg = wal_stream_connection_config(tli.ttid, &donor.pg_connstr, None, None)?;
+    let mut cfg = cfg.to_tokio_postgres_config();
+    // It will make safekeeper give out not committed WAL (up to flush_lsn).
+    cfg.application_name(&format!("safekeeper_{}", conf.my_id));
+    cfg.replication_mode(tokio_postgres::config::ReplicationMode::Physical);
+
+    let connect_timeout = Duration::from_millis(10000);
+    let (client, connection) = match time::timeout(connect_timeout, cfg.connect(postgres::NoTls))
+        .await
+    {
+        Ok(client_and_conn) => client_and_conn?,
+        Err(_elapsed) => {
+            bail!("timed out while waiting {connect_timeout:?} for connection to peer safekeeper to open");
+        }
+    };
+    trace!("connected to {:?}", donor);
+
+    // The connection object performs the actual communication with the
+    // server, spawn it off to run on its own.
+    let ttid = tli.ttid;
+    tokio::spawn(async move {
+        if let Err(e) = connection
+            .instrument(info_span!("recovery task connection poll", ttid = %ttid))
+            .await
+        {
+            // This logging isn't very useful as error is anyway forwarded to client.
+            trace!(
+                "tokio_postgres connection object finished with error: {}",
+                e
+            );
+        }
+    });
+
+    let query = format!(
+        "START_REPLICATION PHYSICAL {} (term='{}')",
+        start_streaming_at, donor.term
+    );
+
+    let copy_stream = client.copy_both_simple(&query).await?;
+    let physical_stream = ReplicationStream::new(copy_stream);
+
+    // As in normal walreceiver, do networking and writing to disk in parallel.
+    let (msg_tx, msg_rx) = channel(MSG_QUEUE_SIZE);
+    let (reply_tx, reply_rx) = channel(REPLY_QUEUE_SIZE);
+    let wa = WalAcceptor::spawn(tli.clone(), msg_rx, reply_tx, None);
+
+    let res = tokio::select! {
+        r = network_io(physical_stream, msg_tx, donor.clone(), tli.clone(), conf.clone()) => r,
+        r = read_replies(reply_rx, donor.term) => r.map(|()| None),
+    };
+
+    // Join the spawned WalAcceptor. At this point chans to/from it passed to
+    // network routines are dropped, so it will exit as soon as it touches them.
+    match wa.await {
+        Ok(Ok(())) => {
+            // WalAcceptor finished normally, termination reason is different
+            match res {
+                Ok(Some(success_desc)) => Ok(success_desc),
+                Ok(None) => bail!("unexpected recovery end without error/success"), // can't happen
+                Err(e) => Err(e), // network error or term change
+            }
+        }
+        Ok(Err(e)) => Err(e), // error while processing message
+        Err(e) => bail!("WalAcceptor panicked: {}", e),
+    }
+}
+
+// Perform network part of streaming: read data and push it to msg_tx, send KA
+// to make sender hear from us. If there is nothing coming for a while, check
+// for termination.
+// Returns
+// - Ok(None) if channel to WalAcceptor closed -- its task should return error.
+// - Ok(Some(String)) if recovery successfully completed.
+// - Err if error happened while reading/writing to socket.
+async fn network_io(
+    physical_stream: ReplicationStream,
+    msg_tx: Sender<ProposerAcceptorMessage>,
+    donor: Donor,
+    tli: Arc<Timeline>,
+    conf: SafeKeeperConf,
+) -> anyhow::Result<Option<String>> {
+    let mut physical_stream = pin!(physical_stream);
+    let mut last_received_lsn = Lsn::INVALID;
+    // tear down connection if no data arrives withing this period
+    let no_data_timeout = Duration::from_millis(30000);
+
+    loop {
+        let msg = match timeout(no_data_timeout, physical_stream.next()).await {
+            Ok(next) => match next {
+                None => bail!("unexpected end of replication stream"),
+                Some(msg) => msg.context("get replication message")?,
+            },
+            Err(_) => bail!("no message received within {:?}", no_data_timeout),
+        };
+
+        match msg {
+            ReplicationMessage::XLogData(xlog_data) => {
+                let ar_hdr = AppendRequestHeader {
+                    term: donor.term,
+                    epoch_start_lsn: Lsn::INVALID, // unused
+                    begin_lsn: Lsn(xlog_data.wal_start()),
+                    end_lsn: Lsn(xlog_data.wal_start()) + xlog_data.data().len() as u64,
+                    commit_lsn: Lsn::INVALID, // do not attempt to advance, peer communication anyway does it
+                    truncate_lsn: Lsn::INVALID, // do not attempt to advance
+                    proposer_uuid: [0; 16],
+                };
+                let ar = AppendRequest {
+                    h: ar_hdr,
+                    wal_data: xlog_data.into_data(),
+                };
+                trace!(
+                    "processing AppendRequest {}-{}, len {}",
+                    ar.h.begin_lsn,
+                    ar.h.end_lsn,
+                    ar.wal_data.len()
+                );
+                last_received_lsn = ar.h.end_lsn;
+                if msg_tx
+                    .send(ProposerAcceptorMessage::AppendRequest(ar))
+                    .await
+                    .is_err()
+                {
+                    return Ok(None); // chan closed, WalAcceptor terminated
+                }
+            }
+            ReplicationMessage::PrimaryKeepAlive(_) => {
+                // keepalive means nothing is being streamed for a while. Check whether we need to stop.
+                let recovery_needed_info = tli.recovery_needed(conf.heartbeat_timeout).await;
+                // do current donors still contain one we currently connected to?
+                if !recovery_needed_info
+                    .donors
+                    .iter()
+                    .any(|d| d.sk_id == donor.sk_id)
+                {
+                    // Most likely it means we are caughtup.
+                    // note: just exiting makes tokio_postgres send CopyFail to the far end.
+                    return Ok(Some(format!(
+                        "terminating at {} as connected safekeeper {} with term {} is not a donor anymore: {}",
+                        last_received_lsn, donor.sk_id, donor.term, recovery_needed_info
+                    )));
+                }
+            }
+            _ => {}
+        }
+        // Send reply to each message to keep connection alive. Ideally we
+        // should do that once in a while instead, but this again requires
+        // stream split or similar workaround, and recovery is anyway not that
+        // performance critical.
+        //
+        // We do not know here real write/flush LSNs (need to take mutex again
+        // or check replies which are read in different future), but neither
+        // sender much cares about them, so just send last received.
+        physical_stream
+            .as_mut()
+            .standby_status_update(
+                PgLsn::from(last_received_lsn.0),
+                PgLsn::from(last_received_lsn.0),
+                PgLsn::from(last_received_lsn.0),
+                SystemTime::now(),
+                0,
+            )
+            .await?;
+    }
+}
+
+// Read replies from WalAcceptor. We are not interested much in sending them to
+// donor safekeeper, so don't route them anywhere. However, we should check if
+// term changes and exit if it does.
+// Returns Ok(()) if channel closed, Err in case of term change.
+async fn read_replies(
+    mut reply_rx: Receiver<AcceptorProposerMessage>,
+    donor_term: Term,
+) -> anyhow::Result<()> {
+    loop {
+        match reply_rx.recv().await {
+            Some(msg) => {
+                if let AcceptorProposerMessage::AppendResponse(ar) = msg {
+                    if ar.term != donor_term {
+                        bail!("donor term changed from {} to {}", donor_term, ar.term);
+                    }
+                }
+            }
+            None => return Ok(()), // chan closed, WalAcceptor terminated
+        }
+    }
+}