Proxy release 2025-06-10 08:56 UTC

## Problem

Looks like our sql-over-http tests get to rely on "trust"
authentication, so the path that made sure the authkeys data was set was
never being hit.

## Summary of changes

Slight refactor to WakeComputeBackends, as well as making sure auth keys
are propagated. Fix tests to ensure passwords are tested.

Cargo.lock

@@ -1445,6 +1445,7 @@ dependencies = [
  "regex",
  "reqwest",
  "safekeeper_api",
+ "safekeeper_client",
  "scopeguard",
  "serde",
  "serde_json",
@@ -2054,6 +2055,7 @@ dependencies = [
  "axum-extra",
  "camino",
  "camino-tempfile",
+ "clap",
  "futures",
  "http-body-util",
  "itertools 0.10.5",
@@ -4236,6 +4238,8 @@ name = "pagebench"
 version = "0.1.0"
 dependencies = [
  "anyhow",
+ "async-trait",
+ "bytes",
  "camino",
  "clap",
  "futures",
@@ -4244,12 +4248,15 @@ dependencies = [
  "humantime-serde",
  "pageserver_api",
  "pageserver_client",
+ "pageserver_page_api",
  "rand 0.8.5",
  "reqwest",
  "serde",
  "serde_json",
  "tokio",
+ "tokio-stream",
  "tokio-util",
+ "tonic 0.13.1",
  "tracing",
  "utils",
  "workspace_hack",
@@ -4305,6 +4312,7 @@ dependencies = [
  "hashlink",
  "hex",
  "hex-literal",
+ "http 1.1.0",
  "http-utils",
  "humantime",
  "humantime-serde",
@@ -4367,6 +4375,7 @@ dependencies = [
  "toml_edit",
  "tonic 0.13.1",
  "tonic-reflection",
+ "tower 0.5.2",
  "tracing",
  "tracing-utils",
  "twox-hash",
@@ -4463,7 +4472,6 @@ dependencies = [
  "pageserver_api",
  "postgres_ffi",
  "prost 0.13.5",
- "smallvec",
  "thiserror 1.0.69",
  "tonic 0.13.1",
  "tonic-build",

Dockerfile

@@ -110,6 +110,19 @@ RUN set -e \
 	# System postgres for use with client libraries (e.g. in storage controller)
         postgresql-15 \
         openssl \
+        unzip \
+        curl \
+    && ARCH=$(uname -m) \
+    && if [ "$ARCH" = "x86_64" ]; then \
+        curl "https://awscli.amazonaws.com/awscli-exe-linux-x86_64.zip" -o "awscliv2.zip"; \
+    elif [ "$ARCH" = "aarch64" ]; then \
+        curl "https://awscli.amazonaws.com/awscli-exe-linux-aarch64.zip" -o "awscliv2.zip"; \
+    else \
+        echo "Unsupported architecture: $ARCH" && exit 1; \
+    fi \
+    && unzip awscliv2.zip \
+    && ./aws/install \
+    && rm -rf aws awscliv2.zip \
     && rm -f /etc/apt/apt.conf.d/80-retries \
     && rm -rf /var/lib/apt/lists/* /tmp/* /var/tmp/* \
     && useradd -d /data neon \

build-tools.Dockerfile

@@ -310,13 +310,13 @@ RUN curl -sSO https://static.rust-lang.org/rustup/dist/$(uname -m)-unknown-linux
     . "$HOME/.cargo/env" && \
     cargo --version && rustup --version && \
     rustup component add llvm-tools rustfmt clippy && \
-    cargo install rustfilt            --version ${RUSTFILT_VERSION} && \
-    cargo install cargo-hakari        --version ${CARGO_HAKARI_VERSION} && \
-    cargo install cargo-deny --locked --version ${CARGO_DENY_VERSION} && \
-    cargo install cargo-hack          --version ${CARGO_HACK_VERSION} && \
-    cargo install cargo-nextest       --version ${CARGO_NEXTEST_VERSION} && \
-    cargo install cargo-chef --locked --version ${CARGO_CHEF_VERSION} && \
-    cargo install diesel_cli          --version ${CARGO_DIESEL_CLI_VERSION} \
+    cargo install rustfilt            --version ${RUSTFILT_VERSION} --locked && \
+    cargo install cargo-hakari        --version ${CARGO_HAKARI_VERSION} --locked && \
+    cargo install cargo-deny          --version ${CARGO_DENY_VERSION} --locked && \
+    cargo install cargo-hack          --version ${CARGO_HACK_VERSION} --locked && \
+    cargo install cargo-nextest       --version ${CARGO_NEXTEST_VERSION} --locked && \
+    cargo install cargo-chef          --version ${CARGO_CHEF_VERSION} --locked && \
+    cargo install diesel_cli          --version ${CARGO_DIESEL_CLI_VERSION} --locked \
                                       --features postgres-bundled --no-default-features && \
     rm -rf /home/nonroot/.cargo/registry && \
     rm -rf /home/nonroot/.cargo/git

compute/compute-node.Dockerfile

@@ -297,6 +297,7 @@ RUN ./autogen.sh && \
     ./configure --with-sfcgal=/usr/local/bin/sfcgal-config && \
     make -j $(getconf _NPROCESSORS_ONLN) && \
     make -j $(getconf _NPROCESSORS_ONLN) install && \
+    make staged-install && \
     cd extensions/postgis && \
     make clean && \
     make -j $(getconf _NPROCESSORS_ONLN) install && \
@@ -602,7 +603,7 @@ RUN case "${PG_VERSION:?}" in \
         ;; \
     esac && \
 	wget https://github.com/knizhnik/online_advisor/archive/refs/tags/1.0.tar.gz -O online_advisor.tar.gz && \
-    echo "059b7d9e5a90013a58bdd22e9505b88406ce05790675eb2d8434e5b215652d54 online_advisor.tar.gz" | sha256sum --check && \
+    echo "37dcadf8f7cc8d6cc1f8831276ee245b44f1b0274f09e511e47a67738ba9ed0f online_advisor.tar.gz" | sha256sum --check && \
     mkdir online_advisor-src && cd online_advisor-src && tar xzf ../online_advisor.tar.gz --strip-components=1 -C .
 
 FROM pg-build AS online_advisor-build
@@ -1180,14 +1181,14 @@ RUN cd exts/rag && \
 RUN cd exts/rag_bge_small_en_v15 && \
     sed -i 's/pgrx = "0.14.1"/pgrx = { version = "0.14.1", features = [ "unsafe-postgres" ] }/g' Cargo.toml && \
     ORT_LIB_LOCATION=/ext-src/onnxruntime-src/build/Linux \
-        REMOTE_ONNX_URL=http://pg-ext-s3-gateway/pgrag-data/bge_small_en_v15.onnx \
+        REMOTE_ONNX_URL=http://pg-ext-s3-gateway.pg-ext-s3-gateway.svc.cluster.local/pgrag-data/bge_small_en_v15.onnx \
         cargo pgrx install --release --features remote_onnx && \
     echo "trusted = true" >> /usr/local/pgsql/share/extension/rag_bge_small_en_v15.control
 
 RUN cd exts/rag_jina_reranker_v1_tiny_en && \
     sed -i 's/pgrx = "0.14.1"/pgrx = { version = "0.14.1", features = [ "unsafe-postgres" ] }/g' Cargo.toml && \
     ORT_LIB_LOCATION=/ext-src/onnxruntime-src/build/Linux \
-        REMOTE_ONNX_URL=http://pg-ext-s3-gateway/pgrag-data/jina_reranker_v1_tiny_en.onnx \
+        REMOTE_ONNX_URL=http://pg-ext-s3-gateway.pg-ext-s3-gateway.svc.cluster.local/pgrag-data/jina_reranker_v1_tiny_en.onnx \
         cargo pgrx install --release --features remote_onnx && \
     echo "trusted = true" >> /usr/local/pgsql/share/extension/rag_jina_reranker_v1_tiny_en.control
 
@@ -1842,10 +1843,25 @@ RUN make PG_VERSION="${PG_VERSION:?}" -C compute
 
 FROM pg-build AS extension-tests
 ARG PG_VERSION
+# This is required for the PostGIS test
+RUN apt-get update && case $DEBIAN_VERSION in \
+      bullseye) \
+        apt-get install -y libproj19 libgdal28 time; \
+      ;; \
+      bookworm) \
+        apt-get install -y libgdal32 libproj25 time; \
+      ;; \
+      *) \
+        echo "Unknown Debian version ${DEBIAN_VERSION}" && exit 1 \
+      ;; \
+    esac
+
 COPY docker-compose/ext-src/ /ext-src/
 
 COPY --from=pg-build /postgres /postgres
-#COPY --from=postgis-src /ext-src/ /ext-src/
+COPY --from=postgis-build /usr/local/pgsql/ /usr/local/pgsql/
+COPY --from=postgis-build /ext-src/postgis-src /ext-src/postgis-src
+COPY --from=postgis-build /sfcgal/* /usr
 COPY --from=plv8-src /ext-src/ /ext-src/
 COPY --from=h3-pg-src /ext-src/h3-pg-src /ext-src/h3-pg-src
 COPY --from=postgresql-unit-src /ext-src/ /ext-src/
@@ -1886,6 +1902,7 @@ COPY compute/patches/pg_repack.patch /ext-src
 RUN cd /ext-src/pg_repack-src && patch -p1 </ext-src/pg_repack.patch && rm -f /ext-src/pg_repack.patch
 
 COPY --chmod=755 docker-compose/run-tests.sh /run-tests.sh
+RUN echo /usr/local/pgsql/lib > /etc/ld.so.conf.d/00-neon.conf && /sbin/ldconfig
 RUN apt-get update && apt-get install -y libtap-parser-sourcehandler-pgtap-perl jq \
    && apt clean && rm -rf /ext-src/*.tar.gz /ext-src/*.patch /var/lib/apt/lists/*
 ENV PATH=/usr/local/pgsql/bin:$PATH

compute/manifest.yaml

@@ -0,0 +1,121 @@
+pg_settings:
+  # Common settings for primaries and replicas of all versions.
+  common:
+    # Check for client disconnection every 1 minute. By default, Postgres will detect the
+    # loss of the connection only at the next interaction with the socket, when it waits
+    # for, receives or sends data, so it will likely waste resources till the end of the
+    # query execution. There should be no drawbacks in setting this for everyone, so enable
+    # it by default. If anyone will complain, we can allow editing it.
+    # https://www.postgresql.org/docs/16/runtime-config-connection.html#GUC-CLIENT-CONNECTION-CHECK-INTERVAL
+    client_connection_check_interval: "60000" # 1 minute
+    # ---- IO ---- 
+    effective_io_concurrency: "20"
+    maintenance_io_concurrency: "100"
+    fsync: "off"
+    hot_standby: "off"
+    # We allow users to change this if needed, but by default we
+    # just don't want to see long-lasting idle transactions, as they
+    # prevent activity monitor from suspending projects.
+    idle_in_transaction_session_timeout: "300000" # 5 minutes
+    listen_addresses: "*"
+    # --- LOGGING ---- helps investigations
+    log_connections: "on"
+    log_disconnections: "on"
+    # 1GB, unit is KB
+    log_temp_files: "1048576"
+    # Disable dumping customer data to logs, both to increase data privacy
+    # and to reduce the amount the logs.
+    log_error_verbosity: "terse"
+    log_min_error_statement: "panic"
+    max_connections: "100"
+    # --- WAL ---
+    # - flush lag is the max amount of WAL that has been generated but not yet stored
+    # to disk in the page server. A smaller value means less delay after a pageserver
+    # restart, but if you set it too small you might again need to slow down writes if the
+    # pageserver cannot flush incoming WAL to disk fast enough. This must be larger
+    # than the pageserver's checkpoint interval, currently 1 GB! Otherwise you get a
+    # a deadlock where the compute node refuses to generate more WAL before the
+    # old WAL has been uploaded to S3, but the pageserver is waiting for more WAL
+    # to be generated before it is uploaded to S3.
+    max_replication_flush_lag: "10GB"
+    max_replication_slots: "10"
+    # Backpressure configuration:
+    # - write lag is the max amount of WAL that has been generated by Postgres but not yet
+    # processed by the page server. Making this smaller reduces the worst case latency
+    # of a GetPage request, if you request a page that was recently modified. On the other
+    # hand, if this is too small, the compute node might need to wait on a write if there is a
+    # hiccup in the network or page server so that the page server has temporarily fallen
+    # behind.
+    #
+    # Previously it was set to 500 MB, but it caused compute being unresponsive under load
+    # https://github.com/neondatabase/neon/issues/2028
+    max_replication_write_lag: "500MB"
+    max_wal_senders: "10"
+    # A Postgres checkpoint is cheap in storage, as doesn't involve any significant amount
+    # of real I/O. Only the SLRU buffers and some other small files are flushed to disk.
+    # However, as long as we have full_page_writes=on, page updates after a checkpoint
+    # include full-page images which bloats the WAL. So may want to bump max_wal_size to
+    # reduce the WAL bloating, but at the same it will increase pg_wal directory size on
+    # compute and can lead to out of disk error on k8s nodes.
+    max_wal_size: "1024"
+    wal_keep_size: "0"
+    wal_level: "replica"
+    # Reduce amount of WAL generated by default.
+    wal_log_hints: "off"
+    # - without wal_sender_timeout set we don't get feedback messages,
+    # required for backpressure.
+    wal_sender_timeout: "10000"
+    # We have some experimental extensions, which we don't want users to install unconsciously.
+    # To install them, users would need to set the `neon.allow_unstable_extensions` setting.
+    # There are two of them currently:
+    # - `pgrag` - https://github.com/neondatabase-labs/pgrag - extension is actually called just `rag`,
+    #                                                          and two dependencies:
+    #                                                          - `rag_bge_small_en_v15`
+    #                                                          - `rag_jina_reranker_v1_tiny_en`
+    # - `pg_mooncake` - https://github.com/Mooncake-Labs/pg_mooncake/  
+    neon.unstable_extensions: "rag,rag_bge_small_en_v15,rag_jina_reranker_v1_tiny_en,pg_mooncake,anon"
+    neon.protocol_version: "3"
+    password_encryption: "scram-sha-256"
+    # This is important to prevent Postgres from trying to perform
+    # a local WAL redo after backend crash. It should exit and let
+    # the systemd or k8s to do a fresh startup with compute_ctl.
+    restart_after_crash: "off"
+    # By default 3. We have the following persistent connections in the VM:
+    # * compute_activity_monitor (from compute_ctl)
+    # * postgres-exporter (metrics collector; it has 2 connections)
+    # * sql_exporter (metrics collector; we have 2 instances [1 for us & users; 1 for autoscaling])
+    # * vm-monitor (to query & change file cache size)
+    # i.e. total of 6. Let's reserve 7, so there's still at least one left over.
+    superuser_reserved_connections: "7"
+    synchronous_standby_names: "walproposer"
+
+  replica:
+    hot_standby: "on"
+
+  per_version:
+    17:
+      common:
+        # PostgreSQL 17 has a new IO system called "read stream", which can combine IOs up to some
+        # size. It still has some issues with readahead, though, so we default to disabled/
+        # "no combining of IOs" to make sure we get the maximum prefetch depth.
+        # See also: https://github.com/neondatabase/neon/pull/9860
+        io_combine_limit: "1"
+      replica:
+        # prefetching of blocks referenced in WAL doesn't make sense for us
+        # Neon hot standby ignores pages that are not in the shared_buffers
+        recovery_prefetch: "off"
+    16:
+      common:
+      replica:
+        # prefetching of blocks referenced in WAL doesn't make sense for us
+        # Neon hot standby ignores pages that are not in the shared_buffers
+        recovery_prefetch: "off"
+    15:
+      common:
+      replica:
+        # prefetching of blocks referenced in WAL doesn't make sense for us
+        # Neon hot standby ignores pages that are not in the shared_buffers
+        recovery_prefetch: "off"
+    14:
+      common:
+      replica:

compute_tools/src/compute.rs

@@ -3,7 +3,7 @@ use chrono::{DateTime, Utc};
 use compute_api::privilege::Privilege;
 use compute_api::responses::{
     ComputeConfig, ComputeCtlConfig, ComputeMetrics, ComputeStatus, LfcOffloadState,
-    LfcPrewarmState,
+    LfcPrewarmState, TlsConfig,
 };
 use compute_api::spec::{
     ComputeAudit, ComputeFeature, ComputeMode, ComputeSpec, ExtVersion, PgIdent,
@@ -396,7 +396,7 @@ impl ComputeNode {
         // because QEMU will already have its memory allocated from the host, and
         // the necessary binaries will already be cached.
         if cli_spec.is_none() {
-            this.prewarm_postgres()?;
+            this.prewarm_postgres_vm_memory()?;
         }
 
         // Set the up metric with Empty status before starting the HTTP server.
@@ -603,6 +603,8 @@ impl ComputeNode {
             });
         }
 
+        let tls_config = self.tls_config(&pspec.spec);
+
         // If there are any remote extensions in shared_preload_libraries, start downloading them
         if pspec.spec.remote_extensions.is_some() {
             let (this, spec) = (self.clone(), pspec.spec.clone());
@@ -659,7 +661,7 @@ impl ComputeNode {
             info!("tuning pgbouncer");
 
             let pgbouncer_settings = pgbouncer_settings.clone();
-            let tls_config = self.compute_ctl_config.tls.clone();
+            let tls_config = tls_config.clone();
 
             // Spawn a background task to do the tuning,
             // so that we don't block the main thread that starts Postgres.
@@ -678,7 +680,10 @@ impl ComputeNode {
 
             // Spawn a background task to do the configuration,
             // so that we don't block the main thread that starts Postgres.
-            let local_proxy = local_proxy.clone();
+
+            let mut local_proxy = local_proxy.clone();
+            local_proxy.tls = tls_config.clone();
+
             let _handle = tokio::spawn(async move {
                 if let Err(err) = local_proxy::configure(&local_proxy) {
                     error!("error while configuring local_proxy: {err:?}");
@@ -779,7 +784,7 @@ impl ComputeNode {
         // Spawn the extension stats background task
         self.spawn_extension_stats_task();
 
-        if pspec.spec.prewarm_lfc_on_startup {
+        if pspec.spec.autoprewarm {
             self.prewarm_lfc();
         }
         Ok(())
@@ -1205,13 +1210,15 @@ impl ComputeNode {
         let spec = &pspec.spec;
         let pgdata_path = Path::new(&self.params.pgdata);
 
+        let tls_config = self.tls_config(&pspec.spec);
+
         // Remove/create an empty pgdata directory and put configuration there.
         self.create_pgdata()?;
         config::write_postgres_conf(
             pgdata_path,
             &pspec.spec,
             self.params.internal_http_port,
-            &self.compute_ctl_config.tls,
+            tls_config,
         )?;
 
         // Syncing safekeepers is only safe with primary nodes: if a primary
@@ -1307,8 +1314,8 @@ impl ComputeNode {
     }
 
     /// Start and stop a postgres process to warm up the VM for startup.
-    pub fn prewarm_postgres(&self) -> Result<()> {
-        info!("prewarming");
+    pub fn prewarm_postgres_vm_memory(&self) -> Result<()> {
+        info!("prewarming VM memory");
 
         // Create pgdata
         let pgdata = &format!("{}.warmup", self.params.pgdata);
@@ -1350,7 +1357,7 @@ impl ComputeNode {
         kill(pm_pid, Signal::SIGQUIT)?;
         info!("sent SIGQUIT signal");
         pg.wait()?;
-        info!("done prewarming");
+        info!("done prewarming vm memory");
 
         // clean up
         let _ok = fs::remove_dir_all(pgdata);
@@ -1536,14 +1543,22 @@ impl ComputeNode {
                 .clone(),
         );
 
+        let mut tls_config = None::<TlsConfig>;
+        if spec.features.contains(&ComputeFeature::TlsExperimental) {
+            tls_config = self.compute_ctl_config.tls.clone();
+        }
+
         let max_concurrent_connections = self.max_service_connections(compute_state, &spec);
 
         // Merge-apply spec & changes to PostgreSQL state.
         self.apply_spec_sql(spec.clone(), conf.clone(), max_concurrent_connections)?;
 
         if let Some(local_proxy) = &spec.clone().local_proxy_config {
+            let mut local_proxy = local_proxy.clone();
+            local_proxy.tls = tls_config.clone();
+
             info!("configuring local_proxy");
-            local_proxy::configure(local_proxy).context("apply_config local_proxy")?;
+            local_proxy::configure(&local_proxy).context("apply_config local_proxy")?;
         }
 
         // Run migrations separately to not hold up cold starts
@@ -1595,11 +1610,13 @@ impl ComputeNode {
     pub fn reconfigure(&self) -> Result<()> {
         let spec = self.state.lock().unwrap().pspec.clone().unwrap().spec;
 
+        let tls_config = self.tls_config(&spec);
+
         if let Some(ref pgbouncer_settings) = spec.pgbouncer_settings {
             info!("tuning pgbouncer");
 
             let pgbouncer_settings = pgbouncer_settings.clone();
-            let tls_config = self.compute_ctl_config.tls.clone();
+            let tls_config = tls_config.clone();
 
             // Spawn a background task to do the tuning,
             // so that we don't block the main thread that starts Postgres.
@@ -1617,7 +1634,7 @@ impl ComputeNode {
             // Spawn a background task to do the configuration,
             // so that we don't block the main thread that starts Postgres.
             let mut local_proxy = local_proxy.clone();
-            local_proxy.tls = self.compute_ctl_config.tls.clone();
+            local_proxy.tls = tls_config.clone();
             tokio::spawn(async move {
                 if let Err(err) = local_proxy::configure(&local_proxy) {
                     error!("error while configuring local_proxy: {err:?}");
@@ -1635,7 +1652,7 @@ impl ComputeNode {
             pgdata_path,
             &spec,
             self.params.internal_http_port,
-            &self.compute_ctl_config.tls,
+            tls_config,
         )?;
 
         if !spec.skip_pg_catalog_updates {
@@ -1755,6 +1772,14 @@ impl ComputeNode {
         }
     }
 
+    pub fn tls_config(&self, spec: &ComputeSpec) -> &Option<TlsConfig> {
+        if spec.features.contains(&ComputeFeature::TlsExperimental) {
+            &self.compute_ctl_config.tls
+        } else {
+            &None::<TlsConfig>
+        }
+    }
+
     /// Update the `last_active` in the shared state, but ensure that it's a more recent one.
     pub fn update_last_active(&self, last_active: Option<DateTime<Utc>>) {
         let mut state = self.state.lock().unwrap();

compute_tools/src/http/mod.rs

@@ -48,11 +48,9 @@ impl JsonResponse {
 
     /// Create an error response related to the compute being in an invalid state
     pub(self) fn invalid_status(status: ComputeStatus) -> Response {
-        Self::create_response(
+        Self::error(
             StatusCode::PRECONDITION_FAILED,
-            &GenericAPIError {
-                error: format!("invalid compute status: {status}"),
-            },
+            format!("invalid compute status: {status}"),
         )
     }
 }

compute_tools/src/http/routes/configure.rs

@@ -22,7 +22,7 @@ pub(in crate::http) async fn configure(
     State(compute): State<Arc<ComputeNode>>,
     request: Json<ConfigurationRequest>,
 ) -> Response {
-    let pspec = match ParsedSpec::try_from(request.spec.clone()) {
+    let pspec = match ParsedSpec::try_from(request.0.spec) {
         Ok(p) => p,
         Err(e) => return JsonResponse::error(StatusCode::BAD_REQUEST, e),
     };

compute_tools/src/monitor.rs

@@ -13,6 +13,12 @@ use crate::metrics::{PG_CURR_DOWNTIME_MS, PG_TOTAL_DOWNTIME_MS};
 
 const MONITOR_CHECK_INTERVAL: Duration = Duration::from_millis(500);
 
+/// Struct to store runtime state of the compute monitor thread.
+/// In theory, this could be a part of `Compute`, but i)
+/// this state is expected to be accessed only by single thread,
+/// so we don't need to care about locking; ii) `Compute` is
+/// already quite big. Thus, it seems to be a good idea to keep
+/// all the activity/health monitoring parts here.
 struct ComputeMonitor {
     compute: Arc<ComputeNode>,
 
@@ -70,12 +76,36 @@ impl ComputeMonitor {
         )
     }
 
+    /// Check if compute is in some terminal or soon-to-be-terminal
+    /// state, then return `true`, signalling the caller that it
+    /// should exit gracefully. Otherwise, return `false`.
+    fn check_interrupts(&mut self) -> bool {
+        let compute_status = self.compute.get_status();
+        if matches!(
+            compute_status,
+            ComputeStatus::Terminated | ComputeStatus::TerminationPending | ComputeStatus::Failed
+        ) {
+            info!(
+                "compute is in {} status, stopping compute monitor",
+                compute_status
+            );
+            return true;
+        }
+
+        false
+    }
+
     /// Spin in a loop and figure out the last activity time in the Postgres.
-    /// Then update it in the shared state. This function never errors out.
+    /// Then update it in the shared state. This function currently never
+    /// errors out explicitly, but there is a graceful termination path.
+    /// Every time we receive an error trying to check Postgres, we use
+    /// [`ComputeMonitor::check_interrupts()`] because it could be that
+    /// compute is being terminated already, then we can exit gracefully
+    /// to not produce errors' noise in the log.
     /// NB: the only expected panic is at `Mutex` unwrap(), all other errors
     /// should be handled gracefully.
     #[instrument(skip_all)]
-    pub fn run(&mut self) {
+    pub fn run(&mut self) -> anyhow::Result<()> {
         // Suppose that `connstr` doesn't change
         let connstr = self.compute.params.connstr.clone();
         let conf = self
@@ -93,6 +123,10 @@ impl ComputeMonitor {
         info!("starting compute monitor for {}", connstr);
 
         loop {
+            if self.check_interrupts() {
+                break;
+            }
+
             match &mut client {
                 Ok(cli) => {
                     if cli.is_closed() {
@@ -100,6 +134,10 @@ impl ComputeMonitor {
                             downtime_info = self.downtime_info(),
                             "connection to Postgres is closed, trying to reconnect"
                         );
+                        if self.check_interrupts() {
+                            break;
+                        }
+
                         self.report_down();
 
                         // Connection is closed, reconnect and try again.
@@ -111,15 +149,19 @@ impl ComputeMonitor {
                                 self.compute.update_last_active(self.last_active);
                             }
                             Err(e) => {
+                                error!(
+                                    downtime_info = self.downtime_info(),
+                                    "could not check Postgres: {}", e
+                                );
+                                if self.check_interrupts() {
+                                    break;
+                                }
+
                                 // Although we have many places where we can return errors in `check()`,
                                 // normally it shouldn't happen. I.e., we will likely return error if
                                 // connection got broken, query timed out, Postgres returned invalid data, etc.
                                 // In all such cases it's suspicious, so let's report this as downtime.
                                 self.report_down();
-                                error!(
-                                    downtime_info = self.downtime_info(),
-                                    "could not check Postgres: {}", e
-                                );
 
                                 // Reconnect to Postgres just in case. During tests, I noticed
                                 // that queries in `check()` can fail with `connection closed`,
@@ -136,6 +178,10 @@ impl ComputeMonitor {
                         downtime_info = self.downtime_info(),
                         "could not connect to Postgres: {}, retrying", e
                     );
+                    if self.check_interrupts() {
+                        break;
+                    }
+
                     self.report_down();
 
                     // Establish a new connection and try again.
@@ -147,6 +193,9 @@ impl ComputeMonitor {
             self.last_checked = Utc::now();
             thread::sleep(MONITOR_CHECK_INTERVAL);
         }
+
+        // Graceful termination path
+        Ok(())
     }
 
     #[instrument(skip_all)]
@@ -429,7 +478,10 @@ pub fn launch_monitor(compute: &Arc<ComputeNode>) -> thread::JoinHandle<()> {
         .spawn(move || {
             let span = span!(Level::INFO, "compute_monitor");
             let _enter = span.enter();
-            monitor.run();
+            match monitor.run() {
+                Ok(_) => info!("compute monitor thread terminated gracefully"),
+                Err(err) => error!("compute monitor thread terminated abnormally {:?}", err),
+            }
         })
         .expect("cannot launch compute monitor thread")
 }

compute_tools/tests/pg_helpers_tests.rs

@@ -30,7 +30,7 @@ mod pg_helpers_tests {
             r#"fsync = off
 wal_level = logical
 hot_standby = on
-prewarm_lfc_on_startup = off
+autoprewarm = off
 neon.safekeepers = '127.0.0.1:6502,127.0.0.1:6503,127.0.0.1:6501'
 wal_log_hints = on
 log_connections = on

control_plane/Cargo.toml

@@ -36,6 +36,7 @@ pageserver_api.workspace = true
 pageserver_client.workspace = true
 postgres_backend.workspace = true
 safekeeper_api.workspace = true
+safekeeper_client.workspace = true
 postgres_connection.workspace = true
 storage_broker.workspace = true
 http-utils.workspace = true

control_plane/src/bin/neon_local.rs

@@ -45,7 +45,7 @@ use pageserver_api::models::{
 use pageserver_api::shard::{DEFAULT_STRIPE_SIZE, ShardCount, ShardStripeSize, TenantShardId};
 use postgres_backend::AuthType;
 use postgres_connection::parse_host_port;
-use safekeeper_api::membership::SafekeeperGeneration;
+use safekeeper_api::membership::{SafekeeperGeneration, SafekeeperId};
 use safekeeper_api::{
     DEFAULT_HTTP_LISTEN_PORT as DEFAULT_SAFEKEEPER_HTTP_PORT,
     DEFAULT_PG_LISTEN_PORT as DEFAULT_SAFEKEEPER_PG_PORT,
@@ -1255,6 +1255,45 @@ async fn handle_timeline(cmd: &TimelineCmd, env: &mut local_env::LocalEnv) -> Re
             pageserver
                 .timeline_import(tenant_id, timeline_id, base, pg_wal, args.pg_version)
                 .await?;
+            if env.storage_controller.timelines_onto_safekeepers {
+                println!("Creating timeline on safekeeper ...");
+                let timeline_info = pageserver
+                    .timeline_info(
+                        TenantShardId::unsharded(tenant_id),
+                        timeline_id,
+                        pageserver_client::mgmt_api::ForceAwaitLogicalSize::No,
+                    )
+                    .await?;
+                let default_sk = SafekeeperNode::from_env(env, env.safekeepers.first().unwrap());
+                let default_host = default_sk
+                    .conf
+                    .listen_addr
+                    .clone()
+                    .unwrap_or_else(|| "localhost".to_string());
+                let mconf = safekeeper_api::membership::Configuration {
+                    generation: SafekeeperGeneration::new(1),
+                    members: safekeeper_api::membership::MemberSet {
+                        m: vec![SafekeeperId {
+                            host: default_host,
+                            id: default_sk.conf.id,
+                            pg_port: default_sk.conf.pg_port,
+                        }],
+                    },
+                    new_members: None,
+                };
+                let pg_version = args.pg_version * 10000;
+                let req = safekeeper_api::models::TimelineCreateRequest {
+                    tenant_id,
+                    timeline_id,
+                    mconf,
+                    pg_version,
+                    system_id: None,
+                    wal_seg_size: None,
+                    start_lsn: timeline_info.last_record_lsn,
+                    commit_lsn: None,
+                };
+                default_sk.create_timeline(&req).await?;
+            }
             env.register_branch_mapping(branch_name.to_string(), tenant_id, timeline_id)?;
             println!("Done");
         }

control_plane/src/endpoint.rs

@@ -747,7 +747,7 @@ impl Endpoint {
                 logs_export_host: None::<String>,
                 endpoint_storage_addr: Some(endpoint_storage_addr),
                 endpoint_storage_token: Some(endpoint_storage_token),
-                prewarm_lfc_on_startup: false,
+                autoprewarm: false,
             };
 
             // this strange code is needed to support respec() in tests

control_plane/src/pageserver.rs

@@ -513,11 +513,6 @@ impl PageServerNode {
                 .map(|x| x.parse::<bool>())
                 .transpose()
                 .context("Failed to parse 'timeline_offloading' as bool")?,
-            wal_receiver_protocol_override: settings
-                .remove("wal_receiver_protocol_override")
-                .map(serde_json::from_str)
-                .transpose()
-                .context("parse `wal_receiver_protocol_override` from json")?,
             rel_size_v2_enabled: settings
                 .remove("rel_size_v2_enabled")
                 .map(|x| x.parse::<bool>())
@@ -640,4 +635,16 @@ impl PageServerNode {
 
         Ok(())
     }
+    pub async fn timeline_info(
+        &self,
+        tenant_shard_id: TenantShardId,
+        timeline_id: TimelineId,
+        force_await_logical_size: mgmt_api::ForceAwaitLogicalSize,
+    ) -> anyhow::Result<TimelineInfo> {
+        let timeline_info = self
+            .http_client
+            .timeline_info(tenant_shard_id, timeline_id, force_await_logical_size)
+            .await?;
+        Ok(timeline_info)
+    }
 }

control_plane/src/safekeeper.rs

@@ -6,7 +6,6 @@
 //!   .neon/safekeepers/<safekeeper id>
 //! ```
 use std::error::Error as _;
-use std::future::Future;
 use std::io::Write;
 use std::path::PathBuf;
 use std::time::Duration;
@@ -14,9 +13,9 @@ use std::{io, result};
 
 use anyhow::Context;
 use camino::Utf8PathBuf;
-use http_utils::error::HttpErrorBody;
 use postgres_connection::PgConnectionConfig;
-use reqwest::{IntoUrl, Method};
+use safekeeper_api::models::TimelineCreateRequest;
+use safekeeper_client::mgmt_api;
 use thiserror::Error;
 use utils::auth::{Claims, Scope};
 use utils::id::NodeId;
@@ -35,25 +34,14 @@ pub enum SafekeeperHttpError {
 
 type Result<T> = result::Result<T, SafekeeperHttpError>;
 
-pub(crate) trait ResponseErrorMessageExt: Sized {
-    fn error_from_body(self) -> impl Future<Output = Result<Self>> + Send;
-}
-
-impl ResponseErrorMessageExt for reqwest::Response {
-    async fn error_from_body(self) -> Result<Self> {
-        let status = self.status();
-        if !(status.is_client_error() || status.is_server_error()) {
-            return Ok(self);
-        }
-
-        // reqwest does not export its error construction utility functions, so let's craft the message ourselves
-        let url = self.url().to_owned();
-        Err(SafekeeperHttpError::Response(
-            match self.json::<HttpErrorBody>().await {
-                Ok(err_body) => format!("Error: {}", err_body.msg),
-                Err(_) => format!("Http error ({}) at {}.", status.as_u16(), url),
-            },
-        ))
+fn err_from_client_err(err: mgmt_api::Error) -> SafekeeperHttpError {
+    use mgmt_api::Error::*;
+    match err {
+        ApiError(_, str) => SafekeeperHttpError::Response(str),
+        Cancelled => SafekeeperHttpError::Response("Cancelled".to_owned()),
+        ReceiveBody(err) => SafekeeperHttpError::Transport(err),
+        ReceiveErrorBody(err) => SafekeeperHttpError::Response(err),
+        Timeout(str) => SafekeeperHttpError::Response(format!("timeout: {str}")),
     }
 }
 
@@ -70,9 +58,8 @@ pub struct SafekeeperNode {
 
     pub pg_connection_config: PgConnectionConfig,
     pub env: LocalEnv,
-    pub http_client: reqwest::Client,
+    pub http_client: mgmt_api::Client,
     pub listen_addr: String,
-    pub http_base_url: String,
 }
 
 impl SafekeeperNode {
@@ -82,13 +69,14 @@ impl SafekeeperNode {
         } else {
             "127.0.0.1".to_string()
         };
+        let jwt = None;
+        let http_base_url = format!("http://{}:{}", listen_addr, conf.http_port);
         SafekeeperNode {
             id: conf.id,
             conf: conf.clone(),
             pg_connection_config: Self::safekeeper_connection_config(&listen_addr, conf.pg_port),
             env: env.clone(),
-            http_client: env.create_http_client(),
-            http_base_url: format!("http://{}:{}/v1", listen_addr, conf.http_port),
+            http_client: mgmt_api::Client::new(env.create_http_client(), http_base_url, jwt),
             listen_addr,
         }
     }
@@ -278,20 +266,19 @@ impl SafekeeperNode {
         )
     }
 
-    fn http_request<U: IntoUrl>(&self, method: Method, url: U) -> reqwest::RequestBuilder {
-        // TODO: authentication
-        //if self.env.auth_type == AuthType::NeonJWT {
-        //    builder = builder.bearer_auth(&self.env.safekeeper_auth_token)
-        //}
-        self.http_client.request(method, url)
+    pub async fn check_status(&self) -> Result<()> {
+        self.http_client
+            .status()
+            .await
+            .map_err(err_from_client_err)?;
+        Ok(())
     }
 
-    pub async fn check_status(&self) -> Result<()> {
-        self.http_request(Method::GET, format!("{}/{}", self.http_base_url, "status"))
-            .send()
-            .await?
-            .error_from_body()
-            .await?;
+    pub async fn create_timeline(&self, req: &TimelineCreateRequest) -> Result<()> {
+        self.http_client
+            .create_timeline(req)
+            .await
+            .map_err(err_from_client_err)?;
         Ok(())
     }
 }

control_plane/storcon_cli/src/main.rs

@@ -61,10 +61,16 @@ enum Command {
         #[arg(long)]
         scheduling: Option<NodeSchedulingPolicy>,
     },
+    // Set a node status as deleted.
     NodeDelete {
         #[arg(long)]
         node_id: NodeId,
     },
+    /// Delete a tombstone of node from the storage controller.
+    NodeDeleteTombstone {
+        #[arg(long)]
+        node_id: NodeId,
+    },
     /// Modify a tenant's policies in the storage controller
     TenantPolicy {
         #[arg(long)]
@@ -82,6 +88,8 @@ enum Command {
     },
     /// List nodes known to the storage controller
     Nodes {},
+    /// List soft deleted nodes known to the storage controller
+    NodeTombstones {},
     /// List tenants known to the storage controller
     Tenants {
         /// If this field is set, it will list the tenants on a specific node
@@ -900,6 +908,39 @@ async fn main() -> anyhow::Result<()> {
                 .dispatch::<(), ()>(Method::DELETE, format!("control/v1/node/{node_id}"), None)
                 .await?;
         }
+        Command::NodeDeleteTombstone { node_id } => {
+            storcon_client
+                .dispatch::<(), ()>(
+                    Method::DELETE,
+                    format!("debug/v1/tombstone/{node_id}"),
+                    None,
+                )
+                .await?;
+        }
+        Command::NodeTombstones {} => {
+            let mut resp = storcon_client
+                .dispatch::<(), Vec<NodeDescribeResponse>>(
+                    Method::GET,
+                    "debug/v1/tombstone".to_string(),
+                    None,
+                )
+                .await?;
+
+            resp.sort_by(|a, b| a.listen_http_addr.cmp(&b.listen_http_addr));
+
+            let mut table = comfy_table::Table::new();
+            table.set_header(["Id", "Hostname", "AZ", "Scheduling", "Availability"]);
+            for node in resp {
+                table.add_row([
+                    format!("{}", node.id),
+                    node.listen_http_addr,
+                    node.availability_zone_id,
+                    format!("{:?}", node.scheduling),
+                    format!("{:?}", node.availability),
+                ]);
+            }
+            println!("{table}");
+        }
         Command::TenantSetTimeBasedEviction {
             tenant_id,
             period,

docker-compose/compute_wrapper/Dockerfile

@@ -13,6 +13,6 @@ RUN echo 'Acquire::Retries "5";' > /etc/apt/apt.conf.d/80-retries && \
                        jq   \
                        netcat-openbsd
 #This is required for the pg_hintplan test
-RUN mkdir -p /ext-src/pg_hint_plan-src /postgres/contrib/file_fdw && chown postgres /ext-src/pg_hint_plan-src /postgres/contrib/file_fdw
+RUN mkdir -p /ext-src/pg_hint_plan-src /postgres/contrib/file_fdw /ext-src/postgis-src/ && chown postgres /ext-src/pg_hint_plan-src /postgres/contrib/file_fdw /ext-src/postgis-src
 
 USER postgres

docker-compose/compute_wrapper/shell/compute.sh

@@ -1,18 +1,18 @@
-#!/bin/bash
+#!/usr/bin/env bash
 set -eux
 
 # Generate a random tenant or timeline ID
 #
 # Takes a variable name as argument. The result is stored in that variable.
 generate_id() {
-    local -n resvar=$1
-    printf -v resvar '%08x%08x%08x%08x' $SRANDOM $SRANDOM $SRANDOM $SRANDOM
+    local -n resvar=${1}
+    printf -v resvar '%08x%08x%08x%08x' ${SRANDOM} ${SRANDOM} ${SRANDOM} ${SRANDOM}
 }
 
 PG_VERSION=${PG_VERSION:-14}
 
-CONFIG_FILE_ORG=/var/db/postgres/configs/config.json
-CONFIG_FILE=/tmp/config.json
+readonly CONFIG_FILE_ORG=/var/db/postgres/configs/config.json
+readonly CONFIG_FILE=/tmp/config.json
 
 # Test that the first library path that the dynamic loader looks in is the path
 # that we use for custom compiled software
@@ -20,17 +20,17 @@ first_path="$(ldconfig --verbose 2>/dev/null \
     | grep --invert-match ^$'\t' \
     | cut --delimiter=: --fields=1 \
     | head --lines=1)"
-test "$first_path" == '/usr/local/lib'
+test "${first_path}" = '/usr/local/lib'
 
 echo "Waiting pageserver become ready."
 while ! nc -z pageserver 6400; do
-     sleep 1;
+     sleep 1
 done
 echo "Page server is ready."
 
-cp ${CONFIG_FILE_ORG} ${CONFIG_FILE}
+cp "${CONFIG_FILE_ORG}" "${CONFIG_FILE}"
 
- if [ -n "${TENANT_ID:-}" ] && [ -n "${TIMELINE_ID:-}" ]; then
+ if [[ -n "${TENANT_ID:-}" && -n "${TIMELINE_ID:-}" ]]; then
    tenant_id=${TENANT_ID}
    timeline_id=${TIMELINE_ID}
 else
@@ -41,7 +41,7 @@ else
        "http://pageserver:9898/v1/tenant"
   )
   tenant_id=$(curl "${PARAMS[@]}" | jq -r .[0].id)
-  if [ -z "${tenant_id}" ] || [ "${tenant_id}" = null ]; then
+  if [[ -z "${tenant_id}" || "${tenant_id}" = null ]]; then
     echo "Create a tenant"
     generate_id tenant_id
     PARAMS=(
@@ -51,7 +51,7 @@ else
         "http://pageserver:9898/v1/tenant/${tenant_id}/location_config"
     )
     result=$(curl "${PARAMS[@]}")
-    echo $result | jq .
+    printf '%s\n' "${result}" | jq .
   fi
 
   echo "Check if a timeline present"
@@ -61,7 +61,7 @@ else
        "http://pageserver:9898/v1/tenant/${tenant_id}/timeline"
   )
   timeline_id=$(curl "${PARAMS[@]}" | jq -r .[0].timeline_id)
-  if [ -z "${timeline_id}" ] || [ "${timeline_id}" = null ]; then
+  if [[ -z "${timeline_id}" || "${timeline_id}" = null ]]; then
     generate_id timeline_id
     PARAMS=(
         -sbf
@@ -71,7 +71,7 @@ else
         "http://pageserver:9898/v1/tenant/${tenant_id}/timeline/"
     )
     result=$(curl "${PARAMS[@]}")
-    echo $result | jq .
+    printf '%s\n' "${result}" | jq .
   fi
 fi
 
@@ -82,10 +82,10 @@ else
 fi
 echo "Adding pgx_ulid"
 shared_libraries=$(jq -r '.spec.cluster.settings[] | select(.name=="shared_preload_libraries").value' ${CONFIG_FILE})
-sed -i "s/${shared_libraries}/${shared_libraries},${ulid_extension}/" ${CONFIG_FILE}
+sed -i "s|${shared_libraries}|${shared_libraries},${ulid_extension}|" ${CONFIG_FILE}
 echo "Overwrite tenant id and timeline id in spec file"
-sed -i "s/TENANT_ID/${tenant_id}/" ${CONFIG_FILE}
-sed -i "s/TIMELINE_ID/${timeline_id}/" ${CONFIG_FILE}
+sed -i "s|TENANT_ID|${tenant_id}|" ${CONFIG_FILE}
+sed -i "s|TIMELINE_ID|${timeline_id}|" ${CONFIG_FILE}
 
 cat ${CONFIG_FILE}
 
@@ -93,5 +93,5 @@ echo "Start compute node"
 /usr/local/bin/compute_ctl --pgdata /var/db/postgres/compute \
      -C "postgresql://cloud_admin@localhost:55433/postgres"  \
      -b /usr/local/bin/postgres                              \
-     --compute-id "compute-$RANDOM"                          \
-     --config "$CONFIG_FILE"
+     --compute-id "compute-${RANDOM}"                          \
+     --config "${CONFIG_FILE}"

docker-compose/docker-compose.yml

@@ -186,13 +186,14 @@ services:
 
   neon-test-extensions:
     profiles: ["test-extensions"]
-    image: ${REPOSITORY:-ghcr.io/neondatabase}/neon-test-extensions-v${PG_TEST_VERSION:-16}:${TEST_EXTENSIONS_TAG:-${TAG:-latest}}
+    image: ${REPOSITORY:-ghcr.io/neondatabase}/neon-test-extensions-v${PG_TEST_VERSION:-${PG_VERSION:-16}}:${TEST_EXTENSIONS_TAG:-${TAG:-latest}}
     environment:
-      - PGPASSWORD=cloud_admin
+      - PGUSER=${PGUSER:-cloud_admin}
+      - PGPASSWORD=${PGPASSWORD:-cloud_admin}
     entrypoint:
       - "/bin/bash"
       - "-c"
     command:
-      - sleep 1800
+      - sleep 3600
     depends_on:
       - compute

docker-compose/docker_compose_test.sh

@@ -54,6 +54,15 @@ for pg_version in ${TEST_VERSION_ONLY-14 15 16 17}; do
         # It cannot be moved to Dockerfile now because the database directory is created after the start of the container
         echo Adding dummy config
         docker compose exec compute touch /var/db/postgres/compute/compute_ctl_temp_override.conf
+        # Prepare for the PostGIS test
+        docker compose exec compute mkdir -p /tmp/pgis_reg/pgis_reg_tmp
+        TMPDIR=$(mktemp -d)
+        docker compose cp neon-test-extensions:/ext-src/postgis-src/raster/test "${TMPDIR}"
+        docker compose cp neon-test-extensions:/ext-src/postgis-src/regress/00-regress-install "${TMPDIR}"
+        docker compose exec compute mkdir -p /ext-src/postgis-src/raster /ext-src/postgis-src/regress /ext-src/postgis-src/regress/00-regress-install
+        docker compose cp "${TMPDIR}/test" compute:/ext-src/postgis-src/raster/test
+        docker compose cp "${TMPDIR}/00-regress-install" compute:/ext-src/postgis-src/regress
+        rm -rf "${TMPDIR}"
         # The following block copies the files for the pg_hintplan test to the compute node for the extension test in an isolated docker-compose environment
         TMPDIR=$(mktemp -d)
         docker compose cp neon-test-extensions:/ext-src/pg_hint_plan-src/data "${TMPDIR}/data"
@@ -68,7 +77,7 @@ for pg_version in ${TEST_VERSION_ONLY-14 15 16 17}; do
         docker compose exec -T neon-test-extensions bash -c "(cd /postgres && patch -p1)" <"../compute/patches/contrib_pg${pg_version}.patch"
         # We are running tests now
         rm -f testout.txt testout_contrib.txt
-        docker compose exec -e USE_PGXS=1 -e SKIP=timescaledb-src,rdkit-src,postgis-src,pg_jsonschema-src,kq_imcx-src,wal2json_2_5-src,rag_jina_reranker_v1_tiny_en-src,rag_bge_small_en_v15-src \
+        docker compose exec -e USE_PGXS=1 -e SKIP=timescaledb-src,rdkit-src,pg_jsonschema-src,kq_imcx-src,wal2json_2_5-src,rag_jina_reranker_v1_tiny_en-src,rag_bge_small_en_v15-src \
         neon-test-extensions /run-tests.sh /ext-src | tee testout.txt && EXT_SUCCESS=1 || EXT_SUCCESS=0
         docker compose exec -e SKIP=start-scripts,postgres_fdw,ltree_plpython,jsonb_plpython,jsonb_plperl,hstore_plpython,hstore_plperl,dblink,bool_plperl \
         neon-test-extensions /run-tests.sh /postgres/contrib | tee testout_contrib.txt && CONTRIB_SUCCESS=1 || CONTRIB_SUCCESS=0

docker-compose/ext-src/postgis-src/README-Neon.md

@@ -0,0 +1,70 @@
+# PostGIS Testing in Neon
+
+This directory contains configuration files and patches for running PostGIS tests in the Neon database environment.
+
+## Overview
+
+PostGIS is a spatial database extension for PostgreSQL that adds support for geographic objects. Testing PostGIS compatibility ensures that Neon's modifications to PostgreSQL don't break compatibility with this critical extension.
+
+## PostGIS Versions
+
+- PostgreSQL v17: PostGIS 3.5.0
+- PostgreSQL v14/v15/v16: PostGIS 3.3.3
+
+## Test Configuration
+
+The test setup includes:
+
+- `postgis-no-upgrade-test.patch`: Disables upgrade tests by removing the upgrade test section from regress/runtest.mk
+- `postgis-regular-v16.patch`: Version-specific patch for PostgreSQL v16
+- `postgis-regular-v17.patch`: Version-specific patch for PostgreSQL v17
+- `regular-test.sh`: Script to run PostGIS tests as a regular user
+- `neon-test.sh`: Script to handle version-specific test configurations
+- `raster_outdb_template.sql`: Template for raster tests with explicit file paths
+
+## Excluded Tests
+
+**Important Note:** The test exclusions listed below are specifically for regular-user tests against staging instances. These exclusions are necessary because staging instances run with limited privileges and cannot perform operations requiring superuser access. Docker-compose based tests are not affected by these exclusions.
+
+### Tests Requiring Superuser Permissions
+
+These tests cannot be run as a regular user:
+- `estimatedextent`
+- `regress/core/legacy`
+- `regress/core/typmod`
+- `regress/loader/TestSkipANALYZE`
+- `regress/loader/TestANALYZE`
+
+### Tests Requiring Filesystem Access
+
+These tests need direct filesystem access that is only possible for superusers:
+- `loader/load_outdb`
+
+### Tests with Flaky Results
+
+These tests have assumptions that don't always hold true:
+- `regress/core/computed_columns` - Assumes computed columns always outperform alternatives, which is not consistently true
+
+### Tests Requiring Tunable Parameter Modifications
+
+These tests attempt to modify the `postgis.gdal_enabled_drivers` parameter, which is only accessible to superusers:
+- `raster/test/regress/rt_wkb`
+- `raster/test/regress/rt_addband`
+- `raster/test/regress/rt_setbandpath`
+- `raster/test/regress/rt_fromgdalraster`
+- `raster/test/regress/rt_asgdalraster`
+- `raster/test/regress/rt_astiff`
+- `raster/test/regress/rt_asjpeg`
+- `raster/test/regress/rt_aspng`
+- `raster/test/regress/permitted_gdal_drivers`
+- Loader tests: `BasicOutDB`, `Tiled10x10`, `Tiled10x10Copy`, `Tiled8x8`, `TiledAuto`, `TiledAutoSkipNoData`, `TiledAutoCopyn`
+
+### Topology Tests (v17 only)
+- `populate_topology_layer`
+- `renametopogeometrycolumn`
+
+## Other Modifications
+
+- Binary.sql tests are modified to use explicit file paths
+- Server-side SQL COPY commands (which require superuser privileges) are converted to client-side `\copy` commands
+- Upgrade tests are disabled

docker-compose/ext-src/postgis-src/neon-test.sh

@@ -0,0 +1,6 @@
+#!/bin/sh
+set -ex
+cd "$(dirname "$0")"
+patch -p1 <"postgis-common-${PG_VERSION}.patch"
+trap 'echo Cleaning up; patch -R -p1 <postgis-common-${PG_VERSION}.patch' EXIT
+make installcheck-base

docker-compose/ext-src/postgis-src/postgis-common-v16.patch

@@ -0,0 +1,37 @@
+diff --git a/regress/core/tests.mk b/regress/core/tests.mk
+index 3abd7bc..64a9254 100644
+--- a/regress/core/tests.mk
++++ b/regress/core/tests.mk
+@@ -144,11 +144,6 @@ TESTS_SLOW = \
+ 	$(top_srcdir)/regress/core/concave_hull_hard \
+ 	$(top_srcdir)/regress/core/knn_recheck
+ 
+-ifeq ($(shell expr "$(POSTGIS_PGSQL_VERSION)" ">=" 120),1)
+-	TESTS += \
+-		$(top_srcdir)/regress/core/computed_columns
+-endif
+-
+ ifeq ($(shell expr "$(POSTGIS_GEOS_VERSION)" ">=" 30700),1)
+ 	# GEOS-3.7 adds:
+ 	# ST_FrechetDistance
+diff --git a/regress/runtest.mk b/regress/runtest.mk
+index c051f03..010e493 100644
+--- a/regress/runtest.mk
++++ b/regress/runtest.mk
+@@ -24,16 +24,6 @@ check-regress:
+ 
+ 	POSTGIS_TOP_BUILD_DIR=$(abs_top_builddir) $(PERL) $(top_srcdir)/regress/run_test.pl $(RUNTESTFLAGS) $(RUNTESTFLAGS_INTERNAL) $(TESTS)
+ 
+-	@if echo "$(RUNTESTFLAGS)" | grep -vq -- --upgrade; then \
+-		echo "Running upgrade test as RUNTESTFLAGS did not contain that"; \
+-		POSTGIS_TOP_BUILD_DIR=$(abs_top_builddir) $(PERL) $(top_srcdir)/regress/run_test.pl \
+-      --upgrade \
+-      $(RUNTESTFLAGS) \
+-      $(RUNTESTFLAGS_INTERNAL) \
+-      $(TESTS); \
+-	else \
+-		echo "Skipping upgrade test as RUNTESTFLAGS already requested upgrades"; \
+-	fi
+ 
+ check-long:
+ 	$(PERL) $(top_srcdir)/regress/run_test.pl $(RUNTESTFLAGS) $(TESTS) $(TESTS_SLOW)

docker-compose/ext-src/postgis-src/postgis-common-v17.patch

@@ -0,0 +1,35 @@
+diff --git a/regress/core/tests.mk b/regress/core/tests.mk
+index 9e05244..90987df 100644
+--- a/regress/core/tests.mk
++++ b/regress/core/tests.mk
+@@ -143,8 +143,7 @@ TESTS += \
+ 	$(top_srcdir)/regress/core/oriented_envelope \
+ 	$(top_srcdir)/regress/core/point_coordinates \
+ 	$(top_srcdir)/regress/core/out_geojson \
+-  $(top_srcdir)/regress/core/wrapx \
+-	$(top_srcdir)/regress/core/computed_columns
++  $(top_srcdir)/regress/core/wrapx
+ 
+ # Slow slow tests
+ TESTS_SLOW = \
+diff --git a/regress/runtest.mk b/regress/runtest.mk
+index 4b95b7e..449d5a2 100644
+--- a/regress/runtest.mk
++++ b/regress/runtest.mk
+@@ -24,16 +24,6 @@ check-regress:
+ 
+ 	@POSTGIS_TOP_BUILD_DIR=$(abs_top_builddir) $(PERL) $(top_srcdir)/regress/run_test.pl $(RUNTESTFLAGS) $(RUNTESTFLAGS_INTERNAL) $(TESTS)
+ 
+-	@if echo "$(RUNTESTFLAGS)" | grep -vq -- --upgrade; then \
+-		echo "Running upgrade test as RUNTESTFLAGS did not contain that"; \
+-		POSTGIS_TOP_BUILD_DIR=$(abs_top_builddir) $(PERL) $(top_srcdir)/regress/run_test.pl \
+-      --upgrade \
+-      $(RUNTESTFLAGS) \
+-      $(RUNTESTFLAGS_INTERNAL) \
+-      $(TESTS); \
+-	else \
+-		echo "Skipping upgrade test as RUNTESTFLAGS already requested upgrades"; \
+-	fi
+ 
+ check-long:
+ 	$(PERL) $(top_srcdir)/regress/run_test.pl $(RUNTESTFLAGS) $(TESTS) $(TESTS_SLOW)

docker-compose/ext-src/postgis-src/postgis-regular-v16.patch

@@ -0,0 +1,186 @@
+diff --git a/raster/test/regress/tests.mk b/raster/test/regress/tests.mk
+index 00918e1..7e2b6cd 100644
+--- a/raster/test/regress/tests.mk
++++ b/raster/test/regress/tests.mk
+@@ -17,9 +17,7 @@ override RUNTESTFLAGS_INTERNAL := \
+   $(RUNTESTFLAGS_INTERNAL) \
+   --after-upgrade-script $(top_srcdir)/raster/test/regress/hooks/hook-after-upgrade-raster.sql
+ 
+-RASTER_TEST_FIRST = \
+-	$(top_srcdir)/raster/test/regress/check_gdal \
+-	$(top_srcdir)/raster/test/regress/loader/load_outdb
++RASTER_TEST_FIRST =
+ 
+ RASTER_TEST_LAST = \
+ 	$(top_srcdir)/raster/test/regress/clean
+@@ -33,9 +31,7 @@ RASTER_TEST_IO = \
+ 
+ RASTER_TEST_BASIC_FUNC = \
+ 	$(top_srcdir)/raster/test/regress/rt_bytea \
+-	$(top_srcdir)/raster/test/regress/rt_wkb \
+ 	$(top_srcdir)/raster/test/regress/box3d \
+-	$(top_srcdir)/raster/test/regress/rt_addband \
+ 	$(top_srcdir)/raster/test/regress/rt_band \
+ 	$(top_srcdir)/raster/test/regress/rt_tile
+ 
+@@ -73,16 +69,10 @@ RASTER_TEST_BANDPROPS = \
+ 	$(top_srcdir)/raster/test/regress/rt_neighborhood \
+ 	$(top_srcdir)/raster/test/regress/rt_nearestvalue \
+ 	$(top_srcdir)/raster/test/regress/rt_pixelofvalue \
+-	$(top_srcdir)/raster/test/regress/rt_polygon \
+-	$(top_srcdir)/raster/test/regress/rt_setbandpath
++	$(top_srcdir)/raster/test/regress/rt_polygon
+ 
+ RASTER_TEST_UTILITY = \
+ 	$(top_srcdir)/raster/test/regress/rt_utility \
+-	$(top_srcdir)/raster/test/regress/rt_fromgdalraster \
+-	$(top_srcdir)/raster/test/regress/rt_asgdalraster \
+-	$(top_srcdir)/raster/test/regress/rt_astiff \
+-	$(top_srcdir)/raster/test/regress/rt_asjpeg \
+-	$(top_srcdir)/raster/test/regress/rt_aspng \
+ 	$(top_srcdir)/raster/test/regress/rt_reclass \
+ 	$(top_srcdir)/raster/test/regress/rt_gdalwarp \
+ 	$(top_srcdir)/raster/test/regress/rt_gdalcontour \
+@@ -120,21 +110,13 @@ RASTER_TEST_SREL = \
+ 
+ RASTER_TEST_BUGS = \
+ 	$(top_srcdir)/raster/test/regress/bug_test_car5 \
+-	$(top_srcdir)/raster/test/regress/permitted_gdal_drivers \
+ 	$(top_srcdir)/raster/test/regress/tickets
+ 
+ RASTER_TEST_LOADER = \
+ 	$(top_srcdir)/raster/test/regress/loader/Basic \
+ 	$(top_srcdir)/raster/test/regress/loader/Projected \
+ 	$(top_srcdir)/raster/test/regress/loader/BasicCopy \
+-	$(top_srcdir)/raster/test/regress/loader/BasicFilename \
+-	$(top_srcdir)/raster/test/regress/loader/BasicOutDB \
+-	$(top_srcdir)/raster/test/regress/loader/Tiled10x10 \
+-	$(top_srcdir)/raster/test/regress/loader/Tiled10x10Copy \
+-	$(top_srcdir)/raster/test/regress/loader/Tiled8x8 \
+-	$(top_srcdir)/raster/test/regress/loader/TiledAuto \
+-	$(top_srcdir)/raster/test/regress/loader/TiledAutoSkipNoData \
+-	$(top_srcdir)/raster/test/regress/loader/TiledAutoCopyn
++	$(top_srcdir)/raster/test/regress/loader/BasicFilename
+ 
+ RASTER_TESTS := $(RASTER_TEST_FIRST) \
+ 	$(RASTER_TEST_METADATA) $(RASTER_TEST_IO) $(RASTER_TEST_BASIC_FUNC) \
+diff --git a/regress/core/binary.sql b/regress/core/binary.sql
+index 7a36b65..ad78fc7 100644
+--- a/regress/core/binary.sql
++++ b/regress/core/binary.sql
+@@ -1,4 +1,5 @@
+ SET client_min_messages TO warning;
++
+ CREATE SCHEMA tm;
+ 
+ CREATE TABLE tm.geoms (id serial, g geometry);
+@@ -31,24 +32,39 @@ SELECT st_force4d(g) FROM tm.geoms WHERE id < 15 ORDER BY id;
+ INSERT INTO tm.geoms(g)
+ SELECT st_setsrid(g,4326) FROM tm.geoms ORDER BY id;
+ 
+-COPY tm.geoms TO :tmpfile WITH BINARY;
++-- define temp file path
++\set tmpfile '/tmp/postgis_binary_test.dat'
++
++-- export
++\set command '\\copy tm.geoms TO ':tmpfile' WITH (FORMAT BINARY)'
++:command
++
++-- import
+ CREATE TABLE tm.geoms_in AS SELECT * FROM tm.geoms LIMIT 0;
+-COPY tm.geoms_in FROM :tmpfile WITH BINARY;
+-SELECT 'geometry', count(*) FROM tm.geoms_in i, tm.geoms o WHERE i.id = o.id
+- AND ST_OrderingEquals(i.g, o.g);
++\set command '\\copy tm.geoms_in FROM ':tmpfile' WITH (FORMAT BINARY)'
++:command
++
++SELECT 'geometry', count(*) FROM tm.geoms_in i, tm.geoms o
++WHERE i.id = o.id AND ST_OrderingEquals(i.g, o.g);
+ 
+ CREATE TABLE tm.geogs AS SELECT id,g::geography FROM tm.geoms
+ WHERE geometrytype(g) NOT LIKE '%CURVE%'
+   AND geometrytype(g) NOT LIKE '%CIRCULAR%'
+   AND geometrytype(g) NOT LIKE '%SURFACE%'
+   AND geometrytype(g) NOT LIKE 'TRIANGLE%'
+-  AND geometrytype(g) NOT LIKE 'TIN%'
+-;
++  AND geometrytype(g) NOT LIKE 'TIN%';
+ 
+-COPY tm.geogs TO :tmpfile WITH BINARY;
++-- export
++\set command '\\copy tm.geogs TO ':tmpfile' WITH (FORMAT BINARY)'
++:command
++
++-- import
+ CREATE TABLE tm.geogs_in AS SELECT * FROM tm.geogs LIMIT 0;
+-COPY tm.geogs_in FROM :tmpfile WITH BINARY;
+-SELECT 'geometry', count(*) FROM tm.geogs_in i, tm.geogs o WHERE i.id = o.id
+- AND ST_OrderingEquals(i.g::geometry, o.g::geometry);
++\set command '\\copy tm.geogs_in FROM ':tmpfile' WITH (FORMAT BINARY)'
++:command
++
++SELECT 'geometry', count(*) FROM tm.geogs_in i, tm.geogs o
++WHERE i.id = o.id AND ST_OrderingEquals(i.g::geometry, o.g::geometry);
+ 
+ DROP SCHEMA tm CASCADE;
++
+diff --git a/regress/core/tests.mk b/regress/core/tests.mk
+index 64a9254..94903c3 100644
+--- a/regress/core/tests.mk
++++ b/regress/core/tests.mk
+@@ -23,7 +23,6 @@ current_dir := $(dir $(abspath $(lastword $(MAKEFILE_LIST))))
+ RUNTESTFLAGS_INTERNAL += \
+   --before-upgrade-script $(top_srcdir)/regress/hooks/hook-before-upgrade.sql \
+   --after-upgrade-script  $(top_srcdir)/regress/hooks/hook-after-upgrade.sql \
+-  --after-create-script   $(top_srcdir)/regress/hooks/hook-after-create.sql \
+   --before-uninstall-script $(top_srcdir)/regress/hooks/hook-before-uninstall.sql
+ 
+ TESTS += \
+@@ -40,7 +39,6 @@ TESTS += \
+ 	$(top_srcdir)/regress/core/dumppoints \
+ 	$(top_srcdir)/regress/core/dumpsegments \
+ 	$(top_srcdir)/regress/core/empty \
+-	$(top_srcdir)/regress/core/estimatedextent \
+ 	$(top_srcdir)/regress/core/forcecurve \
+ 	$(top_srcdir)/regress/core/flatgeobuf \
+ 	$(top_srcdir)/regress/core/geography \
+@@ -55,7 +53,6 @@ TESTS += \
+ 	$(top_srcdir)/regress/core/out_marc21 \
+ 	$(top_srcdir)/regress/core/in_encodedpolyline \
+ 	$(top_srcdir)/regress/core/iscollection \
+-	$(top_srcdir)/regress/core/legacy \
+ 	$(top_srcdir)/regress/core/letters \
+ 	$(top_srcdir)/regress/core/long_xact \
+ 	$(top_srcdir)/regress/core/lwgeom_regress \
+@@ -112,7 +109,6 @@ TESTS += \
+ 	$(top_srcdir)/regress/core/temporal_knn \
+ 	$(top_srcdir)/regress/core/tickets \
+ 	$(top_srcdir)/regress/core/twkb \
+-	$(top_srcdir)/regress/core/typmod \
+ 	$(top_srcdir)/regress/core/wkb \
+ 	$(top_srcdir)/regress/core/wkt \
+ 	$(top_srcdir)/regress/core/wmsservers \
+diff --git a/regress/loader/tests.mk b/regress/loader/tests.mk
+index 1fc77ac..c3cb9de 100644
+--- a/regress/loader/tests.mk
++++ b/regress/loader/tests.mk
+@@ -38,7 +38,5 @@ TESTS += \
+ 	$(top_srcdir)/regress/loader/Latin1 \
+ 	$(top_srcdir)/regress/loader/Latin1-implicit \
+ 	$(top_srcdir)/regress/loader/mfile \
+-	$(top_srcdir)/regress/loader/TestSkipANALYZE \
+-	$(top_srcdir)/regress/loader/TestANALYZE \
+ 	$(top_srcdir)/regress/loader/CharNoWidth
+ 
+diff --git a/regress/run_test.pl b/regress/run_test.pl
+index 0ec5b2d..1c331f4 100755
+--- a/regress/run_test.pl
++++ b/regress/run_test.pl
+@@ -147,7 +147,6 @@ $ENV{"LANG"} = "C";
+ # Add locale info to the psql options
+ # Add pg12 precision suppression
+ my $PGOPTIONS = $ENV{"PGOPTIONS"};
+-$PGOPTIONS .= " -c lc_messages=C";
+ $PGOPTIONS .= " -c client_min_messages=NOTICE";
+ $PGOPTIONS .= " -c extra_float_digits=0";
+ $ENV{"PGOPTIONS"} = $PGOPTIONS;

docker-compose/ext-src/postgis-src/postgis-regular-v17.patch

@@ -0,0 +1,208 @@
+diff --git a/raster/test/regress/tests.mk b/raster/test/regress/tests.mk
+index 00918e1..7e2b6cd 100644
+--- a/raster/test/regress/tests.mk
++++ b/raster/test/regress/tests.mk
+@@ -17,9 +17,7 @@ override RUNTESTFLAGS_INTERNAL := \
+   $(RUNTESTFLAGS_INTERNAL) \
+   --after-upgrade-script $(top_srcdir)/raster/test/regress/hooks/hook-after-upgrade-raster.sql
+ 
+-RASTER_TEST_FIRST = \
+-	$(top_srcdir)/raster/test/regress/check_gdal \
+-	$(top_srcdir)/raster/test/regress/loader/load_outdb
++RASTER_TEST_FIRST =
+ 
+ RASTER_TEST_LAST = \
+ 	$(top_srcdir)/raster/test/regress/clean
+@@ -33,9 +31,7 @@ RASTER_TEST_IO = \
+ 
+ RASTER_TEST_BASIC_FUNC = \
+ 	$(top_srcdir)/raster/test/regress/rt_bytea \
+-	$(top_srcdir)/raster/test/regress/rt_wkb \
+ 	$(top_srcdir)/raster/test/regress/box3d \
+-	$(top_srcdir)/raster/test/regress/rt_addband \
+ 	$(top_srcdir)/raster/test/regress/rt_band \
+ 	$(top_srcdir)/raster/test/regress/rt_tile
+ 
+@@ -73,16 +69,10 @@ RASTER_TEST_BANDPROPS = \
+ 	$(top_srcdir)/raster/test/regress/rt_neighborhood \
+ 	$(top_srcdir)/raster/test/regress/rt_nearestvalue \
+ 	$(top_srcdir)/raster/test/regress/rt_pixelofvalue \
+-	$(top_srcdir)/raster/test/regress/rt_polygon \
+-	$(top_srcdir)/raster/test/regress/rt_setbandpath
++	$(top_srcdir)/raster/test/regress/rt_polygon
+ 
+ RASTER_TEST_UTILITY = \
+ 	$(top_srcdir)/raster/test/regress/rt_utility \
+-	$(top_srcdir)/raster/test/regress/rt_fromgdalraster \
+-	$(top_srcdir)/raster/test/regress/rt_asgdalraster \
+-	$(top_srcdir)/raster/test/regress/rt_astiff \
+-	$(top_srcdir)/raster/test/regress/rt_asjpeg \
+-	$(top_srcdir)/raster/test/regress/rt_aspng \
+ 	$(top_srcdir)/raster/test/regress/rt_reclass \
+ 	$(top_srcdir)/raster/test/regress/rt_gdalwarp \
+ 	$(top_srcdir)/raster/test/regress/rt_gdalcontour \
+@@ -120,21 +110,13 @@ RASTER_TEST_SREL = \
+ 
+ RASTER_TEST_BUGS = \
+ 	$(top_srcdir)/raster/test/regress/bug_test_car5 \
+-	$(top_srcdir)/raster/test/regress/permitted_gdal_drivers \
+ 	$(top_srcdir)/raster/test/regress/tickets
+ 
+ RASTER_TEST_LOADER = \
+ 	$(top_srcdir)/raster/test/regress/loader/Basic \
+ 	$(top_srcdir)/raster/test/regress/loader/Projected \
+ 	$(top_srcdir)/raster/test/regress/loader/BasicCopy \
+-	$(top_srcdir)/raster/test/regress/loader/BasicFilename \
+-	$(top_srcdir)/raster/test/regress/loader/BasicOutDB \
+-	$(top_srcdir)/raster/test/regress/loader/Tiled10x10 \
+-	$(top_srcdir)/raster/test/regress/loader/Tiled10x10Copy \
+-	$(top_srcdir)/raster/test/regress/loader/Tiled8x8 \
+-	$(top_srcdir)/raster/test/regress/loader/TiledAuto \
+-	$(top_srcdir)/raster/test/regress/loader/TiledAutoSkipNoData \
+-	$(top_srcdir)/raster/test/regress/loader/TiledAutoCopyn
++	$(top_srcdir)/raster/test/regress/loader/BasicFilename
+ 
+ RASTER_TESTS := $(RASTER_TEST_FIRST) \
+ 	$(RASTER_TEST_METADATA) $(RASTER_TEST_IO) $(RASTER_TEST_BASIC_FUNC) \
+diff --git a/regress/core/binary.sql b/regress/core/binary.sql
+index 7a36b65..ad78fc7 100644
+--- a/regress/core/binary.sql
++++ b/regress/core/binary.sql
+@@ -1,4 +1,5 @@
+ SET client_min_messages TO warning;
++
+ CREATE SCHEMA tm;
+ 
+ CREATE TABLE tm.geoms (id serial, g geometry);
+@@ -31,24 +32,39 @@ SELECT st_force4d(g) FROM tm.geoms WHERE id < 15 ORDER BY id;
+ INSERT INTO tm.geoms(g)
+ SELECT st_setsrid(g,4326) FROM tm.geoms ORDER BY id;
+ 
+-COPY tm.geoms TO :tmpfile WITH BINARY;
++-- define temp file path
++\set tmpfile '/tmp/postgis_binary_test.dat'
++
++-- export
++\set command '\\copy tm.geoms TO ':tmpfile' WITH (FORMAT BINARY)'
++:command
++
++-- import
+ CREATE TABLE tm.geoms_in AS SELECT * FROM tm.geoms LIMIT 0;
+-COPY tm.geoms_in FROM :tmpfile WITH BINARY;
+-SELECT 'geometry', count(*) FROM tm.geoms_in i, tm.geoms o WHERE i.id = o.id
+- AND ST_OrderingEquals(i.g, o.g);
++\set command '\\copy tm.geoms_in FROM ':tmpfile' WITH (FORMAT BINARY)'
++:command
++
++SELECT 'geometry', count(*) FROM tm.geoms_in i, tm.geoms o
++WHERE i.id = o.id AND ST_OrderingEquals(i.g, o.g);
+ 
+ CREATE TABLE tm.geogs AS SELECT id,g::geography FROM tm.geoms
+ WHERE geometrytype(g) NOT LIKE '%CURVE%'
+   AND geometrytype(g) NOT LIKE '%CIRCULAR%'
+   AND geometrytype(g) NOT LIKE '%SURFACE%'
+   AND geometrytype(g) NOT LIKE 'TRIANGLE%'
+-  AND geometrytype(g) NOT LIKE 'TIN%'
+-;
++  AND geometrytype(g) NOT LIKE 'TIN%';
+ 
+-COPY tm.geogs TO :tmpfile WITH BINARY;
++-- export
++\set command '\\copy tm.geogs TO ':tmpfile' WITH (FORMAT BINARY)'
++:command
++
++-- import
+ CREATE TABLE tm.geogs_in AS SELECT * FROM tm.geogs LIMIT 0;
+-COPY tm.geogs_in FROM :tmpfile WITH BINARY;
+-SELECT 'geometry', count(*) FROM tm.geogs_in i, tm.geogs o WHERE i.id = o.id
+- AND ST_OrderingEquals(i.g::geometry, o.g::geometry);
++\set command '\\copy tm.geogs_in FROM ':tmpfile' WITH (FORMAT BINARY)'
++:command
++
++SELECT 'geometry', count(*) FROM tm.geogs_in i, tm.geogs o
++WHERE i.id = o.id AND ST_OrderingEquals(i.g::geometry, o.g::geometry);
+ 
+ DROP SCHEMA tm CASCADE;
++
+diff --git a/regress/core/tests.mk b/regress/core/tests.mk
+index 90987df..74fe3f1 100644
+--- a/regress/core/tests.mk
++++ b/regress/core/tests.mk
+@@ -16,14 +16,13 @@ POSTGIS_PGSQL_VERSION=170
+ POSTGIS_GEOS_VERSION=31101
+ HAVE_JSON=yes
+ HAVE_SPGIST=yes
+-INTERRUPTTESTS=yes
++INTERRUPTTESTS=no
+ 
+ current_dir := $(dir $(abspath $(lastword $(MAKEFILE_LIST))))
+ 
+ RUNTESTFLAGS_INTERNAL += \
+   --before-upgrade-script $(top_srcdir)/regress/hooks/hook-before-upgrade.sql \
+   --after-upgrade-script  $(top_srcdir)/regress/hooks/hook-after-upgrade.sql \
+-  --after-create-script   $(top_srcdir)/regress/hooks/hook-after-create.sql \
+   --before-uninstall-script $(top_srcdir)/regress/hooks/hook-before-uninstall.sql
+ 
+ TESTS += \
+@@ -40,7 +39,6 @@ TESTS += \
+ 	$(top_srcdir)/regress/core/dumppoints \
+ 	$(top_srcdir)/regress/core/dumpsegments \
+ 	$(top_srcdir)/regress/core/empty \
+-	$(top_srcdir)/regress/core/estimatedextent \
+ 	$(top_srcdir)/regress/core/forcecurve \
+ 	$(top_srcdir)/regress/core/flatgeobuf \
+ 	$(top_srcdir)/regress/core/frechet \
+@@ -60,7 +58,6 @@ TESTS += \
+ 	$(top_srcdir)/regress/core/out_marc21 \
+ 	$(top_srcdir)/regress/core/in_encodedpolyline \
+ 	$(top_srcdir)/regress/core/iscollection \
+-	$(top_srcdir)/regress/core/legacy \
+ 	$(top_srcdir)/regress/core/letters \
+ 	$(top_srcdir)/regress/core/lwgeom_regress \
+ 	$(top_srcdir)/regress/core/measures \
+@@ -119,7 +116,6 @@ TESTS += \
+ 	$(top_srcdir)/regress/core/temporal_knn \
+ 	$(top_srcdir)/regress/core/tickets \
+ 	$(top_srcdir)/regress/core/twkb \
+-	$(top_srcdir)/regress/core/typmod \
+ 	$(top_srcdir)/regress/core/wkb \
+ 	$(top_srcdir)/regress/core/wkt \
+ 	$(top_srcdir)/regress/core/wmsservers \
+diff --git a/regress/loader/tests.mk b/regress/loader/tests.mk
+index ac4f8ad..4bad4fc 100644
+--- a/regress/loader/tests.mk
++++ b/regress/loader/tests.mk
+@@ -38,7 +38,5 @@ TESTS += \
+ 	$(top_srcdir)/regress/loader/Latin1 \
+ 	$(top_srcdir)/regress/loader/Latin1-implicit \
+ 	$(top_srcdir)/regress/loader/mfile \
+-	$(top_srcdir)/regress/loader/TestSkipANALYZE \
+-	$(top_srcdir)/regress/loader/TestANALYZE \
+ 	$(top_srcdir)/regress/loader/CharNoWidth \
+ 
+diff --git a/regress/run_test.pl b/regress/run_test.pl
+index cac4b2e..4c7c82b 100755
+--- a/regress/run_test.pl
++++ b/regress/run_test.pl
+@@ -238,7 +238,6 @@ $ENV{"LANG"} = "C";
+ # Add locale info to the psql options
+ # Add pg12 precision suppression
+ my $PGOPTIONS = $ENV{"PGOPTIONS"};
+-$PGOPTIONS .= " -c lc_messages=C";
+ $PGOPTIONS .= " -c client_min_messages=NOTICE";
+ $PGOPTIONS .= " -c extra_float_digits=0";
+ $ENV{"PGOPTIONS"} = $PGOPTIONS;
+diff --git a/topology/test/tests.mk b/topology/test/tests.mk
+index cbe2633..2c7c18f 100644
+--- a/topology/test/tests.mk
++++ b/topology/test/tests.mk
+@@ -46,9 +46,7 @@ TESTS += \
+ 	$(top_srcdir)/topology/test/regress/legacy_query.sql \
+ 	$(top_srcdir)/topology/test/regress/legacy_validate.sql \
+ 	$(top_srcdir)/topology/test/regress/polygonize.sql \
+-	$(top_srcdir)/topology/test/regress/populate_topology_layer.sql \
+ 	$(top_srcdir)/topology/test/regress/removeunusedprimitives.sql \
+-	$(top_srcdir)/topology/test/regress/renametopogeometrycolumn.sql \
+ 	$(top_srcdir)/topology/test/regress/renametopology.sql \
+ 	$(top_srcdir)/topology/test/regress/share_sequences.sql \
+ 	$(top_srcdir)/topology/test/regress/sqlmm.sql \

docker-compose/ext-src/postgis-src/regular-test.sh

@@ -0,0 +1,17 @@
+#!/bin/bash
+set -ex
+cd "$(dirname "${0}")"
+dropdb --if-exist contrib_regression
+createdb contrib_regression
+psql -d contrib_regression -c "ALTER DATABASE contrib_regression SET TimeZone='UTC'" \
+     -c "ALTER DATABASE contrib_regression SET DateStyle='ISO, MDY'" \
+     -c "CREATE EXTENSION postgis SCHEMA public" \
+     -c "CREATE EXTENSION postgis_topology" \
+     -c "CREATE EXTENSION postgis_tiger_geocoder CASCADE" \
+     -c "CREATE EXTENSION postgis_raster SCHEMA public" \
+     -c "CREATE EXTENSION postgis_sfcgal SCHEMA public"
+patch -p1 <"postgis-common-${PG_VERSION}.patch"
+patch -p1 <"postgis-regular-${PG_VERSION}.patch"
+psql -d contrib_regression -f raster_outdb_template.sql
+trap 'patch -R -p1 <postgis-regular-${PG_VERSION}.patch && patch -R -p1 <"postgis-common-${PG_VERSION}.patch"' EXIT
+POSTGIS_REGRESS_DB=contrib_regression RUNTESTFLAGS=--nocreate make installcheck-base

docker-compose/run-tests.sh

@@ -63,5 +63,9 @@ done
 for d in ${FAILED}; do
   cat "$(find $d -name regression.diffs)"
 done
+for postgis_diff in /tmp/pgis_reg/*_diff; do
+  echo "${postgis_diff}:"
+  cat "${postgis_diff}"
+done
 echo "${FAILED}"
 exit 1

endpoint_storage/Cargo.toml

@@ -8,6 +8,7 @@ anyhow.workspace = true
 axum-extra.workspace = true
 axum.workspace = true
 camino.workspace = true
+clap.workspace = true
 futures.workspace = true
 jsonwebtoken.workspace = true
 prometheus.workspace = true

endpoint_storage/src/main.rs

@@ -4,6 +4,8 @@
 //! for large computes.
 mod app;
 use anyhow::Context;
+use clap::Parser;
+use std::net::{IpAddr, Ipv4Addr, SocketAddr};
 use tracing::info;
 use utils::logging;
 
@@ -12,9 +14,26 @@ const fn max_upload_file_limit() -> usize {
     100 * 1024 * 1024
 }
 
+const fn listen() -> SocketAddr {
+    SocketAddr::new(IpAddr::V4(Ipv4Addr::new(0, 0, 0, 0)), 51243)
+}
+
+#[derive(Parser)]
+struct Args {
+    #[arg(exclusive = true)]
+    config_file: Option<String>,
+    #[arg(long, default_value = "false", requires = "config")]
+    /// to allow testing k8s helm chart where we don't have s3 credentials
+    no_s3_check_on_startup: bool,
+    #[arg(long, value_name = "FILE")]
+    /// inline config mode for k8s helm chart
+    config: Option<String>,
+}
+
 #[derive(serde::Deserialize)]
 #[serde(tag = "type")]
 struct Config {
+    #[serde(default = "listen")]
     listen: std::net::SocketAddr,
     pemfile: camino::Utf8PathBuf,
     #[serde(flatten)]
@@ -31,13 +50,18 @@ async fn main() -> anyhow::Result<()> {
         logging::Output::Stdout,
     )?;
 
-    let config: String = std::env::args().skip(1).take(1).collect();
-    if config.is_empty() {
-        anyhow::bail!("Usage: endpoint_storage config.json")
-    }
-    info!("Reading config from {config}");
-    let config = std::fs::read_to_string(config.clone())?;
-    let config: Config = serde_json::from_str(&config).context("parsing config")?;
+    let args = Args::parse();
+    let config: Config = if let Some(config_path) = args.config_file {
+        info!("Reading config from {config_path}");
+        let config = std::fs::read_to_string(config_path)?;
+        serde_json::from_str(&config).context("parsing config")?
+    } else if let Some(config) = args.config {
+        info!("Reading inline config");
+        serde_json::from_str(&config).context("parsing config")?
+    } else {
+        anyhow::bail!("Supply either config file path or --config=inline-config");
+    };
+
     info!("Reading pemfile from {}", config.pemfile.clone());
     let pemfile = std::fs::read(config.pemfile.clone())?;
     info!("Loading public key from {}", config.pemfile.clone());
@@ -48,7 +72,9 @@ async fn main() -> anyhow::Result<()> {
 
     let storage = remote_storage::GenericRemoteStorage::from_config(&config.storage_config).await?;
     let cancel = tokio_util::sync::CancellationToken::new();
-    app::check_storage_permissions(&storage, cancel.clone()).await?;
+    if !args.no_s3_check_on_startup {
+        app::check_storage_permissions(&storage, cancel.clone()).await?;
+    }
 
     let proxy = std::sync::Arc::new(endpoint_storage::Storage {
         auth,

libs/compute_api/src/spec.rs

@@ -178,9 +178,9 @@ pub struct ComputeSpec {
     /// JWT for authorizing requests to endpoint storage service
     pub endpoint_storage_token: Option<String>,
 
-    /// If true, download LFC state from endpoint_storage and pass it to Postgres on startup
+    /// Download LFC state from endpoint_storage and pass it to Postgres on startup
     #[serde(default)]
-    pub prewarm_lfc_on_startup: bool,
+    pub autoprewarm: bool,
 }
 
 /// Feature flag to signal `compute_ctl` to enable certain experimental functionality.
@@ -192,6 +192,9 @@ pub enum ComputeFeature {
     /// track short-lived connections as user activity.
     ActivityMonitorExperimental,
 
+    /// Enable TLS functionality.
+    TlsExperimental,
+
     /// This is a special feature flag that is used to represent unknown feature flags.
     /// Basically all unknown to enum flags are represented as this one. See unit test
     /// `parse_unknown_features()` for more details.

libs/compute_api/tests/cluster_spec.json

@@ -85,7 +85,7 @@
                 "vartype": "bool"
             },
             {
-                "name": "prewarm_lfc_on_startup",
+                "name": "autoprewarm",
                 "value": "off",
                 "vartype": "bool"
             },

libs/metrics/src/hll.rs

@@ -107,7 +107,7 @@ impl<const N: usize> MetricType for HyperLogLogState<N> {
 }
 
 impl<const N: usize> HyperLogLogState<N> {
-    pub fn measure(&self, item: &impl Hash) {
+    pub fn measure(&self, item: &(impl Hash + ?Sized)) {
         // changing the hasher will break compatibility with previous measurements.
         self.record(BuildHasherDefault::<xxh3::Hash64>::default().hash_one(item));
     }

libs/metrics/src/lib.rs

@@ -27,6 +27,7 @@ pub use prometheus::{
 
 pub mod launch_timestamp;
 mod wrappers;
+pub use prometheus;
 pub use wrappers::{CountedReader, CountedWriter};
 mod hll;
 pub use hll::{HyperLogLog, HyperLogLogState, HyperLogLogVec};

libs/pageserver_api/src/config.rs

@@ -20,7 +20,6 @@ use postgres_backend::AuthType;
 use remote_storage::RemoteStorageConfig;
 use serde_with::serde_as;
 use utils::logging::LogFormat;
-use utils::postgres_client::PostgresClientProtocol;
 
 use crate::models::{ImageCompressionAlgorithm, LsnLease};
 
@@ -181,6 +180,7 @@ pub struct ConfigToml {
     pub virtual_file_io_engine: Option<crate::models::virtual_file::IoEngineKind>,
     pub ingest_batch_size: u64,
     pub max_vectored_read_bytes: MaxVectoredReadBytes,
+    pub max_get_vectored_keys: MaxGetVectoredKeys,
     pub image_compression: ImageCompressionAlgorithm,
     pub timeline_offloading: bool,
     pub ephemeral_bytes_per_memory_kb: usize,
@@ -188,7 +188,6 @@ pub struct ConfigToml {
     pub virtual_file_io_mode: Option<crate::models::virtual_file::IoMode>,
     #[serde(skip_serializing_if = "Option::is_none")]
     pub no_sync: Option<bool>,
-    pub wal_receiver_protocol: PostgresClientProtocol,
     pub page_service_pipelining: PageServicePipeliningConfig,
     pub get_vectored_concurrent_io: GetVectoredConcurrentIo,
     pub enable_read_path_debugging: Option<bool>,
@@ -229,7 +228,7 @@ pub enum PageServicePipeliningConfig {
 }
 #[derive(Debug, Clone, PartialEq, Eq, serde::Serialize, serde::Deserialize)]
 pub struct PageServicePipeliningConfigPipelined {
-    /// Causes runtime errors if larger than max get_vectored batch size.
+    /// Failed config parsing and validation if larger than `max_get_vectored_keys`.
     pub max_batch_size: NonZeroUsize,
     pub execution: PageServiceProtocolPipelinedExecutionStrategy,
     // The default below is such that new versions of the software can start
@@ -329,6 +328,8 @@ pub struct TimelineImportConfig {
     pub import_job_concurrency: NonZeroUsize,
     pub import_job_soft_size_limit: NonZeroUsize,
     pub import_job_checkpoint_threshold: NonZeroUsize,
+    /// Max size of the remote storage partial read done by any job
+    pub import_job_max_byte_range_size: NonZeroUsize,
 }
 
 #[derive(Debug, Clone, PartialEq, Eq, serde::Serialize, serde::Deserialize)]
@@ -403,6 +404,16 @@ impl Default for EvictionOrder {
 #[serde(transparent)]
 pub struct MaxVectoredReadBytes(pub NonZeroUsize);
 
+#[derive(Copy, Clone, Debug, PartialEq, Eq, serde::Serialize, serde::Deserialize)]
+#[serde(transparent)]
+pub struct MaxGetVectoredKeys(NonZeroUsize);
+
+impl MaxGetVectoredKeys {
+    pub fn get(&self) -> usize {
+        self.0.get()
+    }
+}
+
 /// Tenant-level configuration values, used for various purposes.
 #[derive(Debug, Clone, PartialEq, Eq, serde::Serialize, serde::Deserialize)]
 #[serde(default)]
@@ -514,8 +525,6 @@ pub struct TenantConfigToml {
     /// (either this flag or the pageserver-global one need to be set)
     pub timeline_offloading: bool,
 
-    pub wal_receiver_protocol_override: Option<PostgresClientProtocol>,
-
     /// Enable rel_size_v2 for this tenant. Once enabled, the tenant will persist this information into
     /// `index_part.json`, and it cannot be reversed.
     pub rel_size_v2_enabled: bool,
@@ -587,6 +596,8 @@ pub mod defaults {
     /// That is, slightly above 128 kB.
     pub const DEFAULT_MAX_VECTORED_READ_BYTES: usize = 130 * 1024; // 130 KiB
 
+    pub const DEFAULT_MAX_GET_VECTORED_KEYS: usize = 32;
+
     pub const DEFAULT_IMAGE_COMPRESSION: ImageCompressionAlgorithm =
         ImageCompressionAlgorithm::Zstd { level: Some(1) };
 
@@ -594,9 +605,6 @@ pub mod defaults {
 
     pub const DEFAULT_IO_BUFFER_ALIGNMENT: usize = 512;
 
-    pub const DEFAULT_WAL_RECEIVER_PROTOCOL: utils::postgres_client::PostgresClientProtocol =
-        utils::postgres_client::PostgresClientProtocol::Vanilla;
-
     pub const DEFAULT_SSL_KEY_FILE: &str = "server.key";
     pub const DEFAULT_SSL_CERT_FILE: &str = "server.crt";
 }
@@ -685,6 +693,9 @@ impl Default for ConfigToml {
             max_vectored_read_bytes: (MaxVectoredReadBytes(
                 NonZeroUsize::new(DEFAULT_MAX_VECTORED_READ_BYTES).unwrap(),
             )),
+            max_get_vectored_keys: (MaxGetVectoredKeys(
+                NonZeroUsize::new(DEFAULT_MAX_GET_VECTORED_KEYS).unwrap(),
+            )),
             image_compression: (DEFAULT_IMAGE_COMPRESSION),
             timeline_offloading: true,
             ephemeral_bytes_per_memory_kb: (DEFAULT_EPHEMERAL_BYTES_PER_MEMORY_KB),
@@ -692,7 +703,6 @@ impl Default for ConfigToml {
             virtual_file_io_mode: None,
             tenant_config: TenantConfigToml::default(),
             no_sync: None,
-            wal_receiver_protocol: DEFAULT_WAL_RECEIVER_PROTOCOL,
             page_service_pipelining: PageServicePipeliningConfig::Pipelined(
                 PageServicePipeliningConfigPipelined {
                     max_batch_size: NonZeroUsize::new(32).unwrap(),
@@ -713,9 +723,10 @@ impl Default for ConfigToml {
             enable_tls_page_service_api: false,
             dev_mode: false,
             timeline_import_config: TimelineImportConfig {
-                import_job_concurrency: NonZeroUsize::new(128).unwrap(),
-                import_job_soft_size_limit: NonZeroUsize::new(1024 * 1024 * 1024).unwrap(),
-                import_job_checkpoint_threshold: NonZeroUsize::new(128).unwrap(),
+                import_job_concurrency: NonZeroUsize::new(32).unwrap(),
+                import_job_soft_size_limit: NonZeroUsize::new(256 * 1024 * 1024).unwrap(),
+                import_job_checkpoint_threshold: NonZeroUsize::new(32).unwrap(),
+                import_job_max_byte_range_size: NonZeroUsize::new(4 * 1024 * 1024).unwrap(),
             },
             basebackup_cache_config: None,
             posthog_config: None,
@@ -836,7 +847,6 @@ impl Default for TenantConfigToml {
             lsn_lease_length: LsnLease::DEFAULT_LENGTH,
             lsn_lease_length_for_ts: LsnLease::DEFAULT_LENGTH_FOR_TS,
             timeline_offloading: true,
-            wal_receiver_protocol_override: None,
             rel_size_v2_enabled: false,
             gc_compaction_enabled: DEFAULT_GC_COMPACTION_ENABLED,
             gc_compaction_verification: DEFAULT_GC_COMPACTION_VERIFICATION,

libs/pageserver_api/src/controller_api.rs

@@ -344,6 +344,35 @@ impl Default for ShardSchedulingPolicy {
     }
 }
 
+#[derive(Serialize, Deserialize, Clone, Copy, Eq, PartialEq, Debug)]
+pub enum NodeLifecycle {
+    Active,
+    Deleted,
+}
+
+impl FromStr for NodeLifecycle {
+    type Err = anyhow::Error;
+
+    fn from_str(s: &str) -> Result<Self, Self::Err> {
+        match s {
+            "active" => Ok(Self::Active),
+            "deleted" => Ok(Self::Deleted),
+            _ => Err(anyhow::anyhow!("Unknown node lifecycle '{s}'")),
+        }
+    }
+}
+
+impl From<NodeLifecycle> for String {
+    fn from(value: NodeLifecycle) -> String {
+        use NodeLifecycle::*;
+        match value {
+            Active => "active",
+            Deleted => "deleted",
+        }
+        .to_string()
+    }
+}
+
 #[derive(Serialize, Deserialize, Clone, Copy, Eq, PartialEq, Debug)]
 pub enum NodeSchedulingPolicy {
     Active,

libs/pageserver_api/src/models.rs

@@ -20,7 +20,6 @@ use serde_with::serde_as;
 pub use utilization::PageserverUtilization;
 use utils::id::{NodeId, TenantId, TimelineId};
 use utils::lsn::Lsn;
-use utils::postgres_client::PostgresClientProtocol;
 use utils::{completion, serde_system_time};
 
 use crate::config::Ratio;
@@ -622,8 +621,6 @@ pub struct TenantConfigPatch {
     #[serde(skip_serializing_if = "FieldPatch::is_noop")]
     pub timeline_offloading: FieldPatch<bool>,
     #[serde(skip_serializing_if = "FieldPatch::is_noop")]
-    pub wal_receiver_protocol_override: FieldPatch<PostgresClientProtocol>,
-    #[serde(skip_serializing_if = "FieldPatch::is_noop")]
     pub rel_size_v2_enabled: FieldPatch<bool>,
     #[serde(skip_serializing_if = "FieldPatch::is_noop")]
     pub gc_compaction_enabled: FieldPatch<bool>,
@@ -748,9 +745,6 @@ pub struct TenantConfig {
     #[serde(skip_serializing_if = "Option::is_none")]
     pub timeline_offloading: Option<bool>,
 
-    #[serde(skip_serializing_if = "Option::is_none")]
-    pub wal_receiver_protocol_override: Option<PostgresClientProtocol>,
-
     #[serde(skip_serializing_if = "Option::is_none")]
     pub rel_size_v2_enabled: Option<bool>,
 
@@ -812,7 +806,6 @@ impl TenantConfig {
             mut lsn_lease_length,
             mut lsn_lease_length_for_ts,
             mut timeline_offloading,
-            mut wal_receiver_protocol_override,
             mut rel_size_v2_enabled,
             mut gc_compaction_enabled,
             mut gc_compaction_verification,
@@ -905,9 +898,6 @@ impl TenantConfig {
             .map(|v| humantime::parse_duration(&v))?
             .apply(&mut lsn_lease_length_for_ts);
         patch.timeline_offloading.apply(&mut timeline_offloading);
-        patch
-            .wal_receiver_protocol_override
-            .apply(&mut wal_receiver_protocol_override);
         patch.rel_size_v2_enabled.apply(&mut rel_size_v2_enabled);
         patch
             .gc_compaction_enabled
@@ -960,7 +950,6 @@ impl TenantConfig {
             lsn_lease_length,
             lsn_lease_length_for_ts,
             timeline_offloading,
-            wal_receiver_protocol_override,
             rel_size_v2_enabled,
             gc_compaction_enabled,
             gc_compaction_verification,
@@ -1058,9 +1047,6 @@ impl TenantConfig {
             timeline_offloading: self
                 .timeline_offloading
                 .unwrap_or(global_conf.timeline_offloading),
-            wal_receiver_protocol_override: self
-                .wal_receiver_protocol_override
-                .or(global_conf.wal_receiver_protocol_override),
             rel_size_v2_enabled: self
                 .rel_size_v2_enabled
                 .unwrap_or(global_conf.rel_size_v2_enabled),
@@ -1934,7 +1920,7 @@ pub enum PagestreamFeMessage {
 }
 
 // Wrapped in libpq CopyData
-#[derive(strum_macros::EnumProperty)]
+#[derive(Debug, strum_macros::EnumProperty)]
 pub enum PagestreamBeMessage {
     Exists(PagestreamExistsResponse),
     Nblocks(PagestreamNblocksResponse),
@@ -2045,7 +2031,7 @@ pub enum PagestreamProtocolVersion {
 
 pub type RequestId = u64;
 
-#[derive(Debug, PartialEq, Eq, Clone, Copy)]
+#[derive(Debug, Default, PartialEq, Eq, Clone, Copy)]
 pub struct PagestreamRequest {
     pub reqid: RequestId,
     pub request_lsn: Lsn,
@@ -2064,7 +2050,7 @@ pub struct PagestreamNblocksRequest {
     pub rel: RelTag,
 }
 
-#[derive(Debug, PartialEq, Eq, Clone, Copy)]
+#[derive(Debug, Default, PartialEq, Eq, Clone, Copy)]
 pub struct PagestreamGetPageRequest {
     pub hdr: PagestreamRequest,
     pub rel: RelTag,

libs/pageserver_api/src/reltag.rs

@@ -24,7 +24,7 @@ use serde::{Deserialize, Serialize};
 // FIXME: should move 'forknum' as last field to keep this consistent with Postgres.
 // 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, Deserialize)]
+#[derive(Debug, Default, PartialEq, Eq, Hash, Clone, Copy, Serialize, Deserialize)]
 pub struct RelTag {
     pub forknum: u8,
     pub spcnode: Oid,
@@ -184,12 +184,12 @@ pub enum SlruKind {
     MultiXactOffsets,
 }
 
-impl SlruKind {
-    pub fn to_str(&self) -> &'static str {
+impl fmt::Display for SlruKind {
+    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
         match self {
-            Self::Clog => "pg_xact",
-            Self::MultiXactMembers => "pg_multixact/members",
-            Self::MultiXactOffsets => "pg_multixact/offsets",
+            Self::Clog => write!(f, "pg_xact"),
+            Self::MultiXactMembers => write!(f, "pg_multixact/members"),
+            Self::MultiXactOffsets => write!(f, "pg_multixact/offsets"),
         }
     }
 }

libs/posthog_client_lite/src/background_loop.rs

@@ -4,8 +4,9 @@ use std::{sync::Arc, time::Duration};
 
 use arc_swap::ArcSwap;
 use tokio_util::sync::CancellationToken;
+use tracing::{Instrument, info_span};
 
-use crate::{FeatureStore, PostHogClient, PostHogClientConfig};
+use crate::{CaptureEvent, FeatureStore, PostHogClient, PostHogClientConfig};
 
 /// A background loop that fetches feature flags from PostHog and updates the feature store.
 pub struct FeatureResolverBackgroundLoop {
@@ -23,34 +24,61 @@ impl FeatureResolverBackgroundLoop {
         }
     }
 
-    pub fn spawn(self: Arc<Self>, handle: &tokio::runtime::Handle, refresh_period: Duration) {
+    pub fn spawn(
+        self: Arc<Self>,
+        handle: &tokio::runtime::Handle,
+        refresh_period: Duration,
+        fake_tenants: Vec<CaptureEvent>,
+    ) {
         let this = self.clone();
         let cancel = self.cancel.clone();
-        handle.spawn(async move {
-            tracing::info!("Starting PostHog feature resolver");
-            let mut ticker = tokio::time::interval(refresh_period);
-            ticker.set_missed_tick_behavior(tokio::time::MissedTickBehavior::Skip);
-            loop {
-                tokio::select! {
-                    _ = ticker.tick() => {}
-                    _ = cancel.cancelled() => break
-                }
-                let resp = match this
-                    .posthog_client
-                    .get_feature_flags_local_evaluation()
-                    .await
-                {
-                    Ok(resp) => resp,
-                    Err(e) => {
-                        tracing::warn!("Cannot get feature flags: {}", e);
-                        continue;
+
+        // Main loop of updating the feature flags.
+        handle.spawn(
+            async move {
+                tracing::info!("Starting PostHog feature resolver");
+                let mut ticker = tokio::time::interval(refresh_period);
+                ticker.set_missed_tick_behavior(tokio::time::MissedTickBehavior::Skip);
+                loop {
+                    tokio::select! {
+                        _ = ticker.tick() => {}
+                        _ = cancel.cancelled() => break
                     }
-                };
-                let feature_store = FeatureStore::new_with_flags(resp.flags);
-                this.feature_store.store(Arc::new(feature_store));
+                    let resp = match this
+                        .posthog_client
+                        .get_feature_flags_local_evaluation()
+                        .await
+                    {
+                        Ok(resp) => resp,
+                        Err(e) => {
+                            tracing::warn!("Cannot get feature flags: {}", e);
+                            continue;
+                        }
+                    };
+                    let feature_store = FeatureStore::new_with_flags(resp.flags);
+                    this.feature_store.store(Arc::new(feature_store));
+                    tracing::info!("Feature flag updated");
+                }
+                tracing::info!("PostHog feature resolver stopped");
             }
-            tracing::info!("PostHog feature resolver stopped");
-        });
+            .instrument(info_span!("posthog_feature_resolver")),
+        );
+
+        // Report fake tenants to PostHog so that we have the combination of all the properties in the UI.
+        // Do one report per pageserver restart.
+        let this = self.clone();
+        handle.spawn(
+            async move {
+                tracing::info!("Starting PostHog feature reporter");
+                for tenant in &fake_tenants {
+                    tracing::info!("Reporting fake tenant: {:?}", tenant);
+                }
+                if let Err(e) = this.posthog_client.capture_event_batch(&fake_tenants).await {
+                    tracing::warn!("Cannot report fake tenants: {}", e);
+                }
+            }
+            .instrument(info_span!("posthog_feature_reporter")),
+        );
     }
 
     pub fn feature_store(&self) -> Arc<FeatureStore> {

libs/posthog_client_lite/src/lib.rs

@@ -22,6 +22,16 @@ pub enum PostHogEvaluationError {
     Internal(String),
 }
 
+impl PostHogEvaluationError {
+    pub fn as_variant_str(&self) -> &'static str {
+        match self {
+            PostHogEvaluationError::NotAvailable(_) => "not_available",
+            PostHogEvaluationError::NoConditionGroupMatched => "no_condition_group_matched",
+            PostHogEvaluationError::Internal(_) => "internal",
+        }
+    }
+}
+
 #[derive(Deserialize)]
 pub struct LocalEvaluationResponse {
     pub flags: Vec<LocalEvaluationFlag>,
@@ -54,7 +64,7 @@ pub struct LocalEvaluationFlagFilterProperty {
     operator: String,
 }
 
-#[derive(Debug, Serialize, Deserialize)]
+#[derive(Debug, Serialize, Deserialize, Clone)]
 #[serde(untagged)]
 pub enum PostHogFlagFilterPropertyValue {
     String(String),
@@ -448,6 +458,18 @@ impl FeatureStore {
             )))
         }
     }
+
+    /// Infer whether a feature flag is a boolean flag by checking if it has a multivariate filter.
+    pub fn is_feature_flag_boolean(&self, flag_key: &str) -> Result<bool, PostHogEvaluationError> {
+        if let Some(flag_config) = self.flags.get(flag_key) {
+            Ok(flag_config.filters.multivariate.is_none())
+        } else {
+            Err(PostHogEvaluationError::NotAvailable(format!(
+                "Not found in the local evaluation spec: {}",
+                flag_key
+            )))
+        }
+    }
 }
 
 pub struct PostHogClientConfig {
@@ -485,6 +507,13 @@ pub struct PostHogClient {
     client: reqwest::Client,
 }
 
+#[derive(Serialize, Debug)]
+pub struct CaptureEvent {
+    pub event: String,
+    pub distinct_id: String,
+    pub properties: serde_json::Value,
+}
+
 impl PostHogClient {
     pub fn new(config: PostHogClientConfig) -> Self {
         let client = reqwest::Client::new();
@@ -528,7 +557,15 @@ impl PostHogClient {
             .bearer_auth(&self.config.server_api_key)
             .send()
             .await?;
+        let status = response.status();
         let body = response.text().await?;
+        if !status.is_success() {
+            return Err(anyhow::anyhow!(
+                "Failed to get feature flags: {}, {}",
+                status,
+                body
+            ));
+        }
         Ok(serde_json::from_str(&body)?)
     }
 
@@ -540,12 +577,12 @@ impl PostHogClient {
         &self,
         event: &str,
         distinct_id: &str,
-        properties: &HashMap<String, PostHogFlagFilterPropertyValue>,
+        properties: &serde_json::Value,
     ) -> anyhow::Result<()> {
         // PUBLIC_URL/capture/
-        // with bearer token of self.client_api_key
         let url = format!("{}/capture/", self.config.public_api_url);
-        self.client
+        let response = self
+            .client
             .post(url)
             .body(serde_json::to_string(&json!({
                 "api_key": self.config.client_api_key,
@@ -555,6 +592,39 @@ impl PostHogClient {
             }))?)
             .send()
             .await?;
+        let status = response.status();
+        let body = response.text().await?;
+        if !status.is_success() {
+            return Err(anyhow::anyhow!(
+                "Failed to capture events: {}, {}",
+                status,
+                body
+            ));
+        }
+        Ok(())
+    }
+
+    pub async fn capture_event_batch(&self, events: &[CaptureEvent]) -> anyhow::Result<()> {
+        // PUBLIC_URL/batch/
+        let url = format!("{}/batch/", self.config.public_api_url);
+        let response = self
+            .client
+            .post(url)
+            .body(serde_json::to_string(&json!({
+                "api_key": self.config.client_api_key,
+                "batch": events,
+            }))?)
+            .send()
+            .await?;
+        let status = response.status();
+        let body = response.text().await?;
+        if !status.is_success() {
+            return Err(anyhow::anyhow!(
+                "Failed to capture events: {}, {}",
+                status,
+                body
+            ));
+        }
         Ok(())
     }
 }

libs/proxy/tokio-postgres2/src/cancel_query.rs

@@ -10,7 +10,7 @@ use crate::{Error, cancel_query_raw, connect_socket};
 pub(crate) async fn cancel_query<T>(
     config: Option<SocketConfig>,
     ssl_mode: SslMode,
-    mut tls: T,
+    tls: T,
     process_id: i32,
     secret_key: i32,
 ) -> Result<(), Error>

libs/proxy/tokio-postgres2/src/config.rs

@@ -17,7 +17,6 @@ use crate::{Client, Connection, Error};
 
 /// TLS configuration.
 #[derive(Debug, Copy, Clone, PartialEq, Eq, Serialize, Deserialize)]
-#[non_exhaustive]
 pub enum SslMode {
     /// Do not use TLS.
     Disable,
@@ -231,7 +230,7 @@ impl Config {
     /// Requires the `runtime` Cargo feature (enabled by default).
     pub async fn connect<T>(
         &self,
-        tls: T,
+        tls: &T,
     ) -> Result<(Client, Connection<TcpStream, T::Stream>), Error>
     where
         T: MakeTlsConnect<TcpStream>,

libs/proxy/tokio-postgres2/src/connect.rs

@@ -13,7 +13,7 @@ use crate::tls::{MakeTlsConnect, TlsConnect};
 use crate::{Client, Config, Connection, Error, RawConnection};
 
 pub async fn connect<T>(
-    mut tls: T,
+    tls: &T,
     config: &Config,
 ) -> Result<(Client, Connection<TcpStream, T::Stream>), Error>
 where

libs/proxy/tokio-postgres2/src/tls.rs

@@ -47,7 +47,7 @@ pub trait MakeTlsConnect<S> {
     /// Creates a new `TlsConnect`or.
     ///
     /// The domain name is provided for certificate verification and SNI.
-    fn make_tls_connect(&mut self, domain: &str) -> Result<Self::TlsConnect, Self::Error>;
+    fn make_tls_connect(&self, domain: &str) -> Result<Self::TlsConnect, Self::Error>;
 }
 
 /// An asynchronous function wrapping a stream in a TLS session.
@@ -85,7 +85,7 @@ impl<S> MakeTlsConnect<S> for NoTls {
     type TlsConnect = NoTls;
     type Error = NoTlsError;
 
-    fn make_tls_connect(&mut self, _: &str) -> Result<NoTls, NoTlsError> {
+    fn make_tls_connect(&self, _: &str) -> Result<NoTls, NoTlsError> {
         Ok(NoTls)
     }
 }

libs/remote_storage/src/azure_blob.rs

@@ -10,7 +10,7 @@ use std::sync::Arc;
 use std::time::{Duration, SystemTime};
 use std::{env, io};
 
-use anyhow::{Context, Result};
+use anyhow::{Context, Result, anyhow};
 use azure_core::request_options::{IfMatchCondition, MaxResults, Metadata, Range};
 use azure_core::{Continuable, HttpClient, RetryOptions, TransportOptions};
 use azure_storage::StorageCredentials;
@@ -37,6 +37,7 @@ use crate::metrics::{AttemptOutcome, RequestKind, start_measuring_requests};
 use crate::{
     ConcurrencyLimiter, Download, DownloadError, DownloadKind, DownloadOpts, Listing, ListingMode,
     ListingObject, RemotePath, RemoteStorage, StorageMetadata, TimeTravelError, TimeoutOrCancel,
+    Version, VersionKind,
 };
 
 pub struct AzureBlobStorage {
@@ -405,6 +406,39 @@ impl AzureBlobStorage {
     pub fn container_name(&self) -> &str {
         &self.container_name
     }
+
+    async fn list_versions_with_permit(
+        &self,
+        _permit: &tokio::sync::SemaphorePermit<'_>,
+        prefix: Option<&RemotePath>,
+        mode: ListingMode,
+        max_keys: Option<NonZeroU32>,
+        cancel: &CancellationToken,
+    ) -> Result<crate::VersionListing, DownloadError> {
+        let customize_builder = |mut builder: ListBlobsBuilder| {
+            builder = builder.include_versions(true);
+            // We do not return this info back to `VersionListing` yet.
+            builder = builder.include_deleted(true);
+            builder
+        };
+        let kind = RequestKind::ListVersions;
+
+        let mut stream = std::pin::pin!(self.list_streaming_for_fn(
+            prefix,
+            mode,
+            max_keys,
+            cancel,
+            kind,
+            customize_builder
+        ));
+        let mut combined: crate::VersionListing =
+            stream.next().await.expect("At least one item required")?;
+        while let Some(list) = stream.next().await {
+            let list = list?;
+            combined.versions.extend(list.versions.into_iter());
+        }
+        Ok(combined)
+    }
 }
 
 trait ListingCollector {
@@ -488,27 +522,10 @@ impl RemoteStorage for AzureBlobStorage {
         max_keys: Option<NonZeroU32>,
         cancel: &CancellationToken,
     ) -> std::result::Result<crate::VersionListing, DownloadError> {
-        let customize_builder = |mut builder: ListBlobsBuilder| {
-            builder = builder.include_versions(true);
-            builder
-        };
         let kind = RequestKind::ListVersions;
-
-        let mut stream = std::pin::pin!(self.list_streaming_for_fn(
-            prefix,
-            mode,
-            max_keys,
-            cancel,
-            kind,
-            customize_builder
-        ));
-        let mut combined: crate::VersionListing =
-            stream.next().await.expect("At least one item required")?;
-        while let Some(list) = stream.next().await {
-            let list = list?;
-            combined.versions.extend(list.versions.into_iter());
-        }
-        Ok(combined)
+        let permit = self.permit(kind, cancel).await?;
+        self.list_versions_with_permit(&permit, prefix, mode, max_keys, cancel)
+            .await
     }
 
     async fn head_object(
@@ -803,14 +820,158 @@ impl RemoteStorage for AzureBlobStorage {
 
     async fn time_travel_recover(
         &self,
-        _prefix: Option<&RemotePath>,
-        _timestamp: SystemTime,
-        _done_if_after: SystemTime,
-        _cancel: &CancellationToken,
+        prefix: Option<&RemotePath>,
+        timestamp: SystemTime,
+        done_if_after: SystemTime,
+        cancel: &CancellationToken,
     ) -> Result<(), TimeTravelError> {
-        // TODO use Azure point in time recovery feature for this
-        // https://learn.microsoft.com/en-us/azure/storage/blobs/point-in-time-restore-overview
-        Err(TimeTravelError::Unimplemented)
+        let msg = "PLEASE NOTE: Azure Blob storage time-travel recovery may not work as expected "
+            .to_string()
+            + "for some specific files. If a file gets deleted but then overwritten and we want to recover "
+            + "to the time during the file was not present, this functionality will recover the file. Only "
+            + "use the functionality for services that can tolerate this. For example, recovering a state of the "
+            + "pageserver tenants.";
+        tracing::error!("{}", msg);
+
+        let kind = RequestKind::TimeTravel;
+        let permit = self.permit(kind, cancel).await?;
+
+        let mode = ListingMode::NoDelimiter;
+        let version_listing = self
+            .list_versions_with_permit(&permit, prefix, mode, None, cancel)
+            .await
+            .map_err(|err| match err {
+                DownloadError::Other(e) => TimeTravelError::Other(e),
+                DownloadError::Cancelled => TimeTravelError::Cancelled,
+                other => TimeTravelError::Other(other.into()),
+            })?;
+        let versions_and_deletes = version_listing.versions;
+
+        tracing::info!(
+            "Built list for time travel with {} versions and deletions",
+            versions_and_deletes.len()
+        );
+
+        // Work on the list of references instead of the objects directly,
+        // otherwise we get lifetime errors in the sort_by_key call below.
+        let mut versions_and_deletes = versions_and_deletes.iter().collect::<Vec<_>>();
+
+        versions_and_deletes.sort_by_key(|vd| (&vd.key, &vd.last_modified));
+
+        let mut vds_for_key = HashMap::<_, Vec<_>>::new();
+
+        for vd in &versions_and_deletes {
+            let Version { key, .. } = &vd;
+            let version_id = vd.version_id().map(|v| v.0.as_str());
+            if version_id == Some("null") {
+                return Err(TimeTravelError::Other(anyhow!(
+                    "Received ListVersions response for key={key} with version_id='null', \
+                        indicating either disabled versioning, or legacy objects with null version id values"
+                )));
+            }
+            tracing::trace!("Parsing version key={key} kind={:?}", vd.kind);
+
+            vds_for_key.entry(key).or_default().push(vd);
+        }
+
+        let warn_threshold = 3;
+        let max_retries = 10;
+        let is_permanent = |e: &_| matches!(e, TimeTravelError::Cancelled);
+
+        for (key, versions) in vds_for_key {
+            let last_vd = versions.last().unwrap();
+            let key = self.relative_path_to_name(key);
+            if last_vd.last_modified > done_if_after {
+                tracing::debug!("Key {key} has version later than done_if_after, skipping");
+                continue;
+            }
+            // the version we want to restore to.
+            let version_to_restore_to =
+                match versions.binary_search_by_key(&timestamp, |tpl| tpl.last_modified) {
+                    Ok(v) => v,
+                    Err(e) => e,
+                };
+            if version_to_restore_to == versions.len() {
+                tracing::debug!("Key {key} has no changes since timestamp, skipping");
+                continue;
+            }
+            let mut do_delete = false;
+            if version_to_restore_to == 0 {
+                // All versions more recent, so the key didn't exist at the specified time point.
+                tracing::debug!(
+                    "All {} versions more recent for {key}, deleting",
+                    versions.len()
+                );
+                do_delete = true;
+            } else {
+                match &versions[version_to_restore_to - 1] {
+                    Version {
+                        kind: VersionKind::Version(version_id),
+                        ..
+                    } => {
+                        let source_url = format!(
+                            "{}/{}?versionid={}",
+                            self.client
+                                .url()
+                                .map_err(|e| TimeTravelError::Other(anyhow!("{e}")))?,
+                            key,
+                            version_id.0
+                        );
+                        tracing::debug!(
+                            "Promoting old version {} for {key} at {}...",
+                            version_id.0,
+                            source_url
+                        );
+                        backoff::retry(
+                            || async {
+                                let blob_client = self.client.blob_client(key.clone());
+                                let op = blob_client.copy(Url::from_str(&source_url).unwrap());
+                                tokio::select! {
+                                    res = op => res.map_err(|e| TimeTravelError::Other(e.into())),
+                                    _ = cancel.cancelled() => Err(TimeTravelError::Cancelled),
+                                }
+                            },
+                            is_permanent,
+                            warn_threshold,
+                            max_retries,
+                            "copying object version for time_travel_recover",
+                            cancel,
+                        )
+                        .await
+                        .ok_or_else(|| TimeTravelError::Cancelled)
+                        .and_then(|x| x)?;
+                        tracing::info!(?version_id, %key, "Copied old version in Azure blob storage");
+                    }
+                    Version {
+                        kind: VersionKind::DeletionMarker,
+                        ..
+                    } => {
+                        do_delete = true;
+                    }
+                }
+            };
+            if do_delete {
+                if matches!(last_vd.kind, VersionKind::DeletionMarker) {
+                    // Key has since been deleted (but there was some history), no need to do anything
+                    tracing::debug!("Key {key} already deleted, skipping.");
+                } else {
+                    tracing::debug!("Deleting {key}...");
+
+                    self.delete(&RemotePath::from_string(&key).unwrap(), cancel)
+                        .await
+                        .map_err(|e| {
+                            // delete_oid0 will use TimeoutOrCancel
+                            if TimeoutOrCancel::caused_by_cancel(&e) {
+                                TimeTravelError::Cancelled
+                            } else {
+                                TimeTravelError::Other(e)
+                            }
+                        })?;
+                }
+            }
+        }
+
+        Ok(())
     }
 }
 

libs/remote_storage/src/s3_bucket.rs

@@ -1022,6 +1022,7 @@ impl RemoteStorage for S3Bucket {
             let Version { key, .. } = &vd;
             let version_id = vd.version_id().map(|v| v.0.as_str());
             if version_id == Some("null") {
+                // TODO: check the behavior of using the SDK on a non-versioned container
                 return Err(TimeTravelError::Other(anyhow!(
                     "Received ListVersions response for key={key} with version_id='null', \
                     indicating either disabled versioning, or legacy objects with null version id values"

libs/safekeeper_api/src/models.rs

@@ -13,7 +13,7 @@ use utils::pageserver_feedback::PageserverFeedback;
 use crate::membership::Configuration;
 use crate::{ServerInfo, Term};
 
-#[derive(Debug, Serialize)]
+#[derive(Debug, Serialize, Deserialize)]
 pub struct SafekeeperStatus {
     pub id: NodeId,
 }

libs/utils/src/leaky_bucket.rs

@@ -28,6 +28,7 @@ use std::time::Duration;
 use tokio::sync::Notify;
 use tokio::time::Instant;
 
+#[derive(Clone, Copy)]
 pub struct LeakyBucketConfig {
     /// This is the "time cost" of a single request unit.
     /// Should loosely represent how long it takes to handle a request unit in active resource time.

libs/utils/src/lib.rs

@@ -73,6 +73,7 @@ pub mod error;
 /// async timeout helper
 pub mod timeout;
 
+pub mod span;
 pub mod sync;
 
 pub mod failpoint_support;

libs/utils/src/span.rs

@@ -0,0 +1,19 @@
+//! Tracing span helpers.
+
+/// Records the given fields in the current span, as a single call. The fields must already have
+/// been declared for the span (typically with empty values).
+#[macro_export]
+macro_rules! span_record {
+    ($($tokens:tt)*) => {$crate::span_record_in!(::tracing::Span::current(), $($tokens)*)};
+}
+
+/// Records the given fields in the given span, as a single call. The fields must already have been
+/// declared for the span (typically with empty values).
+#[macro_export]
+macro_rules! span_record_in {
+    ($span:expr, $($tokens:tt)*) => {
+        if let Some(meta) = $span.metadata() {
+            $span.record_all(&tracing::valueset!(meta.fields(), $($tokens)*));
+        }
+    };
+}

libs/walproposer/src/api_bindings.rs

@@ -439,6 +439,7 @@ pub fn empty_shmem() -> crate::bindings::WalproposerShmemState {
         currentClusterSize: crate::bindings::pg_atomic_uint64 { value: 0 },
         shard_ps_feedback: [empty_feedback; 128],
         num_shards: 0,
+        replica_promote: false,
         min_ps_feedback: empty_feedback,
     }
 }

pageserver/Cargo.toml

@@ -34,6 +34,7 @@ fail.workspace = true
 futures.workspace = true
 hashlink.workspace = true
 hex.workspace = true
+http.workspace = true
 http-utils.workspace = true
 humantime-serde.workspace = true
 humantime.workspace = true
@@ -93,6 +94,7 @@ tokio-util.workspace = true
 toml_edit = { workspace = true, features = [ "serde" ] }
 tonic.workspace = true
 tonic-reflection.workspace = true
+tower.workspace = true
 tracing.workspace = true
 tracing-utils.workspace = true
 url.workspace = true

pageserver/benches/bench_metrics.rs

@@ -264,10 +264,56 @@ mod propagation_of_cached_label_value {
     }
 }
 
+criterion_group!(histograms, histograms::bench_bucket_scalability);
+mod histograms {
+    use std::time::Instant;
+
+    use criterion::{BenchmarkId, Criterion};
+    use metrics::core::Collector;
+
+    pub fn bench_bucket_scalability(c: &mut Criterion) {
+        let mut g = c.benchmark_group("bucket_scalability");
+
+        for n in [1, 4, 8, 16, 32, 64, 128, 256] {
+            g.bench_with_input(BenchmarkId::new("nbuckets", n), &n, |b, n| {
+                b.iter_custom(|iters| {
+                    let buckets: Vec<f64> = (0..*n).map(|i| i as f64 * 100.0).collect();
+                    let histo = metrics::Histogram::with_opts(
+                        metrics::prometheus::HistogramOpts::new("name", "help")
+                            .buckets(buckets.clone()),
+                    )
+                    .unwrap();
+                    let start = Instant::now();
+                    for i in 0..usize::try_from(iters).unwrap() {
+                        histo.observe(buckets[i % buckets.len()]);
+                    }
+                    let elapsed = start.elapsed();
+                    // self-test
+                    let mfs = histo.collect();
+                    assert_eq!(mfs.len(), 1);
+                    let metrics = mfs[0].get_metric();
+                    assert_eq!(metrics.len(), 1);
+                    let histo = metrics[0].get_histogram();
+                    let buckets = histo.get_bucket();
+                    assert!(
+                        buckets
+                            .iter()
+                            .enumerate()
+                            .all(|(i, b)| b.get_cumulative_count()
+                                >= i as u64 * (iters / buckets.len() as u64))
+                    );
+                    elapsed
+                })
+            });
+        }
+    }
+}
+
 criterion_main!(
     label_values,
     single_metric_multicore_scalability,
-    propagation_of_cached_label_value
+    propagation_of_cached_label_value,
+    histograms,
 );
 
 /*
@@ -290,6 +336,14 @@ propagation_of_cached_label_value__naive/nthreads/8 time:   [211.50 ns 214.44 ns
 propagation_of_cached_label_value__long_lived_reference_per_thread/nthreads/1 time:   [14.135 ns 14.147 ns 14.160 ns]
 propagation_of_cached_label_value__long_lived_reference_per_thread/nthreads/4 time:   [14.243 ns 14.255 ns 14.268 ns]
 propagation_of_cached_label_value__long_lived_reference_per_thread/nthreads/8 time:   [14.470 ns 14.682 ns 14.895 ns]
+bucket_scalability/nbuckets/1     time:   [30.352 ns 30.353 ns 30.354 ns]
+bucket_scalability/nbuckets/4     time:   [30.464 ns 30.465 ns 30.467 ns]
+bucket_scalability/nbuckets/8     time:   [30.569 ns 30.575 ns 30.584 ns]
+bucket_scalability/nbuckets/16      time:   [30.961 ns 30.965 ns 30.969 ns]
+bucket_scalability/nbuckets/32      time:   [35.691 ns 35.707 ns 35.722 ns]
+bucket_scalability/nbuckets/64      time:   [47.829 ns 47.898 ns 47.974 ns]
+bucket_scalability/nbuckets/128     time:   [73.479 ns 73.512 ns 73.545 ns]
+bucket_scalability/nbuckets/256     time:   [127.92 ns 127.94 ns 127.96 ns]
 
 Results on an i3en.3xlarge instance
 
@@ -344,6 +398,14 @@ propagation_of_cached_label_value__naive/nthreads/8     time:   [434.87 ns 456.4
 propagation_of_cached_label_value__long_lived_reference_per_thread/nthreads/1     time:   [3.3767 ns 3.3974 ns 3.4220 ns]
 propagation_of_cached_label_value__long_lived_reference_per_thread/nthreads/4     time:   [3.6105 ns 4.2355 ns 5.1463 ns]
 propagation_of_cached_label_value__long_lived_reference_per_thread/nthreads/8     time:   [4.0889 ns 4.9714 ns 6.0779 ns]
+bucket_scalability/nbuckets/1     time:   [4.8455 ns 4.8542 ns 4.8646 ns]
+bucket_scalability/nbuckets/4     time:   [4.5663 ns 4.5722 ns 4.5787 ns]
+bucket_scalability/nbuckets/8     time:   [4.5531 ns 4.5670 ns 4.5842 ns]
+bucket_scalability/nbuckets/16      time:   [4.6392 ns 4.6524 ns 4.6685 ns]
+bucket_scalability/nbuckets/32      time:   [6.0302 ns 6.0439 ns 6.0589 ns]
+bucket_scalability/nbuckets/64      time:   [10.608 ns 10.644 ns 10.691 ns]
+bucket_scalability/nbuckets/128     time:   [22.178 ns 22.316 ns 22.483 ns]
+bucket_scalability/nbuckets/256     time:   [42.190 ns 42.328 ns 42.492 ns]
 
 Results on a Hetzner AX102 AMD Ryzen 9 7950X3D 16-Core Processor
 
@@ -362,5 +424,13 @@ propagation_of_cached_label_value__naive/nthreads/8     time:   [164.24 ns 170.1
 propagation_of_cached_label_value__long_lived_reference_per_thread/nthreads/1     time:   [2.2915 ns 2.2960 ns 2.3012 ns]
 propagation_of_cached_label_value__long_lived_reference_per_thread/nthreads/4     time:   [2.5726 ns 2.6158 ns 2.6624 ns]
 propagation_of_cached_label_value__long_lived_reference_per_thread/nthreads/8     time:   [2.7068 ns 2.8243 ns 2.9824 ns]
+bucket_scalability/nbuckets/1     time:   [6.3998 ns 6.4288 ns 6.4684 ns]
+bucket_scalability/nbuckets/4     time:   [6.3603 ns 6.3620 ns 6.3637 ns]
+bucket_scalability/nbuckets/8     time:   [6.1646 ns 6.1654 ns 6.1667 ns]
+bucket_scalability/nbuckets/16      time:   [6.1341 ns 6.1391 ns 6.1454 ns]
+bucket_scalability/nbuckets/32      time:   [8.2206 ns 8.2254 ns 8.2301 ns]
+bucket_scalability/nbuckets/64      time:   [13.988 ns 13.994 ns 14.000 ns]
+bucket_scalability/nbuckets/128     time:   [28.180 ns 28.216 ns 28.251 ns]
+bucket_scalability/nbuckets/256     time:   [54.914 ns 54.931 ns 54.951 ns]
 
 */

pageserver/page_api/Cargo.toml

@@ -9,7 +9,6 @@ bytes.workspace = true
 pageserver_api.workspace = true
 postgres_ffi.workspace = true
 prost.workspace = true
-smallvec.workspace = true
 thiserror.workspace = true
 tonic.workspace = true
 utils.workspace = true

pageserver/page_api/src/model.rs

@@ -9,10 +9,16 @@
 //! - Use more precise datatypes, e.g. Lsn and uints shorter than 32 bits.
 //!
 //! - Validate protocol invariants, via try_from() and try_into().
+//!
+//! Validation only happens on the receiver side, i.e. when converting from Protobuf to domain
+//! types. This is where it matters -- the Protobuf types are less strict than the domain types, and
+//! receivers should expect all sorts of junk from senders. This also allows the sender to use e.g.
+//! stream combinators without dealing with errors, and avoids validating the same message twice.
+
+use std::fmt::Display;
 
 use bytes::Bytes;
 use postgres_ffi::Oid;
-use smallvec::SmallVec;
 // TODO: split out Lsn, RelTag, SlruKind, Oid and other basic types to a separate crate, to avoid
 // pulling in all of their other crate dependencies when building the client.
 use utils::lsn::Lsn;
@@ -48,7 +54,8 @@ pub struct ReadLsn {
     pub request_lsn: Lsn,
     /// If given, the caller guarantees that the page has not been modified since this LSN. Must be
     /// smaller than or equal to request_lsn. This allows the Pageserver to serve an old page
-    /// without waiting for the request LSN to arrive. Valid for all request types.
+    /// without waiting for the request LSN to arrive. If not given, the request will read at the
+    /// request_lsn and wait for it to arrive if necessary. Valid for all request types.
     ///
     /// It is undefined behaviour to make a request such that the page was, in fact, modified
     /// between request_lsn and not_modified_since_lsn. The Pageserver might detect it and return an
@@ -58,19 +65,14 @@ pub struct ReadLsn {
     pub not_modified_since_lsn: Option<Lsn>,
 }
 
-impl ReadLsn {
-    /// Validates the ReadLsn.
-    pub fn validate(&self) -> Result<(), ProtocolError> {
-        if self.request_lsn == Lsn::INVALID {
-            return Err(ProtocolError::invalid("request_lsn", self.request_lsn));
+impl Display for ReadLsn {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        let req_lsn = self.request_lsn;
+        if let Some(mod_lsn) = self.not_modified_since_lsn {
+            write!(f, "{req_lsn}>={mod_lsn}")
+        } else {
+            req_lsn.fmt(f)
         }
-        if self.not_modified_since_lsn > Some(self.request_lsn) {
-            return Err(ProtocolError::invalid(
-                "not_modified_since_lsn",
-                self.not_modified_since_lsn,
-            ));
-        }
-        Ok(())
     }
 }
 
@@ -78,27 +80,31 @@ impl TryFrom<proto::ReadLsn> for ReadLsn {
     type Error = ProtocolError;
 
     fn try_from(pb: proto::ReadLsn) -> Result<Self, Self::Error> {
-        let read_lsn = Self {
+        if pb.request_lsn == 0 {
+            return Err(ProtocolError::invalid("request_lsn", pb.request_lsn));
+        }
+        if pb.not_modified_since_lsn > pb.request_lsn {
+            return Err(ProtocolError::invalid(
+                "not_modified_since_lsn",
+                pb.not_modified_since_lsn,
+            ));
+        }
+        Ok(Self {
             request_lsn: Lsn(pb.request_lsn),
             not_modified_since_lsn: match pb.not_modified_since_lsn {
                 0 => None,
                 lsn => Some(Lsn(lsn)),
             },
-        };
-        read_lsn.validate()?;
-        Ok(read_lsn)
+        })
     }
 }
 
-impl TryFrom<ReadLsn> for proto::ReadLsn {
-    type Error = ProtocolError;
-
-    fn try_from(read_lsn: ReadLsn) -> Result<Self, Self::Error> {
-        read_lsn.validate()?;
-        Ok(Self {
+impl From<ReadLsn> for proto::ReadLsn {
+    fn from(read_lsn: ReadLsn) -> Self {
+        Self {
             request_lsn: read_lsn.request_lsn.0,
             not_modified_since_lsn: read_lsn.not_modified_since_lsn.unwrap_or_default().0,
-        })
+        }
     }
 }
 
@@ -153,6 +159,15 @@ impl TryFrom<proto::CheckRelExistsRequest> for CheckRelExistsRequest {
     }
 }
 
+impl From<CheckRelExistsRequest> for proto::CheckRelExistsRequest {
+    fn from(request: CheckRelExistsRequest) -> Self {
+        Self {
+            read_lsn: Some(request.read_lsn.into()),
+            rel: Some(request.rel.into()),
+        }
+    }
+}
+
 pub type CheckRelExistsResponse = bool;
 
 impl From<proto::CheckRelExistsResponse> for CheckRelExistsResponse {
@@ -190,14 +205,12 @@ impl TryFrom<proto::GetBaseBackupRequest> for GetBaseBackupRequest {
     }
 }
 
-impl TryFrom<GetBaseBackupRequest> for proto::GetBaseBackupRequest {
-    type Error = ProtocolError;
-
-    fn try_from(request: GetBaseBackupRequest) -> Result<Self, Self::Error> {
-        Ok(Self {
-            read_lsn: Some(request.read_lsn.try_into()?),
+impl From<GetBaseBackupRequest> for proto::GetBaseBackupRequest {
+    fn from(request: GetBaseBackupRequest) -> Self {
+        Self {
+            read_lsn: Some(request.read_lsn.into()),
             replica: request.replica,
-        })
+        }
     }
 }
 
@@ -214,14 +227,9 @@ impl TryFrom<proto::GetBaseBackupResponseChunk> for GetBaseBackupResponseChunk {
     }
 }
 
-impl TryFrom<GetBaseBackupResponseChunk> for proto::GetBaseBackupResponseChunk {
-    type Error = ProtocolError;
-
-    fn try_from(chunk: GetBaseBackupResponseChunk) -> Result<Self, Self::Error> {
-        if chunk.is_empty() {
-            return Err(ProtocolError::Missing("chunk"));
-        }
-        Ok(Self { chunk })
+impl From<GetBaseBackupResponseChunk> for proto::GetBaseBackupResponseChunk {
+    fn from(chunk: GetBaseBackupResponseChunk) -> Self {
+        Self { chunk }
     }
 }
 
@@ -246,14 +254,12 @@ impl TryFrom<proto::GetDbSizeRequest> for GetDbSizeRequest {
     }
 }
 
-impl TryFrom<GetDbSizeRequest> for proto::GetDbSizeRequest {
-    type Error = ProtocolError;
-
-    fn try_from(request: GetDbSizeRequest) -> Result<Self, Self::Error> {
-        Ok(Self {
-            read_lsn: Some(request.read_lsn.try_into()?),
+impl From<GetDbSizeRequest> for proto::GetDbSizeRequest {
+    fn from(request: GetDbSizeRequest) -> Self {
+        Self {
+            read_lsn: Some(request.read_lsn.into()),
             db_oid: request.db_oid,
-        })
+        }
     }
 }
 
@@ -288,7 +294,7 @@ pub struct GetPageRequest {
     /// Multiple pages will be executed as a single batch by the Pageserver, amortizing layer access
     /// costs and parallelizing them. This may increase the latency of any individual request, but
     /// improves the overall latency and throughput of the batch as a whole.
-    pub block_numbers: SmallVec<[u32; 1]>,
+    pub block_numbers: Vec<u32>,
 }
 
 impl TryFrom<proto::GetPageRequest> for GetPageRequest {
@@ -306,25 +312,20 @@ impl TryFrom<proto::GetPageRequest> for GetPageRequest {
                 .ok_or(ProtocolError::Missing("read_lsn"))?
                 .try_into()?,
             rel: pb.rel.ok_or(ProtocolError::Missing("rel"))?.try_into()?,
-            block_numbers: pb.block_number.into(),
+            block_numbers: pb.block_number,
         })
     }
 }
 
-impl TryFrom<GetPageRequest> for proto::GetPageRequest {
-    type Error = ProtocolError;
-
-    fn try_from(request: GetPageRequest) -> Result<Self, Self::Error> {
-        if request.block_numbers.is_empty() {
-            return Err(ProtocolError::Missing("block_number"));
-        }
-        Ok(Self {
+impl From<GetPageRequest> for proto::GetPageRequest {
+    fn from(request: GetPageRequest) -> Self {
+        Self {
             request_id: request.request_id,
             request_class: request.request_class.into(),
-            read_lsn: Some(request.read_lsn.try_into()?),
+            read_lsn: Some(request.read_lsn.into()),
             rel: Some(request.rel.into()),
-            block_number: request.block_numbers.into_vec(),
-        })
+            block_number: request.block_numbers,
+        }
     }
 }
 
@@ -396,7 +397,7 @@ pub struct GetPageResponse {
     /// A string describing the status, if any.
     pub reason: Option<String>,
     /// The 8KB page images, in the same order as the request. Empty if status != OK.
-    pub page_images: SmallVec<[Bytes; 1]>,
+    pub page_images: Vec<Bytes>,
 }
 
 impl From<proto::GetPageResponse> for GetPageResponse {
@@ -405,7 +406,7 @@ impl From<proto::GetPageResponse> for GetPageResponse {
             request_id: pb.request_id,
             status_code: pb.status_code.into(),
             reason: Some(pb.reason).filter(|r| !r.is_empty()),
-            page_images: pb.page_image.into(),
+            page_images: pb.page_image,
         }
     }
 }
@@ -416,7 +417,7 @@ impl From<GetPageResponse> for proto::GetPageResponse {
             request_id: response.request_id,
             status_code: response.status_code.into(),
             reason: response.reason.unwrap_or_default(),
-            page_image: response.page_images.into_vec(),
+            page_image: response.page_images,
         }
     }
 }
@@ -505,14 +506,12 @@ impl TryFrom<proto::GetRelSizeRequest> for GetRelSizeRequest {
     }
 }
 
-impl TryFrom<GetRelSizeRequest> for proto::GetRelSizeRequest {
-    type Error = ProtocolError;
-
-    fn try_from(request: GetRelSizeRequest) -> Result<Self, Self::Error> {
-        Ok(Self {
-            read_lsn: Some(request.read_lsn.try_into()?),
+impl From<GetRelSizeRequest> for proto::GetRelSizeRequest {
+    fn from(request: GetRelSizeRequest) -> Self {
+        Self {
+            read_lsn: Some(request.read_lsn.into()),
             rel: Some(request.rel.into()),
-        })
+        }
     }
 }
 
@@ -555,15 +554,13 @@ impl TryFrom<proto::GetSlruSegmentRequest> for GetSlruSegmentRequest {
     }
 }
 
-impl TryFrom<GetSlruSegmentRequest> for proto::GetSlruSegmentRequest {
-    type Error = ProtocolError;
-
-    fn try_from(request: GetSlruSegmentRequest) -> Result<Self, Self::Error> {
-        Ok(Self {
-            read_lsn: Some(request.read_lsn.try_into()?),
+impl From<GetSlruSegmentRequest> for proto::GetSlruSegmentRequest {
+    fn from(request: GetSlruSegmentRequest) -> Self {
+        Self {
+            read_lsn: Some(request.read_lsn.into()),
             kind: request.kind as u32,
             segno: request.segno,
-        })
+        }
     }
 }
 
@@ -580,14 +577,9 @@ impl TryFrom<proto::GetSlruSegmentResponse> for GetSlruSegmentResponse {
     }
 }
 
-impl TryFrom<GetSlruSegmentResponse> for proto::GetSlruSegmentResponse {
-    type Error = ProtocolError;
-
-    fn try_from(segment: GetSlruSegmentResponse) -> Result<Self, Self::Error> {
-        if segment.is_empty() {
-            return Err(ProtocolError::Missing("segment"));
-        }
-        Ok(Self { segment })
+impl From<GetSlruSegmentResponse> for proto::GetSlruSegmentResponse {
+    fn from(segment: GetSlruSegmentResponse) -> Self {
+        Self { segment }
     }
 }
 

pageserver/pagebench/Cargo.toml

@@ -8,6 +8,8 @@ license.workspace = true
 
 [dependencies]
 anyhow.workspace = true
+async-trait.workspace = true
+bytes.workspace = true
 camino.workspace = true
 clap.workspace = true
 futures.workspace = true
@@ -15,14 +17,17 @@ hdrhistogram.workspace = true
 humantime.workspace = true
 humantime-serde.workspace = true
 rand.workspace = true
-reqwest.workspace=true
+reqwest.workspace = true
 serde.workspace = true
 serde_json.workspace = true
 tracing.workspace = true
 tokio.workspace = true
+tokio-stream.workspace = true
 tokio-util.workspace = true
+tonic.workspace = true
 
 pageserver_client.workspace = true
 pageserver_api.workspace = true
+pageserver_page_api.workspace = true
 utils = { path = "../../libs/utils/" }
 workspace_hack = { version = "0.1", path = "../../workspace_hack" }

pageserver/pagebench/src/cmd/getpage_latest_lsn.rs

@@ -1,4 +1,4 @@
-use std::collections::{HashSet, VecDeque};
+use std::collections::{HashMap, HashSet, VecDeque};
 use std::future::Future;
 use std::num::NonZeroUsize;
 use std::pin::Pin;
@@ -7,11 +7,15 @@ use std::sync::{Arc, Mutex};
 use std::time::{Duration, Instant};
 
 use anyhow::Context;
+use async_trait::async_trait;
+use bytes::Bytes;
 use camino::Utf8PathBuf;
 use pageserver_api::key::Key;
 use pageserver_api::keyspace::KeySpaceAccum;
 use pageserver_api::models::{PagestreamGetPageRequest, PagestreamRequest};
+use pageserver_api::reltag::RelTag;
 use pageserver_api::shard::TenantShardId;
+use pageserver_page_api::proto;
 use rand::prelude::*;
 use tokio::task::JoinSet;
 use tokio_util::sync::CancellationToken;
@@ -22,6 +26,12 @@ use utils::lsn::Lsn;
 use crate::util::tokio_thread_local_stats::AllThreadLocalStats;
 use crate::util::{request_stats, tokio_thread_local_stats};
 
+#[derive(clap::ValueEnum, Clone, Debug)]
+enum Protocol {
+    Libpq,
+    Grpc,
+}
+
 /// GetPage@LatestLSN, uniformly distributed across the compute-accessible keyspace.
 #[derive(clap::Parser)]
 pub(crate) struct Args {
@@ -35,6 +45,8 @@ pub(crate) struct Args {
     num_clients: NonZeroUsize,
     #[clap(long)]
     runtime: Option<humantime::Duration>,
+    #[clap(long, value_enum, default_value = "libpq")]
+    protocol: Protocol,
     /// Each client sends requests at the given rate.
     ///
     /// If a request takes too long and we should be issuing a new request already,
@@ -65,6 +77,16 @@ pub(crate) struct Args {
     #[clap(long, default_value = "1")]
     queue_depth: NonZeroUsize,
 
+    /// Batch size of contiguous pages generated by each client. This is equivalent to how Postgres
+    /// will request page batches (e.g. prefetches or vectored reads). A batch counts as 1 RPS and
+    /// 1 queue depth.
+    ///
+    /// The libpq protocol does not support client-side batching, and will submit batches as many
+    /// individual requests, in the hope that the server will batch them. Each batch still counts as
+    /// 1 RPS and 1 queue depth.
+    #[clap(long, default_value = "1")]
+    batch_size: NonZeroUsize,
+
     #[clap(long)]
     only_relnode: Option<u32>,
 
@@ -303,7 +325,20 @@ async fn main_impl(
                 .unwrap();
 
         Box::pin(async move {
-            client_libpq(args, worker_id, ss, cancel, rps_period, ranges, weights).await
+            let client: Box<dyn Client> = match args.protocol {
+                Protocol::Libpq => Box::new(
+                    LibpqClient::new(args.page_service_connstring.clone(), worker_id.timeline)
+                        .await
+                        .unwrap(),
+                ),
+
+                Protocol::Grpc => Box::new(
+                    GrpcClient::new(args.page_service_connstring.clone(), worker_id.timeline)
+                        .await
+                        .unwrap(),
+                ),
+            };
+            run_worker(args, client, ss, cancel, rps_period, ranges, weights).await
         })
     };
 
@@ -355,27 +390,28 @@ async fn main_impl(
     anyhow::Ok(())
 }
 
-async fn client_libpq(
+async fn run_worker(
     args: &Args,
-    worker_id: WorkerId,
+    mut client: Box<dyn Client>,
     shared_state: Arc<SharedState>,
     cancel: CancellationToken,
     rps_period: Option<Duration>,
     ranges: Vec<KeyRange>,
     weights: rand::distributions::weighted::WeightedIndex<i128>,
 ) {
-    let client = pageserver_client::page_service::Client::new(args.page_service_connstring.clone())
-        .await
-        .unwrap();
-    let mut client = client
-        .pagestream(worker_id.timeline.tenant_id, worker_id.timeline.timeline_id)
-        .await
-        .unwrap();
-
     shared_state.start_work_barrier.wait().await;
     let client_start = Instant::now();
     let mut ticks_processed = 0;
-    let mut inflight = VecDeque::new();
+    let mut req_id = 0;
+    let batch_size: usize = args.batch_size.into();
+
+    // Track inflight requests by request ID and start time. This times the request duration, and
+    // ensures responses match requests. We don't expect responses back in any particular order.
+    //
+    // NB: this does not check that all requests received a response, because we don't wait for the
+    // inflight requests to complete when the duration elapses.
+    let mut inflight: HashMap<u64, Instant> = HashMap::new();
+
     while !cancel.is_cancelled() {
         // Detect if a request took longer than the RPS rate
         if let Some(period) = &rps_period {
@@ -391,36 +427,72 @@ async fn client_libpq(
         }
 
         while inflight.len() < args.queue_depth.get() {
+            req_id += 1;
             let start = Instant::now();
-            let req = {
+            let (req_lsn, mod_lsn, rel, blks) = {
+                /// Converts a compact i128 key to a relation tag and block number.
+                fn key_to_block(key: i128) -> (RelTag, u32) {
+                    let key = Key::from_i128(key);
+                    assert!(key.is_rel_block_key());
+                    key.to_rel_block()
+                        .expect("we filter non-rel-block keys out above")
+                }
+
+                // Pick a random page from a random relation.
                 let mut rng = rand::thread_rng();
                 let r = &ranges[weights.sample(&mut rng)];
                 let key: i128 = rng.gen_range(r.start..r.end);
-                let key = Key::from_i128(key);
-                assert!(key.is_rel_block_key());
-                let (rel_tag, block_no) = key
-                    .to_rel_block()
-                    .expect("we filter non-rel-block keys out above");
-                PagestreamGetPageRequest {
-                    hdr: PagestreamRequest {
-                        reqid: 0,
-                        request_lsn: if rng.gen_bool(args.req_latest_probability) {
-                            Lsn::MAX
-                        } else {
-                            r.timeline_lsn
-                        },
-                        not_modified_since: r.timeline_lsn,
-                    },
-                    rel: rel_tag,
-                    blkno: block_no,
+                let (rel_tag, block_no) = key_to_block(key);
+
+                let mut blks = VecDeque::with_capacity(batch_size);
+                blks.push_back(block_no);
+
+                // If requested, populate a batch of sequential pages. This is how Postgres will
+                // request page batches (e.g. prefetches). If we hit the end of the relation, we
+                // grow the batch towards the start too.
+                for i in 1..batch_size {
+                    let (r, b) = key_to_block(key + i as i128);
+                    if r != rel_tag {
+                        break; // went outside relation
+                    }
+                    blks.push_back(b)
                 }
+
+                if blks.len() < batch_size {
+                    // Grow batch backwards if needed.
+                    for i in 1..batch_size {
+                        let (r, b) = key_to_block(key - i as i128);
+                        if r != rel_tag {
+                            break; // went outside relation
+                        }
+                        blks.push_front(b)
+                    }
+                }
+
+                // We assume that the entire batch can fit within the relation.
+                assert_eq!(blks.len(), batch_size, "incomplete batch");
+
+                let req_lsn = if rng.gen_bool(args.req_latest_probability) {
+                    Lsn::MAX
+                } else {
+                    r.timeline_lsn
+                };
+                (req_lsn, r.timeline_lsn, rel_tag, blks.into())
             };
-            client.getpage_send(req).await.unwrap();
-            inflight.push_back(start);
+            client
+                .send_get_page(req_id, req_lsn, mod_lsn, rel, blks)
+                .await
+                .unwrap();
+            let old = inflight.insert(req_id, start);
+            assert!(old.is_none(), "duplicate request ID {req_id}");
         }
 
-        let start = inflight.pop_front().unwrap();
-        client.getpage_recv().await.unwrap();
+        let (req_id, pages) = client.recv_get_page().await.unwrap();
+        assert_eq!(pages.len(), batch_size, "unexpected page count");
+        assert!(pages.iter().all(|p| !p.is_empty()), "empty page");
+        let start = inflight
+            .remove(&req_id)
+            .expect("response for unknown request ID");
         let end = Instant::now();
         shared_state.live_stats.request_done();
         ticks_processed += 1;
@@ -442,3 +514,154 @@ async fn client_libpq(
         }
     }
 }
+
+/// A benchmark client, to allow switching out the transport protocol.
+///
+/// For simplicity, this just uses separate asynchronous send/recv methods. The send method could
+/// return a future that resolves when the response is received, but we don't really need it.
+#[async_trait]
+trait Client: Send {
+    /// Sends an asynchronous GetPage request to the pageserver.
+    async fn send_get_page(
+        &mut self,
+        req_id: u64,
+        req_lsn: Lsn,
+        mod_lsn: Lsn,
+        rel: RelTag,
+        blks: Vec<u32>,
+    ) -> anyhow::Result<()>;
+
+    /// Receives the next GetPage response from the pageserver.
+    async fn recv_get_page(&mut self) -> anyhow::Result<(u64, Vec<Bytes>)>;
+}
+
+/// A libpq-based Pageserver client.
+struct LibpqClient {
+    inner: pageserver_client::page_service::PagestreamClient,
+    // Track sent batches, so we know how many responses to expect.
+    batch_sizes: VecDeque<usize>,
+}
+
+impl LibpqClient {
+    async fn new(connstring: String, ttid: TenantTimelineId) -> anyhow::Result<Self> {
+        let inner = pageserver_client::page_service::Client::new(connstring)
+            .await?
+            .pagestream(ttid.tenant_id, ttid.timeline_id)
+            .await?;
+        Ok(Self {
+            inner,
+            batch_sizes: VecDeque::new(),
+        })
+    }
+}
+
+#[async_trait]
+impl Client for LibpqClient {
+    async fn send_get_page(
+        &mut self,
+        req_id: u64,
+        req_lsn: Lsn,
+        mod_lsn: Lsn,
+        rel: RelTag,
+        blks: Vec<u32>,
+    ) -> anyhow::Result<()> {
+        // libpq doesn't support client-side batches, so we send a bunch of individual requests
+        // instead in the hope that the server will batch them for us. We use the same request ID
+        // for all, because we'll return a single batch response.
+        self.batch_sizes.push_back(blks.len());
+        for blkno in blks {
+            let req = PagestreamGetPageRequest {
+                hdr: PagestreamRequest {
+                    reqid: req_id,
+                    request_lsn: req_lsn,
+                    not_modified_since: mod_lsn,
+                },
+                rel,
+                blkno,
+            };
+            self.inner.getpage_send(req).await?;
+        }
+        Ok(())
+    }
+
+    async fn recv_get_page(&mut self) -> anyhow::Result<(u64, Vec<Bytes>)> {
+        let batch_size = self.batch_sizes.pop_front().unwrap();
+        let mut batch = Vec::with_capacity(batch_size);
+        let mut req_id = None;
+        for _ in 0..batch_size {
+            let resp = self.inner.getpage_recv().await?;
+            if req_id.is_none() {
+                req_id = Some(resp.req.hdr.reqid);
+            }
+            assert_eq!(req_id, Some(resp.req.hdr.reqid), "request ID mismatch");
+            batch.push(resp.page);
+        }
+        Ok((req_id.unwrap(), batch))
+    }
+}
+
+/// A gRPC client using the raw, no-frills gRPC client.
+struct GrpcClient {
+    req_tx: tokio::sync::mpsc::Sender<proto::GetPageRequest>,
+    resp_rx: tonic::Streaming<proto::GetPageResponse>,
+}
+
+impl GrpcClient {
+    async fn new(connstring: String, ttid: TenantTimelineId) -> anyhow::Result<Self> {
+        let mut client = pageserver_page_api::proto::PageServiceClient::connect(connstring).await?;
+
+        // The channel has a buffer size of 1, since 0 is not allowed. It does not matter, since the
+        // benchmark will control the queue depth (i.e. in-flight requests) anyway, and requests are
+        // buffered by Tonic and the OS too.
+        let (req_tx, req_rx) = tokio::sync::mpsc::channel(1);
+        let req_stream = tokio_stream::wrappers::ReceiverStream::new(req_rx);
+        let mut req = tonic::Request::new(req_stream);
+        let metadata = req.metadata_mut();
+        metadata.insert("neon-tenant-id", ttid.tenant_id.to_string().try_into()?);
+        metadata.insert("neon-timeline-id", ttid.timeline_id.to_string().try_into()?);
+        metadata.insert("neon-shard-id", "0000".try_into()?);
+
+        let resp = client.get_pages(req).await?;
+        let resp_stream = resp.into_inner();
+
+        Ok(Self {
+            req_tx,
+            resp_rx: resp_stream,
+        })
+    }
+}
+
+#[async_trait]
+impl Client for GrpcClient {
+    async fn send_get_page(
+        &mut self,
+        req_id: u64,
+        req_lsn: Lsn,
+        mod_lsn: Lsn,
+        rel: RelTag,
+        blks: Vec<u32>,
+    ) -> anyhow::Result<()> {
+        let req = proto::GetPageRequest {
+            request_id: req_id,
+            request_class: proto::GetPageClass::Normal as i32,
+            read_lsn: Some(proto::ReadLsn {
+                request_lsn: req_lsn.0,
+                not_modified_since_lsn: mod_lsn.0,
+            }),
+            rel: Some(rel.into()),
+            block_number: blks,
+        };
+        self.req_tx.send(req).await?;
+        Ok(())
+    }
+
+    async fn recv_get_page(&mut self) -> anyhow::Result<(u64, Vec<Bytes>)> {
+        let resp = self.resp_rx.message().await?.unwrap();
+        anyhow::ensure!(
+            resp.status_code == proto::GetPageStatusCode::Ok as i32,
+            "unexpected status code: {}",
+            resp.status_code
+        );
+        Ok((resp.request_id, resp.page_image))
+    }
+}

pageserver/src/basebackup.rs

@@ -65,6 +65,30 @@ impl From<GetVectoredError> for BasebackupError {
     }
 }
 
+impl From<BasebackupError> for postgres_backend::QueryError {
+    fn from(err: BasebackupError) -> Self {
+        use postgres_backend::QueryError;
+        use pq_proto::framed::ConnectionError;
+        match err {
+            BasebackupError::Client(err, _) => QueryError::Disconnected(ConnectionError::Io(err)),
+            BasebackupError::Server(err) => QueryError::Other(err),
+            BasebackupError::Shutdown => QueryError::Shutdown,
+        }
+    }
+}
+
+impl From<BasebackupError> for tonic::Status {
+    fn from(err: BasebackupError) -> Self {
+        use tonic::Code;
+        let code = match &err {
+            BasebackupError::Client(_, _) => Code::Cancelled,
+            BasebackupError::Server(_) => Code::Internal,
+            BasebackupError::Shutdown => Code::Unavailable,
+        };
+        tonic::Status::new(code, err.to_string())
+    }
+}
+
 /// Create basebackup with non-rel data in it.
 /// Only include relational data if 'full_backup' is true.
 ///
@@ -248,7 +272,7 @@ where
     async fn flush(&mut self) -> Result<(), BasebackupError> {
         let nblocks = self.buf.len() / BLCKSZ as usize;
         let (kind, segno) = self.current_segment.take().unwrap();
-        let segname = format!("{}/{:>04X}", kind.to_str(), segno);
+        let segname = format!("{kind}/{segno:>04X}");
         let header = new_tar_header(&segname, self.buf.len() as u64)?;
         self.ar
             .append(&header, self.buf.as_slice())
@@ -347,7 +371,7 @@ where
                 .await?
                 .partition(
                     self.timeline.get_shard_identity(),
-                    Timeline::MAX_GET_VECTORED_KEYS * BLCKSZ as u64,
+                    self.timeline.conf.max_get_vectored_keys.get() as u64 * BLCKSZ as u64,
                 );
 
             let mut slru_builder = SlruSegmentsBuilder::new(&mut self.ar);

pageserver/src/bin/pageserver.rs

@@ -23,6 +23,7 @@ use pageserver::deletion_queue::DeletionQueue;
 use pageserver::disk_usage_eviction_task::{self, launch_disk_usage_global_eviction_task};
 use pageserver::feature_resolver::FeatureResolver;
 use pageserver::metrics::{STARTUP_DURATION, STARTUP_IS_LOADING};
+use pageserver::page_service::GrpcPageServiceHandler;
 use pageserver::task_mgr::{
     BACKGROUND_RUNTIME, COMPUTE_REQUEST_RUNTIME, MGMT_REQUEST_RUNTIME, WALRECEIVER_RUNTIME,
 };
@@ -158,7 +159,6 @@ fn main() -> anyhow::Result<()> {
     // (maybe we should automate this with a visitor?).
     info!(?conf.virtual_file_io_engine, "starting with virtual_file IO engine");
     info!(?conf.virtual_file_io_mode, "starting with virtual_file IO mode");
-    info!(?conf.wal_receiver_protocol, "starting with WAL receiver protocol");
     info!(?conf.validate_wal_contiguity, "starting with WAL contiguity validation");
     info!(?conf.page_service_pipelining, "starting with page service pipelining config");
     info!(?conf.get_vectored_concurrent_io, "starting with get_vectored IO concurrency config");
@@ -804,7 +804,7 @@ fn start_pageserver(
         } else {
             None
         },
-        basebackup_cache.clone(),
+        basebackup_cache,
     );
 
     // Spawn a Pageserver gRPC server task. It will spawn separate tasks for
@@ -815,13 +815,12 @@ fn start_pageserver(
     // necessary?
     let mut page_service_grpc = None;
     if let Some(grpc_listener) = grpc_listener {
-        page_service_grpc = Some(page_service::spawn_grpc(
-            conf,
+        page_service_grpc = Some(GrpcPageServiceHandler::spawn(
             tenant_manager.clone(),
             grpc_auth,
             otel_guard.as_ref().map(|g| g.dispatch.clone()),
+            conf.get_vectored_concurrent_io,
             grpc_listener,
-            basebackup_cache,
         )?);
     }
 

pageserver/src/config.rs

@@ -14,7 +14,10 @@ use std::time::Duration;
 use anyhow::{Context, bail, ensure};
 use camino::{Utf8Path, Utf8PathBuf};
 use once_cell::sync::OnceCell;
-use pageserver_api::config::{DiskUsageEvictionTaskConfig, MaxVectoredReadBytes, PostHogConfig};
+use pageserver_api::config::{
+    DiskUsageEvictionTaskConfig, MaxGetVectoredKeys, MaxVectoredReadBytes,
+    PageServicePipeliningConfig, PageServicePipeliningConfigPipelined, PostHogConfig,
+};
 use pageserver_api::models::ImageCompressionAlgorithm;
 use pageserver_api::shard::TenantShardId;
 use pem::Pem;
@@ -24,7 +27,6 @@ use reqwest::Url;
 use storage_broker::Uri;
 use utils::id::{NodeId, TimelineId};
 use utils::logging::{LogFormat, SecretString};
-use utils::postgres_client::PostgresClientProtocol;
 
 use crate::tenant::storage_layer::inmemory_layer::IndexEntry;
 use crate::tenant::{TENANTS_SEGMENT_NAME, TIMELINES_SEGMENT_NAME};
@@ -185,6 +187,9 @@ pub struct PageServerConf {
 
     pub max_vectored_read_bytes: MaxVectoredReadBytes,
 
+    /// Maximum number of keys to be read in a single get_vectored call.
+    pub max_get_vectored_keys: MaxGetVectoredKeys,
+
     pub image_compression: ImageCompressionAlgorithm,
 
     /// Whether to offload archived timelines automatically
@@ -205,8 +210,6 @@ pub struct PageServerConf {
     /// Optionally disable disk syncs (unsafe!)
     pub no_sync: bool,
 
-    pub wal_receiver_protocol: PostgresClientProtocol,
-
     pub page_service_pipelining: pageserver_api::config::PageServicePipeliningConfig,
 
     pub get_vectored_concurrent_io: pageserver_api::config::GetVectoredConcurrentIo,
@@ -404,6 +407,7 @@ impl PageServerConf {
             secondary_download_concurrency,
             ingest_batch_size,
             max_vectored_read_bytes,
+            max_get_vectored_keys,
             image_compression,
             timeline_offloading,
             ephemeral_bytes_per_memory_kb,
@@ -414,7 +418,6 @@ impl PageServerConf {
             virtual_file_io_engine,
             tenant_config,
             no_sync,
-            wal_receiver_protocol,
             page_service_pipelining,
             get_vectored_concurrent_io,
             enable_read_path_debugging,
@@ -470,13 +473,13 @@ impl PageServerConf {
             secondary_download_concurrency,
             ingest_batch_size,
             max_vectored_read_bytes,
+            max_get_vectored_keys,
             image_compression,
             timeline_offloading,
             ephemeral_bytes_per_memory_kb,
             import_pgdata_upcall_api,
             import_pgdata_upcall_api_token: import_pgdata_upcall_api_token.map(SecretString::from),
             import_pgdata_aws_endpoint_url,
-            wal_receiver_protocol,
             page_service_pipelining,
             get_vectored_concurrent_io,
             tracing,
@@ -598,6 +601,19 @@ impl PageServerConf {
                 )
             })?;
 
+        if let PageServicePipeliningConfig::Pipelined(PageServicePipeliningConfigPipelined {
+            max_batch_size,
+            ..
+        }) = conf.page_service_pipelining
+        {
+            if max_batch_size.get() > conf.max_get_vectored_keys.get() {
+                return Err(anyhow::anyhow!(
+                    "`max_batch_size` ({max_batch_size}) must be less than or equal to `max_get_vectored_keys` ({})",
+                    conf.max_get_vectored_keys.get()
+                ));
+            }
+        };
+
         Ok(conf)
     }
 
@@ -685,6 +701,7 @@ impl ConfigurableSemaphore {
 mod tests {
 
     use camino::Utf8PathBuf;
+    use rstest::rstest;
     use utils::id::NodeId;
 
     use super::PageServerConf;
@@ -724,4 +741,28 @@ mod tests {
         PageServerConf::parse_and_validate(NodeId(0), config_toml, &workdir)
             .expect_err("parse_and_validate should fail for endpoint without scheme");
     }
+
+    #[rstest]
+    #[case(32, 32, true)]
+    #[case(64, 32, false)]
+    #[case(64, 64, true)]
+    #[case(128, 128, true)]
+    fn test_config_max_batch_size_is_valid(
+        #[case] max_batch_size: usize,
+        #[case] max_get_vectored_keys: usize,
+        #[case] is_valid: bool,
+    ) {
+        let input = format!(
+            r#"
+            control_plane_api = "http://localhost:6666"
+            max_get_vectored_keys = {max_get_vectored_keys}
+            page_service_pipelining = {{ mode="pipelined", execution="concurrent-futures", max_batch_size={max_batch_size}, batching="uniform-lsn" }}
+        "#,
+        );
+        let config_toml = toml_edit::de::from_str::<pageserver_api::config::ConfigToml>(&input)
+            .expect("config has valid fields");
+        let workdir = Utf8PathBuf::from("/nonexistent");
+        let result = PageServerConf::parse_and_validate(NodeId(0), config_toml, &workdir);
+        assert_eq!(result.is_ok(), is_valid);
+    }
 }

pageserver/src/feature_resolver.rs

@@ -1,21 +1,29 @@
 use std::{collections::HashMap, sync::Arc, time::Duration};
 
+use pageserver_api::config::NodeMetadata;
 use posthog_client_lite::{
-    FeatureResolverBackgroundLoop, PostHogClientConfig, PostHogEvaluationError,
+    CaptureEvent, FeatureResolverBackgroundLoop, PostHogClientConfig, PostHogEvaluationError,
+    PostHogFlagFilterPropertyValue,
 };
+use remote_storage::RemoteStorageKind;
+use serde_json::json;
 use tokio_util::sync::CancellationToken;
 use utils::id::TenantId;
 
-use crate::config::PageServerConf;
+use crate::{config::PageServerConf, metrics::FEATURE_FLAG_EVALUATION};
 
 #[derive(Clone)]
 pub struct FeatureResolver {
     inner: Option<Arc<FeatureResolverBackgroundLoop>>,
+    internal_properties: Option<Arc<HashMap<String, PostHogFlagFilterPropertyValue>>>,
 }
 
 impl FeatureResolver {
     pub fn new_disabled() -> Self {
-        Self { inner: None }
+        Self {
+            inner: None,
+            internal_properties: None,
+        }
     }
 
     pub fn spawn(
@@ -36,14 +44,142 @@ impl FeatureResolver {
                 shutdown_pageserver,
             );
             let inner = Arc::new(inner);
-            // TODO: make this configurable
-            inner.clone().spawn(handle, Duration::from_secs(60));
-            Ok(FeatureResolver { inner: Some(inner) })
+
+            // The properties shared by all tenants on this pageserver.
+            let internal_properties = {
+                let mut properties = HashMap::new();
+                properties.insert(
+                    "pageserver_id".to_string(),
+                    PostHogFlagFilterPropertyValue::String(conf.id.to_string()),
+                );
+                if let Some(availability_zone) = &conf.availability_zone {
+                    properties.insert(
+                        "availability_zone".to_string(),
+                        PostHogFlagFilterPropertyValue::String(availability_zone.clone()),
+                    );
+                }
+                // Infer region based on the remote storage config.
+                if let Some(remote_storage) = &conf.remote_storage_config {
+                    match &remote_storage.storage {
+                        RemoteStorageKind::AwsS3(config) => {
+                            properties.insert(
+                                "region".to_string(),
+                                PostHogFlagFilterPropertyValue::String(format!(
+                                    "aws-{}",
+                                    config.bucket_region
+                                )),
+                            );
+                        }
+                        RemoteStorageKind::AzureContainer(config) => {
+                            properties.insert(
+                                "region".to_string(),
+                                PostHogFlagFilterPropertyValue::String(format!(
+                                    "azure-{}",
+                                    config.container_region
+                                )),
+                            );
+                        }
+                        RemoteStorageKind::LocalFs { .. } => {
+                            properties.insert(
+                                "region".to_string(),
+                                PostHogFlagFilterPropertyValue::String("local".to_string()),
+                            );
+                        }
+                    }
+                }
+                // TODO: move this to a background task so that we don't block startup in case of slow disk
+                let metadata_path = conf.metadata_path();
+                match std::fs::read_to_string(&metadata_path) {
+                    Ok(metadata_str) => match serde_json::from_str::<NodeMetadata>(&metadata_str) {
+                        Ok(metadata) => {
+                            properties.insert(
+                                "hostname".to_string(),
+                                PostHogFlagFilterPropertyValue::String(metadata.http_host),
+                            );
+                            if let Some(cplane_region) = metadata.other.get("region_id") {
+                                if let Some(cplane_region) = cplane_region.as_str() {
+                                    // This region contains the cell number
+                                    properties.insert(
+                                        "neon_region".to_string(),
+                                        PostHogFlagFilterPropertyValue::String(
+                                            cplane_region.to_string(),
+                                        ),
+                                    );
+                                }
+                            }
+                        }
+                        Err(e) => {
+                            tracing::warn!("Failed to parse metadata.json: {}", e);
+                        }
+                    },
+                    Err(e) => {
+                        tracing::warn!("Failed to read metadata.json: {}", e);
+                    }
+                }
+                Arc::new(properties)
+            };
+            let fake_tenants = {
+                let mut tenants = Vec::new();
+                for i in 0..10 {
+                    let distinct_id = format!(
+                        "fake_tenant_{}_{}_{}",
+                        conf.availability_zone.as_deref().unwrap_or_default(),
+                        conf.id,
+                        i
+                    );
+                    let properties = Self::collect_properties_inner(
+                        distinct_id.clone(),
+                        Some(&internal_properties),
+                    );
+                    tenants.push(CaptureEvent {
+                        event: "initial_tenant_report".to_string(),
+                        distinct_id,
+                        properties: json!({ "$set": properties }), // use `$set` to set the person properties instead of the event properties
+                    });
+                }
+                tenants
+            };
+            // TODO: make refresh period configurable
+            inner
+                .clone()
+                .spawn(handle, Duration::from_secs(60), fake_tenants);
+            Ok(FeatureResolver {
+                inner: Some(inner),
+                internal_properties: Some(internal_properties),
+            })
         } else {
-            Ok(FeatureResolver { inner: None })
+            Ok(FeatureResolver {
+                inner: None,
+                internal_properties: None,
+            })
         }
     }
 
+    fn collect_properties_inner(
+        tenant_id: String,
+        internal_properties: Option<&HashMap<String, PostHogFlagFilterPropertyValue>>,
+    ) -> HashMap<String, PostHogFlagFilterPropertyValue> {
+        let mut properties = HashMap::new();
+        if let Some(internal_properties) = internal_properties {
+            for (key, value) in internal_properties.iter() {
+                properties.insert(key.clone(), value.clone());
+            }
+        }
+        properties.insert(
+            "tenant_id".to_string(),
+            PostHogFlagFilterPropertyValue::String(tenant_id),
+        );
+        properties
+    }
+
+    /// Collect all properties availble for the feature flag evaluation.
+    pub(crate) fn collect_properties(
+        &self,
+        tenant_id: TenantId,
+    ) -> HashMap<String, PostHogFlagFilterPropertyValue> {
+        Self::collect_properties_inner(tenant_id.to_string(), self.internal_properties.as_deref())
+    }
+
     /// Evaluate a multivariate feature flag. Currently, we do not support any properties.
     ///
     /// Error handling: the caller should inspect the error and decide the behavior when a feature flag
@@ -55,11 +191,24 @@ impl FeatureResolver {
         tenant_id: TenantId,
     ) -> Result<String, PostHogEvaluationError> {
         if let Some(inner) = &self.inner {
-            inner.feature_store().evaluate_multivariate(
+            let res = inner.feature_store().evaluate_multivariate(
                 flag_key,
                 &tenant_id.to_string(),
-                &HashMap::new(),
-            )
+                &self.collect_properties(tenant_id),
+            );
+            match &res {
+                Ok(value) => {
+                    FEATURE_FLAG_EVALUATION
+                        .with_label_values(&[flag_key, "ok", value])
+                        .inc();
+                }
+                Err(e) => {
+                    FEATURE_FLAG_EVALUATION
+                        .with_label_values(&[flag_key, "error", e.as_variant_str()])
+                        .inc();
+                }
+            }
+            res
         } else {
             Err(PostHogEvaluationError::NotAvailable(
                 "PostHog integration is not enabled".to_string(),
@@ -80,11 +229,34 @@ impl FeatureResolver {
         tenant_id: TenantId,
     ) -> Result<(), PostHogEvaluationError> {
         if let Some(inner) = &self.inner {
-            inner.feature_store().evaluate_boolean(
+            let res = inner.feature_store().evaluate_boolean(
                 flag_key,
                 &tenant_id.to_string(),
-                &HashMap::new(),
-            )
+                &self.collect_properties(tenant_id),
+            );
+            match &res {
+                Ok(()) => {
+                    FEATURE_FLAG_EVALUATION
+                        .with_label_values(&[flag_key, "ok", "true"])
+                        .inc();
+                }
+                Err(e) => {
+                    FEATURE_FLAG_EVALUATION
+                        .with_label_values(&[flag_key, "error", e.as_variant_str()])
+                        .inc();
+                }
+            }
+            res
+        } else {
+            Err(PostHogEvaluationError::NotAvailable(
+                "PostHog integration is not enabled".to_string(),
+            ))
+        }
+    }
+
+    pub fn is_feature_flag_boolean(&self, flag_key: &str) -> Result<bool, PostHogEvaluationError> {
+        if let Some(inner) = &self.inner {
+            inner.feature_store().is_feature_flag_boolean(flag_key)
         } else {
             Err(PostHogEvaluationError::NotAvailable(
                 "PostHog integration is not enabled".to_string(),

pageserver/src/http/routes.rs

@@ -43,6 +43,7 @@ use pageserver_api::models::{
 use pageserver_api::shard::{ShardCount, TenantShardId};
 use remote_storage::{DownloadError, GenericRemoteStorage, TimeTravelError};
 use scopeguard::defer;
+use serde_json::json;
 use tenant_size_model::svg::SvgBranchKind;
 use tenant_size_model::{SizeResult, StorageModel};
 use tokio::time::Instant;
@@ -3663,6 +3664,47 @@ async fn read_tar_eof(mut reader: (impl tokio::io::AsyncRead + Unpin)) -> anyhow
     Ok(())
 }
 
+async fn tenant_evaluate_feature_flag(
+    request: Request<Body>,
+    _cancel: CancellationToken,
+) -> Result<Response<Body>, ApiError> {
+    let tenant_shard_id: TenantShardId = parse_request_param(&request, "tenant_shard_id")?;
+    check_permission(&request, Some(tenant_shard_id.tenant_id))?;
+
+    let flag: String = must_parse_query_param(&request, "flag")?;
+    let as_type: String = must_parse_query_param(&request, "as")?;
+
+    let state = get_state(&request);
+
+    async {
+        let tenant = state
+            .tenant_manager
+            .get_attached_tenant_shard(tenant_shard_id)?;
+        let properties = tenant.feature_resolver.collect_properties(tenant_shard_id.tenant_id);
+        if as_type == "boolean" {
+            let result = tenant.feature_resolver.evaluate_boolean(&flag, tenant_shard_id.tenant_id);
+            let result = result.map(|_| true).map_err(|e| e.to_string());
+            json_response(StatusCode::OK, json!({ "result": result, "properties": properties }))
+        } else if as_type == "multivariate" {
+            let result = tenant.feature_resolver.evaluate_multivariate(&flag, tenant_shard_id.tenant_id).map_err(|e| e.to_string());
+            json_response(StatusCode::OK, json!({ "result": result, "properties": properties }))
+        } else {
+            // Auto infer the type of the feature flag.
+            let is_boolean = tenant.feature_resolver.is_feature_flag_boolean(&flag).map_err(|e| ApiError::InternalServerError(anyhow::anyhow!("{e}")))?;
+            if is_boolean {
+                let result = tenant.feature_resolver.evaluate_boolean(&flag, tenant_shard_id.tenant_id);
+                let result = result.map(|_| true).map_err(|e| e.to_string());
+                json_response(StatusCode::OK, json!({ "result": result, "properties": properties }))
+            } else {
+                let result = tenant.feature_resolver.evaluate_multivariate(&flag, tenant_shard_id.tenant_id).map_err(|e| e.to_string());
+                json_response(StatusCode::OK, json!({ "result": result, "properties": properties }))
+            }
+        }
+    }
+    .instrument(info_span!("tenant_evaluate_feature_flag", tenant_id = %tenant_shard_id.tenant_id, shard_id = %tenant_shard_id.shard_slug()))
+    .await
+}
+
 /// Common functionality of all the HTTP API handlers.
 ///
 /// - Adds a tracing span to each request (by `request_span`)
@@ -4039,5 +4081,8 @@ pub fn make_router(
             "/v1/tenant/:tenant_shard_id/timeline/:timeline_id/activate_post_import",
             |r| api_handler(r, activate_post_import_handler),
         )
+        .get("/v1/tenant/:tenant_shard_id/feature_flag", |r| {
+            api_handler(r, tenant_evaluate_feature_flag)
+        })
         .any(handler_404))
 }

pageserver/src/metrics.rs

@@ -15,6 +15,7 @@ use metrics::{
     register_int_gauge, register_int_gauge_vec, register_uint_gauge, register_uint_gauge_vec,
 };
 use once_cell::sync::Lazy;
+use pageserver_api::config::defaults::DEFAULT_MAX_GET_VECTORED_KEYS;
 use pageserver_api::config::{
     PageServicePipeliningConfig, PageServicePipeliningConfigPipelined,
     PageServiceProtocolPipelinedBatchingStrategy, PageServiceProtocolPipelinedExecutionStrategy,
@@ -32,7 +33,6 @@ use crate::config::PageServerConf;
 use crate::context::{PageContentKind, RequestContext};
 use crate::pgdatadir_mapping::DatadirModificationStats;
 use crate::task_mgr::TaskKind;
-use crate::tenant::Timeline;
 use crate::tenant::layer_map::LayerMap;
 use crate::tenant::mgr::TenantSlot;
 use crate::tenant::storage_layer::{InMemoryLayer, PersistentLayerDesc};
@@ -446,6 +446,15 @@ static PAGE_CACHE_ERRORS: Lazy<IntCounterVec> = Lazy::new(|| {
     .expect("failed to define a metric")
 });
 
+pub(crate) static FEATURE_FLAG_EVALUATION: Lazy<CounterVec> = Lazy::new(|| {
+    register_counter_vec!(
+        "pageserver_feature_flag_evaluation",
+        "Number of times a feature flag is evaluated",
+        &["flag_key", "status", "value"],
+    )
+    .unwrap()
+});
+
 #[derive(IntoStaticStr)]
 #[strum(serialize_all = "kebab_case")]
 pub(crate) enum PageCacheErrorKind {
@@ -1044,6 +1053,15 @@ pub(crate) static TENANT_STATE_METRIC: Lazy<UIntGaugeVec> = Lazy::new(|| {
     .expect("Failed to register pageserver_tenant_states_count metric")
 });
 
+pub(crate) static TIMELINE_STATE_METRIC: Lazy<UIntGaugeVec> = Lazy::new(|| {
+    register_uint_gauge_vec!(
+        "pageserver_timeline_states_count",
+        "Count of timelines per state",
+        &["state"]
+    )
+    .expect("Failed to register pageserver_timeline_states_count metric")
+});
+
 /// A set of broken tenants.
 ///
 /// These are expected to be so rare that a set is fine. Set as in a new timeseries per each broken
@@ -1312,11 +1330,44 @@ impl EvictionsWithLowResidenceDuration {
 //
 // Roughly logarithmic scale.
 const STORAGE_IO_TIME_BUCKETS: &[f64] = &[
-    0.000030, // 30 usec
-    0.001000, // 1000 usec
-    0.030,    // 30 ms
-    1.000,    // 1000 ms
-    30.000,   // 30000 ms
+    0.00005,  // 50us
+    0.00006,  // 60us
+    0.00007,  // 70us
+    0.00008,  // 80us
+    0.00009,  // 90us
+    0.0001,   // 100us
+    0.000110, // 110us
+    0.000120, // 120us
+    0.000130, // 130us
+    0.000140, // 140us
+    0.000150, // 150us
+    0.000160, // 160us
+    0.000170, // 170us
+    0.000180, // 180us
+    0.000190, // 190us
+    0.000200, // 200us
+    0.000210, // 210us
+    0.000220, // 220us
+    0.000230, // 230us
+    0.000240, // 240us
+    0.000250, // 250us
+    0.000300, // 300us
+    0.000350, // 350us
+    0.000400, // 400us
+    0.000450, // 450us
+    0.000500, // 500us
+    0.000600, // 600us
+    0.000700, // 700us
+    0.000800, // 800us
+    0.000900, // 900us
+    0.001000, // 1ms
+    0.002000, // 2ms
+    0.003000, // 3ms
+    0.004000, // 4ms
+    0.005000, // 5ms
+    0.01000,  // 10ms
+    0.02000,  // 20ms
+    0.05000,  // 50ms
 ];
 
 /// VirtualFile fs operation variants.
@@ -1906,7 +1957,7 @@ static SMGR_QUERY_TIME_GLOBAL: Lazy<HistogramVec> = Lazy::new(|| {
 });
 
 static PAGE_SERVICE_BATCH_SIZE_BUCKETS_GLOBAL: Lazy<Vec<f64>> = Lazy::new(|| {
-    (1..=u32::try_from(Timeline::MAX_GET_VECTORED_KEYS).unwrap())
+    (1..=u32::try_from(DEFAULT_MAX_GET_VECTORED_KEYS).unwrap())
         .map(|v| v.into())
         .collect()
 });
@@ -1924,7 +1975,7 @@ static PAGE_SERVICE_BATCH_SIZE_BUCKETS_PER_TIMELINE: Lazy<Vec<f64>> = Lazy::new(
     let mut buckets = Vec::new();
     for i in 0.. {
         let bucket = 1 << i;
-        if bucket > u32::try_from(Timeline::MAX_GET_VECTORED_KEYS).unwrap() {
+        if bucket > u32::try_from(DEFAULT_MAX_GET_VECTORED_KEYS).unwrap() {
             break;
         }
         buckets.push(bucket.into());
@@ -2813,7 +2864,6 @@ pub(crate) struct WalIngestMetrics {
     pub(crate) records_received: IntCounter,
     pub(crate) records_observed: IntCounter,
     pub(crate) records_committed: IntCounter,
-    pub(crate) records_filtered: IntCounter,
     pub(crate) values_committed_metadata_images: IntCounter,
     pub(crate) values_committed_metadata_deltas: IntCounter,
     pub(crate) values_committed_data_images: IntCounter,
@@ -2869,11 +2919,6 @@ pub(crate) static WAL_INGEST: Lazy<WalIngestMetrics> = Lazy::new(|| {
         "Number of WAL records which resulted in writes to pageserver storage"
     )
     .expect("failed to define a metric"),
-    records_filtered: register_int_counter!(
-        "pageserver_wal_ingest_records_filtered",
-        "Number of WAL records filtered out due to sharding"
-    )
-    .expect("failed to define a metric"),
     values_committed_metadata_images: values_committed.with_label_values(&["metadata", "image"]),
     values_committed_metadata_deltas: values_committed.with_label_values(&["metadata", "delta"]),
     values_committed_data_images: values_committed.with_label_values(&["data", "image"]),
@@ -3289,6 +3334,8 @@ impl TimelineMetrics {
                 &timeline_id,
             );
 
+        TIMELINE_STATE_METRIC.with_label_values(&["active"]).inc();
+
         TimelineMetrics {
             tenant_id,
             shard_id,
@@ -3443,6 +3490,8 @@ impl TimelineMetrics {
             return;
         }
 
+        TIMELINE_STATE_METRIC.with_label_values(&["active"]).dec();
+
         let tenant_id = &self.tenant_id;
         let timeline_id = &self.timeline_id;
         let shard_id = &self.shard_id;

pageserver/src/pgdatadir_mapping.rs

@@ -431,10 +431,10 @@ impl Timeline {
                         GetVectoredError::InvalidLsn(e) => {
                             Err(anyhow::anyhow!("invalid LSN: {e:?}").into())
                         }
-                        // NB: this should never happen in practice because we limit MAX_GET_VECTORED_KEYS
+                        // NB: this should never happen in practice because we limit batch size to be smaller than max_get_vectored_keys
                         // TODO: we can prevent this error class by moving this check into the type system
-                        GetVectoredError::Oversized(err) => {
-                            Err(anyhow::anyhow!("batching oversized: {err:?}").into())
+                        GetVectoredError::Oversized(err, max) => {
+                            Err(anyhow::anyhow!("batching oversized: {err} > {max}").into())
                         }
                     };
 
@@ -471,8 +471,19 @@ impl Timeline {
 
         let rels = self.list_rels(spcnode, dbnode, version, ctx).await?;
 
+        if rels.is_empty() {
+            return Ok(0);
+        }
+
+        // Pre-deserialize the rel directory to avoid duplicated work in `get_relsize_cached`.
+        let reldir_key = rel_dir_to_key(spcnode, dbnode);
+        let buf = version.get(self, reldir_key, ctx).await?;
+        let reldir = RelDirectory::des(&buf)?;
+
         for rel in rels {
-            let n_blocks = self.get_rel_size(rel, version, ctx).await?;
+            let n_blocks = self
+                .get_rel_size_in_reldir(rel, version, Some((reldir_key, &reldir)), ctx)
+                .await?;
             total_blocks += n_blocks as usize;
         }
         Ok(total_blocks)
@@ -487,6 +498,19 @@ impl Timeline {
         tag: RelTag,
         version: Version<'_>,
         ctx: &RequestContext,
+    ) -> Result<BlockNumber, PageReconstructError> {
+        self.get_rel_size_in_reldir(tag, version, None, ctx).await
+    }
+
+    /// Get size of a relation file. The relation must exist, otherwise an error is returned.
+    ///
+    /// See [`Self::get_rel_exists_in_reldir`] on why we need `deserialized_reldir_v1`.
+    pub(crate) async fn get_rel_size_in_reldir(
+        &self,
+        tag: RelTag,
+        version: Version<'_>,
+        deserialized_reldir_v1: Option<(Key, &RelDirectory)>,
+        ctx: &RequestContext,
     ) -> Result<BlockNumber, PageReconstructError> {
         if tag.relnode == 0 {
             return Err(PageReconstructError::Other(
@@ -499,7 +523,9 @@ impl Timeline {
         }
 
         if (tag.forknum == FSM_FORKNUM || tag.forknum == VISIBILITYMAP_FORKNUM)
-            && !self.get_rel_exists(tag, version, ctx).await?
+            && !self
+                .get_rel_exists_in_reldir(tag, version, deserialized_reldir_v1, ctx)
+                .await?
         {
             // FIXME: Postgres sometimes calls smgrcreate() to create
             // FSM, and smgrnblocks() on it immediately afterwards,
@@ -521,11 +547,28 @@ impl Timeline {
     ///
     /// Only shard 0 has a full view of the relations. Other shards only know about relations that
     /// the shard stores pages for.
+    ///
     pub(crate) async fn get_rel_exists(
         &self,
         tag: RelTag,
         version: Version<'_>,
         ctx: &RequestContext,
+    ) -> Result<bool, PageReconstructError> {
+        self.get_rel_exists_in_reldir(tag, version, None, ctx).await
+    }
+
+    /// Does the relation exist? With a cached deserialized `RelDirectory`.
+    ///
+    /// There are some cases where the caller loops across all relations. In that specific case,
+    /// the caller should obtain the deserialized `RelDirectory` first and then call this function
+    /// to avoid duplicated work of deserliazation. This is a hack and should be removed by introducing
+    /// a new API (e.g., `get_rel_exists_batched`).
+    pub(crate) async fn get_rel_exists_in_reldir(
+        &self,
+        tag: RelTag,
+        version: Version<'_>,
+        deserialized_reldir_v1: Option<(Key, &RelDirectory)>,
+        ctx: &RequestContext,
     ) -> Result<bool, PageReconstructError> {
         if tag.relnode == 0 {
             return Err(PageReconstructError::Other(
@@ -568,6 +611,17 @@ impl Timeline {
         // fetch directory listing (old)
 
         let key = rel_dir_to_key(tag.spcnode, tag.dbnode);
+
+        if let Some((cached_key, dir)) = deserialized_reldir_v1 {
+            if cached_key == key {
+                return Ok(dir.rels.contains(&(tag.relnode, tag.forknum)));
+            } else if cfg!(test) || cfg!(feature = "testing") {
+                panic!("cached reldir key mismatch: {cached_key} != {key}");
+            } else {
+                warn!("cached reldir key mismatch: {cached_key} != {key}");
+            }
+            // Fallback to reading the directory from the datadir.
+        }
         let buf = version.get(self, key, ctx).await?;
 
         let dir = RelDirectory::des(&buf)?;
@@ -665,7 +719,7 @@ impl Timeline {
 
         let batches = keyspace.partition(
             self.get_shard_identity(),
-            Timeline::MAX_GET_VECTORED_KEYS * BLCKSZ as u64,
+            self.conf.max_get_vectored_keys.get() as u64 * BLCKSZ as u64,
         );
 
         let io_concurrency = IoConcurrency::spawn_from_conf(
@@ -905,7 +959,7 @@ impl Timeline {
 
             let batches = keyspace.partition(
                 self.get_shard_identity(),
-                Timeline::MAX_GET_VECTORED_KEYS * BLCKSZ as u64,
+                self.conf.max_get_vectored_keys.get() as u64 * BLCKSZ as u64,
             );
 
             let io_concurrency = IoConcurrency::spawn_from_conf(

pageserver/src/tenant.rs

@@ -89,7 +89,8 @@ use crate::l0_flush::L0FlushGlobalState;
 use crate::metrics::{
     BROKEN_TENANTS_SET, CIRCUIT_BREAKERS_BROKEN, CIRCUIT_BREAKERS_UNBROKEN, CONCURRENT_INITDBS,
     INITDB_RUN_TIME, INITDB_SEMAPHORE_ACQUISITION_TIME, TENANT, TENANT_OFFLOADED_TIMELINES,
-    TENANT_STATE_METRIC, TENANT_SYNTHETIC_SIZE_METRIC, remove_tenant_metrics,
+    TENANT_STATE_METRIC, TENANT_SYNTHETIC_SIZE_METRIC, TIMELINE_STATE_METRIC,
+    remove_tenant_metrics,
 };
 use crate::task_mgr::TaskKind;
 use crate::tenant::config::LocationMode;
@@ -383,7 +384,7 @@ pub struct TenantShard {
 
     l0_flush_global_state: L0FlushGlobalState,
 
-    feature_resolver: FeatureResolver,
+    pub(crate) feature_resolver: FeatureResolver,
 }
 impl std::fmt::Debug for TenantShard {
     fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
@@ -544,6 +545,28 @@ pub struct OffloadedTimeline {
 
     /// Part of the `OffloadedTimeline` object's lifecycle: this needs to be set before we drop it
     pub deleted_from_ancestor: AtomicBool,
+
+    _metrics_guard: OffloadedTimelineMetricsGuard,
+}
+
+/// Increases the offloaded timeline count metric when created, and decreases when dropped.
+struct OffloadedTimelineMetricsGuard;
+
+impl OffloadedTimelineMetricsGuard {
+    fn new() -> Self {
+        TIMELINE_STATE_METRIC
+            .with_label_values(&["offloaded"])
+            .inc();
+        Self
+    }
+}
+
+impl Drop for OffloadedTimelineMetricsGuard {
+    fn drop(&mut self) {
+        TIMELINE_STATE_METRIC
+            .with_label_values(&["offloaded"])
+            .dec();
+    }
 }
 
 impl OffloadedTimeline {
@@ -576,6 +599,8 @@ impl OffloadedTimeline {
 
             delete_progress: timeline.delete_progress.clone(),
             deleted_from_ancestor: AtomicBool::new(false),
+
+            _metrics_guard: OffloadedTimelineMetricsGuard::new(),
         })
     }
     fn from_manifest(tenant_shard_id: TenantShardId, manifest: &OffloadedTimelineManifest) -> Self {
@@ -595,6 +620,7 @@ impl OffloadedTimeline {
             archived_at,
             delete_progress: TimelineDeleteProgress::default(),
             deleted_from_ancestor: AtomicBool::new(false),
+            _metrics_guard: OffloadedTimelineMetricsGuard::new(),
         }
     }
     fn manifest(&self) -> OffloadedTimelineManifest {
@@ -5832,6 +5858,7 @@ pub(crate) mod harness {
         pub conf: &'static PageServerConf,
         pub tenant_conf: pageserver_api::models::TenantConfig,
         pub tenant_shard_id: TenantShardId,
+        pub shard_identity: ShardIdentity,
         pub generation: Generation,
         pub shard: ShardIndex,
         pub remote_storage: GenericRemoteStorage,
@@ -5899,6 +5926,7 @@ pub(crate) mod harness {
                 conf,
                 tenant_conf,
                 tenant_shard_id,
+                shard_identity,
                 generation,
                 shard,
                 remote_storage,
@@ -5960,8 +5988,7 @@ pub(crate) mod harness {
                     &ShardParameters::default(),
                 ))
                 .unwrap(),
-                // This is a legacy/test code path: sharding isn't supported here.
-                ShardIdentity::unsharded(),
+                self.shard_identity,
                 Some(walredo_mgr),
                 self.tenant_shard_id,
                 self.remote_storage.clone(),
@@ -6083,6 +6110,7 @@ mod tests {
     use timeline::compaction::{KeyHistoryRetention, KeyLogAtLsn};
     use timeline::{CompactOptions, DeltaLayerTestDesc, VersionedKeySpaceQuery};
     use utils::id::TenantId;
+    use utils::shard::{ShardCount, ShardNumber};
 
     use super::*;
     use crate::DEFAULT_PG_VERSION;
@@ -7195,7 +7223,7 @@ mod tests {
             let end = desc
                 .key_range
                 .start
-                .add(Timeline::MAX_GET_VECTORED_KEYS.try_into().unwrap());
+                .add(tenant.conf.max_get_vectored_keys.get() as u32);
             reads.push(KeySpace {
                 ranges: vec![start..end],
             });
@@ -9418,6 +9446,77 @@ mod tests {
         Ok(())
     }
 
+    #[tokio::test]
+    async fn test_failed_flush_should_not_update_disk_consistent_lsn() -> anyhow::Result<()> {
+        //
+        // Setup
+        //
+        let harness = TenantHarness::create_custom(
+            "test_failed_flush_should_not_upload_disk_consistent_lsn",
+            pageserver_api::models::TenantConfig::default(),
+            TenantId::generate(),
+            ShardIdentity::new(ShardNumber(0), ShardCount(4), ShardStripeSize(128)).unwrap(),
+            Generation::new(1),
+        )
+        .await?;
+        let (tenant, ctx) = harness.load().await;
+
+        let timeline = tenant
+            .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
+            .await?;
+        assert_eq!(timeline.get_shard_identity().count, ShardCount(4));
+        let mut writer = timeline.writer().await;
+        writer
+            .put(
+                *TEST_KEY,
+                Lsn(0x20),
+                &Value::Image(test_img("foo at 0x20")),
+                &ctx,
+            )
+            .await?;
+        writer.finish_write(Lsn(0x20));
+        drop(writer);
+        timeline.freeze_and_flush().await.unwrap();
+
+        timeline.remote_client.wait_completion().await.unwrap();
+        let disk_consistent_lsn = timeline.get_disk_consistent_lsn();
+        let remote_consistent_lsn = timeline.get_remote_consistent_lsn_projected();
+        assert_eq!(Some(disk_consistent_lsn), remote_consistent_lsn);
+
+        //
+        // Test
+        //
+
+        let mut writer = timeline.writer().await;
+        writer
+            .put(
+                *TEST_KEY,
+                Lsn(0x30),
+                &Value::Image(test_img("foo at 0x30")),
+                &ctx,
+            )
+            .await?;
+        writer.finish_write(Lsn(0x30));
+        drop(writer);
+
+        fail::cfg(
+            "flush-layer-before-update-remote-consistent-lsn",
+            "return()",
+        )
+        .unwrap();
+
+        let flush_res = timeline.freeze_and_flush().await;
+        // if flush failed, the disk/remote consistent LSN should not be updated
+        assert!(flush_res.is_err());
+        assert_eq!(disk_consistent_lsn, timeline.get_disk_consistent_lsn());
+        assert_eq!(
+            remote_consistent_lsn,
+            timeline.get_remote_consistent_lsn_projected()
+        );
+
+        Ok(())
+    }
+
     #[cfg(feature = "testing")]
     #[tokio::test]
     async fn test_simple_bottom_most_compaction_deltas_1() -> anyhow::Result<()> {
@@ -11187,11 +11286,11 @@ mod tests {
                 let mut keyspaces_at_lsn: HashMap<Lsn, KeySpaceRandomAccum> = HashMap::default();
                 let mut used_keys: HashSet<Key> = HashSet::default();
 
-                while used_keys.len() < Timeline::MAX_GET_VECTORED_KEYS as usize {
+                while used_keys.len() < tenant.conf.max_get_vectored_keys.get() {
                     let selected_lsn = interesting_lsns.choose(&mut random).expect("not empty");
                     let mut selected_key = start_key.add(random.gen_range(0..KEY_DIMENSION_SIZE));
 
-                    while used_keys.len() < Timeline::MAX_GET_VECTORED_KEYS as usize {
+                    while used_keys.len() < tenant.conf.max_get_vectored_keys.get() {
                         if used_keys.contains(&selected_key)
                             || selected_key >= start_key.add(KEY_DIMENSION_SIZE)
                         {

pageserver/src/tenant/mgr.rs

@@ -1671,7 +1671,12 @@ impl TenantManager {
             }
         }
 
-        // Phase 5: Shut down the parent shard, and erase it from disk
+        // Phase 5: Shut down the parent shard. We leave it on disk in case the split fails and we
+        // have to roll back to the parent shard, avoiding a cold start. It will be cleaned up once
+        // the storage controller commits the split, or if all else fails, on the next restart.
+        //
+        // TODO: We don't flush the ephemeral layer here, because the split is likely to succeed and
+        // catching up the parent should be reasonably quick. Consider using FreezeAndFlush instead.
         let (_guard, progress) = completion::channel();
         match parent.shutdown(progress, ShutdownMode::Hard).await {
             Ok(()) => {}
@@ -1679,11 +1684,6 @@ impl TenantManager {
                 other.wait().await;
             }
         }
-        let local_tenant_directory = self.conf.tenant_path(&tenant_shard_id);
-        let tmp_path = safe_rename_tenant_dir(&local_tenant_directory)
-            .await
-            .with_context(|| format!("local tenant directory {local_tenant_directory:?} rename"))?;
-        self.background_purges.spawn(tmp_path);
 
         fail::fail_point!("shard-split-pre-finish", |_| Err(anyhow::anyhow!(
             "failpoint"
@@ -1846,42 +1846,70 @@ impl TenantManager {
         shutdown_all_tenants0(self.tenants).await
     }
 
+    /// Detaches a tenant, and removes its local files asynchronously.
+    ///
+    /// File removal is idempotent: even if the tenant has already been removed, this will still
+    /// remove any local files. This is used during shard splits, where we leave the parent shard's
+    /// files around in case we have to roll back the split.
     pub(crate) async fn detach_tenant(
         &self,
         conf: &'static PageServerConf,
         tenant_shard_id: TenantShardId,
         deletion_queue_client: &DeletionQueueClient,
     ) -> Result<(), TenantStateError> {
-        let tmp_path = self
+        if let Some(tmp_path) = self
             .detach_tenant0(conf, tenant_shard_id, deletion_queue_client)
-            .await?;
-        self.background_purges.spawn(tmp_path);
+            .await?
+        {
+            self.background_purges.spawn(tmp_path);
+        }
 
         Ok(())
     }
 
+    /// Detaches a tenant. This renames the tenant directory to a temporary path and returns it,
+    /// allowing the caller to delete it asynchronously. Returns None if the dir is already removed.
     async fn detach_tenant0(
         &self,
         conf: &'static PageServerConf,
         tenant_shard_id: TenantShardId,
         deletion_queue_client: &DeletionQueueClient,
-    ) -> Result<Utf8PathBuf, TenantStateError> {
+    ) -> Result<Option<Utf8PathBuf>, TenantStateError> {
         let tenant_dir_rename_operation = |tenant_id_to_clean: TenantShardId| async move {
             let local_tenant_directory = conf.tenant_path(&tenant_id_to_clean);
+            if !tokio::fs::try_exists(&local_tenant_directory).await? {
+                // If the tenant directory doesn't exist, it's already cleaned up.
+                return Ok(None);
+            }
             safe_rename_tenant_dir(&local_tenant_directory)
                 .await
                 .with_context(|| {
                     format!("local tenant directory {local_tenant_directory:?} rename")
                 })
+                .map(Some)
         };
 
-        let removal_result = remove_tenant_from_memory(
+        let mut removal_result = remove_tenant_from_memory(
             self.tenants,
             tenant_shard_id,
             tenant_dir_rename_operation(tenant_shard_id),
         )
         .await;
 
+        // If the tenant was not found, it was likely already removed. Attempt to remove the tenant
+        // directory on disk anyway. For example, during shard splits, we shut down and remove the
+        // parent shard, but leave its directory on disk in case we have to roll back the split.
+        //
+        // TODO: it would be better to leave the parent shard attached until the split is committed.
+        // This will be needed by the gRPC page service too, such that a compute can continue to
+        // read from the parent shard until it's notified about the new child shards. See:
+        // <https://github.com/neondatabase/neon/issues/11728>.
+        if let Err(TenantStateError::SlotError(TenantSlotError::NotFound(_))) = removal_result {
+            removal_result = tenant_dir_rename_operation(tenant_shard_id)
+                .await
+                .map_err(TenantStateError::Other);
+        }
+
         // Flush pending deletions, so that they have a good chance of passing validation
         // before this tenant is potentially re-attached elsewhere.
         deletion_queue_client.flush_advisory();

pageserver/src/tenant/timeline.rs

@@ -817,8 +817,8 @@ pub(crate) enum GetVectoredError {
     #[error("timeline shutting down")]
     Cancelled,
 
-    #[error("requested too many keys: {0} > {}", Timeline::MAX_GET_VECTORED_KEYS)]
-    Oversized(u64),
+    #[error("requested too many keys: {0} > {1}")]
+    Oversized(u64, u64),
 
     #[error("requested at invalid LSN: {0}")]
     InvalidLsn(Lsn),
@@ -950,6 +950,18 @@ pub(crate) enum WaitLsnError {
     Timeout(String),
 }
 
+impl From<WaitLsnError> for tonic::Status {
+    fn from(err: WaitLsnError) -> Self {
+        use tonic::Code;
+        let code = match &err {
+            WaitLsnError::Timeout(_) => Code::Internal,
+            WaitLsnError::BadState(_) => Code::Internal,
+            WaitLsnError::Shutdown => Code::Unavailable,
+        };
+        tonic::Status::new(code, err.to_string())
+    }
+}
+
 // The impls below achieve cancellation mapping for errors.
 // Perhaps there's a way of achieving this with less cruft.
 
@@ -1007,7 +1019,7 @@ impl From<GetVectoredError> for PageReconstructError {
         match e {
             GetVectoredError::Cancelled => PageReconstructError::Cancelled,
             GetVectoredError::InvalidLsn(_) => PageReconstructError::Other(anyhow!("Invalid LSN")),
-            err @ GetVectoredError::Oversized(_) => PageReconstructError::Other(err.into()),
+            err @ GetVectoredError::Oversized(_, _) => PageReconstructError::Other(err.into()),
             GetVectoredError::MissingKey(err) => PageReconstructError::MissingKey(err),
             GetVectoredError::GetReadyAncestorError(err) => PageReconstructError::from(err),
             GetVectoredError::Other(err) => PageReconstructError::Other(err),
@@ -1043,8 +1055,8 @@ pub(crate) enum WaitLsnWaiter<'a> {
 /// Argument to [`Timeline::shutdown`].
 #[derive(Debug, Clone, Copy)]
 pub(crate) enum ShutdownMode {
-    /// Graceful shutdown, may do a lot of I/O as we flush any open layers to disk and then
-    /// also to remote storage.  This method can easily take multiple seconds for a busy timeline.
+    /// Graceful shutdown, may do a lot of I/O as we flush any open layers to disk. This method can
+    /// take multiple seconds for a busy timeline.
     ///
     /// While we are flushing, we continue to accept read I/O for LSNs ingested before
     /// the call to [`Timeline::shutdown`].
@@ -1187,7 +1199,6 @@ impl Timeline {
         }
     }
 
-    pub(crate) const MAX_GET_VECTORED_KEYS: u64 = 32;
     pub(crate) const LAYERS_VISITED_WARN_THRESHOLD: u32 = 100;
 
     /// Look up multiple page versions at a given LSN
@@ -1202,9 +1213,12 @@ impl Timeline {
     ) -> Result<BTreeMap<Key, Result<Bytes, PageReconstructError>>, GetVectoredError> {
         let total_keyspace = query.total_keyspace();
 
-        let key_count = total_keyspace.total_raw_size().try_into().unwrap();
-        if key_count > Timeline::MAX_GET_VECTORED_KEYS {
-            return Err(GetVectoredError::Oversized(key_count));
+        let key_count = total_keyspace.total_raw_size();
+        if key_count > self.conf.max_get_vectored_keys.get() {
+            return Err(GetVectoredError::Oversized(
+                key_count as u64,
+                self.conf.max_get_vectored_keys.get() as u64,
+            ));
         }
 
         for range in &total_keyspace.ranges {
@@ -2492,6 +2506,13 @@ impl Timeline {
             // Preparing basebackup doesn't make sense for shards other than shard zero.
             return;
         }
+        if !self.is_active() {
+            // May happen during initial timeline creation.
+            // Such timeline is not in the global timeline map yet,
+            // so basebackup cache will not be able to find it.
+            // TODO(diko): We can prepare such timelines in finish_creation().
+            return;
+        }
 
         let res = self
             .basebackup_prepare_sender
@@ -2831,21 +2852,6 @@ impl Timeline {
             )
     }
 
-    /// Resolve the effective WAL receiver protocol to use for this tenant.
-    ///
-    /// Priority order is:
-    /// 1. Tenant config override
-    /// 2. Default value for tenant config override
-    /// 3. Pageserver config override
-    /// 4. Pageserver config default
-    pub fn resolve_wal_receiver_protocol(&self) -> PostgresClientProtocol {
-        let tenant_conf = self.tenant_conf.load().tenant_conf.clone();
-        tenant_conf
-            .wal_receiver_protocol_override
-            .or(self.conf.default_tenant_conf.wal_receiver_protocol_override)
-            .unwrap_or(self.conf.wal_receiver_protocol)
-    }
-
     pub(super) fn tenant_conf_updated(&self, new_conf: &AttachedTenantConf) {
         // NB: Most tenant conf options are read by background loops, so,
         // changes will automatically be picked up.
@@ -3201,10 +3207,16 @@ impl Timeline {
             guard.is_none(),
             "multiple launches / re-launches of WAL receiver are not supported"
         );
+
+        let protocol = PostgresClientProtocol::Interpreted {
+            format: utils::postgres_client::InterpretedFormat::Protobuf,
+            compression: Some(utils::postgres_client::Compression::Zstd { level: 1 }),
+        };
+
         *guard = Some(WalReceiver::start(
             Arc::clone(self),
             WalReceiverConf {
-                protocol: self.resolve_wal_receiver_protocol(),
+                protocol,
                 wal_connect_timeout,
                 lagging_wal_timeout,
                 max_lsn_wal_lag,
@@ -4767,7 +4779,10 @@ impl Timeline {
                     || !flushed_to_lsn.is_valid()
             );
 
-            if flushed_to_lsn < frozen_to_lsn && self.shard_identity.count.count() > 1 {
+            if flushed_to_lsn < frozen_to_lsn
+                && self.shard_identity.count.count() > 1
+                && result.is_ok()
+            {
                 // If our layer flushes didn't carry disk_consistent_lsn up to the `to_lsn` advertised
                 // to us via layer_flush_start_rx, then advance it here.
                 //
@@ -4946,6 +4961,10 @@ impl Timeline {
             return Err(FlushLayerError::Cancelled);
         }
 
+        fail_point!("flush-layer-before-update-remote-consistent-lsn", |_| {
+            Err(FlushLayerError::Other(anyhow!("failpoint").into()))
+        });
+
         let disk_consistent_lsn = Lsn(lsn_range.end.0 - 1);
 
         // The new on-disk layers are now in the layer map. We can remove the
@@ -5251,7 +5270,7 @@ impl Timeline {
                 key = key.next();
 
                 // Maybe flush `key_rest_accum`
-                if key_request_accum.raw_size() >= Timeline::MAX_GET_VECTORED_KEYS
+                if key_request_accum.raw_size() >= self.conf.max_get_vectored_keys.get() as u64
                     || (last_key_in_range && key_request_accum.raw_size() > 0)
                 {
                     let query =

pageserver/src/tenant/timeline/import_pgdata.rs

@@ -106,6 +106,8 @@ pub async fn doit(
                 );
             }
 
+            tracing::info!("Import plan executed. Flushing remote changes and notifying storcon");
+
             timeline
                 .remote_client
                 .schedule_index_upload_for_file_changes()?;
@@ -199,8 +201,8 @@ async fn prepare_import(
         .await;
         match res {
             Ok(_) => break,
-            Err(err) => {
-                info!(?err, "indefinitely waiting for pgdata to finish");
+            Err(_err) => {
+                info!("indefinitely waiting for pgdata to finish");
                 if tokio::time::timeout(std::time::Duration::from_secs(10), cancel.cancelled())
                     .await
                     .is_ok()

pageserver/src/tenant/timeline/import_pgdata/flow.rs

@@ -11,19 +11,7 @@
 //! - => S3 as the source for the PGDATA instead of local filesystem
 //!
 //! TODOs before productionization:
-//! - ChunkProcessingJob size / ImportJob::total_size does not account for sharding.
-//!   => produced image layers likely too small.
 //! - ChunkProcessingJob should cut up an ImportJob to hit exactly target image layer size.
-//! - asserts / unwraps need to be replaced with errors
-//! - don't trust remote objects will be small (=prevent OOMs in those cases)
-//!     - limit all in-memory buffers in size, or download to disk and read from there
-//! - limit task concurrency
-//! - generally play nice with other tenants in the system
-//!   - importbucket is different bucket than main pageserver storage, so, should be fine wrt S3 rate limits
-//!   - but concerns like network bandwidth, local disk write bandwidth, local disk capacity, etc
-//! - integrate with layer eviction system
-//! - audit for Tenant::cancel nor Timeline::cancel responsivity
-//! - audit for Tenant/Timeline gate holding (we spawn tokio tasks during this flow!)
 //!
 //! An incomplete set of TODOs from the Hackathon:
 //! - version-specific CheckPointData (=> pgv abstraction, already exists for regular walingest)
@@ -44,7 +32,7 @@ use pageserver_api::key::{
     rel_dir_to_key, rel_size_to_key, relmap_file_key, slru_block_to_key, slru_dir_to_key,
     slru_segment_size_to_key,
 };
-use pageserver_api::keyspace::{contiguous_range_len, is_contiguous_range, singleton_range};
+use pageserver_api::keyspace::{ShardedRange, singleton_range};
 use pageserver_api::models::{ShardImportProgress, ShardImportProgressV1, ShardImportStatus};
 use pageserver_api::reltag::{RelTag, SlruKind};
 use pageserver_api::shard::ShardIdentity;
@@ -112,6 +100,7 @@ async fn run_v1(
                         .unwrap(),
                     import_job_concurrency: base.import_job_concurrency,
                     import_job_checkpoint_threshold: base.import_job_checkpoint_threshold,
+                    import_job_max_byte_range_size: base.import_job_max_byte_range_size,
                 }
             }
             None => timeline.conf.timeline_import_config.clone(),
@@ -142,7 +131,15 @@ async fn run_v1(
 
     pausable_failpoint!("import-timeline-pre-execute-pausable");
 
+    let jobs_count = import_progress.as_ref().map(|p| p.jobs);
     let start_from_job_idx = import_progress.map(|progress| progress.completed);
+
+    tracing::info!(
+        start_from_job_idx=?start_from_job_idx,
+        jobs=?jobs_count,
+        "Executing import plan"
+    );
+
     plan.execute(timeline, start_from_job_idx, plan_hash, &import_config, ctx)
         .await
 }
@@ -167,6 +164,7 @@ impl Planner {
     /// This function is and must remain pure: given the same input, it will generate the same import plan.
     async fn plan(mut self, import_config: &TimelineImportConfig) -> anyhow::Result<Plan> {
         let pgdata_lsn = Lsn(self.control_file.control_file_data().checkPoint).align();
+        anyhow::ensure!(pgdata_lsn.is_valid());
 
         let datadir = PgDataDir::new(&self.storage).await?;
 
@@ -249,14 +247,22 @@ impl Planner {
         });
 
         // Assigns parts of key space to later parallel jobs
+        // Note: The image layers produced here may have gaps, meaning,
+        //       there is not an image for each key in the layer's key range.
+        //       The read path stops traversal at the first image layer, regardless
+        //       of whether a base image has been found for a key or not.
+        //       (Concept of sparse image layers doesn't exist.)
+        //       This behavior is exactly right for the base image layers we're producing here.
+        //       But, since no other place in the code currently produces image layers with gaps,
+        //       it seems noteworthy.
         let mut last_end_key = Key::MIN;
         let mut current_chunk = Vec::new();
         let mut current_chunk_size: usize = 0;
         let mut jobs = Vec::new();
         for task in std::mem::take(&mut self.tasks).into_iter() {
-            if current_chunk_size + task.total_size()
-                > import_config.import_job_soft_size_limit.into()
-            {
+            let task_size = task.total_size(&self.shard);
+            let projected_chunk_size = current_chunk_size.saturating_add(task_size);
+            if projected_chunk_size > import_config.import_job_soft_size_limit.into() {
                 let key_range = last_end_key..task.key_range().start;
                 jobs.push(ChunkProcessingJob::new(
                     key_range.clone(),
@@ -266,7 +272,7 @@ impl Planner {
                 last_end_key = key_range.end;
                 current_chunk_size = 0;
             }
-            current_chunk_size += task.total_size();
+            current_chunk_size = current_chunk_size.saturating_add(task_size);
             current_chunk.push(task);
         }
         jobs.push(ChunkProcessingJob::new(
@@ -436,6 +442,7 @@ impl Plan {
 
         let mut last_completed_job_idx = start_after_job_idx.unwrap_or(0);
         let checkpoint_every: usize = import_config.import_job_checkpoint_threshold.into();
+        let max_byte_range_size: usize = import_config.import_job_max_byte_range_size.into();
 
         // Run import jobs concurrently up to the limit specified by the pageserver configuration.
         // Note that we process completed futures in the oreder of insertion. This will be the
@@ -451,7 +458,7 @@ impl Plan {
 
                     work.push_back(tokio::task::spawn(async move {
                         let _permit = permit;
-                        let res = job.run(job_timeline, &ctx).await;
+                        let res = job.run(job_timeline, max_byte_range_size, &ctx).await;
                         (job_idx, res)
                     }));
                 },
@@ -466,6 +473,8 @@ impl Plan {
                             last_completed_job_idx = job_idx;
 
                             if last_completed_job_idx % checkpoint_every == 0 {
+                                tracing::info!(last_completed_job_idx, jobs=%jobs_in_plan, "Checkpointing import status");
+
                                 let progress = ShardImportProgressV1 {
                                     jobs: jobs_in_plan,
                                     completed: last_completed_job_idx,
@@ -604,18 +613,18 @@ impl PgDataDirDb {
                 };
 
                 let path = datadir_path.join(rel_tag.to_segfile_name(segno));
-                assert!(filesize % BLCKSZ as usize == 0); // TODO: this should result in an error
+                anyhow::ensure!(filesize % BLCKSZ as usize == 0);
                 let nblocks = filesize / BLCKSZ as usize;
 
-                PgDataDirDbFile {
+                Ok(PgDataDirDbFile {
                     path,
                     filesize,
                     rel_tag,
                     segno,
                     nblocks: Some(nblocks), // first non-cummulative sizes
-                }
+                })
             })
-            .collect();
+            .collect::<anyhow::Result<_, _>>()?;
 
         // Set cummulative sizes. Do all of that math here, so that later we could easier
         // parallelize over segments and know with which segments we need to write relsize
@@ -650,18 +659,29 @@ impl PgDataDirDb {
 trait ImportTask {
     fn key_range(&self) -> Range<Key>;
 
-    fn total_size(&self) -> usize {
-        // TODO: revisit this
-        if is_contiguous_range(&self.key_range()) {
-            contiguous_range_len(&self.key_range()) as usize * 8192
+    fn total_size(&self, shard_identity: &ShardIdentity) -> usize {
+        let range = ShardedRange::new(self.key_range(), shard_identity);
+        let page_count = range.page_count();
+        if page_count == u32::MAX {
+            tracing::warn!(
+                "Import task has non contiguous key range: {}..{}",
+                self.key_range().start,
+                self.key_range().end
+            );
+
+            // Tasks should operate on contiguous ranges. It is unexpected for
+            // ranges to violate this assumption. Calling code handles this by mapping
+            // any task on a non contiguous range to its own image layer.
+            usize::MAX
         } else {
-            u32::MAX as usize
+            page_count as usize * 8192
         }
     }
 
     async fn doit(
         self,
         layer_writer: &mut ImageLayerWriter,
+        max_byte_range_size: usize,
         ctx: &RequestContext,
     ) -> anyhow::Result<usize>;
 }
@@ -698,6 +718,7 @@ impl ImportTask for ImportSingleKeyTask {
     async fn doit(
         self,
         layer_writer: &mut ImageLayerWriter,
+        _max_byte_range_size: usize,
         ctx: &RequestContext,
     ) -> anyhow::Result<usize> {
         layer_writer.put_image(self.key, self.buf, ctx).await?;
@@ -751,6 +772,7 @@ impl ImportTask for ImportRelBlocksTask {
     async fn doit(
         self,
         layer_writer: &mut ImageLayerWriter,
+        max_byte_range_size: usize,
         ctx: &RequestContext,
     ) -> anyhow::Result<usize> {
         debug!("Importing relation file");
@@ -777,7 +799,7 @@ impl ImportTask for ImportRelBlocksTask {
                 assert_eq!(key.len(), 1);
                 assert!(!acc.is_empty());
                 assert!(acc_end > acc_start);
-                if acc_end == start /* TODO additional max range check here, to limit memory consumption per task to X */ {
+                if acc_end == start && end - acc_start <= max_byte_range_size {
                     acc.push(key.pop().unwrap());
                     Ok((acc, acc_start, end))
                 } else {
@@ -792,8 +814,8 @@ impl ImportTask for ImportRelBlocksTask {
                 .get_range(&self.path, range_start.into_u64(), range_end.into_u64())
                 .await?;
             let mut buf = Bytes::from(range_buf);
-            // TODO: batched writes
             for key in keys {
+                // The writer buffers writes internally
                 let image = buf.split_to(8192);
                 layer_writer.put_image(key, image, ctx).await?;
                 nimages += 1;
@@ -841,11 +863,15 @@ impl ImportTask for ImportSlruBlocksTask {
     async fn doit(
         self,
         layer_writer: &mut ImageLayerWriter,
+        _max_byte_range_size: usize,
         ctx: &RequestContext,
     ) -> anyhow::Result<usize> {
         debug!("Importing SLRU segment file {}", self.path);
         let buf = self.storage.get(&self.path).await?;
 
+        // TODO(vlad): Does timestamp to LSN work for imported timelines?
+        // Probably not since we don't append the `xact_time` to it as in
+        // [`WalIngest::ingest_xact_record`].
         let (kind, segno, start_blk) = self.key_range.start.to_slru_block()?;
         let (_kind, _segno, end_blk) = self.key_range.end.to_slru_block()?;
         let mut blknum = start_blk;
@@ -884,12 +910,13 @@ impl ImportTask for AnyImportTask {
     async fn doit(
         self,
         layer_writer: &mut ImageLayerWriter,
+        max_byte_range_size: usize,
         ctx: &RequestContext,
     ) -> anyhow::Result<usize> {
         match self {
-            Self::SingleKey(t) => t.doit(layer_writer, ctx).await,
-            Self::RelBlocks(t) => t.doit(layer_writer, ctx).await,
-            Self::SlruBlocks(t) => t.doit(layer_writer, ctx).await,
+            Self::SingleKey(t) => t.doit(layer_writer, max_byte_range_size, ctx).await,
+            Self::RelBlocks(t) => t.doit(layer_writer, max_byte_range_size, ctx).await,
+            Self::SlruBlocks(t) => t.doit(layer_writer, max_byte_range_size, ctx).await,
         }
     }
 }
@@ -930,7 +957,12 @@ impl ChunkProcessingJob {
         }
     }
 
-    async fn run(self, timeline: Arc<Timeline>, ctx: &RequestContext) -> anyhow::Result<()> {
+    async fn run(
+        self,
+        timeline: Arc<Timeline>,
+        max_byte_range_size: usize,
+        ctx: &RequestContext,
+    ) -> anyhow::Result<()> {
         let mut writer = ImageLayerWriter::new(
             timeline.conf,
             timeline.timeline_id,
@@ -945,7 +977,7 @@ impl ChunkProcessingJob {
 
         let mut nimages = 0;
         for task in self.tasks {
-            nimages += task.doit(&mut writer, ctx).await?;
+            nimages += task.doit(&mut writer, max_byte_range_size, ctx).await?;
         }
 
         let resident_layer = if nimages > 0 {

pageserver/src/tenant/timeline/import_pgdata/importbucket_client.rs

@@ -6,7 +6,7 @@ use bytes::Bytes;
 use postgres_ffi::ControlFileData;
 use remote_storage::{
     Download, DownloadError, DownloadKind, DownloadOpts, GenericRemoteStorage, Listing,
-    ListingObject, RemotePath,
+    ListingObject, RemotePath, RemoteStorageConfig,
 };
 use serde::de::DeserializeOwned;
 use tokio_util::sync::CancellationToken;
@@ -22,11 +22,9 @@ pub async fn new(
     location: &index_part_format::Location,
     cancel: CancellationToken,
 ) -> Result<RemoteStorageWrapper, anyhow::Error> {
-    // FIXME: we probably want some timeout, and we might be able to assume the max file
-    // size on S3 is 1GiB (postgres segment size). But the problem is that the individual
-    // downloaders don't know enough about concurrent downloads to make a guess on the
-    // expected bandwidth and resulting best timeout.
-    let timeout = std::time::Duration::from_secs(24 * 60 * 60);
+    // Downloads should be reasonably sized. We do ranged reads for relblock raw data
+    // and full reads for SLRU segments which are bounded by Postgres.
+    let timeout = RemoteStorageConfig::DEFAULT_TIMEOUT;
     let location_storage = match location {
         #[cfg(feature = "testing")]
         index_part_format::Location::LocalFs { path } => {
@@ -50,9 +48,12 @@ pub async fn new(
                             .import_pgdata_aws_endpoint_url
                             .clone()
                             .map(|url| url.to_string()), //  by specifying None here, remote_storage/aws-sdk-rust will infer from env
-                        concurrency_limit: 100.try_into().unwrap(), // TODO: think about this
-                        max_keys_per_list_response: Some(1000),     // TODO: think about this
-                        upload_storage_class: None,                 // irrelevant
+                        // This matches the default import job concurrency. This is managed
+                        // separately from the usual S3 client, but the concern here is bandwidth
+                        // usage.
+                        concurrency_limit: 128.try_into().unwrap(),
+                        max_keys_per_list_response: Some(1000),
+                        upload_storage_class: None, // irrelevant
                     },
                     timeout,
                 )

pageserver/src/tenant/timeline/walreceiver.rs

@@ -113,7 +113,7 @@ impl WalReceiver {
                 }
                 connection_manager_state.shutdown().await;
                 *loop_status.write().unwrap() = None;
-                debug!("task exits");
+                info!("task exits");
             }
             .instrument(info_span!(parent: None, "wal_connection_manager", tenant_id = %tenant_shard_id.tenant_id, shard_id = %tenant_shard_id.shard_slug(), timeline_id = %timeline_id))
         });

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

@@ -32,9 +32,7 @@ use utils::backoff::{
 };
 use utils::id::{NodeId, TenantTimelineId};
 use utils::lsn::Lsn;
-use utils::postgres_client::{
-    ConnectionConfigArgs, PostgresClientProtocol, wal_stream_connection_config,
-};
+use utils::postgres_client::{ConnectionConfigArgs, wal_stream_connection_config};
 
 use super::walreceiver_connection::{WalConnectionStatus, WalReceiverError};
 use super::{TaskEvent, TaskHandle, TaskStateUpdate, WalReceiverConf};
@@ -991,19 +989,12 @@ impl ConnectionManagerState {
                     return None; // no connection string, ignore sk
                 }
 
-                let (shard_number, shard_count, shard_stripe_size) = match self.conf.protocol {
-                    PostgresClientProtocol::Vanilla => {
-                        (None, None, None)
-                    },
-                    PostgresClientProtocol::Interpreted { .. } => {
-                        let shard_identity = self.timeline.get_shard_identity();
-                        (
-                            Some(shard_identity.number.0),
-                            Some(shard_identity.count.0),
-                            Some(shard_identity.stripe_size.0),
-                        )
-                    }
-                };
+                let shard_identity = self.timeline.get_shard_identity();
+                let (shard_number, shard_count, shard_stripe_size) = (
+                    Some(shard_identity.number.0),
+                    Some(shard_identity.count.0),
+                    Some(shard_identity.stripe_size.0),
+                );
 
                 let connection_conf_args = ConnectionConfigArgs {
                     protocol: self.conf.protocol,
@@ -1120,8 +1111,8 @@ impl ReconnectReason {
 
 #[cfg(test)]
 mod tests {
-    use pageserver_api::config::defaults::DEFAULT_WAL_RECEIVER_PROTOCOL;
     use url::Host;
+    use utils::postgres_client::PostgresClientProtocol;
 
     use super::*;
     use crate::tenant::harness::{TIMELINE_ID, TenantHarness};
@@ -1552,6 +1543,11 @@ mod tests {
             .await
             .expect("Failed to create an empty timeline for dummy wal connection manager");
 
+        let protocol = PostgresClientProtocol::Interpreted {
+            format: utils::postgres_client::InterpretedFormat::Protobuf,
+            compression: Some(utils::postgres_client::Compression::Zstd { level: 1 }),
+        };
+
         ConnectionManagerState {
             id: TenantTimelineId {
                 tenant_id: harness.tenant_shard_id.tenant_id,
@@ -1560,7 +1556,7 @@ mod tests {
             timeline,
             cancel: CancellationToken::new(),
             conf: WalReceiverConf {
-                protocol: DEFAULT_WAL_RECEIVER_PROTOCOL,
+                protocol,
                 wal_connect_timeout: Duration::from_secs(1),
                 lagging_wal_timeout: Duration::from_secs(1),
                 max_lsn_wal_lag: NonZeroU64::new(1024 * 1024).unwrap(),

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

@@ -15,7 +15,7 @@ use postgres_backend::is_expected_io_error;
 use postgres_connection::PgConnectionConfig;
 use postgres_ffi::WAL_SEGMENT_SIZE;
 use postgres_ffi::v14::xlog_utils::normalize_lsn;
-use postgres_ffi::waldecoder::{WalDecodeError, WalStreamDecoder};
+use postgres_ffi::waldecoder::WalDecodeError;
 use postgres_protocol::message::backend::ReplicationMessage;
 use postgres_types::PgLsn;
 use tokio::sync::watch;
@@ -31,7 +31,7 @@ use utils::lsn::Lsn;
 use utils::pageserver_feedback::PageserverFeedback;
 use utils::postgres_client::PostgresClientProtocol;
 use utils::sync::gate::GateError;
-use wal_decoder::models::{FlushUncommittedRecords, InterpretedWalRecord, InterpretedWalRecords};
+use wal_decoder::models::{FlushUncommittedRecords, InterpretedWalRecords};
 use wal_decoder::wire_format::FromWireFormat;
 
 use super::TaskStateUpdate;
@@ -275,8 +275,6 @@ pub(super) async fn handle_walreceiver_connection(
     let copy_stream = replication_client.copy_both_simple(&query).await?;
     let mut physical_stream = pin!(ReplicationStream::new(copy_stream));
 
-    let mut waldecoder = WalStreamDecoder::new(startpoint, timeline.pg_version);
-
     let mut walingest = WalIngest::new(timeline.as_ref(), startpoint, &ctx)
         .await
         .map_err(|e| match e.kind {
@@ -284,19 +282,22 @@ pub(super) async fn handle_walreceiver_connection(
             _ => WalReceiverError::Other(e.into()),
         })?;
 
-    let shard = vec![*timeline.get_shard_identity()];
-
-    let interpreted_proto_config = match protocol {
-        PostgresClientProtocol::Vanilla => None,
+    let (format, compression) = match protocol {
         PostgresClientProtocol::Interpreted {
             format,
             compression,
-        } => Some((format, compression)),
+        } => (format, compression),
+        PostgresClientProtocol::Vanilla => {
+            return Err(WalReceiverError::Other(anyhow!(
+                "Vanilla WAL receiver protocol is no longer supported for ingest"
+            )));
+        }
     };
 
     let mut expected_wal_start = startpoint;
     while let Some(replication_message) = {
         select! {
+            biased;
             _ = cancellation.cancelled() => {
                 debug!("walreceiver interrupted");
                 None
@@ -312,16 +313,6 @@ pub(super) async fn handle_walreceiver_connection(
         // Update the connection status before processing the message. If the message processing
         // fails (e.g. in walingest), we still want to know latests LSNs from the safekeeper.
         match &replication_message {
-            ReplicationMessage::XLogData(xlog_data) => {
-                connection_status.latest_connection_update = now;
-                connection_status.commit_lsn = Some(Lsn::from(xlog_data.wal_end()));
-                connection_status.streaming_lsn = Some(Lsn::from(
-                    xlog_data.wal_start() + xlog_data.data().len() as u64,
-                ));
-                if !xlog_data.data().is_empty() {
-                    connection_status.latest_wal_update = now;
-                }
-            }
             ReplicationMessage::PrimaryKeepAlive(keepalive) => {
                 connection_status.latest_connection_update = now;
                 connection_status.commit_lsn = Some(Lsn::from(keepalive.wal_end()));
@@ -352,7 +343,6 @@ pub(super) async fn handle_walreceiver_connection(
                 // were interpreted.
                 let streaming_lsn = Lsn::from(raw.streaming_lsn());
 
-                let (format, compression) = interpreted_proto_config.unwrap();
                 let batch = InterpretedWalRecords::from_wire(raw.data(), format, compression)
                     .await
                     .with_context(|| {
@@ -508,138 +498,6 @@ pub(super) async fn handle_walreceiver_connection(
                 Some(streaming_lsn)
             }
 
-            ReplicationMessage::XLogData(xlog_data) => {
-                async fn commit(
-                    modification: &mut DatadirModification<'_>,
-                    uncommitted: &mut u64,
-                    filtered: &mut u64,
-                    ctx: &RequestContext,
-                ) -> anyhow::Result<()> {
-                    let stats = modification.stats();
-                    modification.commit(ctx).await?;
-                    WAL_INGEST
-                        .records_committed
-                        .inc_by(*uncommitted - *filtered);
-                    WAL_INGEST.inc_values_committed(&stats);
-                    *uncommitted = 0;
-                    *filtered = 0;
-                    Ok(())
-                }
-
-                // Pass the WAL data to the decoder, and see if we can decode
-                // more records as a result.
-                let data = xlog_data.data();
-                let startlsn = Lsn::from(xlog_data.wal_start());
-                let endlsn = startlsn + data.len() as u64;
-
-                trace!("received XLogData between {startlsn} and {endlsn}");
-
-                WAL_INGEST.bytes_received.inc_by(data.len() as u64);
-                waldecoder.feed_bytes(data);
-
-                {
-                    let mut modification = timeline.begin_modification(startlsn);
-                    let mut uncommitted_records = 0;
-                    let mut filtered_records = 0;
-
-                    while let Some((next_record_lsn, recdata)) = waldecoder.poll_decode()? {
-                        // It is important to deal with the aligned records as lsn in getPage@LSN is
-                        // aligned and can be several bytes bigger. Without this alignment we are
-                        // at risk of hitting a deadlock.
-                        if !next_record_lsn.is_aligned() {
-                            return Err(WalReceiverError::Other(anyhow!("LSN not aligned")));
-                        }
-
-                        // Deserialize and interpret WAL record
-                        let interpreted = InterpretedWalRecord::from_bytes_filtered(
-                            recdata,
-                            &shard,
-                            next_record_lsn,
-                            modification.tline.pg_version,
-                        )?
-                        .remove(timeline.get_shard_identity())
-                        .unwrap();
-
-                        if matches!(interpreted.flush_uncommitted, FlushUncommittedRecords::Yes)
-                            && uncommitted_records > 0
-                        {
-                            // Special case: legacy PG database creations operate by reading pages from a 'template' database:
-                            // these are the only kinds of WAL record that require reading data blocks while ingesting.  Ensure
-                            // all earlier writes of data blocks are visible by committing any modification in flight.
-                            commit(
-                                &mut modification,
-                                &mut uncommitted_records,
-                                &mut filtered_records,
-                                &ctx,
-                            )
-                            .await?;
-                        }
-
-                        // Ingest the records without immediately committing them.
-                        timeline.metrics.wal_records_received.inc();
-                        let ingested = walingest
-                            .ingest_record(interpreted, &mut modification, &ctx)
-                            .await
-                            .with_context(|| {
-                                format!("could not ingest record at {next_record_lsn}")
-                            })
-                            .inspect_err(|err| {
-                                // TODO: we can't differentiate cancellation errors with
-                                // anyhow::Error, so just ignore it if we're cancelled.
-                                if !cancellation.is_cancelled() && !timeline.is_stopping() {
-                                    critical!("{err:?}")
-                                }
-                            })?;
-                        if !ingested {
-                            tracing::debug!("ingest: filtered out record @ LSN {next_record_lsn}");
-                            WAL_INGEST.records_filtered.inc();
-                            filtered_records += 1;
-                        }
-
-                        // FIXME: this cannot be made pausable_failpoint without fixing the
-                        // failpoint library; in tests, the added amount of debugging will cause us
-                        // to timeout the tests.
-                        fail_point!("walreceiver-after-ingest");
-
-                        last_rec_lsn = next_record_lsn;
-
-                        // Commit every ingest_batch_size records. Even if we filtered out
-                        // all records, we still need to call commit to advance the LSN.
-                        uncommitted_records += 1;
-                        if uncommitted_records >= ingest_batch_size
-                            || modification.approx_pending_bytes()
-                                > DatadirModification::MAX_PENDING_BYTES
-                        {
-                            commit(
-                                &mut modification,
-                                &mut uncommitted_records,
-                                &mut filtered_records,
-                                &ctx,
-                            )
-                            .await?;
-                        }
-                    }
-
-                    // Commit the remaining records.
-                    if uncommitted_records > 0 {
-                        commit(
-                            &mut modification,
-                            &mut uncommitted_records,
-                            &mut filtered_records,
-                            &ctx,
-                        )
-                        .await?;
-                    }
-                }
-
-                if !caught_up && endlsn >= end_of_wal {
-                    info!("caught up at LSN {endlsn}");
-                    caught_up = true;
-                }
-
-                Some(endlsn)
-            }
-
             ReplicationMessage::PrimaryKeepAlive(keepalive) => {
                 let wal_end = keepalive.wal_end();
                 let timestamp = keepalive.timestamp();

pgxn/neon/neon.c

@@ -16,6 +16,7 @@
 #if PG_MAJORVERSION_NUM >= 15
 #include "access/xlogrecovery.h"
 #endif
+#include "executor/instrument.h"
 #include "replication/logical.h"
 #include "replication/logicallauncher.h"
 #include "replication/slot.h"
@@ -33,6 +34,7 @@
 #include "file_cache.h"
 #include "neon.h"
 #include "neon_lwlsncache.h"
+#include "neon_perf_counters.h"
 #include "control_plane_connector.h"
 #include "logical_replication_monitor.h"
 #include "unstable_extensions.h"
@@ -46,6 +48,13 @@ void		_PG_init(void);
 
 
 static int  running_xacts_overflow_policy;
+static bool monitor_query_exec_time = false;
+
+static ExecutorStart_hook_type prev_ExecutorStart = NULL;
+static ExecutorEnd_hook_type prev_ExecutorEnd = NULL;
+
+static void neon_ExecutorStart(QueryDesc *queryDesc, int eflags);
+static void neon_ExecutorEnd(QueryDesc *queryDesc);
 
 #if PG_MAJORVERSION_NUM >= 16
 static shmem_startup_hook_type prev_shmem_startup_hook;
@@ -470,6 +479,16 @@ _PG_init(void)
 							0,
 							NULL, NULL, NULL);
 
+	DefineCustomBoolVariable(
+							"neon.monitor_query_exec_time",
+							"Collect infortmation about query execution time",
+							NULL,
+							&monitor_query_exec_time,
+							false,
+							PGC_USERSET,
+							0,
+							NULL, NULL, NULL);
+
 	DefineCustomBoolVariable(
 							"neon.allow_replica_misconfig",
 							"Allow replica startup when some critical GUCs have smaller value than on primary node",
@@ -508,6 +527,11 @@ _PG_init(void)
 	EmitWarningsOnPlaceholders("neon");
 
 	ReportSearchPath();
+
+	prev_ExecutorStart = ExecutorStart_hook;
+	ExecutorStart_hook = neon_ExecutorStart;
+	prev_ExecutorEnd = ExecutorEnd_hook;
+	ExecutorEnd_hook = neon_ExecutorEnd;
 }
 
 PG_FUNCTION_INFO_V1(pg_cluster_size);
@@ -581,3 +605,55 @@ neon_shmem_startup_hook(void)
 #endif
 }
 #endif
+
+/*
+ * ExecutorStart hook: start up tracking if needed
+ */
+static void
+neon_ExecutorStart(QueryDesc *queryDesc, int eflags)
+{
+	if (prev_ExecutorStart)
+		prev_ExecutorStart(queryDesc, eflags);
+	else
+		standard_ExecutorStart(queryDesc, eflags);
+
+	if (monitor_query_exec_time)
+	{
+		/*
+		 * Set up to track total elapsed time in ExecutorRun.  Make sure the
+		 * space is allocated in the per-query context so it will go away at
+		 * ExecutorEnd.
+		 */
+		if (queryDesc->totaltime == NULL)
+		{
+			MemoryContext oldcxt;
+
+			oldcxt = MemoryContextSwitchTo(queryDesc->estate->es_query_cxt);
+			queryDesc->totaltime = InstrAlloc(1, INSTRUMENT_TIMER, false);
+			MemoryContextSwitchTo(oldcxt);
+		}
+	}
+}
+
+/*
+ * ExecutorEnd hook: store results if needed
+ */
+static void
+neon_ExecutorEnd(QueryDesc *queryDesc)
+{
+	if (monitor_query_exec_time && queryDesc->totaltime)
+	{
+		/*
+		 * Make sure stats accumulation is done.  (Note: it's okay if several
+		 * levels of hook all do this.)
+		 */
+		InstrEndLoop(queryDesc->totaltime);
+
+		inc_query_time(queryDesc->totaltime->total*1000000); /* convert to usec */
+	}
+
+	if (prev_ExecutorEnd)
+		prev_ExecutorEnd(queryDesc);
+	else
+		standard_ExecutorEnd(queryDesc);
+}

pgxn/neon/neon_perf_counters.c

@@ -71,6 +71,27 @@ inc_iohist(IOHistogram hist, uint64 latency_us)
 	hist->wait_us_count++;
 }
 
+static inline void
+inc_qthist(QTHistogram hist, uint64 elapsed_us)
+{
+	int			lo = 0;
+	int			hi = NUM_QT_BUCKETS - 1;
+
+	/* Find the right bucket with binary search */
+	while (lo < hi)
+	{
+		int			mid = (lo + hi) / 2;
+
+		if (elapsed_us < qt_bucket_thresholds[mid])
+			hi = mid;
+		else
+			lo = mid + 1;
+	}
+	hist->elapsed_us_bucket[lo]++;
+	hist->elapsed_us_sum += elapsed_us;
+	hist->elapsed_us_count++;
+}
+
 /*
  * Count a GetPage wait operation.
  */
@@ -98,6 +119,13 @@ inc_page_cache_write_wait(uint64 latency)
 	inc_iohist(&MyNeonCounters->file_cache_write_hist, latency);
 }
 
+
+void
+inc_query_time(uint64 elapsed)
+{
+	inc_qthist(&MyNeonCounters->query_time_hist, elapsed);
+}
+
 /*
  * Support functions for the views, neon_backend_perf_counters and
  * neon_perf_counters.
@@ -112,11 +140,11 @@ typedef struct
 } metric_t;
 
 static int
-histogram_to_metrics(IOHistogram histogram,
-					 metric_t *metrics,
-					 const char *count,
-					 const char *sum,
-					 const char *bucket)
+io_histogram_to_metrics(IOHistogram histogram,
+						metric_t *metrics,
+						const char *count,
+						const char *sum,
+						const char *bucket)
 {
 	int		i = 0;
 	uint64	bucket_accum = 0;
@@ -145,10 +173,44 @@ histogram_to_metrics(IOHistogram histogram,
 	return i;
 }
 
+static int
+qt_histogram_to_metrics(QTHistogram histogram,
+						metric_t *metrics,
+						const char *count,
+						const char *sum,
+						const char *bucket)
+{
+	int		i = 0;
+	uint64	bucket_accum = 0;
+
+	metrics[i].name = count;
+	metrics[i].is_bucket = false;
+	metrics[i].value = (double) histogram->elapsed_us_count;
+	i++;
+	metrics[i].name = sum;
+	metrics[i].is_bucket = false;
+	metrics[i].value = (double) histogram->elapsed_us_sum / 1000000.0;
+	i++;
+	for (int bucketno = 0; bucketno < NUM_QT_BUCKETS; bucketno++)
+	{
+		uint64		threshold = qt_bucket_thresholds[bucketno];
+
+		bucket_accum += histogram->elapsed_us_bucket[bucketno];
+
+		metrics[i].name = bucket;
+		metrics[i].is_bucket = true;
+		metrics[i].bucket_le = (threshold == UINT64_MAX) ? INFINITY : ((double) threshold) / 1000000.0;
+		metrics[i].value = (double) bucket_accum;
+		i++;
+	}
+
+	return i;
+}
+
 static metric_t *
 neon_perf_counters_to_metrics(neon_per_backend_counters *counters)
 {
-#define NUM_METRICS ((2 + NUM_IO_WAIT_BUCKETS) * 3 + 12)
+#define NUM_METRICS ((2 + NUM_IO_WAIT_BUCKETS) * 3 + (2 + NUM_QT_BUCKETS) + 12)
 	metric_t   *metrics = palloc((NUM_METRICS + 1) * sizeof(metric_t));
 	int			i = 0;
 
@@ -159,10 +221,10 @@ neon_perf_counters_to_metrics(neon_per_backend_counters *counters)
 		i++; \
 	} while (false)
 
-	i += histogram_to_metrics(&counters->getpage_hist, &metrics[i],
-							  "getpage_wait_seconds_count",
-							  "getpage_wait_seconds_sum",
-							  "getpage_wait_seconds_bucket");
+	i += io_histogram_to_metrics(&counters->getpage_hist, &metrics[i],
+								 "getpage_wait_seconds_count",
+								 "getpage_wait_seconds_sum",
+								 "getpage_wait_seconds_bucket");
 
 	APPEND_METRIC(getpage_prefetch_requests_total);
 	APPEND_METRIC(getpage_sync_requests_total);
@@ -178,14 +240,19 @@ neon_perf_counters_to_metrics(neon_per_backend_counters *counters)
 
 	APPEND_METRIC(file_cache_hits_total);
 
-	i += histogram_to_metrics(&counters->file_cache_read_hist, &metrics[i],
-							  "file_cache_read_wait_seconds_count",
-							  "file_cache_read_wait_seconds_sum",
-							  "file_cache_read_wait_seconds_bucket");
-	i += histogram_to_metrics(&counters->file_cache_write_hist, &metrics[i],
-							  "file_cache_write_wait_seconds_count",
-							  "file_cache_write_wait_seconds_sum",
-							  "file_cache_write_wait_seconds_bucket");
+	i += io_histogram_to_metrics(&counters->file_cache_read_hist, &metrics[i],
+								 "file_cache_read_wait_seconds_count",
+								 "file_cache_read_wait_seconds_sum",
+								 "file_cache_read_wait_seconds_bucket");
+	i += io_histogram_to_metrics(&counters->file_cache_write_hist, &metrics[i],
+								 "file_cache_write_wait_seconds_count",
+								 "file_cache_write_wait_seconds_sum",
+								 "file_cache_write_wait_seconds_bucket");
+
+	i += qt_histogram_to_metrics(&counters->query_time_hist, &metrics[i],
+								 "query_time_seconds_count",
+								 "query_time_seconds_sum",
+								 "query_time_seconds_bucket");
 
 	Assert(i == NUM_METRICS);
 
@@ -257,7 +324,7 @@ neon_get_backend_perf_counters(PG_FUNCTION_ARGS)
 }
 
 static inline void
-histogram_merge_into(IOHistogram into, IOHistogram from)
+io_histogram_merge_into(IOHistogram into, IOHistogram from)
 {
 	into->wait_us_count += from->wait_us_count;
 	into->wait_us_sum += from->wait_us_sum;
@@ -265,6 +332,15 @@ histogram_merge_into(IOHistogram into, IOHistogram from)
 		into->wait_us_bucket[bucketno] += from->wait_us_bucket[bucketno];
 }
 
+static inline void
+qt_histogram_merge_into(QTHistogram into, QTHistogram from)
+{
+	into->elapsed_us_count += from->elapsed_us_count;
+	into->elapsed_us_sum += from->elapsed_us_sum;
+	for (int bucketno = 0; bucketno < NUM_QT_BUCKETS; bucketno++)
+		into->elapsed_us_bucket[bucketno] += from->elapsed_us_bucket[bucketno];
+}
+
 PG_FUNCTION_INFO_V1(neon_get_perf_counters);
 Datum
 neon_get_perf_counters(PG_FUNCTION_ARGS)
@@ -283,7 +359,7 @@ neon_get_perf_counters(PG_FUNCTION_ARGS)
 	{
 		neon_per_backend_counters *counters = &neon_per_backend_counters_shared[procno];
 
-		histogram_merge_into(&totals.getpage_hist, &counters->getpage_hist);
+		io_histogram_merge_into(&totals.getpage_hist, &counters->getpage_hist);
 		totals.getpage_prefetch_requests_total += counters->getpage_prefetch_requests_total;
 		totals.getpage_sync_requests_total += counters->getpage_sync_requests_total;
 		totals.getpage_prefetch_misses_total += counters->getpage_prefetch_misses_total;
@@ -294,13 +370,13 @@ neon_get_perf_counters(PG_FUNCTION_ARGS)
 		totals.pageserver_open_requests += counters->pageserver_open_requests;
 		totals.getpage_prefetches_buffered += counters->getpage_prefetches_buffered;
 		totals.file_cache_hits_total += counters->file_cache_hits_total;
-		histogram_merge_into(&totals.file_cache_read_hist, &counters->file_cache_read_hist);
-		histogram_merge_into(&totals.file_cache_write_hist, &counters->file_cache_write_hist);
-
 		totals.compute_getpage_stuck_requests_total += counters->compute_getpage_stuck_requests_total;
 		totals.compute_getpage_max_inflight_stuck_time_ms = Max(
 			totals.compute_getpage_max_inflight_stuck_time_ms,
 			counters->compute_getpage_max_inflight_stuck_time_ms);
+		io_histogram_merge_into(&totals.file_cache_read_hist, &counters->file_cache_read_hist);
+		io_histogram_merge_into(&totals.file_cache_write_hist, &counters->file_cache_write_hist);
+		qt_histogram_merge_into(&totals.query_time_hist, &counters->query_time_hist);
 	}
 
 	metrics = neon_perf_counters_to_metrics(&totals);

pgxn/neon/neon_perf_counters.h

@@ -36,6 +36,28 @@ typedef struct IOHistogramData
 
 typedef IOHistogramData *IOHistogram;
 
+static const uint64 qt_bucket_thresholds[] = {
+	       2,        3,        6,        10,  /* 0 us   - 10 us */
+	      20,       30,       60,       100,  /* 10 us  - 100 us */
+	     200,      300,      600,	   1000,  /* 100 us - 1 ms */
+	    2000,     3000,     6000,     10000,  /* 1 ms   - 10 ms */
+	   20000,    30000,    60000,    100000,  /* 10 ms  - 100 ms */
+	  200000,   300000,   600000,   1000000,  /* 100 ms - 1 s */
+	 2000000,  3000000,  6000000,  10000000,  /* 1 s - 10 s */
+	20000000, 30000000, 60000000, 100000000,  /* 10 s - 100 s */
+	UINT64_MAX,
+};
+#define NUM_QT_BUCKETS (lengthof(qt_bucket_thresholds))
+
+typedef struct QTHistogramData
+{
+	uint64		elapsed_us_count;
+	uint64		elapsed_us_sum;
+	uint64		elapsed_us_bucket[NUM_QT_BUCKETS];
+} QTHistogramData;
+
+typedef QTHistogramData *QTHistogram;
+
 typedef struct
 {
 	/*
@@ -127,6 +149,11 @@ typedef struct
 	/* LFC I/O time buckets */
 	IOHistogramData file_cache_read_hist;
 	IOHistogramData file_cache_write_hist;
+
+	/*
+	 * Histogram of query execution time.
+	 */
+	QTHistogramData query_time_hist;
 } neon_per_backend_counters;
 
 /* Pointer to the shared memory array of neon_per_backend_counters structs */
@@ -149,6 +176,7 @@ extern neon_per_backend_counters *neon_per_backend_counters_shared;
 extern void inc_getpage_wait(uint64 latency);
 extern void inc_page_cache_read_wait(uint64 latency);
 extern void inc_page_cache_write_wait(uint64 latency);
+extern void inc_query_time(uint64 elapsed);
 
 extern Size NeonPerfCountersShmemSize(void);
 extern void NeonPerfCountersShmemInit(void);

pgxn/neon/neon_pgversioncompat.c

@@ -5,6 +5,7 @@
 
 #include "funcapi.h"
 #include "miscadmin.h"
+#include "access/xlog.h"
 #include "utils/tuplestore.h"
 
 #include "neon_pgversioncompat.h"
@@ -41,5 +42,12 @@ InitMaterializedSRF(FunctionCallInfo fcinfo, bits32 flags)
 	rsinfo->setDesc = stored_tupdesc;
 	MemoryContextSwitchTo(old_context);
 }
+
+TimeLineID GetWALInsertionTimeLine(void)
+{
+	return ThisTimeLineID + 1;
+}
+
+
 #endif
 

pgxn/neon/neon_pgversioncompat.h

@@ -162,6 +162,7 @@ InitBufferTag(BufferTag *tag, const RelFileNode *rnode,
 
 #if PG_MAJORVERSION_NUM < 15
 extern void InitMaterializedSRF(FunctionCallInfo fcinfo, bits32 flags);
+extern TimeLineID GetWALInsertionTimeLine(void);
 #endif
 
 #endif							/* NEON_PGVERSIONCOMPAT_H */

pgxn/neon/neon_walreader.c

@@ -69,6 +69,7 @@ struct NeonWALReader
 	WALSegmentContext segcxt;
 	WALOpenSegment seg;
 	int			wre_errno;
+	TimeLineID	local_active_tlid;
 	/* Explains failure to read, static for simplicity. */
 	char		err_msg[NEON_WALREADER_ERR_MSG_LEN];
 
@@ -106,7 +107,7 @@ struct NeonWALReader
 
 /* palloc and initialize NeonWALReader */
 NeonWALReader *
-NeonWALReaderAllocate(int wal_segment_size, XLogRecPtr available_lsn, char *log_prefix)
+NeonWALReaderAllocate(int wal_segment_size, XLogRecPtr available_lsn, char *log_prefix, TimeLineID tlid)
 {
 	NeonWALReader *reader;
 
@@ -118,6 +119,7 @@ NeonWALReaderAllocate(int wal_segment_size, XLogRecPtr available_lsn, char *log_
 		MemoryContextAllocZero(TopMemoryContext, sizeof(NeonWALReader));
 
 	reader->available_lsn = available_lsn;
+	reader->local_active_tlid = tlid;
 	reader->seg.ws_file = -1;
 	reader->seg.ws_segno = 0;
 	reader->seg.ws_tli = 0;
@@ -577,6 +579,17 @@ NeonWALReaderIsRemConnEstablished(NeonWALReader *state)
 	return state->rem_state == RS_ESTABLISHED;
 }
 
+/*
+ * Whether remote connection is established. Once this is done, until successful
+ * local read or error socket is stable and user can update socket events
+ * instead of readding it each time.
+ */
+TimeLineID
+NeonWALReaderLocalActiveTimeLineID(NeonWALReader *state)
+{
+	return state->local_active_tlid;
+}
+
 /*
  * Returns events user should wait on connection socket or 0 if remote
  * connection is not active.

pgxn/neon/neon_walreader.h

@@ -19,9 +19,10 @@ typedef enum
 	NEON_WALREAD_ERROR,
 } NeonWALReadResult;
 
-extern NeonWALReader *NeonWALReaderAllocate(int wal_segment_size, XLogRecPtr available_lsn, char *log_prefix);
+extern NeonWALReader *NeonWALReaderAllocate(int wal_segment_size, XLogRecPtr available_lsn, char *log_prefix, TimeLineID tlid);
 extern void NeonWALReaderFree(NeonWALReader *state);
 extern void NeonWALReaderResetRemote(NeonWALReader *state);
+extern TimeLineID NeonWALReaderLocalActiveTimeLineID(NeonWALReader *state);
 extern NeonWALReadResult NeonWALRead(NeonWALReader *state, char *buf, XLogRecPtr startptr, Size count, TimeLineID tli);
 extern pgsocket NeonWALReaderSocket(NeonWALReader *state);
 extern uint32 NeonWALReaderEvents(NeonWALReader *state);

pgxn/neon/walproposer.c

@@ -98,6 +98,7 @@ WalProposerCreate(WalProposerConfig *config, walproposer_api api)
 	wp = palloc0(sizeof(WalProposer));
 	wp->config = config;
 	wp->api = api;
+	wp->localTimeLineID = config->pgTimeline;
 	wp->state = WPS_COLLECTING_TERMS;
 	wp->mconf.generation = INVALID_GENERATION;
 	wp->mconf.members.len = 0;
@@ -119,6 +120,10 @@ WalProposerCreate(WalProposerConfig *config, walproposer_api api)
 		{
 			wp_log(FATAL, "failed to parse neon.safekeepers generation number: %m");
 		}
+		if (*endptr != ':')
+		{
+			wp_log(FATAL, "failed to parse neon.safekeepers: no colon after generation");
+		}
 		/* Skip past : to the first hostname. */
 		host = endptr + 1;
 	}
@@ -1130,7 +1135,7 @@ VotesCollectedMset(WalProposer *wp, MemberSet *mset, Safekeeper **msk, StringInf
 				wp->propTermStartLsn = sk->voteResponse.flushLsn;
 				wp->donor = sk;
 			}
-			wp->truncateLsn = Max(wp->safekeeper[i].voteResponse.truncateLsn, wp->truncateLsn);
+			wp->truncateLsn = Max(sk->voteResponse.truncateLsn, wp->truncateLsn);
 
 			if (n_votes > 0)
 				appendStringInfoString(s, ", ");
@@ -1380,7 +1385,7 @@ ProcessPropStartPos(WalProposer *wp)
 	 * we must bail out, as clog and other non rel data is inconsistent.
 	 */
 	walprop_shared = wp->api.get_shmem_state(wp);
-	if (!wp->config->syncSafekeepers)
+	if (!wp->config->syncSafekeepers && !walprop_shared->replica_promote)
 	{
 		/*
 		 * Basebackup LSN always points to the beginning of the record (not

pgxn/neon/walproposer.h

@@ -391,6 +391,7 @@ typedef struct WalproposerShmemState
 	/* last feedback from each shard */
 	PageserverFeedback shard_ps_feedback[MAX_SHARDS];
 	int			num_shards;
+	bool		replica_promote;
 
 	/* aggregated feedback with min LSNs across shards */
 	PageserverFeedback min_ps_feedback;
@@ -806,6 +807,9 @@ typedef struct WalProposer
 	/* Safekeepers walproposer is connecting to. */
 	Safekeeper	safekeeper[MAX_SAFEKEEPERS];
 
+	/* Current local TimeLineId in use */
+	TimeLineID	localTimeLineID;
+
 	/* WAL has been generated up to this point */
 	XLogRecPtr	availableLsn;
 

pgxn/neon/walproposer_pg.c

@@ -35,6 +35,7 @@
 #include "storage/proc.h"
 #include "storage/ipc.h"
 #include "storage/lwlock.h"
+#include "storage/pg_shmem.h"
 #include "storage/shmem.h"
 #include "storage/spin.h"
 #include "tcop/tcopprot.h"
@@ -159,12 +160,19 @@ WalProposerMain(Datum main_arg)
 {
 	WalProposer *wp;
 
+	if (*wal_acceptors_list == '\0')
+	{
+		wpg_log(WARNING, "Safekeepers list is empty");
+		return;
+	}
+
 	init_walprop_config(false);
 	walprop_pg_init_bgworker();
 	am_walproposer = true;
 	walprop_pg_load_libpqwalreceiver();
 
 	wp = WalProposerCreate(&walprop_config, walprop_pg);
+	wp->localTimeLineID = GetWALInsertionTimeLine();
 	wp->last_reconnect_attempt = walprop_pg_get_current_timestamp(wp);
 
 	walprop_pg_init_walsender();
@@ -272,6 +280,30 @@ split_safekeepers_list(char *safekeepers_list, char *safekeepers[])
 	return n_safekeepers;
 }
 
+static char *split_off_safekeepers_generation(char *safekeepers_list, uint32 *generation)
+{
+	char	   *endptr;
+
+	if (strncmp(safekeepers_list, "g#", 2) != 0)
+	{
+		return safekeepers_list;
+	}
+	else
+	{
+		errno = 0;
+		*generation = strtoul(safekeepers_list + 2, &endptr, 10);
+		if (errno != 0)
+		{
+			wp_log(FATAL, "failed to parse neon.safekeepers generation number: %m");
+		}
+		if (*endptr != ':')
+		{
+			wp_log(FATAL, "failed to parse neon.safekeepers: no colon after generation");
+		}
+		return endptr + 1;
+	}
+}
+
 /*
  * Accept two coma-separated strings with list of safekeeper host:port addresses.
  * Split them into arrays and return false if two sets do not match, ignoring the order.
@@ -283,6 +315,16 @@ safekeepers_cmp(char *old, char *new)
 	char	   *safekeepers_new[MAX_SAFEKEEPERS];
 	int			len_old = 0;
 	int			len_new = 0;
+	uint32		gen_old = INVALID_GENERATION;
+	uint32		gen_new = INVALID_GENERATION;
+
+	old = split_off_safekeepers_generation(old, &gen_old);
+	new = split_off_safekeepers_generation(new, &gen_new);
+
+	if (gen_old != gen_new)
+	{
+		return false;
+	}
 
 	len_old = split_safekeepers_list(old, safekeepers_old);
 	len_new = split_safekeepers_list(new, safekeepers_new);
@@ -316,6 +358,9 @@ assign_neon_safekeepers(const char *newval, void *extra)
 	char	   *newval_copy;
 	char	   *oldval;
 
+	if (newval && *newval != '\0' && UsedShmemSegAddr && walprop_shared && RecoveryInProgress())
+		walprop_shared->replica_promote = true;
+
 	if (!am_walproposer)
 		return;
 
@@ -506,16 +551,15 @@ BackpressureThrottlingTime(void)
 
 /*
  * Register a background worker proposing WAL to wal acceptors.
+ * We start walproposer bgworker even for replicas in order to support possible replica promotion.
+ * When pg_promote() function is called, then walproposer bgworker registered with BgWorkerStart_RecoveryFinished
+ * is automatically launched when promotion is completed.
  */
 static void
 walprop_register_bgworker(void)
 {
 	BackgroundWorker bgw;
 
-	/* If no wal acceptors are specified, don't start the background worker. */
-	if (*wal_acceptors_list == '\0')
-		return;
-
 	memset(&bgw, 0, sizeof(bgw));
 	bgw.bgw_flags = BGWORKER_SHMEM_ACCESS;
 	bgw.bgw_start_time = BgWorkerStart_RecoveryFinished;
@@ -1292,9 +1336,7 @@ StartProposerReplication(WalProposer *wp, StartReplicationCmd *cmd)
 
 #if PG_VERSION_NUM < 150000
 	if (ThisTimeLineID == 0)
-		ereport(ERROR,
-				(errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
-				 errmsg("IDENTIFY_SYSTEM has not been run before START_REPLICATION")));
+		ThisTimeLineID = 1;
 #endif
 
 	/*
@@ -1508,7 +1550,7 @@ walprop_pg_wal_reader_allocate(Safekeeper *sk)
 
 	snprintf(log_prefix, sizeof(log_prefix), WP_LOG_PREFIX "sk %s:%s nwr: ", sk->host, sk->port);
 	Assert(!sk->xlogreader);
-	sk->xlogreader = NeonWALReaderAllocate(wal_segment_size, sk->wp->propTermStartLsn, log_prefix);
+	sk->xlogreader = NeonWALReaderAllocate(wal_segment_size, sk->wp->propTermStartLsn, log_prefix, sk->wp->localTimeLineID);
 	if (sk->xlogreader == NULL)
 		wpg_log(FATAL, "failed to allocate xlog reader");
 }
@@ -1522,7 +1564,7 @@ walprop_pg_wal_read(Safekeeper *sk, char *buf, XLogRecPtr startptr, Size count,
 					  buf,
 					  startptr,
 					  count,
-					  walprop_pg_get_timeline_id());
+					  sk->wp->localTimeLineID);
 
 	if (res == NEON_WALREAD_SUCCESS)
 	{

pgxn/neon/walsender_hooks.c

@@ -111,7 +111,7 @@ NeonWALPageRead(
 											readBuf,
 											targetPagePtr,
 											count,
-											walprop_pg_get_timeline_id());
+											NeonWALReaderLocalActiveTimeLineID(wal_reader));
 
 		if (res == NEON_WALREAD_SUCCESS)
 		{
@@ -202,7 +202,7 @@ NeonOnDemandXLogReaderRoutines(XLogReaderRoutine *xlr)
 		{
 			elog(ERROR, "unable to start walsender when basebackupLsn is 0");
 		}
-		wal_reader = NeonWALReaderAllocate(wal_segment_size, basebackupLsn, "[walsender] ");
+		wal_reader = NeonWALReaderAllocate(wal_segment_size, basebackupLsn, "[walsender] ", 1);
 	}
 	xlr->page_read = NeonWALPageRead;
 	xlr->segment_open = NeonWALReadSegmentOpen;

proxy/src/auth/backend/classic.rs

@@ -17,35 +17,18 @@ pub(super) async fn authenticate(
     config: &'static AuthenticationConfig,
     secret: AuthSecret,
 ) -> auth::Result<ComputeCredentials> {
-    let flow = AuthFlow::new(client);
     let scram_keys = match secret {
-        #[cfg(any(test, feature = "testing"))]
-        AuthSecret::Md5(_) => {
-            debug!("auth endpoint chooses MD5");
-            return Err(auth::AuthError::bad_auth_method("MD5"));
-        }
         AuthSecret::Scram(secret) => {
             debug!("auth endpoint chooses SCRAM");
-            let scram = auth::Scram(&secret, ctx);
 
             let auth_outcome = tokio::time::timeout(
                 config.scram_protocol_timeout,
-                async {
-
-                    flow.begin(scram).await.map_err(|error| {
-                        warn!(?error, "error sending scram acknowledgement");
-                        error
-                    })?.authenticate().await.map_err(|error| {
-                        warn!(?error, "error processing scram messages");
-                        error
-                    })
-                }
+                AuthFlow::new(client, auth::Scram(&secret, ctx)).authenticate(),
             )
             .await
-            .map_err(|e| {
-                warn!("error processing scram messages error = authentication timed out, execution time exceeded {} seconds", config.scram_protocol_timeout.as_secs());
-                auth::AuthError::user_timeout(e)
-            })??;
+            .inspect_err(|_| warn!("error processing scram messages error = authentication timed out, execution time exceeded {} seconds", config.scram_protocol_timeout.as_secs()))
+            .map_err(auth::AuthError::user_timeout)?
+            .inspect_err(|error| warn!(?error, "error processing scram messages"))?;
 
             let client_key = match auth_outcome {
                 sasl::Outcome::Success(key) => key,

proxy/src/auth/backend/console_redirect.rs

@@ -2,22 +2,21 @@ use std::fmt;
 
 use async_trait::async_trait;
 use postgres_client::config::SslMode;
-use pq_proto::BeMessage as Be;
 use thiserror::Error;
 use tokio::io::{AsyncRead, AsyncWrite};
 use tracing::{info, info_span};
 
-use super::ComputeCredentialKeys;
-use crate::auth::IpPattern;
 use crate::auth::backend::ComputeUserInfo;
 use crate::cache::Cached;
+use crate::compute::AuthInfo;
 use crate::config::AuthenticationConfig;
 use crate::context::RequestContext;
 use crate::control_plane::client::cplane_proxy_v1;
 use crate::control_plane::{self, CachedNodeInfo, NodeInfo};
 use crate::error::{ReportableError, UserFacingError};
+use crate::pglb::connect_compute::WakeComputeBackend;
+use crate::pqproto::BeMessage;
 use crate::proxy::NeonOptions;
-use crate::proxy::connect_compute::ComputeConnectBackend;
 use crate::stream::PqStream;
 use crate::types::RoleName;
 use crate::{auth, compute, waiters};
@@ -98,15 +97,11 @@ impl ConsoleRedirectBackend {
         ctx: &RequestContext,
         auth_config: &'static AuthenticationConfig,
         client: &mut PqStream<impl AsyncRead + AsyncWrite + Unpin>,
-    ) -> auth::Result<(
-        ConsoleRedirectNodeInfo,
-        ComputeUserInfo,
-        Option<Vec<IpPattern>>,
-    )> {
+    ) -> auth::Result<(ConsoleRedirectNodeInfo, AuthInfo, ComputeUserInfo)> {
         authenticate(ctx, auth_config, &self.console_uri, client)
             .await
-            .map(|(node_info, user_info, ip_allowlist)| {
-                (ConsoleRedirectNodeInfo(node_info), user_info, ip_allowlist)
+            .map(|(node_info, auth_info, user_info)| {
+                (ConsoleRedirectNodeInfo(node_info), auth_info, user_info)
             })
     }
 }
@@ -114,17 +109,13 @@ impl ConsoleRedirectBackend {
 pub struct ConsoleRedirectNodeInfo(pub(super) NodeInfo);
 
 #[async_trait]
-impl ComputeConnectBackend for ConsoleRedirectNodeInfo {
+impl WakeComputeBackend for ConsoleRedirectNodeInfo {
     async fn wake_compute(
         &self,
         _ctx: &RequestContext,
     ) -> Result<CachedNodeInfo, control_plane::errors::WakeComputeError> {
         Ok(Cached::new_uncached(self.0.clone()))
     }
-
-    fn get_keys(&self) -> &ComputeCredentialKeys {
-        &ComputeCredentialKeys::None
-    }
 }
 
 async fn authenticate(
@@ -132,7 +123,7 @@ async fn authenticate(
     auth_config: &'static AuthenticationConfig,
     link_uri: &reqwest::Url,
     client: &mut PqStream<impl AsyncRead + AsyncWrite + Unpin>,
-) -> auth::Result<(NodeInfo, ComputeUserInfo, Option<Vec<IpPattern>>)> {
+) -> auth::Result<(NodeInfo, AuthInfo, ComputeUserInfo)> {
     ctx.set_auth_method(crate::context::AuthMethod::ConsoleRedirect);
 
     // registering waiter can fail if we get unlucky with rng.
@@ -154,11 +145,13 @@ async fn authenticate(
 
     // Give user a URL to spawn a new database.
     info!(parent: &span, "sending the auth URL to the user");
-    client
-        .write_message_noflush(&Be::AuthenticationOk)?
-        .write_message_noflush(&Be::CLIENT_ENCODING)?
-        .write_message(&Be::NoticeResponse(&greeting))
-        .await?;
+    client.write_message(BeMessage::AuthenticationOk);
+    client.write_message(BeMessage::ParameterStatus {
+        name: b"client_encoding",
+        value: b"UTF8",
+    });
+    client.write_message(BeMessage::NoticeResponse(&greeting));
+    client.flush().await?;
 
     // Wait for console response via control plane (see `mgmt`).
     info!(parent: &span, "waiting for console's reply...");
@@ -188,12 +181,26 @@ async fn authenticate(
         }
     }
 
-    client.write_message_noflush(&Be::NoticeResponse("Connecting to database."))?;
+    client.write_message(BeMessage::NoticeResponse("Connecting to database."));
 
-    // This config should be self-contained, because we won't
-    // take username or dbname from client's startup message.
-    let mut config = compute::ConnCfg::new(db_info.host.to_string(), db_info.port);
-    config.dbname(&db_info.dbname).user(&db_info.user);
+    // Backwards compatibility. pg_sni_proxy uses "--" in domain names
+    // while direct connections do not. Once we migrate to pg_sni_proxy
+    // everywhere, we can remove this.
+    let ssl_mode = if db_info.host.contains("--") {
+        // we need TLS connection with SNI info to properly route it
+        SslMode::Require
+    } else {
+        SslMode::Disable
+    };
+
+    let conn_info = compute::ConnectInfo {
+        host: db_info.host.into(),
+        port: db_info.port,
+        ssl_mode,
+        host_addr: None,
+    };
+    let auth_info =
+        AuthInfo::for_console_redirect(&db_info.dbname, &db_info.user, db_info.password.as_deref());
 
     let user: RoleName = db_info.user.into();
     let user_info = ComputeUserInfo {
@@ -207,26 +214,12 @@ async fn authenticate(
     ctx.set_project(db_info.aux.clone());
     info!("woken up a compute node");
 
-    // Backwards compatibility. pg_sni_proxy uses "--" in domain names
-    // while direct connections do not. Once we migrate to pg_sni_proxy
-    // everywhere, we can remove this.
-    if db_info.host.contains("--") {
-        // we need TLS connection with SNI info to properly route it
-        config.ssl_mode(SslMode::Require);
-    } else {
-        config.ssl_mode(SslMode::Disable);
-    }
-
-    if let Some(password) = db_info.password {
-        config.password(password.as_ref());
-    }
-
     Ok((
         NodeInfo {
-            config,
+            conn_info,
             aux: db_info.aux,
         },
+        auth_info,
         user_info,
-        db_info.allowed_ips,
     ))
 }

proxy/src/auth/backend/hacks.rs

@@ -24,23 +24,25 @@ pub(crate) async fn authenticate_cleartext(
     debug!("cleartext auth flow override is enabled, proceeding");
     ctx.set_auth_method(crate::context::AuthMethod::Cleartext);
 
-    // pause the timer while we communicate with the client
-    let paused = ctx.latency_timer_pause(crate::metrics::Waiting::Client);
-
     let ep = EndpointIdInt::from(&info.endpoint);
 
-    let auth_flow = AuthFlow::new(client)
-        .begin(auth::CleartextPassword {
+    let auth_flow = AuthFlow::new(
+        client,
+        auth::CleartextPassword {
             secret,
             endpoint: ep,
             pool: config.thread_pool.clone(),
-        })
-        .await?;
-    drop(paused);
-    // cleartext auth is only allowed to the ws/http protocol.
-    // If we're here, we already received the password in the first message.
-    // Scram protocol will be executed on the proxy side.
-    let auth_outcome = auth_flow.authenticate().await?;
+        },
+    );
+    let auth_outcome = {
+        // pause the timer while we communicate with the client
+        let _paused = ctx.latency_timer_pause(crate::metrics::Waiting::Client);
+
+        // cleartext auth is only allowed to the ws/http protocol.
+        // If we're here, we already received the password in the first message.
+        // Scram protocol will be executed on the proxy side.
+        auth_flow.authenticate().await?
+    };
 
     let keys = match auth_outcome {
         sasl::Outcome::Success(key) => key,
@@ -67,9 +69,7 @@ pub(crate) async fn password_hack_no_authentication(
     // pause the timer while we communicate with the client
     let _paused = ctx.latency_timer_pause(crate::metrics::Waiting::Client);
 
-    let payload = AuthFlow::new(client)
-        .begin(auth::PasswordHack)
-        .await?
+    let payload = AuthFlow::new(client, auth::PasswordHack)
         .get_password()
         .await?;
 

proxy/src/auth/backend/local.rs

@@ -1,11 +1,12 @@
 use std::net::SocketAddr;
 
 use arc_swap::ArcSwapOption;
+use postgres_client::config::SslMode;
 use tokio::sync::Semaphore;
 
 use super::jwt::{AuthRule, FetchAuthRules};
 use crate::auth::backend::jwt::FetchAuthRulesError;
-use crate::compute::ConnCfg;
+use crate::compute::ConnectInfo;
 use crate::compute_ctl::ComputeCtlApi;
 use crate::context::RequestContext;
 use crate::control_plane::NodeInfo;
@@ -29,7 +30,12 @@ impl LocalBackend {
                 api: http::Endpoint::new(compute_ctl, http::new_client()),
             },
             node_info: NodeInfo {
-                config: ConnCfg::new(postgres_addr.ip().to_string(), postgres_addr.port()),
+                conn_info: ConnectInfo {
+                    host_addr: Some(postgres_addr.ip()),
+                    host: postgres_addr.ip().to_string().into(),
+                    port: postgres_addr.port(),
+                    ssl_mode: SslMode::Disable,
+                },
                 // TODO(conrad): make this better reflect compute info rather than endpoint info.
                 aux: MetricsAuxInfo {
                     endpoint_id: EndpointIdTag::get_interner().get_or_intern("local"),

proxy/src/auth/backend/mod.rs

@@ -4,37 +4,31 @@ mod hacks;
 pub mod jwt;
 pub mod local;
 
-use std::net::IpAddr;
 use std::sync::Arc;
 
 pub use console_redirect::ConsoleRedirectBackend;
 pub(crate) use console_redirect::ConsoleRedirectError;
-use ipnet::{Ipv4Net, Ipv6Net};
 use local::LocalBackend;
 use postgres_client::config::AuthKeys;
 use serde::{Deserialize, Serialize};
 use tokio::io::{AsyncRead, AsyncWrite};
-use tracing::{debug, info, warn};
+use tracing::{debug, info};
 
-use crate::auth::credentials::check_peer_addr_is_in_list;
-use crate::auth::{
-    self, AuthError, ComputeUserInfoMaybeEndpoint, IpPattern, validate_password_and_exchange,
-};
+use crate::auth::{self, AuthError, ComputeUserInfoMaybeEndpoint, validate_password_and_exchange};
 use crate::cache::Cached;
 use crate::config::AuthenticationConfig;
 use crate::context::RequestContext;
 use crate::control_plane::client::ControlPlaneClient;
 use crate::control_plane::errors::GetAuthInfoError;
 use crate::control_plane::{
-    self, AccessBlockerFlags, AuthSecret, CachedAccessBlockerFlags, CachedAllowedIps,
-    CachedAllowedVpcEndpointIds, CachedNodeInfo, CachedRoleSecret, ControlPlaneApi,
+    self, AccessBlockerFlags, AuthSecret, CachedNodeInfo, ControlPlaneApi, EndpointAccessControl,
+    RoleAccessControl,
 };
 use crate::intern::EndpointIdInt;
-use crate::metrics::Metrics;
-use crate::protocol2::ConnectionInfoExtra;
+use crate::pglb::connect_compute::WakeComputeBackend;
+use crate::pqproto::BeMessage;
 use crate::proxy::NeonOptions;
-use crate::proxy::connect_compute::ComputeConnectBackend;
-use crate::rate_limiter::{BucketRateLimiter, EndpointRateLimiter};
+use crate::rate_limiter::EndpointRateLimiter;
 use crate::stream::Stream;
 use crate::types::{EndpointCacheKey, EndpointId, RoleName};
 use crate::{scram, stream};
@@ -174,8 +168,6 @@ impl ComputeUserInfo {
 
 #[cfg_attr(test, derive(Debug))]
 pub(crate) enum ComputeCredentialKeys {
-    #[cfg(any(test, feature = "testing"))]
-    Password(Vec<u8>),
     AuthKeys(AuthKeys),
     JwtPayload(Vec<u8>),
     None,
@@ -200,78 +192,6 @@ impl TryFrom<ComputeUserInfoMaybeEndpoint> for ComputeUserInfo {
     }
 }
 
-#[derive(PartialEq, PartialOrd, Hash, Eq, Ord, Debug, Copy, Clone)]
-pub struct MaskedIp(IpAddr);
-
-impl MaskedIp {
-    fn new(value: IpAddr, prefix: u8) -> Self {
-        match value {
-            IpAddr::V4(v4) => Self(IpAddr::V4(
-                Ipv4Net::new(v4, prefix).map_or(v4, |x| x.trunc().addr()),
-            )),
-            IpAddr::V6(v6) => Self(IpAddr::V6(
-                Ipv6Net::new(v6, prefix).map_or(v6, |x| x.trunc().addr()),
-            )),
-        }
-    }
-}
-
-// This can't be just per IP because that would limit some PaaS that share IP addresses
-pub type AuthRateLimiter = BucketRateLimiter<(EndpointIdInt, MaskedIp)>;
-
-impl AuthenticationConfig {
-    pub(crate) fn check_rate_limit(
-        &self,
-        ctx: &RequestContext,
-        secret: AuthSecret,
-        endpoint: &EndpointId,
-        is_cleartext: bool,
-    ) -> auth::Result<AuthSecret> {
-        // we have validated the endpoint exists, so let's intern it.
-        let endpoint_int = EndpointIdInt::from(endpoint.normalize());
-
-        // only count the full hash count if password hack or websocket flow.
-        // in other words, if proxy needs to run the hashing
-        let password_weight = if is_cleartext {
-            match &secret {
-                #[cfg(any(test, feature = "testing"))]
-                AuthSecret::Md5(_) => 1,
-                AuthSecret::Scram(s) => s.iterations + 1,
-            }
-        } else {
-            // validating scram takes just 1 hmac_sha_256 operation.
-            1
-        };
-
-        let limit_not_exceeded = self.rate_limiter.check(
-            (
-                endpoint_int,
-                MaskedIp::new(ctx.peer_addr(), self.rate_limit_ip_subnet),
-            ),
-            password_weight,
-        );
-
-        if !limit_not_exceeded {
-            warn!(
-                enabled = self.rate_limiter_enabled,
-                "rate limiting authentication"
-            );
-            Metrics::get().proxy.requests_auth_rate_limits_total.inc();
-            Metrics::get()
-                .proxy
-                .endpoints_auth_rate_limits
-                .get_metric()
-                .measure(endpoint);
-
-            if self.rate_limiter_enabled {
-                return Err(auth::AuthError::too_many_connections());
-            }
-        }
-
-        Ok(secret)
-    }
-}
-
 /// True to its name, this function encapsulates our current auth trade-offs.
 /// Here, we choose the appropriate auth flow based on circumstances.
 ///
@@ -284,7 +204,7 @@ async fn auth_quirks(
     allow_cleartext: bool,
     config: &'static AuthenticationConfig,
     endpoint_rate_limiter: Arc<EndpointRateLimiter>,
-) -> auth::Result<(ComputeCredentials, Option<Vec<IpPattern>>)> {
+) -> auth::Result<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.
@@ -300,55 +220,27 @@ async fn auth_quirks(
 
     debug!("fetching authentication info and allowlists");
 
-    // check allowed list
-    let allowed_ips = if config.ip_allowlist_check_enabled {
-        let allowed_ips = api.get_allowed_ips(ctx, &info).await?;
-        if !check_peer_addr_is_in_list(&ctx.peer_addr(), &allowed_ips) {
-            return Err(auth::AuthError::ip_address_not_allowed(ctx.peer_addr()));
-        }
-        allowed_ips
-    } else {
-        Cached::new_uncached(Arc::new(vec![]))
-    };
+    let access_controls = api
+        .get_endpoint_access_control(ctx, &info.endpoint, &info.user)
+        .await?;
 
-    // check if a VPC endpoint ID is coming in and if yes, if it's allowed
-    let access_blocks = api.get_block_public_or_vpc_access(ctx, &info).await?;
-    if config.is_vpc_acccess_proxy {
-        if access_blocks.vpc_access_blocked {
-            return Err(AuthError::NetworkNotAllowed);
-        }
+    access_controls.check(
+        ctx,
+        config.ip_allowlist_check_enabled,
+        config.is_vpc_acccess_proxy,
+    )?;
 
-        let incoming_vpc_endpoint_id = match ctx.extra() {
-            None => return Err(AuthError::MissingEndpointName),
-            Some(ConnectionInfoExtra::Aws { vpce_id }) => vpce_id.to_string(),
-            Some(ConnectionInfoExtra::Azure { link_id }) => link_id.to_string(),
-        };
-        let allowed_vpc_endpoint_ids = api.get_allowed_vpc_endpoint_ids(ctx, &info).await?;
-        // TODO: For now an empty VPC endpoint ID list means all are allowed. We should replace that.
-        if !allowed_vpc_endpoint_ids.is_empty()
-            && !allowed_vpc_endpoint_ids.contains(&incoming_vpc_endpoint_id)
-        {
-            return Err(AuthError::vpc_endpoint_id_not_allowed(
-                incoming_vpc_endpoint_id,
-            ));
-        }
-    } else if access_blocks.public_access_blocked {
-        return Err(AuthError::NetworkNotAllowed);
-    }
-
-    if !endpoint_rate_limiter.check(info.endpoint.clone().into(), 1) {
+    let endpoint = EndpointIdInt::from(&info.endpoint);
+    let rate_limit_config = None;
+    if !endpoint_rate_limiter.check(endpoint, rate_limit_config, 1) {
         return Err(AuthError::too_many_connections());
     }
-    let cached_secret = api.get_role_secret(ctx, &info).await?;
-    let (cached_entry, secret) = cached_secret.take_value();
+    let role_access = api
+        .get_role_access_control(ctx, &info.endpoint, &info.user)
+        .await?;
 
-    let secret = if let Some(secret) = secret {
-        config.check_rate_limit(
-            ctx,
-            secret,
-            &info.endpoint,
-            unauthenticated_password.is_some() || allow_cleartext,
-        )?
+    let secret = if let Some(secret) = role_access.secret {
+        secret
     } else {
         // If we don't have an authentication secret, we mock one to
         // prevent malicious probing (possible due to missing protocol steps).
@@ -368,14 +260,8 @@ async fn auth_quirks(
     )
     .await
     {
-        Ok(keys) => Ok((keys, Some(allowed_ips.as_ref().clone()))),
-        Err(e) => {
-            if e.is_password_failed() {
-                // The password could have been changed, so we invalidate the cache.
-                cached_entry.invalidate();
-            }
-            Err(e)
-        }
+        Ok(keys) => Ok(keys),
+        Err(e) => Err(e),
     }
 }
 
@@ -402,7 +288,7 @@ async fn authenticate_with_secret(
         };
 
         // we have authenticated the password
-        client.write_message_noflush(&pq_proto::BeMessage::AuthenticationOk)?;
+        client.write_message(BeMessage::AuthenticationOk);
 
         return Ok(ComputeCredentials { info, keys });
     }
@@ -438,7 +324,7 @@ impl<'a> Backend<'a, ComputeUserInfoMaybeEndpoint> {
         allow_cleartext: bool,
         config: &'static AuthenticationConfig,
         endpoint_rate_limiter: Arc<EndpointRateLimiter>,
-    ) -> auth::Result<(Backend<'a, ComputeCredentials>, Option<Vec<IpPattern>>)> {
+    ) -> auth::Result<Backend<'a, ComputeCredentials>> {
         let res = match self {
             Self::ControlPlane(api, user_info) => {
                 debug!(
@@ -447,17 +333,35 @@ impl<'a> Backend<'a, ComputeUserInfoMaybeEndpoint> {
                     "performing authentication using the console"
                 );
 
-                let (credentials, ip_allowlist) = auth_quirks(
+                let auth_res = auth_quirks(
                     ctx,
                     &*api,
-                    user_info,
+                    user_info.clone(),
                     client,
                     allow_cleartext,
                     config,
                     endpoint_rate_limiter,
                 )
-                .await?;
-                Ok((Backend::ControlPlane(api, credentials), ip_allowlist))
+                .await;
+                match auth_res {
+                    Ok(credentials) => Ok(Backend::ControlPlane(api, credentials)),
+                    Err(e) => {
+                        // The password could have been changed, so we invalidate the cache.
+                        // We should only invalidate the cache if the TTL might have expired.
+                        if e.is_password_failed() {
+                            #[allow(irrefutable_let_patterns)]
+                            if let ControlPlaneClient::ProxyV1(api) = &*api {
+                                if let Some(ep) = &user_info.endpoint_id {
+                                    api.caches
+                                        .project_info
+                                        .maybe_invalidate_role_secret(ep, &user_info.user);
+                                }
+                            }
+                        }
+
+                        Err(e)
+                    }
+                }
             }
             Self::Local(_) => {
                 return Err(auth::AuthError::bad_auth_method("invalid for local proxy"));
@@ -474,75 +378,52 @@ impl Backend<'_, ComputeUserInfo> {
     pub(crate) async fn get_role_secret(
         &self,
         ctx: &RequestContext,
-    ) -> Result<CachedRoleSecret, GetAuthInfoError> {
-        match self {
-            Self::ControlPlane(api, user_info) => api.get_role_secret(ctx, user_info).await,
-            Self::Local(_) => Ok(Cached::new_uncached(None)),
-        }
-    }
-
-    pub(crate) async fn get_allowed_ips(
-        &self,
-        ctx: &RequestContext,
-    ) -> Result<CachedAllowedIps, GetAuthInfoError> {
-        match self {
-            Self::ControlPlane(api, user_info) => api.get_allowed_ips(ctx, user_info).await,
-            Self::Local(_) => Ok(Cached::new_uncached(Arc::new(vec![]))),
-        }
-    }
-
-    pub(crate) async fn get_allowed_vpc_endpoint_ids(
-        &self,
-        ctx: &RequestContext,
-    ) -> Result<CachedAllowedVpcEndpointIds, GetAuthInfoError> {
+    ) -> Result<RoleAccessControl, GetAuthInfoError> {
         match self {
             Self::ControlPlane(api, user_info) => {
-                api.get_allowed_vpc_endpoint_ids(ctx, user_info).await
+                api.get_role_access_control(ctx, &user_info.endpoint, &user_info.user)
+                    .await
             }
-            Self::Local(_) => Ok(Cached::new_uncached(Arc::new(vec![]))),
+            Self::Local(_) => Ok(RoleAccessControl { secret: None }),
         }
     }
 
-    pub(crate) async fn get_block_public_or_vpc_access(
+    pub(crate) async fn get_endpoint_access_control(
         &self,
         ctx: &RequestContext,
-    ) -> Result<CachedAccessBlockerFlags, GetAuthInfoError> {
+    ) -> Result<EndpointAccessControl, GetAuthInfoError> {
         match self {
             Self::ControlPlane(api, user_info) => {
-                api.get_block_public_or_vpc_access(ctx, user_info).await
+                api.get_endpoint_access_control(ctx, &user_info.endpoint, &user_info.user)
+                    .await
             }
-            Self::Local(_) => Ok(Cached::new_uncached(AccessBlockerFlags::default())),
+            Self::Local(_) => Ok(EndpointAccessControl {
+                allowed_ips: Arc::new(vec![]),
+                allowed_vpce: Arc::new(vec![]),
+                flags: AccessBlockerFlags::default(),
+            }),
         }
     }
 }
 
 #[async_trait::async_trait]
-impl ComputeConnectBackend for Backend<'_, ComputeCredentials> {
+impl WakeComputeBackend for Backend<'_, ComputeUserInfo> {
     async fn wake_compute(
         &self,
         ctx: &RequestContext,
     ) -> Result<CachedNodeInfo, control_plane::errors::WakeComputeError> {
         match self {
-            Self::ControlPlane(api, creds) => api.wake_compute(ctx, &creds.info).await,
+            Self::ControlPlane(api, info) => api.wake_compute(ctx, info).await,
             Self::Local(local) => Ok(Cached::new_uncached(local.node_info.clone())),
         }
     }
-
-    fn get_keys(&self) -> &ComputeCredentialKeys {
-        match self {
-            Self::ControlPlane(_, creds) => &creds.keys,
-            Self::Local(_) => &ComputeCredentialKeys::None,
-        }
-    }
 }
 
 #[cfg(test)]
 mod tests {
     #![allow(clippy::unimplemented, clippy::unwrap_used)]
 
-    use std::net::IpAddr;
     use std::sync::Arc;
-    use std::time::Duration;
 
     use bytes::BytesMut;
     use control_plane::AuthSecret;
@@ -553,18 +434,16 @@ mod tests {
     use postgres_protocol::message::frontend;
     use tokio::io::{AsyncRead, AsyncReadExt, AsyncWriteExt};
 
+    use super::auth_quirks;
     use super::jwt::JwkCache;
-    use super::{AuthRateLimiter, auth_quirks};
-    use crate::auth::backend::MaskedIp;
     use crate::auth::{ComputeUserInfoMaybeEndpoint, IpPattern};
     use crate::config::AuthenticationConfig;
     use crate::context::RequestContext;
     use crate::control_plane::{
-        self, AccessBlockerFlags, CachedAccessBlockerFlags, CachedAllowedIps,
-        CachedAllowedVpcEndpointIds, CachedNodeInfo, CachedRoleSecret,
+        self, AccessBlockerFlags, CachedNodeInfo, EndpointAccessControl, RoleAccessControl,
     };
     use crate::proxy::NeonOptions;
-    use crate::rate_limiter::{EndpointRateLimiter, RateBucketInfo};
+    use crate::rate_limiter::EndpointRateLimiter;
     use crate::scram::ServerSecret;
     use crate::scram::threadpool::ThreadPool;
     use crate::stream::{PqStream, Stream};
@@ -577,46 +456,34 @@ mod tests {
     }
 
     impl control_plane::ControlPlaneApi for Auth {
-        async fn get_role_secret(
+        async fn get_role_access_control(
             &self,
             _ctx: &RequestContext,
-            _user_info: &super::ComputeUserInfo,
-        ) -> Result<CachedRoleSecret, control_plane::errors::GetAuthInfoError> {
-            Ok(CachedRoleSecret::new_uncached(Some(self.secret.clone())))
+            _endpoint: &crate::types::EndpointId,
+            _role: &crate::types::RoleName,
+        ) -> Result<RoleAccessControl, control_plane::errors::GetAuthInfoError> {
+            Ok(RoleAccessControl {
+                secret: Some(self.secret.clone()),
+            })
         }
 
-        async fn get_allowed_ips(
+        async fn get_endpoint_access_control(
             &self,
             _ctx: &RequestContext,
-            _user_info: &super::ComputeUserInfo,
-        ) -> Result<CachedAllowedIps, control_plane::errors::GetAuthInfoError> {
-            Ok(CachedAllowedIps::new_uncached(Arc::new(self.ips.clone())))
-        }
-
-        async fn get_allowed_vpc_endpoint_ids(
-            &self,
-            _ctx: &RequestContext,
-            _user_info: &super::ComputeUserInfo,
-        ) -> Result<CachedAllowedVpcEndpointIds, control_plane::errors::GetAuthInfoError> {
-            Ok(CachedAllowedVpcEndpointIds::new_uncached(Arc::new(
-                self.vpc_endpoint_ids.clone(),
-            )))
-        }
-
-        async fn get_block_public_or_vpc_access(
-            &self,
-            _ctx: &RequestContext,
-            _user_info: &super::ComputeUserInfo,
-        ) -> Result<CachedAccessBlockerFlags, control_plane::errors::GetAuthInfoError> {
-            Ok(CachedAccessBlockerFlags::new_uncached(
-                self.access_blocker_flags.clone(),
-            ))
+            _endpoint: &crate::types::EndpointId,
+            _role: &crate::types::RoleName,
+        ) -> Result<EndpointAccessControl, control_plane::errors::GetAuthInfoError> {
+            Ok(EndpointAccessControl {
+                allowed_ips: Arc::new(self.ips.clone()),
+                allowed_vpce: Arc::new(self.vpc_endpoint_ids.clone()),
+                flags: self.access_blocker_flags,
+            })
         }
 
         async fn get_endpoint_jwks(
             &self,
             _ctx: &RequestContext,
-            _endpoint: crate::types::EndpointId,
+            _endpoint: &crate::types::EndpointId,
         ) -> Result<Vec<super::jwt::AuthRule>, control_plane::errors::GetEndpointJwksError>
         {
             unimplemented!()
@@ -635,9 +502,6 @@ mod tests {
         jwks_cache: JwkCache::default(),
         thread_pool: ThreadPool::new(1),
         scram_protocol_timeout: std::time::Duration::from_secs(5),
-        rate_limiter_enabled: true,
-        rate_limiter: AuthRateLimiter::new(&RateBucketInfo::DEFAULT_AUTH_SET),
-        rate_limit_ip_subnet: 64,
         ip_allowlist_check_enabled: true,
         is_vpc_acccess_proxy: false,
         is_auth_broker: false,
@@ -654,55 +518,10 @@ mod tests {
         }
     }
 
-    #[test]
-    fn masked_ip() {
-        let ip_a = IpAddr::V4([127, 0, 0, 1].into());
-        let ip_b = IpAddr::V4([127, 0, 0, 2].into());
-        let ip_c = IpAddr::V4([192, 168, 1, 101].into());
-        let ip_d = IpAddr::V4([192, 168, 1, 102].into());
-        let ip_e = IpAddr::V6("abcd:abcd:abcd:abcd:abcd:abcd:abcd:abcd".parse().unwrap());
-        let ip_f = IpAddr::V6("abcd:abcd:abcd:abcd:1234:abcd:abcd:abcd".parse().unwrap());
-
-        assert_ne!(MaskedIp::new(ip_a, 64), MaskedIp::new(ip_b, 64));
-        assert_ne!(MaskedIp::new(ip_a, 32), MaskedIp::new(ip_b, 32));
-        assert_eq!(MaskedIp::new(ip_a, 30), MaskedIp::new(ip_b, 30));
-        assert_eq!(MaskedIp::new(ip_c, 30), MaskedIp::new(ip_d, 30));
-
-        assert_ne!(MaskedIp::new(ip_e, 128), MaskedIp::new(ip_f, 128));
-        assert_eq!(MaskedIp::new(ip_e, 64), MaskedIp::new(ip_f, 64));
-    }
-
-    #[test]
-    fn test_default_auth_rate_limit_set() {
-        // these values used to exceed u32::MAX
-        assert_eq!(
-            RateBucketInfo::DEFAULT_AUTH_SET,
-            [
-                RateBucketInfo {
-                    interval: Duration::from_secs(1),
-                    max_rpi: 1000 * 4096,
-                },
-                RateBucketInfo {
-                    interval: Duration::from_secs(60),
-                    max_rpi: 600 * 4096 * 60,
-                },
-                RateBucketInfo {
-                    interval: Duration::from_secs(600),
-                    max_rpi: 300 * 4096 * 600,
-                }
-            ]
-        );
-
-        for x in RateBucketInfo::DEFAULT_AUTH_SET {
-            let y = x.to_string().parse().unwrap();
-            assert_eq!(x, y);
-        }
-    }
-
     #[tokio::test]
     async fn auth_quirks_scram() {
         let (mut client, server) = tokio::io::duplex(1024);
-        let mut stream = PqStream::new(Stream::from_raw(server));
+        let mut stream = PqStream::new_skip_handshake(Stream::from_raw(server));
 
         let ctx = RequestContext::test();
         let api = Auth {
@@ -784,7 +603,7 @@ mod tests {
     #[tokio::test]
     async fn auth_quirks_cleartext() {
         let (mut client, server) = tokio::io::duplex(1024);
-        let mut stream = PqStream::new(Stream::from_raw(server));
+        let mut stream = PqStream::new_skip_handshake(Stream::from_raw(server));
 
         let ctx = RequestContext::test();
         let api = Auth {
@@ -838,7 +657,7 @@ mod tests {
     #[tokio::test]
     async fn auth_quirks_password_hack() {
         let (mut client, server) = tokio::io::duplex(1024);
-        let mut stream = PqStream::new(Stream::from_raw(server));
+        let mut stream = PqStream::new_skip_handshake(Stream::from_raw(server));
 
         let ctx = RequestContext::test();
         let api = Auth {
@@ -887,7 +706,7 @@ mod tests {
         .await
         .unwrap();
 
-        assert_eq!(creds.0.info.endpoint, "my-endpoint");
+        assert_eq!(creds.info.endpoint, "my-endpoint");
 
         handle.await.unwrap();
     }

proxy/src/auth/credentials.rs

@@ -5,7 +5,6 @@ use std::net::IpAddr;
 use std::str::FromStr;
 
 use itertools::Itertools;
-use pq_proto::StartupMessageParams;
 use thiserror::Error;
 use tracing::{debug, warn};
 
@@ -13,6 +12,7 @@ use crate::auth::password_hack::parse_endpoint_param;
 use crate::context::RequestContext;
 use crate::error::{ReportableError, UserFacingError};
 use crate::metrics::{Metrics, SniGroup, SniKind};
+use crate::pqproto::StartupMessageParams;
 use crate::proxy::NeonOptions;
 use crate::serverless::{AUTH_BROKER_SNI, SERVERLESS_DRIVER_SNI};
 use crate::types::{EndpointId, RoleName};

proxy/src/auth/flow.rs

@@ -1,10 +1,8 @@
 //! Main authentication flow.
 
-use std::io;
 use std::sync::Arc;
 
 use postgres_protocol::authentication::sasl::{SCRAM_SHA_256, SCRAM_SHA_256_PLUS};
-use pq_proto::{BeAuthenticationSaslMessage, BeMessage, BeMessage as Be};
 use tokio::io::{AsyncRead, AsyncWrite};
 use tracing::info;
 
@@ -13,35 +11,26 @@ use super::{AuthError, PasswordHackPayload};
 use crate::context::RequestContext;
 use crate::control_plane::AuthSecret;
 use crate::intern::EndpointIdInt;
+use crate::pqproto::{BeAuthenticationSaslMessage, BeMessage};
 use crate::sasl;
 use crate::scram::threadpool::ThreadPool;
 use crate::scram::{self};
 use crate::stream::{PqStream, Stream};
 use crate::tls::TlsServerEndPoint;
 
-/// Every authentication selector is supposed to implement this trait.
-pub(crate) trait AuthMethod {
-    /// Any authentication selector should provide initial backend message
-    /// containing auth method name and parameters, e.g. md5 salt.
-    fn first_message(&self, channel_binding: bool) -> BeMessage<'_>;
-}
-
-/// Initial state of [`AuthFlow`].
-pub(crate) struct Begin;
-
 /// Use [SCRAM](crate::scram)-based auth in [`AuthFlow`].
 pub(crate) struct Scram<'a>(
     pub(crate) &'a scram::ServerSecret,
     pub(crate) &'a RequestContext,
 );
 
-impl AuthMethod for Scram<'_> {
+impl Scram<'_> {
     #[inline(always)]
     fn first_message(&self, channel_binding: bool) -> BeMessage<'_> {
         if channel_binding {
-            Be::AuthenticationSasl(BeAuthenticationSaslMessage::Methods(scram::METHODS))
+            BeMessage::AuthenticationSasl(BeAuthenticationSaslMessage::Methods(scram::METHODS))
         } else {
-            Be::AuthenticationSasl(BeAuthenticationSaslMessage::Methods(
+            BeMessage::AuthenticationSasl(BeAuthenticationSaslMessage::Methods(
                 scram::METHODS_WITHOUT_PLUS,
             ))
         }
@@ -52,13 +41,6 @@ impl AuthMethod for Scram<'_> {
 /// <https://github.com/neondatabase/cloud/issues/1620#issuecomment-1165332290>.
 pub(crate) struct PasswordHack;
 
-impl AuthMethod for PasswordHack {
-    #[inline(always)]
-    fn first_message(&self, _channel_binding: bool) -> BeMessage<'_> {
-        Be::AuthenticationCleartextPassword
-    }
-}
-
 /// Use clear-text password auth called `password` in docs
 /// <https://www.postgresql.org/docs/current/auth-password.html>
 pub(crate) struct CleartextPassword {
@@ -67,53 +49,37 @@ pub(crate) struct CleartextPassword {
     pub(crate) secret: AuthSecret,
 }
 
-impl AuthMethod for CleartextPassword {
-    #[inline(always)]
-    fn first_message(&self, _channel_binding: bool) -> BeMessage<'_> {
-        Be::AuthenticationCleartextPassword
-    }
-}
-
 /// This wrapper for [`PqStream`] performs client authentication.
 #[must_use]
 pub(crate) struct AuthFlow<'a, S, State> {
     /// The underlying stream which implements libpq's protocol.
     stream: &'a mut PqStream<Stream<S>>,
-    /// State might contain ancillary data (see [`Self::begin`]).
+    /// State might contain ancillary data.
     state: State,
     tls_server_end_point: TlsServerEndPoint,
 }
 
 /// Initial state of the stream wrapper.
-impl<'a, S: AsyncRead + AsyncWrite + Unpin> AuthFlow<'a, S, Begin> {
+impl<'a, S: AsyncRead + AsyncWrite + Unpin, M> AuthFlow<'a, S, M> {
     /// Create a new wrapper for client authentication.
-    pub(crate) fn new(stream: &'a mut PqStream<Stream<S>>) -> Self {
+    pub(crate) fn new(stream: &'a mut PqStream<Stream<S>>, method: M) -> Self {
         let tls_server_end_point = stream.get_ref().tls_server_end_point();
 
         Self {
             stream,
-            state: Begin,
+            state: method,
             tls_server_end_point,
         }
     }
-
-    /// Move to the next step by sending auth method's name & params to client.
-    pub(crate) async fn begin<M: AuthMethod>(self, method: M) -> io::Result<AuthFlow<'a, S, M>> {
-        self.stream
-            .write_message(&method.first_message(self.tls_server_end_point.supported()))
-            .await?;
-
-        Ok(AuthFlow {
-            stream: self.stream,
-            state: method,
-            tls_server_end_point: self.tls_server_end_point,
-        })
-    }
 }
 
 impl<S: AsyncRead + AsyncWrite + Unpin> AuthFlow<'_, S, PasswordHack> {
     /// Perform user authentication. Raise an error in case authentication failed.
     pub(crate) async fn get_password(self) -> super::Result<PasswordHackPayload> {
+        self.stream
+            .write_message(BeMessage::AuthenticationCleartextPassword);
+        self.stream.flush().await?;
+
         let msg = self.stream.read_password_message().await?;
         let password = msg
             .strip_suffix(&[0])
@@ -133,6 +99,10 @@ impl<S: AsyncRead + AsyncWrite + Unpin> AuthFlow<'_, S, PasswordHack> {
 impl<S: AsyncRead + AsyncWrite + Unpin> AuthFlow<'_, S, CleartextPassword> {
     /// Perform user authentication. Raise an error in case authentication failed.
     pub(crate) async fn authenticate(self) -> super::Result<sasl::Outcome<ComputeCredentialKeys>> {
+        self.stream
+            .write_message(BeMessage::AuthenticationCleartextPassword);
+        self.stream.flush().await?;
+
         let msg = self.stream.read_password_message().await?;
         let password = msg
             .strip_suffix(&[0])
@@ -147,7 +117,7 @@ impl<S: AsyncRead + AsyncWrite + Unpin> AuthFlow<'_, S, CleartextPassword> {
         .await?;
 
         if let sasl::Outcome::Success(_) = &outcome {
-            self.stream.write_message_noflush(&Be::AuthenticationOk)?;
+            self.stream.write_message(BeMessage::AuthenticationOk);
         }
 
         Ok(outcome)
@@ -159,42 +129,36 @@ impl<S: AsyncRead + AsyncWrite + Unpin> AuthFlow<'_, S, Scram<'_>> {
     /// Perform user authentication. Raise an error in case authentication failed.
     pub(crate) async fn authenticate(self) -> super::Result<sasl::Outcome<scram::ScramKey>> {
         let Scram(secret, ctx) = self.state;
+        let channel_binding = self.tls_server_end_point;
 
-        // pause the timer while we communicate with the client
-        let _paused = ctx.latency_timer_pause(crate::metrics::Waiting::Client);
+        // send sasl message.
+        {
+            // pause the timer while we communicate with the client
+            let _paused = ctx.latency_timer_pause(crate::metrics::Waiting::Client);
 
-        // Initial client message contains the chosen auth method's name.
-        let msg = self.stream.read_password_message().await?;
-        let sasl = sasl::FirstMessage::parse(&msg)
-            .ok_or(AuthError::MalformedPassword("bad sasl message"))?;
-
-        // Currently, the only supported SASL method is SCRAM.
-        if !scram::METHODS.contains(&sasl.method) {
-            return Err(super::AuthError::bad_auth_method(sasl.method));
+            let sasl = self.state.first_message(channel_binding.supported());
+            self.stream.write_message(sasl);
+            self.stream.flush().await?;
         }
 
-        match sasl.method {
-            SCRAM_SHA_256 => ctx.set_auth_method(crate::context::AuthMethod::ScramSha256),
-            SCRAM_SHA_256_PLUS => ctx.set_auth_method(crate::context::AuthMethod::ScramSha256Plus),
-            _ => {}
-        }
+        // complete sasl handshake.
+        sasl::authenticate(ctx, self.stream, |method| {
+            // Currently, the only supported SASL method is SCRAM.
+            match method {
+                SCRAM_SHA_256 => ctx.set_auth_method(crate::context::AuthMethod::ScramSha256),
+                SCRAM_SHA_256_PLUS => {
+                    ctx.set_auth_method(crate::context::AuthMethod::ScramSha256Plus);
+                }
+                method => return Err(sasl::Error::BadAuthMethod(method.into())),
+            }
 
-        // TODO: make this a metric instead
-        info!("client chooses {}", sasl.method);
+            // TODO: make this a metric instead
+            info!("client chooses {}", method);
 
-        let outcome = sasl::SaslStream::new(self.stream, sasl.message)
-            .authenticate(scram::Exchange::new(
-                secret,
-                rand::random,
-                self.tls_server_end_point,
-            ))
-            .await?;
-
-        if let sasl::Outcome::Success(_) = &outcome {
-            self.stream.write_message_noflush(&Be::AuthenticationOk)?;
-        }
-
-        Ok(outcome)
+            Ok(scram::Exchange::new(secret, rand::random, channel_binding))
+        })
+        .await
+        .map_err(AuthError::Sasl)
     }
 }
 
@@ -205,13 +169,6 @@ pub(crate) async fn validate_password_and_exchange(
     secret: AuthSecret,
 ) -> super::Result<sasl::Outcome<ComputeCredentialKeys>> {
     match secret {
-        #[cfg(any(test, feature = "testing"))]
-        AuthSecret::Md5(_) => {
-            // test only
-            Ok(sasl::Outcome::Success(ComputeCredentialKeys::Password(
-                password.to_owned(),
-            )))
-        }
         // perform scram authentication as both client and server to validate the keys
         AuthSecret::Scram(scram_secret) => {
             let outcome = crate::scram::exchange(pool, endpoint, &scram_secret, password).await?;

proxy/src/binary/local_proxy.rs

@@ -32,9 +32,7 @@ use crate::ext::TaskExt;
 use crate::http::health_server::AppMetrics;
 use crate::intern::RoleNameInt;
 use crate::metrics::{Metrics, ThreadPoolMetrics};
-use crate::rate_limiter::{
-    BucketRateLimiter, EndpointRateLimiter, LeakyBucketConfig, RateBucketInfo,
-};
+use crate::rate_limiter::{EndpointRateLimiter, LeakyBucketConfig, RateBucketInfo};
 use crate::scram::threadpool::ThreadPool;
 use crate::serverless::cancel_set::CancelSet;
 use crate::serverless::{self, GlobalConnPoolOptions};
@@ -69,15 +67,6 @@ struct LocalProxyCliArgs {
     /// Can be given multiple times for different bucket sizes.
     #[clap(long, default_values_t = RateBucketInfo::DEFAULT_ENDPOINT_SET)]
     user_rps_limit: Vec<RateBucketInfo>,
-    /// Whether the auth rate limiter actually takes effect (for testing)
-    #[clap(long, default_value_t = false, value_parser = clap::builder::BoolishValueParser::new(), action = clap::ArgAction::Set)]
-    auth_rate_limit_enabled: bool,
-    /// Authentication rate limiter max number of hashes per second.
-    #[clap(long, default_values_t = RateBucketInfo::DEFAULT_AUTH_SET)]
-    auth_rate_limit: Vec<RateBucketInfo>,
-    /// The IP subnet to use when considering whether two IP addresses are considered the same.
-    #[clap(long, default_value_t = 64)]
-    auth_rate_limit_ip_subnet: u8,
     /// Whether to retry the connection to the compute node
     #[clap(long, default_value = config::RetryConfig::CONNECT_TO_COMPUTE_DEFAULT_VALUES)]
     connect_to_compute_retry: String,
@@ -282,9 +271,6 @@ fn build_config(args: &LocalProxyCliArgs) -> anyhow::Result<&'static ProxyConfig
             jwks_cache: JwkCache::default(),
             thread_pool: ThreadPool::new(0),
             scram_protocol_timeout: Duration::from_secs(10),
-            rate_limiter_enabled: false,
-            rate_limiter: BucketRateLimiter::new(vec![]),
-            rate_limit_ip_subnet: 64,
             ip_allowlist_check_enabled: true,
             is_vpc_acccess_proxy: false,
             is_auth_broker: false,

proxy/src/binary/pg_sni_router.rs

@@ -4,8 +4,9 @@
 //! This allows connecting to pods/services running in the same Kubernetes cluster from
 //! the outside. Similar to an ingress controller for HTTPS.
 
+use std::net::SocketAddr;
 use std::path::Path;
-use std::{net::SocketAddr, sync::Arc};
+use std::sync::Arc;
 
 use anyhow::{Context, anyhow, bail, ensure};
 use clap::Arg;
@@ -17,6 +18,7 @@ use rustls::pki_types::{DnsName, PrivateKeyDer};
 use tokio::io::{AsyncRead, AsyncReadExt, AsyncWrite, AsyncWriteExt};
 use tokio::net::TcpListener;
 use tokio_rustls::TlsConnector;
+use tokio_rustls::server::TlsStream;
 use tokio_util::sync::CancellationToken;
 use tracing::{Instrument, error, info};
 use utils::project_git_version;
@@ -24,10 +26,12 @@ use utils::sentry_init::init_sentry;
 
 use crate::context::RequestContext;
 use crate::metrics::{Metrics, ThreadPoolMetrics};
+use crate::pqproto::FeStartupPacket;
 use crate::protocol2::ConnectionInfo;
-use crate::proxy::{ErrorSource, copy_bidirectional_client_compute, run_until_cancelled};
+use crate::proxy::{
+    ErrorSource, TlsRequired, copy_bidirectional_client_compute, run_until_cancelled,
+};
 use crate::stream::{PqStream, Stream};
-use crate::tls::TlsServerEndPoint;
 
 project_git_version!(GIT_VERSION);
 
@@ -84,7 +88,7 @@ pub async fn run() -> anyhow::Result<()> {
         .parse()?;
 
     // Configure TLS
-    let (tls_config, tls_server_end_point): (Arc<rustls::ServerConfig>, TlsServerEndPoint) = match (
+    let tls_config = match (
         args.get_one::<String>("tls-key"),
         args.get_one::<String>("tls-cert"),
     ) {
@@ -117,7 +121,6 @@ pub async fn run() -> anyhow::Result<()> {
         dest.clone(),
         tls_config.clone(),
         None,
-        tls_server_end_point,
         proxy_listener,
         cancellation_token.clone(),
     ))
@@ -127,7 +130,6 @@ pub async fn run() -> anyhow::Result<()> {
         dest,
         tls_config,
         Some(compute_tls_config),
-        tls_server_end_point,
         proxy_listener_compute_tls,
         cancellation_token.clone(),
     ))
@@ -154,7 +156,7 @@ pub async fn run() -> anyhow::Result<()> {
 pub(super) fn parse_tls(
     key_path: &Path,
     cert_path: &Path,
-) -> anyhow::Result<(Arc<rustls::ServerConfig>, TlsServerEndPoint)> {
+) -> anyhow::Result<Arc<rustls::ServerConfig>> {
     let key = {
         let key_bytes = std::fs::read(key_path).context("TLS key file")?;
 
@@ -187,10 +189,6 @@ pub(super) fn parse_tls(
             })?
     };
 
-    // needed for channel bindings
-    let first_cert = cert_chain.first().context("missing certificate")?;
-    let tls_server_end_point = TlsServerEndPoint::new(first_cert)?;
-
     let tls_config =
         rustls::ServerConfig::builder_with_provider(Arc::new(ring::default_provider()))
             .with_protocol_versions(&[&rustls::version::TLS13, &rustls::version::TLS12])
@@ -199,14 +197,13 @@ pub(super) fn parse_tls(
             .with_single_cert(cert_chain, key)?
             .into();
 
-    Ok((tls_config, tls_server_end_point))
+    Ok(tls_config)
 }
 
 pub(super) async fn task_main(
     dest_suffix: Arc<String>,
     tls_config: Arc<rustls::ServerConfig>,
     compute_tls_config: Option<Arc<rustls::ClientConfig>>,
-    tls_server_end_point: TlsServerEndPoint,
     listener: tokio::net::TcpListener,
     cancellation_token: CancellationToken,
 ) -> anyhow::Result<()> {
@@ -242,15 +239,7 @@ pub(super) async fn task_main(
                     crate::metrics::Protocol::SniRouter,
                     "sni",
                 );
-                handle_client(
-                    ctx,
-                    dest_suffix,
-                    tls_config,
-                    compute_tls_config,
-                    tls_server_end_point,
-                    socket,
-                )
-                .await
+                handle_client(ctx, dest_suffix, tls_config, compute_tls_config, socket).await
             }
             .unwrap_or_else(|e| {
                 // Acknowledge that the task has finished with an error.
@@ -269,55 +258,26 @@ pub(super) async fn task_main(
     Ok(())
 }
 
-const ERR_INSECURE_CONNECTION: &str = "connection is insecure (try using `sslmode=require`)";
-
 async fn ssl_handshake<S: AsyncRead + AsyncWrite + Unpin>(
     ctx: &RequestContext,
     raw_stream: S,
     tls_config: Arc<rustls::ServerConfig>,
-    tls_server_end_point: TlsServerEndPoint,
-) -> anyhow::Result<Stream<S>> {
-    let mut stream = PqStream::new(Stream::from_raw(raw_stream));
-
-    let msg = stream.read_startup_packet().await?;
-    use pq_proto::FeStartupPacket::SslRequest;
-
+) -> anyhow::Result<TlsStream<S>> {
+    let (mut stream, msg) = PqStream::parse_startup(Stream::from_raw(raw_stream)).await?;
     match msg {
-        SslRequest { direct: false } => {
-            stream
-                .write_message(&pq_proto::BeMessage::EncryptionResponse(true))
-                .await?;
+        FeStartupPacket::SslRequest { direct: None } => {
+            let raw = stream.accept_tls().await?;
 
-            // Upgrade raw stream into a secure TLS-backed stream.
-            // NOTE: We've consumed `tls`; this fact will be used later.
-
-            let (raw, read_buf) = stream.into_inner();
-            // TODO: Normally, client doesn't send any data before
-            // server says TLS handshake is ok and read_buf is empty.
-            // However, you could imagine pipelining of postgres
-            // SSLRequest + TLS ClientHello in one hunk similar to
-            // pipelining in our node js driver. We should probably
-            // support that by chaining read_buf with the stream.
-            if !read_buf.is_empty() {
-                bail!("data is sent before server replied with EncryptionResponse");
-            }
-
-            Ok(Stream::Tls {
-                tls: Box::new(
-                    raw.upgrade(tls_config, !ctx.has_private_peer_addr())
-                        .await?,
-                ),
-                tls_server_end_point,
-            })
+            Ok(raw
+                .upgrade(tls_config, !ctx.has_private_peer_addr())
+                .await?)
         }
         unexpected => {
             info!(
                 ?unexpected,
                 "unexpected startup packet, rejecting connection"
             );
-            stream
-                .throw_error_str(ERR_INSECURE_CONNECTION, crate::error::ErrorKind::User, None)
-                .await?
+            Err(stream.throw_error(TlsRequired, None).await)?
         }
     }
 }
@@ -327,15 +287,18 @@ async fn handle_client(
     dest_suffix: Arc<String>,
     tls_config: Arc<rustls::ServerConfig>,
     compute_tls_config: Option<Arc<rustls::ClientConfig>>,
-    tls_server_end_point: TlsServerEndPoint,
     stream: impl AsyncRead + AsyncWrite + Unpin,
 ) -> anyhow::Result<()> {
-    let mut tls_stream = ssl_handshake(&ctx, stream, tls_config, tls_server_end_point).await?;
+    let mut tls_stream = ssl_handshake(&ctx, stream, tls_config).await?;
 
     // Cut off first part of the SNI domain
     // We receive required destination details in the format of
     //   `{k8s_service_name}--{k8s_namespace}--{port}.non-sni-domain`
-    let sni = tls_stream.sni_hostname().ok_or(anyhow!("SNI missing"))?;
+    let sni = tls_stream
+        .get_ref()
+        .1
+        .server_name()
+        .ok_or(anyhow!("SNI missing"))?;
     let dest: Vec<&str> = sni
         .split_once('.')
         .context("invalid SNI")?

proxy/src/binary/proxy.rs

@@ -20,7 +20,7 @@ use utils::sentry_init::init_sentry;
 use utils::{project_build_tag, project_git_version};
 
 use crate::auth::backend::jwt::JwkCache;
-use crate::auth::backend::{AuthRateLimiter, ConsoleRedirectBackend, MaybeOwned};
+use crate::auth::backend::{ConsoleRedirectBackend, MaybeOwned};
 use crate::cancellation::{CancellationHandler, handle_cancel_messages};
 use crate::config::{
     self, AuthenticationConfig, CacheOptions, ComputeConfig, HttpConfig, ProjectInfoCacheOptions,
@@ -29,9 +29,7 @@ use crate::config::{
 use crate::context::parquet::ParquetUploadArgs;
 use crate::http::health_server::AppMetrics;
 use crate::metrics::Metrics;
-use crate::rate_limiter::{
-    EndpointRateLimiter, LeakyBucketConfig, RateBucketInfo, WakeComputeRateLimiter,
-};
+use crate::rate_limiter::{EndpointRateLimiter, RateBucketInfo, WakeComputeRateLimiter};
 use crate::redis::connection_with_credentials_provider::ConnectionWithCredentialsProvider;
 use crate::redis::kv_ops::RedisKVClient;
 use crate::redis::{elasticache, notifications};
@@ -154,15 +152,6 @@ struct ProxyCliArgs {
     /// Wake compute rate limiter max number of requests per second.
     #[clap(long, default_values_t = RateBucketInfo::DEFAULT_SET)]
     wake_compute_limit: Vec<RateBucketInfo>,
-    /// Whether the auth rate limiter actually takes effect (for testing)
-    #[clap(long, default_value_t = false, value_parser = clap::builder::BoolishValueParser::new(), action = clap::ArgAction::Set)]
-    auth_rate_limit_enabled: bool,
-    /// Authentication rate limiter max number of hashes per second.
-    #[clap(long, default_values_t = RateBucketInfo::DEFAULT_AUTH_SET)]
-    auth_rate_limit: Vec<RateBucketInfo>,
-    /// The IP subnet to use when considering whether two IP addresses are considered the same.
-    #[clap(long, default_value_t = 64)]
-    auth_rate_limit_ip_subnet: u8,
     /// Redis rate limiter max number of requests per second.
     #[clap(long, default_values_t = RateBucketInfo::DEFAULT_REDIS_SET)]
     redis_rps_limit: Vec<RateBucketInfo>,
@@ -232,8 +221,7 @@ struct ProxyCliArgs {
     is_private_access_proxy: bool,
 
     /// Configure whether all incoming requests have a Proxy Protocol V2 packet.
-    // TODO(conradludgate): switch default to rejected or required once we've updated all deployments
-    #[clap(value_enum, long, default_value_t = ProxyProtocolV2::Supported)]
+    #[clap(value_enum, long, default_value_t = ProxyProtocolV2::Rejected)]
     proxy_protocol_v2: ProxyProtocolV2,
 
     /// Time the proxy waits for the webauth session to be confirmed by the control plane.
@@ -410,22 +398,9 @@ pub async fn run() -> anyhow::Result<()> {
         Some(tx_cancel),
     ));
 
-    // bit of a hack - find the min rps and max rps supported and turn it into
-    // leaky bucket config instead
-    let max = args
-        .endpoint_rps_limit
-        .iter()
-        .map(|x| x.rps())
-        .max_by(f64::total_cmp)
-        .unwrap_or(EndpointRateLimiter::DEFAULT.max);
-    let rps = args
-        .endpoint_rps_limit
-        .iter()
-        .map(|x| x.rps())
-        .min_by(f64::total_cmp)
-        .unwrap_or(EndpointRateLimiter::DEFAULT.rps);
     let endpoint_rate_limiter = Arc::new(EndpointRateLimiter::new_with_shards(
-        LeakyBucketConfig { rps, max },
+        RateBucketInfo::to_leaky_bucket(&args.endpoint_rps_limit)
+            .unwrap_or(EndpointRateLimiter::DEFAULT),
         64,
     ));
 
@@ -476,8 +451,7 @@ pub async fn run() -> anyhow::Result<()> {
         let key_path = args.tls_key.expect("already asserted it is set");
         let cert_path = args.tls_cert.expect("already asserted it is set");
 
-        let (tls_config, tls_server_end_point) =
-            super::pg_sni_router::parse_tls(&key_path, &cert_path)?;
+        let tls_config = super::pg_sni_router::parse_tls(&key_path, &cert_path)?;
 
         let dest = Arc::new(dest);
 
@@ -485,7 +459,6 @@ pub async fn run() -> anyhow::Result<()> {
             dest.clone(),
             tls_config.clone(),
             None,
-            tls_server_end_point,
             listen,
             cancellation_token.clone(),
         ));
@@ -494,7 +467,6 @@ pub async fn run() -> anyhow::Result<()> {
             dest,
             tls_config,
             Some(config.connect_to_compute.tls.clone()),
-            tls_server_end_point,
             listen_tls,
             cancellation_token.clone(),
         ));
@@ -681,9 +653,6 @@ fn build_config(args: &ProxyCliArgs) -> anyhow::Result<&'static ProxyConfig> {
         jwks_cache: JwkCache::default(),
         thread_pool,
         scram_protocol_timeout: args.scram_protocol_timeout,
-        rate_limiter_enabled: args.auth_rate_limit_enabled,
-        rate_limiter: AuthRateLimiter::new(args.auth_rate_limit.clone()),
-        rate_limit_ip_subnet: args.auth_rate_limit_ip_subnet,
         ip_allowlist_check_enabled: !args.is_private_access_proxy,
         is_vpc_acccess_proxy: args.is_private_access_proxy,
         is_auth_broker: args.is_auth_broker,

proxy/src/cache/project_info.rs

@@ -1,30 +1,26 @@
-use std::collections::HashSet;
+use std::collections::{HashMap, HashSet, hash_map};
 use std::convert::Infallible;
-use std::sync::Arc;
 use std::sync::atomic::AtomicU64;
 use std::time::Duration;
 
 use async_trait::async_trait;
 use clashmap::ClashMap;
+use clashmap::mapref::one::Ref;
 use rand::{Rng, thread_rng};
-use smol_str::SmolStr;
 use tokio::sync::Mutex;
 use tokio::time::Instant;
 use tracing::{debug, info};
 
-use super::{Cache, Cached};
-use crate::auth::IpPattern;
 use crate::config::ProjectInfoCacheOptions;
-use crate::control_plane::{AccessBlockerFlags, AuthSecret};
+use crate::control_plane::{EndpointAccessControl, RoleAccessControl};
 use crate::intern::{AccountIdInt, EndpointIdInt, ProjectIdInt, RoleNameInt};
 use crate::types::{EndpointId, RoleName};
 
 #[async_trait]
 pub(crate) trait ProjectInfoCache {
-    fn invalidate_allowed_ips_for_project(&self, project_id: ProjectIdInt);
-    fn invalidate_allowed_vpc_endpoint_ids_for_projects(&self, project_ids: Vec<ProjectIdInt>);
-    fn invalidate_allowed_vpc_endpoint_ids_for_org(&self, account_id: AccountIdInt);
-    fn invalidate_block_public_or_vpc_access_for_project(&self, project_id: ProjectIdInt);
+    fn invalidate_endpoint_access(&self, endpoint_id: EndpointIdInt);
+    fn invalidate_endpoint_access_for_project(&self, project_id: ProjectIdInt);
+    fn invalidate_endpoint_access_for_org(&self, account_id: AccountIdInt);
     fn invalidate_role_secret_for_project(&self, project_id: ProjectIdInt, role_name: RoleNameInt);
     async fn decrement_active_listeners(&self);
     async fn increment_active_listeners(&self);
@@ -42,6 +38,10 @@ impl<T> Entry<T> {
             value,
         }
     }
+
+    pub(crate) fn get(&self, valid_since: Instant) -> Option<&T> {
+        (valid_since < self.created_at).then_some(&self.value)
+    }
 }
 
 impl<T> From<T> for Entry<T> {
@@ -50,101 +50,32 @@ impl<T> From<T> for Entry<T> {
     }
 }
 
-#[derive(Default)]
 struct EndpointInfo {
-    secret: std::collections::HashMap<RoleNameInt, Entry<Option<AuthSecret>>>,
-    allowed_ips: Option<Entry<Arc<Vec<IpPattern>>>>,
-    block_public_or_vpc_access: Option<Entry<AccessBlockerFlags>>,
-    allowed_vpc_endpoint_ids: Option<Entry<Arc<Vec<String>>>>,
+    role_controls: HashMap<RoleNameInt, Entry<RoleAccessControl>>,
+    controls: Option<Entry<EndpointAccessControl>>,
 }
 
 impl EndpointInfo {
-    fn check_ignore_cache(ignore_cache_since: Option<Instant>, created_at: Instant) -> bool {
-        match ignore_cache_since {
-            None => false,
-            Some(t) => t < created_at,
-        }
-    }
     pub(crate) fn get_role_secret(
         &self,
         role_name: RoleNameInt,
         valid_since: Instant,
-        ignore_cache_since: Option<Instant>,
-    ) -> Option<(Option<AuthSecret>, bool)> {
-        if let Some(secret) = self.secret.get(&role_name) {
-            if valid_since < secret.created_at {
-                return Some((
-                    secret.value.clone(),
-                    Self::check_ignore_cache(ignore_cache_since, secret.created_at),
-                ));
-            }
-        }
-        None
+    ) -> Option<RoleAccessControl> {
+        let controls = self.role_controls.get(&role_name)?;
+        controls.get(valid_since).cloned()
     }
 
-    pub(crate) fn get_allowed_ips(
-        &self,
-        valid_since: Instant,
-        ignore_cache_since: Option<Instant>,
-    ) -> Option<(Arc<Vec<IpPattern>>, bool)> {
-        if let Some(allowed_ips) = &self.allowed_ips {
-            if valid_since < allowed_ips.created_at {
-                return Some((
-                    allowed_ips.value.clone(),
-                    Self::check_ignore_cache(ignore_cache_since, allowed_ips.created_at),
-                ));
-            }
-        }
-        None
-    }
-    pub(crate) fn get_allowed_vpc_endpoint_ids(
-        &self,
-        valid_since: Instant,
-        ignore_cache_since: Option<Instant>,
-    ) -> Option<(Arc<Vec<String>>, bool)> {
-        if let Some(allowed_vpc_endpoint_ids) = &self.allowed_vpc_endpoint_ids {
-            if valid_since < allowed_vpc_endpoint_ids.created_at {
-                return Some((
-                    allowed_vpc_endpoint_ids.value.clone(),
-                    Self::check_ignore_cache(
-                        ignore_cache_since,
-                        allowed_vpc_endpoint_ids.created_at,
-                    ),
-                ));
-            }
-        }
-        None
-    }
-    pub(crate) fn get_block_public_or_vpc_access(
-        &self,
-        valid_since: Instant,
-        ignore_cache_since: Option<Instant>,
-    ) -> Option<(AccessBlockerFlags, bool)> {
-        if let Some(block_public_or_vpc_access) = &self.block_public_or_vpc_access {
-            if valid_since < block_public_or_vpc_access.created_at {
-                return Some((
-                    block_public_or_vpc_access.value.clone(),
-                    Self::check_ignore_cache(
-                        ignore_cache_since,
-                        block_public_or_vpc_access.created_at,
-                    ),
-                ));
-            }
-        }
-        None
+    pub(crate) fn get_controls(&self, valid_since: Instant) -> Option<EndpointAccessControl> {
+        let controls = self.controls.as_ref()?;
+        controls.get(valid_since).cloned()
     }
 
-    pub(crate) fn invalidate_allowed_ips(&mut self) {
-        self.allowed_ips = None;
-    }
-    pub(crate) fn invalidate_allowed_vpc_endpoint_ids(&mut self) {
-        self.allowed_vpc_endpoint_ids = None;
-    }
-    pub(crate) fn invalidate_block_public_or_vpc_access(&mut self) {
-        self.block_public_or_vpc_access = None;
+    pub(crate) fn invalidate_endpoint(&mut self) {
+        self.controls = None;
     }
+
     pub(crate) fn invalidate_role_secret(&mut self, role_name: RoleNameInt) {
-        self.secret.remove(&role_name);
+        self.role_controls.remove(&role_name);
     }
 }
 
@@ -170,34 +101,29 @@ pub struct ProjectInfoCacheImpl {
 
 #[async_trait]
 impl ProjectInfoCache for ProjectInfoCacheImpl {
-    fn invalidate_allowed_vpc_endpoint_ids_for_projects(&self, project_ids: Vec<ProjectIdInt>) {
-        info!(
-            "invalidating allowed vpc endpoint ids for projects `{}`",
-            project_ids
-                .iter()
-                .map(|id| id.to_string())
-                .collect::<Vec<_>>()
-                .join(", ")
-        );
-        for project_id in project_ids {
-            let endpoints = self
-                .project2ep
-                .get(&project_id)
-                .map(|kv| kv.value().clone())
-                .unwrap_or_default();
-            for endpoint_id in endpoints {
-                if let Some(mut endpoint_info) = self.cache.get_mut(&endpoint_id) {
-                    endpoint_info.invalidate_allowed_vpc_endpoint_ids();
-                }
+    fn invalidate_endpoint_access(&self, endpoint_id: EndpointIdInt) {
+        info!("invalidating endpoint access for `{endpoint_id}`");
+        if let Some(mut endpoint_info) = self.cache.get_mut(&endpoint_id) {
+            endpoint_info.invalidate_endpoint();
+        }
+    }
+
+    fn invalidate_endpoint_access_for_project(&self, project_id: ProjectIdInt) {
+        info!("invalidating endpoint access for project `{project_id}`");
+        let endpoints = self
+            .project2ep
+            .get(&project_id)
+            .map(|kv| kv.value().clone())
+            .unwrap_or_default();
+        for endpoint_id in endpoints {
+            if let Some(mut endpoint_info) = self.cache.get_mut(&endpoint_id) {
+                endpoint_info.invalidate_endpoint();
             }
         }
     }
 
-    fn invalidate_allowed_vpc_endpoint_ids_for_org(&self, account_id: AccountIdInt) {
-        info!(
-            "invalidating allowed vpc endpoint ids for org `{}`",
-            account_id
-        );
+    fn invalidate_endpoint_access_for_org(&self, account_id: AccountIdInt) {
+        info!("invalidating endpoint access for org `{account_id}`");
         let endpoints = self
             .account2ep
             .get(&account_id)
@@ -205,41 +131,11 @@ impl ProjectInfoCache for ProjectInfoCacheImpl {
             .unwrap_or_default();
         for endpoint_id in endpoints {
             if let Some(mut endpoint_info) = self.cache.get_mut(&endpoint_id) {
-                endpoint_info.invalidate_allowed_vpc_endpoint_ids();
+                endpoint_info.invalidate_endpoint();
             }
         }
     }
 
-    fn invalidate_block_public_or_vpc_access_for_project(&self, project_id: ProjectIdInt) {
-        info!(
-            "invalidating block public or vpc access for project `{}`",
-            project_id
-        );
-        let endpoints = self
-            .project2ep
-            .get(&project_id)
-            .map(|kv| kv.value().clone())
-            .unwrap_or_default();
-        for endpoint_id in endpoints {
-            if let Some(mut endpoint_info) = self.cache.get_mut(&endpoint_id) {
-                endpoint_info.invalidate_block_public_or_vpc_access();
-            }
-        }
-    }
-
-    fn invalidate_allowed_ips_for_project(&self, project_id: ProjectIdInt) {
-        info!("invalidating allowed ips for project `{}`", project_id);
-        let endpoints = self
-            .project2ep
-            .get(&project_id)
-            .map(|kv| kv.value().clone())
-            .unwrap_or_default();
-        for endpoint_id in endpoints {
-            if let Some(mut endpoint_info) = self.cache.get_mut(&endpoint_id) {
-                endpoint_info.invalidate_allowed_ips();
-            }
-        }
-    }
     fn invalidate_role_secret_for_project(&self, project_id: ProjectIdInt, role_name: RoleNameInt) {
         info!(
             "invalidating role secret for project_id `{}` and role_name `{}`",
@@ -256,6 +152,7 @@ impl ProjectInfoCache for ProjectInfoCacheImpl {
             }
         }
     }
+
     async fn decrement_active_listeners(&self) {
         let mut listeners_guard = self.active_listeners_lock.lock().await;
         if *listeners_guard == 0 {
@@ -293,155 +190,71 @@ impl ProjectInfoCacheImpl {
         }
     }
 
+    fn get_endpoint_cache(
+        &self,
+        endpoint_id: &EndpointId,
+    ) -> Option<Ref<'_, EndpointIdInt, EndpointInfo>> {
+        let endpoint_id = EndpointIdInt::get(endpoint_id)?;
+        self.cache.get(&endpoint_id)
+    }
+
     pub(crate) fn get_role_secret(
         &self,
         endpoint_id: &EndpointId,
         role_name: &RoleName,
-    ) -> Option<Cached<&Self, Option<AuthSecret>>> {
-        let endpoint_id = EndpointIdInt::get(endpoint_id)?;
+    ) -> Option<RoleAccessControl> {
+        let valid_since = self.get_cache_times();
         let role_name = RoleNameInt::get(role_name)?;
-        let (valid_since, ignore_cache_since) = self.get_cache_times();
-        let endpoint_info = self.cache.get(&endpoint_id)?;
-        let (value, ignore_cache) =
-            endpoint_info.get_role_secret(role_name, valid_since, ignore_cache_since)?;
-        if !ignore_cache {
-            let cached = Cached {
-                token: Some((
-                    self,
-                    CachedLookupInfo::new_role_secret(endpoint_id, role_name),
-                )),
-                value,
-            };
-            return Some(cached);
-        }
-        Some(Cached::new_uncached(value))
-    }
-    pub(crate) fn get_allowed_ips(
-        &self,
-        endpoint_id: &EndpointId,
-    ) -> Option<Cached<&Self, Arc<Vec<IpPattern>>>> {
-        let endpoint_id = EndpointIdInt::get(endpoint_id)?;
-        let (valid_since, ignore_cache_since) = self.get_cache_times();
-        let endpoint_info = self.cache.get(&endpoint_id)?;
-        let value = endpoint_info.get_allowed_ips(valid_since, ignore_cache_since);
-        let (value, ignore_cache) = value?;
-        if !ignore_cache {
-            let cached = Cached {
-                token: Some((self, CachedLookupInfo::new_allowed_ips(endpoint_id))),
-                value,
-            };
-            return Some(cached);
-        }
-        Some(Cached::new_uncached(value))
-    }
-    pub(crate) fn get_allowed_vpc_endpoint_ids(
-        &self,
-        endpoint_id: &EndpointId,
-    ) -> Option<Cached<&Self, Arc<Vec<String>>>> {
-        let endpoint_id = EndpointIdInt::get(endpoint_id)?;
-        let (valid_since, ignore_cache_since) = self.get_cache_times();
-        let endpoint_info = self.cache.get(&endpoint_id)?;
-        let value = endpoint_info.get_allowed_vpc_endpoint_ids(valid_since, ignore_cache_since);
-        let (value, ignore_cache) = value?;
-        if !ignore_cache {
-            let cached = Cached {
-                token: Some((
-                    self,
-                    CachedLookupInfo::new_allowed_vpc_endpoint_ids(endpoint_id),
-                )),
-                value,
-            };
-            return Some(cached);
-        }
-        Some(Cached::new_uncached(value))
-    }
-    pub(crate) fn get_block_public_or_vpc_access(
-        &self,
-        endpoint_id: &EndpointId,
-    ) -> Option<Cached<&Self, AccessBlockerFlags>> {
-        let endpoint_id = EndpointIdInt::get(endpoint_id)?;
-        let (valid_since, ignore_cache_since) = self.get_cache_times();
-        let endpoint_info = self.cache.get(&endpoint_id)?;
-        let value = endpoint_info.get_block_public_or_vpc_access(valid_since, ignore_cache_since);
-        let (value, ignore_cache) = value?;
-        if !ignore_cache {
-            let cached = Cached {
-                token: Some((
-                    self,
-                    CachedLookupInfo::new_block_public_or_vpc_access(endpoint_id),
-                )),
-                value,
-            };
-            return Some(cached);
-        }
-        Some(Cached::new_uncached(value))
+        let endpoint_info = self.get_endpoint_cache(endpoint_id)?;
+        endpoint_info.get_role_secret(role_name, valid_since)
     }
 
-    pub(crate) fn insert_role_secret(
+    pub(crate) fn get_endpoint_access(
         &self,
-        project_id: ProjectIdInt,
-        endpoint_id: EndpointIdInt,
-        role_name: RoleNameInt,
-        secret: Option<AuthSecret>,
-    ) {
-        if self.cache.len() >= self.config.size {
-            // If there are too many entries, wait until the next gc cycle.
-            return;
-        }
-        self.insert_project2endpoint(project_id, endpoint_id);
-        let mut entry = self.cache.entry(endpoint_id).or_default();
-        if entry.secret.len() < self.config.max_roles {
-            entry.secret.insert(role_name, secret.into());
-        }
+        endpoint_id: &EndpointId,
+    ) -> Option<EndpointAccessControl> {
+        let valid_since = self.get_cache_times();
+        let endpoint_info = self.get_endpoint_cache(endpoint_id)?;
+        endpoint_info.get_controls(valid_since)
     }
-    pub(crate) fn insert_allowed_ips(
-        &self,
-        project_id: ProjectIdInt,
-        endpoint_id: EndpointIdInt,
-        allowed_ips: Arc<Vec<IpPattern>>,
-    ) {
-        if self.cache.len() >= self.config.size {
-            // If there are too many entries, wait until the next gc cycle.
-            return;
-        }
-        self.insert_project2endpoint(project_id, endpoint_id);
-        self.cache.entry(endpoint_id).or_default().allowed_ips = Some(allowed_ips.into());
-    }
-    pub(crate) fn insert_allowed_vpc_endpoint_ids(
+
+    pub(crate) fn insert_endpoint_access(
         &self,
         account_id: Option<AccountIdInt>,
         project_id: ProjectIdInt,
         endpoint_id: EndpointIdInt,
-        allowed_vpc_endpoint_ids: Arc<Vec<String>>,
+        role_name: RoleNameInt,
+        controls: EndpointAccessControl,
+        role_controls: RoleAccessControl,
     ) {
-        if self.cache.len() >= self.config.size {
-            // If there are too many entries, wait until the next gc cycle.
-            return;
-        }
         if let Some(account_id) = account_id {
             self.insert_account2endpoint(account_id, endpoint_id);
         }
         self.insert_project2endpoint(project_id, endpoint_id);
-        self.cache
-            .entry(endpoint_id)
-            .or_default()
-            .allowed_vpc_endpoint_ids = Some(allowed_vpc_endpoint_ids.into());
-    }
-    pub(crate) fn insert_block_public_or_vpc_access(
-        &self,
-        project_id: ProjectIdInt,
-        endpoint_id: EndpointIdInt,
-        access_blockers: AccessBlockerFlags,
-    ) {
+
         if self.cache.len() >= self.config.size {
             // If there are too many entries, wait until the next gc cycle.
             return;
         }
-        self.insert_project2endpoint(project_id, endpoint_id);
-        self.cache
-            .entry(endpoint_id)
-            .or_default()
-            .block_public_or_vpc_access = Some(access_blockers.into());
+
+        let controls = Entry::from(controls);
+        let role_controls = Entry::from(role_controls);
+
+        match self.cache.entry(endpoint_id) {
+            clashmap::Entry::Vacant(e) => {
+                e.insert(EndpointInfo {
+                    role_controls: HashMap::from_iter([(role_name, role_controls)]),
+                    controls: Some(controls),
+                });
+            }
+            clashmap::Entry::Occupied(mut e) => {
+                let ep = e.get_mut();
+                ep.controls = Some(controls);
+                if ep.role_controls.len() < self.config.max_roles {
+                    ep.role_controls.insert(role_name, role_controls);
+                }
+            }
+        }
     }
 
     fn insert_project2endpoint(&self, project_id: ProjectIdInt, endpoint_id: EndpointIdInt) {
@@ -452,6 +265,7 @@ impl ProjectInfoCacheImpl {
                 .insert(project_id, HashSet::from([endpoint_id]));
         }
     }
+
     fn insert_account2endpoint(&self, account_id: AccountIdInt, endpoint_id: EndpointIdInt) {
         if let Some(mut endpoints) = self.account2ep.get_mut(&account_id) {
             endpoints.insert(endpoint_id);
@@ -460,21 +274,57 @@ impl ProjectInfoCacheImpl {
                 .insert(account_id, HashSet::from([endpoint_id]));
         }
     }
-    fn get_cache_times(&self) -> (Instant, Option<Instant>) {
-        let mut valid_since = Instant::now() - self.config.ttl;
-        // Only ignore cache if ttl is disabled.
+
+    fn ignore_ttl_since(&self) -> Option<Instant> {
         let ttl_disabled_since_us = self
             .ttl_disabled_since_us
             .load(std::sync::atomic::Ordering::Relaxed);
-        let ignore_cache_since = if ttl_disabled_since_us == u64::MAX {
-            None
-        } else {
-            let ignore_cache_since = self.start_time + Duration::from_micros(ttl_disabled_since_us);
+
+        if ttl_disabled_since_us == u64::MAX {
+            return None;
+        }
+
+        Some(self.start_time + Duration::from_micros(ttl_disabled_since_us))
+    }
+
+    fn get_cache_times(&self) -> Instant {
+        let mut valid_since = Instant::now() - self.config.ttl;
+        if let Some(ignore_ttl_since) = self.ignore_ttl_since() {
             // We are fine if entry is not older than ttl or was added before we are getting notifications.
-            valid_since = valid_since.min(ignore_cache_since);
-            Some(ignore_cache_since)
+            valid_since = valid_since.min(ignore_ttl_since);
+        }
+        valid_since
+    }
+
+    pub fn maybe_invalidate_role_secret(&self, endpoint_id: &EndpointId, role_name: &RoleName) {
+        let Some(endpoint_id) = EndpointIdInt::get(endpoint_id) else {
+            return;
         };
-        (valid_since, ignore_cache_since)
+        let Some(role_name) = RoleNameInt::get(role_name) else {
+            return;
+        };
+
+        let Some(mut endpoint_info) = self.cache.get_mut(&endpoint_id) else {
+            return;
+        };
+
+        let entry = endpoint_info.role_controls.entry(role_name);
+        let hash_map::Entry::Occupied(role_controls) = entry else {
+            return;
+        };
+
+        let created_at = role_controls.get().created_at;
+        let expire = match self.ignore_ttl_since() {
+            // if ignoring TTL, we should still try and roll the password if it's old
+            // and we the client gave an incorrect password. There could be some lag on the redis channel.
+            Some(_) => created_at + self.config.ttl < Instant::now(),
+            // edge case: redis is down, let's be generous and invalidate the cache immediately.
+            None => true,
+        };
+
+        if expire {
+            role_controls.remove();
+        }
     }
 
     pub async fn gc_worker(&self) -> anyhow::Result<Infallible> {
@@ -509,84 +359,12 @@ impl ProjectInfoCacheImpl {
     }
 }
 
-/// Lookup info for project info cache.
-/// This is used to invalidate cache entries.
-pub(crate) struct CachedLookupInfo {
-    /// Search by this key.
-    endpoint_id: EndpointIdInt,
-    lookup_type: LookupType,
-}
-
-impl CachedLookupInfo {
-    pub(self) fn new_role_secret(endpoint_id: EndpointIdInt, role_name: RoleNameInt) -> Self {
-        Self {
-            endpoint_id,
-            lookup_type: LookupType::RoleSecret(role_name),
-        }
-    }
-    pub(self) fn new_allowed_ips(endpoint_id: EndpointIdInt) -> Self {
-        Self {
-            endpoint_id,
-            lookup_type: LookupType::AllowedIps,
-        }
-    }
-    pub(self) fn new_allowed_vpc_endpoint_ids(endpoint_id: EndpointIdInt) -> Self {
-        Self {
-            endpoint_id,
-            lookup_type: LookupType::AllowedVpcEndpointIds,
-        }
-    }
-    pub(self) fn new_block_public_or_vpc_access(endpoint_id: EndpointIdInt) -> Self {
-        Self {
-            endpoint_id,
-            lookup_type: LookupType::BlockPublicOrVpcAccess,
-        }
-    }
-}
-
-enum LookupType {
-    RoleSecret(RoleNameInt),
-    AllowedIps,
-    AllowedVpcEndpointIds,
-    BlockPublicOrVpcAccess,
-}
-
-impl Cache for ProjectInfoCacheImpl {
-    type Key = SmolStr;
-    // Value is not really used here, but we need to specify it.
-    type Value = SmolStr;
-
-    type LookupInfo<Key> = CachedLookupInfo;
-
-    fn invalidate(&self, key: &Self::LookupInfo<SmolStr>) {
-        match &key.lookup_type {
-            LookupType::RoleSecret(role_name) => {
-                if let Some(mut endpoint_info) = self.cache.get_mut(&key.endpoint_id) {
-                    endpoint_info.invalidate_role_secret(*role_name);
-                }
-            }
-            LookupType::AllowedIps => {
-                if let Some(mut endpoint_info) = self.cache.get_mut(&key.endpoint_id) {
-                    endpoint_info.invalidate_allowed_ips();
-                }
-            }
-            LookupType::AllowedVpcEndpointIds => {
-                if let Some(mut endpoint_info) = self.cache.get_mut(&key.endpoint_id) {
-                    endpoint_info.invalidate_allowed_vpc_endpoint_ids();
-                }
-            }
-            LookupType::BlockPublicOrVpcAccess => {
-                if let Some(mut endpoint_info) = self.cache.get_mut(&key.endpoint_id) {
-                    endpoint_info.invalidate_block_public_or_vpc_access();
-                }
-            }
-        }
-    }
-}
-
 #[cfg(test)]
 mod tests {
+    use std::sync::Arc;
+
     use super::*;
+    use crate::control_plane::{AccessBlockerFlags, AuthSecret};
     use crate::scram::ServerSecret;
     use crate::types::ProjectId;
 
@@ -601,6 +379,8 @@ mod tests {
         });
         let project_id: ProjectId = "project".into();
         let endpoint_id: EndpointId = "endpoint".into();
+        let account_id: Option<AccountIdInt> = None;
+
         let user1: RoleName = "user1".into();
         let user2: RoleName = "user2".into();
         let secret1 = Some(AuthSecret::Scram(ServerSecret::mock([1; 32])));
@@ -609,183 +389,73 @@ mod tests {
             "127.0.0.1".parse().unwrap(),
             "127.0.0.2".parse().unwrap(),
         ]);
-        cache.insert_role_secret(
+
+        cache.insert_endpoint_access(
+            account_id,
             (&project_id).into(),
             (&endpoint_id).into(),
             (&user1).into(),
-            secret1.clone(),
+            EndpointAccessControl {
+                allowed_ips: allowed_ips.clone(),
+                allowed_vpce: Arc::new(vec![]),
+                flags: AccessBlockerFlags::default(),
+            },
+            RoleAccessControl {
+                secret: secret1.clone(),
+            },
         );
-        cache.insert_role_secret(
+
+        cache.insert_endpoint_access(
+            account_id,
             (&project_id).into(),
             (&endpoint_id).into(),
             (&user2).into(),
-            secret2.clone(),
-        );
-        cache.insert_allowed_ips(
-            (&project_id).into(),
-            (&endpoint_id).into(),
-            allowed_ips.clone(),
+            EndpointAccessControl {
+                allowed_ips: allowed_ips.clone(),
+                allowed_vpce: Arc::new(vec![]),
+                flags: AccessBlockerFlags::default(),
+            },
+            RoleAccessControl {
+                secret: secret2.clone(),
+            },
         );
 
         let cached = cache.get_role_secret(&endpoint_id, &user1).unwrap();
-        assert!(cached.cached());
-        assert_eq!(cached.value, secret1);
+        assert_eq!(cached.secret, secret1);
+
         let cached = cache.get_role_secret(&endpoint_id, &user2).unwrap();
-        assert!(cached.cached());
-        assert_eq!(cached.value, secret2);
+        assert_eq!(cached.secret, secret2);
 
         // Shouldn't add more than 2 roles.
         let user3: RoleName = "user3".into();
         let secret3 = Some(AuthSecret::Scram(ServerSecret::mock([3; 32])));
-        cache.insert_role_secret(
+
+        cache.insert_endpoint_access(
+            account_id,
             (&project_id).into(),
             (&endpoint_id).into(),
             (&user3).into(),
-            secret3.clone(),
+            EndpointAccessControl {
+                allowed_ips: allowed_ips.clone(),
+                allowed_vpce: Arc::new(vec![]),
+                flags: AccessBlockerFlags::default(),
+            },
+            RoleAccessControl {
+                secret: secret3.clone(),
+            },
         );
+
         assert!(cache.get_role_secret(&endpoint_id, &user3).is_none());
 
-        let cached = cache.get_allowed_ips(&endpoint_id).unwrap();
-        assert!(cached.cached());
-        assert_eq!(cached.value, allowed_ips);
+        let cached = cache.get_endpoint_access(&endpoint_id).unwrap();
+        assert_eq!(cached.allowed_ips, allowed_ips);
 
         tokio::time::advance(Duration::from_secs(2)).await;
         let cached = cache.get_role_secret(&endpoint_id, &user1);
         assert!(cached.is_none());
         let cached = cache.get_role_secret(&endpoint_id, &user2);
         assert!(cached.is_none());
-        let cached = cache.get_allowed_ips(&endpoint_id);
+        let cached = cache.get_endpoint_access(&endpoint_id);
         assert!(cached.is_none());
     }
-
-    #[tokio::test]
-    async fn test_project_info_cache_invalidations() {
-        tokio::time::pause();
-        let cache = Arc::new(ProjectInfoCacheImpl::new(ProjectInfoCacheOptions {
-            size: 2,
-            max_roles: 2,
-            ttl: Duration::from_secs(1),
-            gc_interval: Duration::from_secs(600),
-        }));
-        cache.clone().increment_active_listeners().await;
-        tokio::time::advance(Duration::from_secs(2)).await;
-
-        let project_id: ProjectId = "project".into();
-        let endpoint_id: EndpointId = "endpoint".into();
-        let user1: RoleName = "user1".into();
-        let user2: RoleName = "user2".into();
-        let secret1 = Some(AuthSecret::Scram(ServerSecret::mock([1; 32])));
-        let secret2 = Some(AuthSecret::Scram(ServerSecret::mock([2; 32])));
-        let allowed_ips = Arc::new(vec![
-            "127.0.0.1".parse().unwrap(),
-            "127.0.0.2".parse().unwrap(),
-        ]);
-        cache.insert_role_secret(
-            (&project_id).into(),
-            (&endpoint_id).into(),
-            (&user1).into(),
-            secret1.clone(),
-        );
-        cache.insert_role_secret(
-            (&project_id).into(),
-            (&endpoint_id).into(),
-            (&user2).into(),
-            secret2.clone(),
-        );
-        cache.insert_allowed_ips(
-            (&project_id).into(),
-            (&endpoint_id).into(),
-            allowed_ips.clone(),
-        );
-
-        tokio::time::advance(Duration::from_secs(2)).await;
-        // Nothing should be invalidated.
-
-        let cached = cache.get_role_secret(&endpoint_id, &user1).unwrap();
-        // TTL is disabled, so it should be impossible to invalidate this value.
-        assert!(!cached.cached());
-        assert_eq!(cached.value, secret1);
-
-        cached.invalidate(); // Shouldn't do anything.
-        let cached = cache.get_role_secret(&endpoint_id, &user1).unwrap();
-        assert_eq!(cached.value, secret1);
-
-        let cached = cache.get_role_secret(&endpoint_id, &user2).unwrap();
-        assert!(!cached.cached());
-        assert_eq!(cached.value, secret2);
-
-        // The only way to invalidate this value is to invalidate via the api.
-        cache.invalidate_role_secret_for_project((&project_id).into(), (&user2).into());
-        assert!(cache.get_role_secret(&endpoint_id, &user2).is_none());
-
-        let cached = cache.get_allowed_ips(&endpoint_id).unwrap();
-        assert!(!cached.cached());
-        assert_eq!(cached.value, allowed_ips);
-    }
-
-    #[tokio::test]
-    async fn test_increment_active_listeners_invalidate_added_before() {
-        tokio::time::pause();
-        let cache = Arc::new(ProjectInfoCacheImpl::new(ProjectInfoCacheOptions {
-            size: 2,
-            max_roles: 2,
-            ttl: Duration::from_secs(1),
-            gc_interval: Duration::from_secs(600),
-        }));
-
-        let project_id: ProjectId = "project".into();
-        let endpoint_id: EndpointId = "endpoint".into();
-        let user1: RoleName = "user1".into();
-        let user2: RoleName = "user2".into();
-        let secret1 = Some(AuthSecret::Scram(ServerSecret::mock([1; 32])));
-        let secret2 = Some(AuthSecret::Scram(ServerSecret::mock([2; 32])));
-        let allowed_ips = Arc::new(vec![
-            "127.0.0.1".parse().unwrap(),
-            "127.0.0.2".parse().unwrap(),
-        ]);
-        cache.insert_role_secret(
-            (&project_id).into(),
-            (&endpoint_id).into(),
-            (&user1).into(),
-            secret1.clone(),
-        );
-        cache.clone().increment_active_listeners().await;
-        tokio::time::advance(Duration::from_millis(100)).await;
-        cache.insert_role_secret(
-            (&project_id).into(),
-            (&endpoint_id).into(),
-            (&user2).into(),
-            secret2.clone(),
-        );
-
-        // Added before ttl was disabled + ttl should be still cached.
-        let cached = cache.get_role_secret(&endpoint_id, &user1).unwrap();
-        assert!(cached.cached());
-        let cached = cache.get_role_secret(&endpoint_id, &user2).unwrap();
-        assert!(cached.cached());
-
-        tokio::time::advance(Duration::from_secs(1)).await;
-        // Added before ttl was disabled + ttl should expire.
-        assert!(cache.get_role_secret(&endpoint_id, &user1).is_none());
-        assert!(cache.get_role_secret(&endpoint_id, &user2).is_none());
-
-        // Added after ttl was disabled + ttl should not be cached.
-        cache.insert_allowed_ips(
-            (&project_id).into(),
-            (&endpoint_id).into(),
-            allowed_ips.clone(),
-        );
-        let cached = cache.get_allowed_ips(&endpoint_id).unwrap();
-        assert!(!cached.cached());
-
-        tokio::time::advance(Duration::from_secs(1)).await;
-        // Added before ttl was disabled + ttl still should expire.
-        assert!(cache.get_role_secret(&endpoint_id, &user1).is_none());
-        assert!(cache.get_role_secret(&endpoint_id, &user2).is_none());
-        // Shouldn't be invalidated.
-
-        let cached = cache.get_allowed_ips(&endpoint_id).unwrap();
-        assert!(!cached.cached());
-        assert_eq!(cached.value, allowed_ips);
-    }
 }

proxy/src/cancellation.rs

@@ -5,7 +5,6 @@ use anyhow::{Context, anyhow};
 use ipnet::{IpNet, Ipv4Net, Ipv6Net};
 use postgres_client::CancelToken;
 use postgres_client::tls::MakeTlsConnect;
-use pq_proto::CancelKeyData;
 use redis::{Cmd, FromRedisValue, Value};
 use serde::{Deserialize, Serialize};
 use thiserror::Error;
@@ -13,19 +12,18 @@ use tokio::net::TcpStream;
 use tokio::sync::{mpsc, oneshot};
 use tracing::{debug, error, info, warn};
 
+use crate::auth::AuthError;
 use crate::auth::backend::ComputeUserInfo;
-use crate::auth::{AuthError, check_peer_addr_is_in_list};
 use crate::config::ComputeConfig;
 use crate::context::RequestContext;
 use crate::control_plane::ControlPlaneApi;
 use crate::error::ReportableError;
 use crate::ext::LockExt;
 use crate::metrics::{CancelChannelSizeGuard, CancellationRequest, Metrics, RedisMsgKind};
-use crate::protocol2::ConnectionInfoExtra;
+use crate::pqproto::CancelKeyData;
 use crate::rate_limiter::LeakyBucketRateLimiter;
 use crate::redis::keys::KeyPrefix;
 use crate::redis::kv_ops::RedisKVClient;
-use crate::tls::postgres_rustls::MakeRustlsConnect;
 
 type IpSubnetKey = IpNet;
 
@@ -272,13 +270,7 @@ pub(crate) enum CancelError {
     #[error("rate limit exceeded")]
     RateLimit,
 
-    #[error("IP is not allowed")]
-    IpNotAllowed,
-
-    #[error("VPC endpoint id is not allowed to connect")]
-    VpcEndpointIdNotAllowed,
-
-    #[error("Authentication backend error")]
+    #[error("Authentication error")]
     AuthError(#[from] AuthError),
 
     #[error("key not found")]
@@ -297,10 +289,7 @@ impl ReportableError for CancelError {
             }
             CancelError::Postgres(_) => crate::error::ErrorKind::Compute,
             CancelError::RateLimit => crate::error::ErrorKind::RateLimit,
-            CancelError::IpNotAllowed
-            | CancelError::VpcEndpointIdNotAllowed
-            | CancelError::NotFound => crate::error::ErrorKind::User,
-            CancelError::AuthError(_) => crate::error::ErrorKind::ControlPlane,
+            CancelError::NotFound | CancelError::AuthError(_) => crate::error::ErrorKind::User,
             CancelError::InternalError => crate::error::ErrorKind::Service,
         }
     }
@@ -422,7 +411,13 @@ impl CancellationHandler {
             IpAddr::V4(ip) => IpNet::V4(Ipv4Net::new_assert(ip, 24).trunc()), // use defaut mask here
             IpAddr::V6(ip) => IpNet::V6(Ipv6Net::new_assert(ip, 64).trunc()),
         };
-        if !self.limiter.lock_propagate_poison().check(subnet_key, 1) {
+
+        let allowed = {
+            let rate_limit_config = None;
+            let limiter = self.limiter.lock_propagate_poison();
+            limiter.check(subnet_key, rate_limit_config, 1)
+        };
+        if !allowed {
             // log only the subnet part of the IP address to know which subnet is rate limited
             tracing::warn!("Rate limit exceeded. Skipping cancellation message, {subnet_key}");
             Metrics::get()
@@ -450,52 +445,13 @@ impl CancellationHandler {
             return Err(CancelError::NotFound);
         };
 
-        if check_ip_allowed {
-            let ip_allowlist = auth_backend
-                .get_allowed_ips(&ctx, &cancel_closure.user_info)
-                .await
-                .map_err(|e| CancelError::AuthError(e.into()))?;
-
-            if !check_peer_addr_is_in_list(&ctx.peer_addr(), &ip_allowlist) {
-                // log it here since cancel_session could be spawned in a task
-                tracing::warn!(
-                    "IP is not allowed to cancel the query: {key}, address: {}",
-                    ctx.peer_addr()
-                );
-                return Err(CancelError::IpNotAllowed);
-            }
-        }
-
-        // check if a VPC endpoint ID is coming in and if yes, if it's allowed
-        let access_blocks = auth_backend
-            .get_block_public_or_vpc_access(&ctx, &cancel_closure.user_info)
+        let info = &cancel_closure.user_info;
+        let access_controls = auth_backend
+            .get_endpoint_access_control(&ctx, &info.endpoint, &info.user)
             .await
             .map_err(|e| CancelError::AuthError(e.into()))?;
 
-        if check_vpc_allowed {
-            if access_blocks.vpc_access_blocked {
-                return Err(CancelError::AuthError(AuthError::NetworkNotAllowed));
-            }
-
-            let incoming_vpc_endpoint_id = match ctx.extra() {
-                None => return Err(CancelError::AuthError(AuthError::MissingVPCEndpointId)),
-                Some(ConnectionInfoExtra::Aws { vpce_id }) => vpce_id.to_string(),
-                Some(ConnectionInfoExtra::Azure { link_id }) => link_id.to_string(),
-            };
-
-            let allowed_vpc_endpoint_ids = auth_backend
-                .get_allowed_vpc_endpoint_ids(&ctx, &cancel_closure.user_info)
-                .await
-                .map_err(|e| CancelError::AuthError(e.into()))?;
-            // TODO: For now an empty VPC endpoint ID list means all are allowed. We should replace that.
-            if !allowed_vpc_endpoint_ids.is_empty()
-                && !allowed_vpc_endpoint_ids.contains(&incoming_vpc_endpoint_id)
-            {
-                return Err(CancelError::VpcEndpointIdNotAllowed);
-            }
-        } else if access_blocks.public_access_blocked {
-            return Err(CancelError::VpcEndpointIdNotAllowed);
-        }
+        access_controls.check(&ctx, check_ip_allowed, check_vpc_allowed)?;
 
         Metrics::get()
             .proxy
@@ -540,10 +496,8 @@ impl CancelClosure {
     ) -> Result<(), CancelError> {
         let socket = TcpStream::connect(self.socket_addr).await?;
 
-        let mut mk_tls =
-            crate::tls::postgres_rustls::MakeRustlsConnect::new(compute_config.tls.clone());
-        let tls = <MakeRustlsConnect as MakeTlsConnect<tokio::net::TcpStream>>::make_tls_connect(
-            &mut mk_tls,
+        let tls = <_ as MakeTlsConnect<tokio::net::TcpStream>>::make_tls_connect(
+            compute_config,
             &self.hostname,
         )
         .map_err(|e| CancelError::IO(std::io::Error::other(e.to_string())))?;

proxy/src/compute/mod.rs

@@ -1,22 +1,24 @@
+mod tls;
+
 use std::fmt::Debug;
 use std::io;
-use std::net::SocketAddr;
-use std::time::Duration;
+use std::net::{IpAddr, SocketAddr};
 
 use futures::{FutureExt, TryFutureExt};
 use itertools::Itertools;
+use postgres_client::config::{AuthKeys, SslMode};
+use postgres_client::maybe_tls_stream::MaybeTlsStream;
 use postgres_client::tls::MakeTlsConnect;
-use postgres_client::{CancelToken, RawConnection};
+use postgres_client::{CancelToken, NoTls, RawConnection};
 use postgres_protocol::message::backend::NoticeResponseBody;
-use pq_proto::StartupMessageParams;
-use rustls::pki_types::InvalidDnsNameError;
 use thiserror::Error;
 use tokio::net::{TcpStream, lookup_host};
 use tracing::{debug, error, info, warn};
 
-use crate::auth::backend::ComputeUserInfo;
+use crate::auth::backend::{ComputeCredentialKeys, ComputeUserInfo};
 use crate::auth::parse_endpoint_param;
 use crate::cancellation::CancelClosure;
+use crate::compute::tls::TlsError;
 use crate::config::ComputeConfig;
 use crate::context::RequestContext;
 use crate::control_plane::client::ApiLockError;
@@ -24,8 +26,8 @@ use crate::control_plane::errors::WakeComputeError;
 use crate::control_plane::messages::MetricsAuxInfo;
 use crate::error::{ReportableError, UserFacingError};
 use crate::metrics::{Metrics, NumDbConnectionsGuard};
+use crate::pqproto::StartupMessageParams;
 use crate::proxy::neon_option;
-use crate::tls::postgres_rustls::MakeRustlsConnect;
 use crate::types::Host;
 
 pub const COULD_NOT_CONNECT: &str = "Couldn't connect to compute node";
@@ -38,10 +40,7 @@ pub(crate) enum ConnectionError {
     Postgres(#[from] postgres_client::Error),
 
     #[error("{COULD_NOT_CONNECT}: {0}")]
-    CouldNotConnect(#[from] io::Error),
-
-    #[error("{COULD_NOT_CONNECT}: {0}")]
-    TlsError(#[from] InvalidDnsNameError),
+    TlsError(#[from] TlsError),
 
     #[error("{COULD_NOT_CONNECT}: {0}")]
     WakeComputeError(#[from] WakeComputeError),
@@ -73,7 +72,7 @@ impl UserFacingError for ConnectionError {
             ConnectionError::TooManyConnectionAttempts(_) => {
                 "Failed to acquire permit to connect to the database. Too many database connection attempts are currently ongoing.".to_owned()
             }
-            _ => COULD_NOT_CONNECT.to_owned(),
+            ConnectionError::TlsError(_) => COULD_NOT_CONNECT.to_owned(),
         }
     }
 }
@@ -85,7 +84,6 @@ impl ReportableError for ConnectionError {
                 crate::error::ErrorKind::Postgres
             }
             ConnectionError::Postgres(_) => crate::error::ErrorKind::Compute,
-            ConnectionError::CouldNotConnect(_) => crate::error::ErrorKind::Compute,
             ConnectionError::TlsError(_) => crate::error::ErrorKind::Compute,
             ConnectionError::WakeComputeError(e) => e.get_error_kind(),
             ConnectionError::TooManyConnectionAttempts(e) => e.get_error_kind(),
@@ -96,34 +94,85 @@ impl ReportableError for ConnectionError {
 /// A pair of `ClientKey` & `ServerKey` for `SCRAM-SHA-256`.
 pub(crate) type ScramKeys = postgres_client::config::ScramKeys<32>;
 
-/// A config for establishing a connection to compute node.
-/// Eventually, `postgres_client` will be replaced with something better.
-/// Newtype allows us to implement methods on top of it.
 #[derive(Clone)]
-pub(crate) struct ConnCfg(Box<postgres_client::Config>);
+pub enum Auth {
+    /// Only used during console-redirect.
+    Password(Vec<u8>),
+    /// Used by sql-over-http, ws, tcp.
+    Scram(Box<ScramKeys>),
+}
+
+/// A config for authenticating to the compute node.
+pub(crate) struct AuthInfo {
+    /// None for local-proxy, as we use trust-based localhost auth.
+    /// Some for sql-over-http, ws, tcp, and in most cases for console-redirect.
+    /// Might be None for console-redirect, but that's only a consequence of testing environments ATM.
+    auth: Option<Auth>,
+    server_params: StartupMessageParams,
+
+    /// Console redirect sets user and database, we shouldn't re-use those from the params.
+    skip_db_user: bool,
+}
+
+/// Contains only the data needed to establish a secure connection to compute.
+#[derive(Clone)]
+pub struct ConnectInfo {
+    pub host_addr: Option<IpAddr>,
+    pub host: Host,
+    pub port: u16,
+    pub ssl_mode: SslMode,
+}
 
 /// Creation and initialization routines.
-impl ConnCfg {
-    pub(crate) fn new(host: String, port: u16) -> Self {
-        Self(Box::new(postgres_client::Config::new(host, port)))
-    }
-
-    /// Reuse password or auth keys from the other config.
-    pub(crate) fn reuse_password(&mut self, other: Self) {
-        if let Some(password) = other.get_password() {
-            self.password(password);
-        }
-
-        if let Some(keys) = other.get_auth_keys() {
-            self.auth_keys(keys);
+impl AuthInfo {
+    pub(crate) fn for_console_redirect(db: &str, user: &str, pw: Option<&str>) -> Self {
+        let mut server_params = StartupMessageParams::default();
+        server_params.insert("database", db);
+        server_params.insert("user", user);
+        Self {
+            auth: pw.map(|pw| Auth::Password(pw.as_bytes().to_owned())),
+            server_params,
+            skip_db_user: true,
         }
     }
 
-    pub(crate) fn get_host(&self) -> Host {
-        match self.0.get_host() {
-            postgres_client::config::Host::Tcp(s) => s.into(),
+    pub(crate) fn with_auth_keys(keys: ComputeCredentialKeys) -> Self {
+        Self {
+            auth: match keys {
+                ComputeCredentialKeys::AuthKeys(AuthKeys::ScramSha256(auth_keys)) => {
+                    Some(Auth::Scram(Box::new(auth_keys)))
+                }
+                ComputeCredentialKeys::JwtPayload(_) | ComputeCredentialKeys::None => None,
+            },
+            server_params: StartupMessageParams::default(),
+            skip_db_user: false,
         }
     }
+}
+
+impl ConnectInfo {
+    pub fn to_postgres_client_config(&self) -> postgres_client::Config {
+        let mut config = postgres_client::Config::new(self.host.to_string(), self.port);
+        config.ssl_mode(self.ssl_mode);
+        if let Some(host_addr) = self.host_addr {
+            config.set_host_addr(host_addr);
+        }
+        config
+    }
+}
+
+impl AuthInfo {
+    fn enrich(&self, mut config: postgres_client::Config) -> postgres_client::Config {
+        match &self.auth {
+            Some(Auth::Scram(keys)) => config.auth_keys(AuthKeys::ScramSha256(**keys)),
+            Some(Auth::Password(pw)) => config.password(pw),
+            None => &mut config,
+        };
+        for (k, v) in self.server_params.iter() {
+            config.set_param(k, v);
+        }
+        config
+    }
 
     /// Apply startup message params to the connection config.
     pub(crate) fn set_startup_params(
@@ -132,27 +181,26 @@ impl ConnCfg {
         arbitrary_params: bool,
     ) {
         if !arbitrary_params {
-            self.set_param("client_encoding", "UTF8");
+            self.server_params.insert("client_encoding", "UTF8");
         }
         for (k, v) in params.iter() {
             match k {
                 // Only set `user` if it's not present in the config.
                 // Console redirect auth flow takes username from the console's response.
-                "user" if self.user_is_set() => {}
-                "database" if self.db_is_set() => {}
+                "user" | "database" if self.skip_db_user => {}
                 "options" => {
                     if let Some(options) = filtered_options(v) {
-                        self.set_param(k, &options);
+                        self.server_params.insert(k, &options);
                     }
                 }
                 "user" | "database" | "application_name" | "replication" => {
-                    self.set_param(k, v);
+                    self.server_params.insert(k, v);
                 }
 
                 // if we allow arbitrary params, then we forward them through.
                 // this is a flag for a period of backwards compatibility
                 k if arbitrary_params => {
-                    self.set_param(k, v);
+                    self.server_params.insert(k, v);
                 }
                 _ => {}
             }
@@ -160,25 +208,13 @@ impl ConnCfg {
     }
 }
 
-impl std::ops::Deref for ConnCfg {
-    type Target = postgres_client::Config;
-
-    fn deref(&self) -> &Self::Target {
-        &self.0
-    }
-}
-
-/// For now, let's make it easier to setup the config.
-impl std::ops::DerefMut for ConnCfg {
-    fn deref_mut(&mut self) -> &mut Self::Target {
-        &mut self.0
-    }
-}
-
-impl ConnCfg {
-    /// Establish a raw TCP connection to the compute node.
-    async fn connect_raw(&self, timeout: Duration) -> io::Result<(SocketAddr, TcpStream, &str)> {
-        use postgres_client::config::Host;
+impl ConnectInfo {
+    /// Establish a raw TCP+TLS connection to the compute node.
+    async fn connect_raw(
+        &self,
+        config: &ComputeConfig,
+    ) -> Result<(SocketAddr, MaybeTlsStream<TcpStream, RustlsStream>), TlsError> {
+        let timeout = config.timeout;
 
         // wrap TcpStream::connect with timeout
         let connect_with_timeout = |addrs| {
@@ -208,34 +244,32 @@ impl ConnCfg {
         // We can't reuse connection establishing logic from `postgres_client` here,
         // because it has no means for extracting the underlying socket which we
         // require for our business.
-        let port = self.0.get_port();
-        let host = self.0.get_host();
+        let port = self.port;
+        let host = &*self.host;
 
-        let host = match host {
-            Host::Tcp(host) => host.as_str(),
-        };
-
-        let addrs = match self.0.get_host_addr() {
+        let addrs = match self.host_addr {
             Some(addr) => vec![SocketAddr::new(addr, port)],
             None => lookup_host((host, port)).await?.collect(),
         };
 
         match connect_once(&*addrs).await {
-            Ok((sockaddr, stream)) => Ok((sockaddr, stream, host)),
+            Ok((sockaddr, stream)) => Ok((
+                sockaddr,
+                tls::connect_tls(stream, self.ssl_mode, config, host).await?,
+            )),
             Err(err) => {
                 warn!("couldn't connect to compute node at {host}:{port}: {err}");
-                Err(err)
+                Err(TlsError::Connection(err))
             }
         }
     }
 }
 
-type RustlsStream = <MakeRustlsConnect as MakeTlsConnect<tokio::net::TcpStream>>::Stream;
+type RustlsStream = <ComputeConfig as MakeTlsConnect<tokio::net::TcpStream>>::Stream;
 
 pub(crate) struct PostgresConnection {
     /// Socket connected to a compute node.
-    pub(crate) stream:
-        postgres_client::maybe_tls_stream::MaybeTlsStream<tokio::net::TcpStream, RustlsStream>,
+    pub(crate) stream: MaybeTlsStream<tokio::net::TcpStream, RustlsStream>,
     /// PostgreSQL connection parameters.
     pub(crate) params: std::collections::HashMap<String, String>,
     /// Query cancellation token.
@@ -248,28 +282,23 @@ pub(crate) struct PostgresConnection {
     _guage: NumDbConnectionsGuard<'static>,
 }
 
-impl ConnCfg {
+impl ConnectInfo {
     /// Connect to a corresponding compute node.
     pub(crate) async fn connect(
         &self,
         ctx: &RequestContext,
         aux: MetricsAuxInfo,
+        auth: &AuthInfo,
         config: &ComputeConfig,
         user_info: ComputeUserInfo,
     ) -> Result<PostgresConnection, ConnectionError> {
-        let pause = ctx.latency_timer_pause(crate::metrics::Waiting::Compute);
-        let (socket_addr, stream, host) = self.connect_raw(config.timeout).await?;
-        drop(pause);
+        let mut tmp_config = auth.enrich(self.to_postgres_client_config());
+        // we setup SSL early in `ConnectInfo::connect_raw`.
+        tmp_config.ssl_mode(SslMode::Disable);
 
-        let mut mk_tls = crate::tls::postgres_rustls::MakeRustlsConnect::new(config.tls.clone());
-        let tls = <MakeRustlsConnect as MakeTlsConnect<tokio::net::TcpStream>>::make_tls_connect(
-            &mut mk_tls,
-            host,
-        )?;
-
-        // connect_raw() will not use TLS if sslmode is "disable"
         let pause = ctx.latency_timer_pause(crate::metrics::Waiting::Compute);
-        let connection = self.0.connect_raw(stream, tls).await?;
+        let (socket_addr, stream) = self.connect_raw(config).await?;
+        let connection = tmp_config.connect_raw(stream, NoTls).await?;
         drop(pause);
 
         let RawConnection {
@@ -282,13 +311,14 @@ impl ConnCfg {
 
         tracing::Span::current().record("pid", tracing::field::display(process_id));
         tracing::Span::current().record("compute_id", tracing::field::display(&aux.compute_id));
-        let stream = stream.into_inner();
+        let MaybeTlsStream::Raw(stream) = stream.into_inner();
 
         // TODO: lots of useful info but maybe we can move it elsewhere (eg traces?)
         info!(
             cold_start_info = ctx.cold_start_info().as_str(),
-            "connected to compute node at {host} ({socket_addr}) sslmode={:?}, latency={}, query_id={}",
-            self.0.get_ssl_mode(),
+            "connected to compute node at {} ({socket_addr}) sslmode={:?}, latency={}, query_id={}",
+            self.host,
+            self.ssl_mode,
             ctx.get_proxy_latency(),
             ctx.get_testodrome_id().unwrap_or_default(),
         );
@@ -299,11 +329,11 @@ impl ConnCfg {
             socket_addr,
             CancelToken {
                 socket_config: None,
-                ssl_mode: self.0.get_ssl_mode(),
+                ssl_mode: self.ssl_mode,
                 process_id,
                 secret_key,
             },
-            host.to_string(),
+            self.host.to_string(),
             user_info,
         );
 

proxy/src/compute/tls.rs

@@ -0,0 +1,63 @@
+use futures::FutureExt;
+use postgres_client::config::SslMode;
+use postgres_client::maybe_tls_stream::MaybeTlsStream;
+use postgres_client::tls::{MakeTlsConnect, TlsConnect};
+use rustls::pki_types::InvalidDnsNameError;
+use thiserror::Error;
+use tokio::io::{AsyncRead, AsyncWrite};
+
+use crate::pqproto::request_tls;
+use crate::proxy::retry::CouldRetry;
+
+#[derive(Debug, Error)]
+pub enum TlsError {
+    #[error(transparent)]
+    Dns(#[from] InvalidDnsNameError),
+    #[error(transparent)]
+    Connection(#[from] std::io::Error),
+    #[error("TLS required but not provided")]
+    Required,
+}
+
+impl CouldRetry for TlsError {
+    fn could_retry(&self) -> bool {
+        match self {
+            TlsError::Dns(_) => false,
+            TlsError::Connection(err) => err.could_retry(),
+            // perhaps compute didn't realise it supports TLS?
+            TlsError::Required => true,
+        }
+    }
+}
+
+pub async fn connect_tls<S, T>(
+    mut stream: S,
+    mode: SslMode,
+    tls: &T,
+    host: &str,
+) -> Result<MaybeTlsStream<S, T::Stream>, TlsError>
+where
+    S: AsyncRead + AsyncWrite + Unpin + Send,
+    T: MakeTlsConnect<
+            S,
+            Error = InvalidDnsNameError,
+            TlsConnect: TlsConnect<S, Error = std::io::Error, Future: Send>,
+        >,
+{
+    match mode {
+        SslMode::Disable => return Ok(MaybeTlsStream::Raw(stream)),
+        SslMode::Prefer | SslMode::Require => {}
+    }
+
+    if !request_tls(&mut stream).await? {
+        if SslMode::Require == mode {
+            return Err(TlsError::Required);
+        }
+
+        return Ok(MaybeTlsStream::Raw(stream));
+    }
+
+    Ok(MaybeTlsStream::Tls(
+        tls.make_tls_connect(host)?.connect(stream).boxed().await?,
+    ))
+}