Run proxy-bench directly on self-hosted runner

## Problem
```
postgres=> select neon.prewarm_local_cache('\xfcfcfcfc01000000ffffffff070000000000000000000000000000000000000000000000000000000000000000000000000000ff', 1);
WARNING:  terminating connection because of crash of another server process
DETAIL:  The postmaster has commanded this server process to roll back the current transaction and exit, because another server process exited abnormally and possibly corrupted shared memory.
HINT:  In a moment you should be able to reconnect to the database and repeat your command.
FATAL:  server conn crashed?
```

The function takes a bytea argument and casts it to a C struct, without
validating the contents.

## Summary of changes

Added validation for number of pages to be prefetched and for the chunks
as well.

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

@@ -146,7 +146,9 @@ jobs:
         with:
           file: build-tools/Dockerfile
           context: .
-          provenance: false
+          attests: |
+            type=provenance,mode=max
+            type=sbom,generator=docker.io/docker/buildkit-syft-scanner:1
           push: true
           pull: true
           build-args: |

.github/workflows/build_and_test.yml

@@ -634,7 +634,9 @@ jobs:
             DEBIAN_VERSION=bookworm
           secrets: |
             SUBZERO_ACCESS_TOKEN=${{ secrets.CI_ACCESS_TOKEN }}
-          provenance: false
+          attests: |
+            type=provenance,mode=max
+            type=sbom,generator=docker.io/docker/buildkit-syft-scanner:1
           push: true
           pull: true
           file: Dockerfile
@@ -747,7 +749,9 @@ jobs:
             PG_VERSION=${{ matrix.version.pg }}
             BUILD_TAG=${{ needs.meta.outputs.release-tag || needs.meta.outputs.build-tag }}
             DEBIAN_VERSION=${{ matrix.version.debian }}
-          provenance: false
+          attests: |
+            type=provenance,mode=max
+            type=sbom,generator=docker.io/docker/buildkit-syft-scanner:1
           push: true
           pull: true
           file: compute/compute-node.Dockerfile
@@ -766,7 +770,9 @@ jobs:
             PG_VERSION=${{ matrix.version.pg }}
             BUILD_TAG=${{ needs.meta.outputs.release-tag || needs.meta.outputs.build-tag }}
             DEBIAN_VERSION=${{ matrix.version.debian }}
-          provenance: false
+          attests: |
+            type=provenance,mode=max
+            type=sbom,generator=docker.io/docker/buildkit-syft-scanner:1
           push: true
           pull: true
           file: compute/compute-node.Dockerfile

.github/workflows/pg-clients.yml

@@ -72,9 +72,10 @@ jobs:
       options: --init --user root
     services:
       clickhouse:
-        image: clickhouse/clickhouse-server:24.8
+        image: clickhouse/clickhouse-server:25.6
         env:
           CLICKHOUSE_PASSWORD: ${{ needs.generate-ch-tmppw.outputs.tmp_val }}
+          PGSSLCERT: /tmp/postgresql.crt
         ports:
           - 9000:9000
           - 8123:8123

.github/workflows/proxy-benchmark.yml

@@ -13,6 +13,12 @@ on:
     #        │ │ │ │ ┌───────────── day of the week (0 - 6 or SUN-SAT)
     # - cron: '0 5 * * *' # Runs at 5 UTC once a day
   workflow_dispatch: # adds an ability to run this manually
+    inputs:
+      commit_hash:
+        type: string
+        description: 'The long neon repo commit hash for the system under test (proxy) to be tested.'
+        required: false
+        default: ''
 
 defaults:
   run:
@@ -33,20 +39,14 @@ jobs:
       contents: write
       pull-requests: write
     runs-on: [ self-hosted, unit-perf-aws-arm ]
-    timeout-minutes: 60  # 1h timeout
     container:
       image: ghcr.io/neondatabase/build-tools:pinned-bookworm
       credentials:
         username: ${{ github.actor }}
         password: ${{ secrets.GITHUB_TOKEN }}
       options: --init
+    timeout-minutes: 60  # 1h timeout
     steps:
-    - name: Checkout proxy-bench Repo
-      uses: actions/checkout@v4
-      with:
-        repository: neondatabase/proxy-bench
-        path: proxy-bench
-
     - name: Set up the environment which depends on $RUNNER_TEMP on nvme drive
       id: set-env
       shell: bash -euxo pipefail {0}
@@ -54,19 +54,64 @@ jobs:
         PROXY_BENCH_PATH=$(realpath ./proxy-bench)
         {
           echo "PROXY_BENCH_PATH=$PROXY_BENCH_PATH"
-          echo "NEON_DIR=${RUNNER_TEMP}/neon"
-          echo "NEON_PROXY_PATH=${RUNNER_TEMP}/neon/bin/proxy"
+          echo "NEON_DIR=${GITHUB_WORKSPACE}/"
+          echo "NEON_PROXY_PATH=${GITHUB_WORKSPACE}/bin/proxy"
           echo "TEST_OUTPUT=${PROXY_BENCH_PATH}/test_output"
+          echo "DOCKER_COMPOSE_FILE=${PROXY_BENCH_PATH}/docker-compose.yml"
           echo ""
         } >> "$GITHUB_ENV"
 
+    - name: Determine commit hash
+      id: commit_hash
+      shell: bash -euxo pipefail {0}
+      env:
+        INPUT_COMMIT_HASH: ${{ github.event.inputs.commit_hash }}
+      run: |
+        if [[ -z "${INPUT_COMMIT_HASH}" ]]; then
+          COMMIT_HASH=$(curl -s https://api.github.com/repos/neondatabase/neon/commits/main | jq -r '.sha')
+          echo "COMMIT_HASH=$COMMIT_HASH" >> $GITHUB_ENV
+          echo "commit_hash=$COMMIT_HASH" >> "$GITHUB_OUTPUT"
+          echo "COMMIT_HASH_TYPE=latest" >> $GITHUB_ENV
+        else
+          COMMIT_HASH="${INPUT_COMMIT_HASH}"
+          echo "COMMIT_HASH=$COMMIT_HASH" >> $GITHUB_ENV
+          echo "commit_hash=$COMMIT_HASH" >> "$GITHUB_OUTPUT"
+          echo "COMMIT_HASH_TYPE=manual" >> $GITHUB_ENV
+        fi
+    - name: Checkout the neon repository at given commit hash
+      uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
+      with:
+        ref: ${{ steps.commit_hash.outputs.commit_hash }}
+
+    - name: Print GITHUB_WORKSPACE
+      run: echo "$GITHUB_WORKSPACE"
+
+    - name: List parent dir of workspace
+      run: ls -la /__w/neon
+
+    - name: List all env vars
+      run: env
+
+    - name: Checkout proxy-bench Repo
+      uses: actions/checkout@v4
+      with:
+        repository: neondatabase/proxy-bench
+        path: proxy-bench
+
     - name: Cache poetry deps
       uses: actions/cache@v4
       with:
         path: ~/.cache/pypoetry/virtualenvs
         key: v2-${{ runner.os }}-${{ runner.arch }}-python-deps-bookworm-${{ hashFiles('poetry.lock') }}
 
+    - name: DEBUG List files for debugging
+      working-directory: ${{ env.NEON_DIR }}
+      run: |
+        pwd
+        ls -la
+
     - name: Install Python deps
+      working-directory: ${{ env.NEON_DIR }}
       shell: bash -euxo pipefail {0}
       run: ./scripts/pysync
 
@@ -77,14 +122,14 @@ jobs:
 
     - name: Run proxy-bench
       working-directory: ${{ env.PROXY_BENCH_PATH }}
-      run: ./run.sh --with-grafana --bare-metal
+      run: ./run.sh --bare-metal
 
     - name: Ingest Bench Results
       if: always()
       working-directory: ${{ env.NEON_DIR }}
       run: |
         mkdir -p $TEST_OUTPUT
-        python $NEON_DIR/scripts/proxy_bench_results_ingest.py --out $TEST_OUTPUT
+        python3 ./scripts/proxy_bench_results_ingest.py --out $TEST_OUTPUT
 
     - name: Push Metrics to Proxy perf database
       shell: bash -euxo pipefail {0}
@@ -109,4 +154,4 @@ jobs:
         fi
         if [[ -d "${PROXY_BENCH_PATH}/test_output" ]]; then
           rm -rf ${PROXY_BENCH_PATH}/test_output
-        fi
+        fi

Cargo.lock

@@ -5077,8 +5077,6 @@ dependencies = [
  "crc32c",
  "criterion",
  "env_logger",
- "log",
- "memoffset 0.9.0",
  "once_cell",
  "postgres",
  "postgres_ffi_types",
@@ -5519,6 +5517,7 @@ dependencies = [
  "workspace_hack",
  "x509-cert",
  "zerocopy 0.8.24",
+ "zeroize",
 ]
 
 [[package]]

Cargo.toml

@@ -135,7 +135,6 @@ lock_api = "0.4.13"
 md5 = "0.7.0"
 measured = { version = "0.0.22", features=["lasso"] }
 measured-process = { version = "0.0.22" }
-memoffset = "0.9"
 moka = { version = "0.12", features = ["sync"] }
 nix = { version = "0.30.1", features = ["dir", "fs", "mman", "process", "socket", "signal", "poll"] }
 # Do not update to >= 7.0.0, at least. The update will have a significant impact
@@ -234,9 +233,10 @@ uuid = { version = "1.6.1", features = ["v4", "v7", "serde"] }
 walkdir = "2.3.2"
 rustls-native-certs = "0.8"
 whoami = "1.5.1"
-zerocopy = { version = "0.8", features = ["derive", "simd"] }
 json-structural-diff = { version = "0.2.0" }
 x509-cert = { version = "0.2.5" }
+zerocopy = { version = "0.8", features = ["derive", "simd"] }
+zeroize = "1.8"
 
 ## TODO replace this with tracing
 env_logger = "0.11"

Dockerfile

@@ -103,7 +103,7 @@ RUN  --mount=type=secret,uid=1000,id=SUBZERO_ACCESS_TOKEN \
     && if [ -s /run/secrets/SUBZERO_ACCESS_TOKEN ]; then \
         export CARGO_FEATURES="rest_broker"; \
     fi \
-    && RUSTFLAGS="-Clinker=clang -Clink-arg=-fuse-ld=mold -Clink-arg=-Wl,--no-rosegment -Cforce-frame-pointers=yes ${ADDITIONAL_RUSTFLAGS}" cargo build \
+    && RUSTFLAGS="-Clinker=clang -Clink-arg=-fuse-ld=mold -Clink-arg=-Wl,--no-rosegment -Cforce-frame-pointers=yes ${ADDITIONAL_RUSTFLAGS}" cargo auditable build \
       --features $CARGO_FEATURES \
       --bin pg_sni_router  \
       --bin pageserver  \

build-tools/Dockerfile

@@ -299,6 +299,7 @@ WORKDIR /home/nonroot
 ENV RUSTC_VERSION=1.88.0
 ENV RUSTUP_HOME="/home/nonroot/.rustup"
 ENV PATH="/home/nonroot/.cargo/bin:${PATH}"
+ARG CARGO_AUDITABLE_VERSION=0.7.0
 ARG RUSTFILT_VERSION=0.2.1
 ARG CARGO_HAKARI_VERSION=0.9.36
 ARG CARGO_DENY_VERSION=0.18.2
@@ -314,14 +315,16 @@ RUN curl -sSO https://static.rust-lang.org/rustup/dist/$(uname -m)-unknown-linux
     . "$HOME/.cargo/env" && \
     cargo --version && rustup --version && \
     rustup component add llvm-tools rustfmt clippy && \
-    cargo install rustfilt      --locked --version "${RUSTFILT_VERSION}" && \
-    cargo install cargo-hakari  --locked --version "${CARGO_HAKARI_VERSION}" && \
-    cargo install cargo-deny    --locked --version "${CARGO_DENY_VERSION}" && \
-    cargo install cargo-hack    --locked --version "${CARGO_HACK_VERSION}" && \
-    cargo install cargo-nextest --locked --version "${CARGO_NEXTEST_VERSION}" && \
-    cargo install cargo-chef    --locked --version "${CARGO_CHEF_VERSION}" && \
-    cargo install diesel_cli    --locked --version "${CARGO_DIESEL_CLI_VERSION}" \
-                                --features postgres-bundled --no-default-features && \
+    cargo install cargo-auditable           --locked --version "${CARGO_AUDITABLE_VERSION}" && \
+    cargo auditable install cargo-auditable --locked --version "${CARGO_AUDITABLE_VERSION}" --force && \
+    cargo auditable install rustfilt                 --version "${RUSTFILT_VERSION}" && \
+    cargo auditable install cargo-hakari    --locked --version "${CARGO_HAKARI_VERSION}" && \
+    cargo auditable install cargo-deny      --locked --version "${CARGO_DENY_VERSION}" && \
+    cargo auditable install cargo-hack      --locked --version "${CARGO_HACK_VERSION}" && \
+    cargo auditable install cargo-nextest   --locked --version "${CARGO_NEXTEST_VERSION}" && \
+    cargo auditable install cargo-chef      --locked --version "${CARGO_CHEF_VERSION}" && \
+    cargo auditable install diesel_cli      --locked --version "${CARGO_DIESEL_CLI_VERSION}" \
+                                            --features postgres-bundled --no-default-features && \
     rm -rf /home/nonroot/.cargo/registry && \
     rm -rf /home/nonroot/.cargo/git
 

compute/patches/pg_repack.patch

@@ -1,5 +1,11 @@
+commit 5eb393810cf7c7bafa4e394dad2e349e2a8cb2cb
+Author: Alexey Masterov <alexey.masterov@databricks.com>
+Date:   Mon Jul 28 18:11:02 2025 +0200
+
+    Patch for pg_repack
+
 diff --git a/regress/Makefile b/regress/Makefile
-index bf6edcb..89b4c7f 100644
+index bf6edcb..110e734 100644
 --- a/regress/Makefile
 +++ b/regress/Makefile
 @@ -17,7 +17,7 @@ INTVERSION := $(shell echo $$(($$(echo $(VERSION).0 | sed 's/\([[:digit:]]\{1,\}
@@ -7,18 +13,36 @@ index bf6edcb..89b4c7f 100644
  #
  
 -REGRESS := init-extension repack-setup repack-run error-on-invalid-idx no-error-on-invalid-idx after-schema repack-check nosuper tablespace get_order_by trigger
-+REGRESS := init-extension repack-setup repack-run error-on-invalid-idx no-error-on-invalid-idx after-schema repack-check nosuper get_order_by trigger
++REGRESS := init-extension noautovacuum repack-setup repack-run error-on-invalid-idx no-error-on-invalid-idx after-schema repack-check nosuper get_order_by trigger autovacuum
  
  USE_PGXS = 1	# use pgxs if not in contrib directory
  PGXS := $(shell $(PG_CONFIG) --pgxs)
-diff --git a/regress/expected/init-extension.out b/regress/expected/init-extension.out
-index 9f2e171..f6e4f8d 100644
---- a/regress/expected/init-extension.out
-+++ b/regress/expected/init-extension.out
-@@ -1,3 +1,2 @@
- SET client_min_messages = warning;
- CREATE EXTENSION pg_repack;
--RESET client_min_messages;
+diff --git a/regress/expected/autovacuum.out b/regress/expected/autovacuum.out
+new file mode 100644
+index 0000000..e7f2363
+--- /dev/null
++++ b/regress/expected/autovacuum.out
+@@ -0,0 +1,7 @@
++ALTER SYSTEM SET autovacuum='on';
++SELECT pg_reload_conf();
++ pg_reload_conf 
++----------------
++ t
++(1 row)
++
+diff --git a/regress/expected/noautovacuum.out b/regress/expected/noautovacuum.out
+new file mode 100644
+index 0000000..fc7978e
+--- /dev/null
++++ b/regress/expected/noautovacuum.out
+@@ -0,0 +1,7 @@
++ALTER SYSTEM SET autovacuum='off';
++SELECT pg_reload_conf();
++ pg_reload_conf 
++----------------
++ t
++(1 row)
++
 diff --git a/regress/expected/nosuper.out b/regress/expected/nosuper.out
 index 8d0a94e..63b68bf 100644
 --- a/regress/expected/nosuper.out
@@ -50,14 +74,22 @@ index 8d0a94e..63b68bf 100644
  INFO: repacking table "public.tbl_cluster"
  ERROR: query failed: ERROR:  current transaction is aborted, commands ignored until end of transaction block
  DETAIL: query was: RESET lock_timeout
-diff --git a/regress/sql/init-extension.sql b/regress/sql/init-extension.sql
-index 9f2e171..f6e4f8d 100644
---- a/regress/sql/init-extension.sql
-+++ b/regress/sql/init-extension.sql
-@@ -1,3 +1,2 @@
- SET client_min_messages = warning;
- CREATE EXTENSION pg_repack;
--RESET client_min_messages;
+diff --git a/regress/sql/autovacuum.sql b/regress/sql/autovacuum.sql
+new file mode 100644
+index 0000000..a8eda63
+--- /dev/null
++++ b/regress/sql/autovacuum.sql
+@@ -0,0 +1,2 @@
++ALTER SYSTEM SET autovacuum='on';
++SELECT pg_reload_conf();
+diff --git a/regress/sql/noautovacuum.sql b/regress/sql/noautovacuum.sql
+new file mode 100644
+index 0000000..13d4836
+--- /dev/null
++++ b/regress/sql/noautovacuum.sql
+@@ -0,0 +1,2 @@
++ALTER SYSTEM SET autovacuum='off';
++SELECT pg_reload_conf();
 diff --git a/regress/sql/nosuper.sql b/regress/sql/nosuper.sql
 index 072f0fa..dbe60f8 100644
 --- a/regress/sql/nosuper.sql

compute_tools/src/bin/compute_ctl.rs

@@ -82,6 +82,15 @@ struct Cli {
     #[arg(long, default_value_t = 3081)]
     pub internal_http_port: u16,
 
+    /// Backwards-compatible --http-port for Hadron deployments. Functionally the
+    /// same as --external-http-port.
+    #[arg(
+        long,
+        conflicts_with = "external_http_port",
+        conflicts_with = "internal_http_port"
+    )]
+    pub http_port: Option<u16>,
+
     #[arg(short = 'D', long, value_name = "DATADIR")]
     pub pgdata: String,
 
@@ -181,6 +190,26 @@ impl Cli {
     }
 }
 
+// Hadron helpers to get compatible compute_ctl http ports from Cli. The old `--http-port`
+// arg is used and acts the same as `--external-http-port`. The internal http port is defined
+// to be http_port + 1. Hadron runs in the dblet environment which uses the host network, so
+// we need to be careful with the ports to choose.
+fn get_external_http_port(cli: &Cli) -> u16 {
+    if cli.lakebase_mode {
+        return cli.http_port.unwrap_or(cli.external_http_port);
+    }
+    cli.external_http_port
+}
+fn get_internal_http_port(cli: &Cli) -> u16 {
+    if cli.lakebase_mode {
+        return cli
+            .http_port
+            .map(|p| p + 1)
+            .unwrap_or(cli.internal_http_port);
+    }
+    cli.internal_http_port
+}
+
 fn main() -> Result<()> {
     let cli = Cli::parse();
 
@@ -205,13 +234,18 @@ fn main() -> Result<()> {
     // enable core dumping for all child processes
     setrlimit(Resource::CORE, rlimit::INFINITY, rlimit::INFINITY)?;
 
-    installed_extensions::initialize_metrics();
-    hadron_metrics::initialize_metrics();
+    if cli.lakebase_mode {
+        installed_extensions::initialize_metrics();
+        hadron_metrics::initialize_metrics();
+    }
 
     let connstr = Url::parse(&cli.connstr).context("cannot parse connstr as a URL")?;
 
     let config = get_config(&cli)?;
 
+    let external_http_port = get_external_http_port(&cli);
+    let internal_http_port = get_internal_http_port(&cli);
+
     let compute_node = ComputeNode::new(
         ComputeNodeParams {
             compute_id: cli.compute_id,
@@ -220,8 +254,8 @@ fn main() -> Result<()> {
             pgdata: cli.pgdata.clone(),
             pgbin: cli.pgbin.clone(),
             pgversion: get_pg_version_string(&cli.pgbin),
-            external_http_port: cli.external_http_port,
-            internal_http_port: cli.internal_http_port,
+            external_http_port,
+            internal_http_port,
             remote_ext_base_url: cli.remote_ext_base_url.clone(),
             resize_swap_on_bind: cli.resize_swap_on_bind,
             set_disk_quota_for_fs: cli.set_disk_quota_for_fs,

compute_tools/src/compute.rs

@@ -6,7 +6,8 @@ use compute_api::responses::{
     LfcPrewarmState, PromoteState, TlsConfig,
 };
 use compute_api::spec::{
-    ComputeAudit, ComputeFeature, ComputeMode, ComputeSpec, ExtVersion, PageserverProtocol, PgIdent,
+    ComputeAudit, ComputeFeature, ComputeMode, ComputeSpec, ExtVersion, GenericOption,
+    PageserverConnectionInfo, PageserverProtocol, PgIdent, Role,
 };
 use futures::StreamExt;
 use futures::future::join_all;
@@ -37,7 +38,7 @@ use utils::id::{TenantId, TimelineId};
 use utils::lsn::Lsn;
 use utils::measured_stream::MeasuredReader;
 use utils::pid_file;
-use utils::shard::{ShardCount, ShardIndex, ShardNumber};
+use utils::shard::{ShardIndex, ShardNumber, ShardStripeSize};
 
 use crate::configurator::launch_configurator;
 use crate::disk_quota::set_disk_quota;
@@ -248,7 +249,7 @@ pub struct ParsedSpec {
     pub spec: ComputeSpec,
     pub tenant_id: TenantId,
     pub timeline_id: TimelineId,
-    pub pageserver_connstr: String,
+    pub pageserver_conninfo: PageserverConnectionInfo,
     pub safekeeper_connstrings: Vec<String>,
     pub storage_auth_token: Option<String>,
     /// k8s dns name and port
@@ -296,25 +297,47 @@ impl ParsedSpec {
 }
 
 impl TryFrom<ComputeSpec> for ParsedSpec {
-    type Error = String;
-    fn try_from(spec: ComputeSpec) -> Result<Self, String> {
+    type Error = anyhow::Error;
+    fn try_from(spec: ComputeSpec) -> Result<Self, anyhow::Error> {
         // Extract the options from the spec file that are needed to connect to
         // the storage system.
         //
-        // For backwards-compatibility, the top-level fields in the spec file
-        // may be empty. In that case, we need to dig them from the GUCs in the
-        // cluster.settings field.
-        let pageserver_connstr = spec
-            .pageserver_connstring
-            .clone()
-            .or_else(|| spec.cluster.settings.find("neon.pageserver_connstring"))
-            .ok_or("pageserver connstr should be provided")?;
+        // In compute specs generated by old control plane versions, the spec file might
+        // be missing the `pageserver_connection_info` field. In that case, we need to dig
+        // the pageserver connection info from the `pageserver_connstr` field instead, or
+        // if that's missing too, from the GUC in the cluster.settings field.
+        let mut pageserver_conninfo = spec.pageserver_connection_info.clone();
+        if pageserver_conninfo.is_none() {
+            if let Some(pageserver_connstr_field) = &spec.pageserver_connstring {
+                pageserver_conninfo = Some(PageserverConnectionInfo::from_connstr(
+                    pageserver_connstr_field,
+                    spec.shard_stripe_size,
+                )?);
+            }
+        }
+        if pageserver_conninfo.is_none() {
+            if let Some(guc) = spec.cluster.settings.find("neon.pageserver_connstring") {
+                let stripe_size = if let Some(guc) = spec.cluster.settings.find("neon.stripe_size")
+                {
+                    Some(ShardStripeSize(u32::from_str(&guc)?))
+                } else {
+                    None
+                };
+                pageserver_conninfo =
+                    Some(PageserverConnectionInfo::from_connstr(&guc, stripe_size)?);
+            }
+        }
+        let pageserver_conninfo = pageserver_conninfo.ok_or(anyhow::anyhow!(
+            "pageserver connection information should be provided"
+        ))?;
+
+        // Similarly for safekeeper connection strings
         let safekeeper_connstrings = if spec.safekeeper_connstrings.is_empty() {
             if matches!(spec.mode, ComputeMode::Primary) {
                 spec.cluster
                     .settings
                     .find("neon.safekeepers")
-                    .ok_or("safekeeper connstrings should be provided")?
+                    .ok_or(anyhow::anyhow!("safekeeper connstrings should be provided"))?
                     .split(',')
                     .map(|str| str.to_string())
                     .collect()
@@ -329,22 +352,22 @@ impl TryFrom<ComputeSpec> for ParsedSpec {
         let tenant_id: TenantId = if let Some(tenant_id) = spec.tenant_id {
             tenant_id
         } else {
-            spec.cluster
+            let guc = spec
+                .cluster
                 .settings
                 .find("neon.tenant_id")
-                .ok_or("tenant id should be provided")
-                .map(|s| TenantId::from_str(&s))?
-                .or(Err("invalid tenant id"))?
+                .ok_or(anyhow::anyhow!("tenant id should be provided"))?;
+            TenantId::from_str(&guc).context("invalid tenant id")?
         };
         let timeline_id: TimelineId = if let Some(timeline_id) = spec.timeline_id {
             timeline_id
         } else {
-            spec.cluster
+            let guc = spec
+                .cluster
                 .settings
                 .find("neon.timeline_id")
-                .ok_or("timeline id should be provided")
-                .map(|s| TimelineId::from_str(&s))?
-                .or(Err("invalid timeline id"))?
+                .ok_or(anyhow::anyhow!("timeline id should be provided"))?;
+            TimelineId::from_str(&guc).context(anyhow::anyhow!("invalid timeline id"))?
         };
 
         let endpoint_storage_addr: Option<String> = spec
@@ -358,7 +381,7 @@ impl TryFrom<ComputeSpec> for ParsedSpec {
 
         let res = ParsedSpec {
             spec,
-            pageserver_connstr,
+            pageserver_conninfo,
             safekeeper_connstrings,
             storage_auth_token,
             tenant_id,
@@ -368,7 +391,7 @@ impl TryFrom<ComputeSpec> for ParsedSpec {
         };
 
         // Now check validity of the parsed specification
-        res.validate()?;
+        res.validate().map_err(anyhow::Error::msg)?;
         Ok(res)
     }
 }
@@ -413,6 +436,66 @@ struct StartVmMonitorResult {
     vm_monitor: Option<JoinHandle<Result<()>>>,
 }
 
+// BEGIN_HADRON
+/// This function creates roles that are used by Databricks.
+/// These roles are not needs to be botostrapped at PG Compute provisioning time.
+/// The auth method for these roles are configured in databricks_pg_hba.conf in universe repository.
+pub(crate) fn create_databricks_roles() -> Vec<String> {
+    let roles = vec![
+        // Role for prometheus_stats_exporter
+        Role {
+            name: "databricks_monitor".to_string(),
+            // This uses "local" connection and auth method for that is "trust", so no password is needed.
+            encrypted_password: None,
+            options: Some(vec![GenericOption {
+                name: "IN ROLE pg_monitor".to_string(),
+                value: None,
+                vartype: "string".to_string(),
+            }]),
+        },
+        // Role for brickstore control plane
+        Role {
+            name: "databricks_control_plane".to_string(),
+            // Certificate user does not need password.
+            encrypted_password: None,
+            options: Some(vec![GenericOption {
+                name: "SUPERUSER".to_string(),
+                value: None,
+                vartype: "string".to_string(),
+            }]),
+        },
+        // Role for brickstore httpgateway.
+        Role {
+            name: "databricks_gateway".to_string(),
+            // Certificate user does not need password.
+            encrypted_password: None,
+            options: None,
+        },
+    ];
+
+    roles
+        .into_iter()
+        .map(|role| {
+            let query = format!(
+                r#"
+                DO $$
+                    BEGIN
+                        IF NOT EXISTS (
+                            SELECT FROM pg_catalog.pg_roles WHERE rolname = '{}')
+                        THEN
+                            CREATE ROLE {} {};
+                        END IF;
+                    END
+                $$;"#,
+                role.name,
+                role.name.pg_quote(),
+                role.to_pg_options(),
+            );
+            query
+        })
+        .collect()
+}
+
 /// Databricks-specific environment variables to be passed to the `postgres` sub-process.
 pub struct DatabricksEnvVars {
     /// The Databricks "endpoint ID" of the compute instance. Used by `postgres` to check
@@ -421,14 +504,27 @@ pub struct DatabricksEnvVars {
     /// Hostname of the Databricks workspace URL this compute instance belongs to.
     /// Used by postgres to verify Databricks PAT tokens.
     pub workspace_host: String,
+
+    pub lakebase_mode: bool,
 }
 
 impl DatabricksEnvVars {
-    pub fn new(compute_spec: &ComputeSpec, compute_id: Option<&String>) -> Self {
-        // compute_id is a string format of "{endpoint_id}/{compute_idx}"
-        // endpoint_id is a uuid. We only need to pass down endpoint_id to postgres.
-        // Panics if compute_id is not set or not in the expected format.
-        let endpoint_id = compute_id.unwrap().split('/').next().unwrap().to_string();
+    pub fn new(
+        compute_spec: &ComputeSpec,
+        compute_id: Option<&String>,
+        instance_id: Option<String>,
+        lakebase_mode: bool,
+    ) -> Self {
+        let endpoint_id = if let Some(instance_id) = instance_id {
+            // Use instance_id as endpoint_id if it is set. This code path is for PuPr model.
+            instance_id
+        } else {
+            // Use compute_id as endpoint_id if instance_id is not set. The code path is for PrPr model.
+            // compute_id is a string format of "{endpoint_id}/{compute_idx}"
+            // endpoint_id is a uuid. We only need to pass down endpoint_id to postgres.
+            // Panics if compute_id is not set or not in the expected format.
+            compute_id.unwrap().split('/').next().unwrap().to_string()
+        };
         let workspace_host = compute_spec
             .databricks_settings
             .as_ref()
@@ -437,6 +533,7 @@ impl DatabricksEnvVars {
         Self {
             endpoint_id,
             workspace_host,
+            lakebase_mode,
         }
     }
 
@@ -446,6 +543,10 @@ impl DatabricksEnvVars {
 
     /// Convert DatabricksEnvVars to a list of string pairs that can be passed as env vars. Consumes `self`.
     pub fn to_env_var_list(self) -> Vec<(String, String)> {
+        if !self.lakebase_mode {
+            // In neon env, we don't need to pass down the env vars to postgres.
+            return vec![];
+        }
         vec![
             (
                 Self::DATABRICKS_ENDPOINT_ID_ENVVAR.to_string(),
@@ -495,7 +596,11 @@ impl ComputeNode {
         let mut new_state = ComputeState::new();
         if let Some(spec) = config.spec {
             let pspec = ParsedSpec::try_from(spec).map_err(|msg| anyhow::anyhow!(msg))?;
-            new_state.pspec = Some(pspec);
+            if params.lakebase_mode {
+                ComputeNode::set_spec(&params, &mut new_state, pspec);
+            } else {
+                new_state.pspec = Some(pspec);
+            }
         }
 
         Ok(ComputeNode {
@@ -1093,7 +1198,14 @@ impl ComputeNode {
         // If it is something different then create_dir() will error out anyway.
         let pgdata = &self.params.pgdata;
         let _ok = fs::remove_dir_all(pgdata);
-        fs::create_dir(pgdata)?;
+        if self.params.lakebase_mode {
+            // Ignore creation errors if the directory already exists (e.g. mounting it ahead of time).
+            // If it is something different then PG startup will error out anyway.
+            let _ok = fs::create_dir(pgdata);
+        } else {
+            fs::create_dir(pgdata)?;
+        }
+
         fs::set_permissions(pgdata, fs::Permissions::from_mode(0o700))?;
 
         Ok(())
@@ -1105,12 +1217,10 @@ impl ComputeNode {
     fn try_get_basebackup(&self, compute_state: &ComputeState, lsn: Lsn) -> Result<()> {
         let spec = compute_state.pspec.as_ref().expect("spec must be set");
 
-        let shard0_connstr = spec.pageserver_connstr.split(',').next().unwrap();
         let started = Instant::now();
-
-        let (connected, size) = match PageserverProtocol::from_connstring(shard0_connstr)? {
-            PageserverProtocol::Libpq => self.try_get_basebackup_libpq(spec, lsn)?,
+        let (connected, size) = match spec.pageserver_conninfo.prefer_protocol {
             PageserverProtocol::Grpc => self.try_get_basebackup_grpc(spec, lsn)?,
+            PageserverProtocol::Libpq => self.try_get_basebackup_libpq(spec, lsn)?,
         };
 
         self.fix_zenith_signal_neon_signal()?;
@@ -1148,23 +1258,20 @@ impl ComputeNode {
     /// Fetches a basebackup via gRPC. The connstring must use grpc://. Returns the timestamp when
     /// the connection was established, and the (compressed) size of the basebackup.
     fn try_get_basebackup_grpc(&self, spec: &ParsedSpec, lsn: Lsn) -> Result<(Instant, usize)> {
-        let shard0_connstr = spec
-            .pageserver_connstr
-            .split(',')
-            .next()
-            .unwrap()
-            .to_string();
-        let shard_index = match spec.pageserver_connstr.split(',').count() as u8 {
-            0 | 1 => ShardIndex::unsharded(),
-            count => ShardIndex::new(ShardNumber(0), ShardCount(count)),
+        let shard0_index = ShardIndex {
+            shard_number: ShardNumber(0),
+            shard_count: spec.pageserver_conninfo.shard_count,
         };
-
+        let shard0_url = spec
+            .pageserver_conninfo
+            .shard_url(ShardNumber(0), PageserverProtocol::Grpc)?
+            .to_owned();
         let (reader, connected) = tokio::runtime::Handle::current().block_on(async move {
             let mut client = page_api::Client::connect(
-                shard0_connstr,
+                shard0_url,
                 spec.tenant_id,
                 spec.timeline_id,
-                shard_index,
+                shard0_index,
                 spec.storage_auth_token.clone(),
                 None, // NB: base backups use payload compression
             )
@@ -1196,7 +1303,9 @@ impl ComputeNode {
     /// Fetches a basebackup via libpq. The connstring must use postgresql://. Returns the timestamp
     /// when the connection was established, and the (compressed) size of the basebackup.
     fn try_get_basebackup_libpq(&self, spec: &ParsedSpec, lsn: Lsn) -> Result<(Instant, usize)> {
-        let shard0_connstr = spec.pageserver_connstr.split(',').next().unwrap();
+        let shard0_connstr = spec
+            .pageserver_conninfo
+            .shard_url(ShardNumber(0), PageserverProtocol::Libpq)?;
         let mut config = postgres::Config::from_str(shard0_connstr)?;
 
         // Use the storage auth token from the config file, if given.
@@ -1283,10 +1392,7 @@ impl ComputeNode {
                     return result;
                 }
                 Err(ref e) if attempts < max_attempts => {
-                    warn!(
-                        "Failed to get basebackup: {} (attempt {}/{})",
-                        e, attempts, max_attempts
-                    );
+                    warn!("Failed to get basebackup: {e:?} (attempt {attempts}/{max_attempts})");
                     std::thread::sleep(std::time::Duration::from_millis(retry_period_ms as u64));
                     retry_period_ms *= 1.5;
                 }
@@ -1499,16 +1605,8 @@ impl ComputeNode {
             }
         };
 
-        info!(
-            "getting basebackup@{} from pageserver {}",
-            lsn, &pspec.pageserver_connstr
-        );
-        self.get_basebackup(compute_state, lsn).with_context(|| {
-            format!(
-                "failed to get basebackup@{} from pageserver {}",
-                lsn, &pspec.pageserver_connstr
-            )
-        })?;
+        self.get_basebackup(compute_state, lsn)
+            .with_context(|| format!("failed to get basebackup@{lsn}"))?;
 
         if let Some(settings) = databricks_settings {
             copy_tls_certificates(
@@ -1572,7 +1670,7 @@ impl ComputeNode {
         // symlink doesn't affect anything.
         //
         // See https://github.com/neondatabase/autoscaling/issues/800
-        std::fs::remove_dir(pgdata_path.join("pg_dynshmem"))?;
+        std::fs::remove_dir_all(pgdata_path.join("pg_dynshmem"))?;
         symlink("/dev/shm/", pgdata_path.join("pg_dynshmem"))?;
 
         match spec.mode {
@@ -1587,6 +1685,12 @@ impl ComputeNode {
 
     /// Start and stop a postgres process to warm up the VM for startup.
     pub fn prewarm_postgres_vm_memory(&self) -> Result<()> {
+        if self.params.lakebase_mode {
+            // We are running in Hadron mode. Disabling this prewarming step for now as it could run
+            // into dblet port conflicts and also doesn't add much value with our current infra.
+            info!("Skipping postgres prewarming in Hadron mode");
+            return Ok(());
+        }
         info!("prewarming VM memory");
 
         // Create pgdata
@@ -1648,7 +1752,12 @@ impl ComputeNode {
             let databricks_env_vars = {
                 let state = self.state.lock().unwrap();
                 let spec = &state.pspec.as_ref().unwrap().spec;
-                DatabricksEnvVars::new(spec, Some(&self.params.compute_id))
+                DatabricksEnvVars::new(
+                    spec,
+                    Some(&self.params.compute_id),
+                    self.params.instance_id.clone(),
+                    self.params.lakebase_mode,
+                )
             };
 
             info!(
@@ -1820,7 +1929,15 @@ impl ComputeNode {
     /// Do initial configuration of the already started Postgres.
     #[instrument(skip_all)]
     pub fn apply_config(&self, compute_state: &ComputeState) -> Result<()> {
-        let conf = self.get_tokio_conn_conf(Some("compute_ctl:apply_config"));
+        let mut conf = self.get_tokio_conn_conf(Some("compute_ctl:apply_config"));
+
+        if self.params.lakebase_mode {
+            // Set a 2-minute statement_timeout for the session applying config. The individual SQL statements
+            // used in apply_spec_sql() should not take long (they are just creating users and installing
+            // extensions). If any of them are stuck for an extended period of time it usually indicates a
+            // pageserver connectivity problem and we should bail out.
+            conf.options("-c statement_timeout=2min");
+        }
 
         let conf = Arc::new(conf);
         let spec = Arc::new(
@@ -2138,7 +2255,17 @@ impl ComputeNode {
     pub fn check_for_core_dumps(&self) -> Result<()> {
         let core_dump_dir = match std::env::consts::OS {
             "macos" => Path::new("/cores/"),
-            _ => Path::new(&self.params.pgdata),
+            // BEGIN HADRON
+            // NB: Read core dump files from a fixed location outside of
+            // the data directory since `compute_ctl` wipes the data directory
+            // across container restarts.
+            _ => {
+                if self.params.lakebase_mode {
+                    Path::new("/databricks/logs/brickstore")
+                } else {
+                    Path::new(&self.params.pgdata)
+                }
+            } // END HADRON
         };
 
         // Collect core dump paths if any
@@ -2451,7 +2578,7 @@ LIMIT 100",
         if let Some(libs) = spec.cluster.settings.find("shared_preload_libraries") {
             libs_vec = libs
                 .split(&[',', '\'', ' '])
-                .filter(|s| *s != "neon" && !s.is_empty())
+                .filter(|s| *s != "neon" && *s != "databricks_auth" && !s.is_empty())
                 .map(str::to_string)
                 .collect();
         }
@@ -2470,7 +2597,7 @@ LIMIT 100",
             if let Some(libs) = shared_preload_libraries_line.split("='").nth(1) {
                 preload_libs_vec = libs
                     .split(&[',', '\'', ' '])
-                    .filter(|s| *s != "neon" && !s.is_empty())
+                    .filter(|s| *s != "neon" && *s != "databricks_auth" && !s.is_empty())
                     .map(str::to_string)
                     .collect();
             }
@@ -2523,22 +2650,22 @@ LIMIT 100",
     /// The operation will time out after a specified duration.
     pub fn wait_timeout_while_pageserver_connstr_unchanged(&self, duration: Duration) {
         let state = self.state.lock().unwrap();
-        let old_pageserver_connstr = state
+        let old_pageserver_conninfo = state
             .pspec
             .as_ref()
             .expect("spec must be set")
-            .pageserver_connstr
+            .pageserver_conninfo
             .clone();
         let mut unchanged = true;
         let _ = self
             .state_changed
             .wait_timeout_while(state, duration, |s| {
-                let pageserver_connstr = &s
+                let pageserver_conninfo = &s
                     .pspec
                     .as_ref()
                     .expect("spec must be set")
-                    .pageserver_connstr;
-                unchanged = pageserver_connstr == &old_pageserver_connstr;
+                    .pageserver_conninfo;
+                unchanged = pageserver_conninfo == &old_pageserver_conninfo;
                 unchanged
             })
             .unwrap();
@@ -2796,7 +2923,10 @@ mod tests {
 
         match ParsedSpec::try_from(spec.clone()) {
             Ok(_p) => panic!("Failed to detect duplicate entry"),
-            Err(e) => assert!(e.starts_with("duplicate entry in safekeeper_connstrings:")),
+            Err(e) => assert!(
+                e.to_string()
+                    .starts_with("duplicate entry in safekeeper_connstrings:")
+            ),
         };
     }
 }

compute_tools/src/config.rs

@@ -18,6 +18,8 @@ use crate::pg_helpers::{
 };
 use crate::tls::{self, SERVER_CRT, SERVER_KEY};
 
+use utils::shard::{ShardIndex, ShardNumber};
+
 /// Check that `line` is inside a text file and put it there if it is not.
 /// Create file if it doesn't exist.
 pub fn line_in_file(path: &Path, line: &str) -> Result<bool> {
@@ -69,9 +71,75 @@ pub fn write_postgres_conf(
     }
     // Add options for connecting to storage
     writeln!(file, "# Neon storage settings")?;
-    if let Some(s) = &spec.pageserver_connstring {
-        writeln!(file, "neon.pageserver_connstring={}", escape_conf_value(s))?;
+    writeln!(file)?;
+    if let Some(conninfo) = &spec.pageserver_connection_info {
+        let mut libpq_urls: Option<Vec<String>> = Some(Vec::new());
+        let num_shards = if conninfo.shard_count.0 == 0 {
+            1 // unsharded, treat it as a single shard
+        } else {
+            conninfo.shard_count.0
+        };
+
+        for shard_number in 0..num_shards {
+            let shard_index = ShardIndex {
+                shard_number: ShardNumber(shard_number),
+                shard_count: conninfo.shard_count,
+            };
+            let info = conninfo.shards.get(&shard_index).ok_or_else(|| {
+                anyhow::anyhow!(
+                    "shard {shard_index} missing from pageserver_connection_info shard map"
+                )
+            })?;
+
+            let first_pageserver = info
+                .pageservers
+                .first()
+                .expect("must have at least one pageserver");
+
+            // Add the libpq URL to the array, or if the URL is missing, reset the array
+            // forgetting any previous entries. All servers must have a libpq URL, or none
+            // at all.
+            if let Some(url) = &first_pageserver.libpq_url {
+                if let Some(ref mut urls) = libpq_urls {
+                    urls.push(url.clone());
+                }
+            } else {
+                libpq_urls = None
+            }
+        }
+        if let Some(libpq_urls) = libpq_urls {
+            writeln!(
+                file,
+                "# derived from compute spec's pageserver_conninfo field"
+            )?;
+            writeln!(
+                file,
+                "neon.pageserver_connstring={}",
+                escape_conf_value(&libpq_urls.join(","))
+            )?;
+        } else {
+            writeln!(file, "# no neon.pageserver_connstring")?;
+        }
+
+        if let Some(stripe_size) = conninfo.stripe_size {
+            writeln!(
+                file,
+                "# from compute spec's pageserver_conninfo.stripe_size field"
+            )?;
+            writeln!(file, "neon.stripe_size={stripe_size}")?;
+        }
+    } else {
+        if let Some(s) = &spec.pageserver_connstring {
+            writeln!(file, "# from compute spec's pageserver_connstring field")?;
+            writeln!(file, "neon.pageserver_connstring={}", escape_conf_value(s))?;
+        }
+
+        if let Some(stripe_size) = spec.shard_stripe_size {
+            writeln!(file, "# from compute spec's shard_stripe_size field")?;
+            writeln!(file, "neon.stripe_size={stripe_size}")?;
+        }
     }
+
     if !spec.safekeeper_connstrings.is_empty() {
         let mut neon_safekeepers_value = String::new();
         tracing::info!(

compute_tools/src/configurator.rs

@@ -122,8 +122,11 @@ fn configurator_main_loop(compute: &Arc<ComputeNode>) {
                         // into the type system.
                         assert_eq!(state.status, ComputeStatus::RefreshConfiguration);
 
-                        if state.pspec.as_ref().map(|ps| ps.pageserver_connstr.clone())
-                            == Some(pspec.pageserver_connstr.clone())
+                        if state
+                            .pspec
+                            .as_ref()
+                            .map(|ps| ps.pageserver_conninfo.clone())
+                            == Some(pspec.pageserver_conninfo.clone())
                         {
                             info!(
                                 "Refresh configuration: Retrieved spec is the same as the current spec. Waiting for control plane to update the spec before attempting reconfiguration."

compute_tools/src/lsn_lease.rs

@@ -4,14 +4,13 @@ use std::thread;
 use std::time::{Duration, SystemTime};
 
 use anyhow::{Result, bail};
-use compute_api::spec::{ComputeMode, PageserverProtocol};
-use itertools::Itertools as _;
+use compute_api::spec::{ComputeMode, PageserverConnectionInfo, PageserverProtocol};
 use pageserver_page_api as page_api;
 use postgres::{NoTls, SimpleQueryMessage};
 use tracing::{info, warn};
 use utils::id::{TenantId, TimelineId};
 use utils::lsn::Lsn;
-use utils::shard::{ShardCount, ShardNumber, TenantShardId};
+use utils::shard::TenantShardId;
 
 use crate::compute::ComputeNode;
 
@@ -78,17 +77,16 @@ fn acquire_lsn_lease_with_retry(
 
     loop {
         // Note: List of pageservers is dynamic, need to re-read configs before each attempt.
-        let (connstrings, auth) = {
+        let (conninfo, auth) = {
             let state = compute.state.lock().unwrap();
             let spec = state.pspec.as_ref().expect("spec must be set");
             (
-                spec.pageserver_connstr.clone(),
+                spec.pageserver_conninfo.clone(),
                 spec.storage_auth_token.clone(),
             )
         };
 
-        let result =
-            try_acquire_lsn_lease(&connstrings, auth.as_deref(), tenant_id, timeline_id, lsn);
+        let result = try_acquire_lsn_lease(conninfo, auth.as_deref(), tenant_id, timeline_id, lsn);
         match result {
             Ok(Some(res)) => {
                 return Ok(res);
@@ -112,35 +110,44 @@ fn acquire_lsn_lease_with_retry(
 
 /// Tries to acquire LSN leases on all Pageserver shards.
 fn try_acquire_lsn_lease(
-    connstrings: &str,
+    conninfo: PageserverConnectionInfo,
     auth: Option<&str>,
     tenant_id: TenantId,
     timeline_id: TimelineId,
     lsn: Lsn,
 ) -> Result<Option<SystemTime>> {
-    let connstrings = connstrings.split(',').collect_vec();
-    let shard_count = connstrings.len();
     let mut leases = Vec::new();
 
-    for (shard_number, &connstring) in connstrings.iter().enumerate() {
-        let tenant_shard_id = match shard_count {
-            0 | 1 => TenantShardId::unsharded(tenant_id),
-            shard_count => TenantShardId {
-                tenant_id,
-                shard_number: ShardNumber(shard_number as u8),
-                shard_count: ShardCount::new(shard_count as u8),
-            },
+    for (shard_index, shard) in conninfo.shards.into_iter() {
+        let tenant_shard_id = TenantShardId {
+            tenant_id,
+            shard_number: shard_index.shard_number,
+            shard_count: shard_index.shard_count,
         };
 
-        let lease = match PageserverProtocol::from_connstring(connstring)? {
-            PageserverProtocol::Libpq => {
-                acquire_lsn_lease_libpq(connstring, auth, tenant_shard_id, timeline_id, lsn)?
-            }
-            PageserverProtocol::Grpc => {
-                acquire_lsn_lease_grpc(connstring, auth, tenant_shard_id, timeline_id, lsn)?
-            }
-        };
-        leases.push(lease);
+        // XXX: If there are more than pageserver for the one shard, do we need to get a
+        // leas on all of them? Currently, that's what we assume, but this is hypothetical
+        // as of this writing, as we never pass the info for more than one pageserver per
+        // shard.
+        for pageserver in shard.pageservers {
+            let lease = match conninfo.prefer_protocol {
+                PageserverProtocol::Grpc => acquire_lsn_lease_grpc(
+                    &pageserver.grpc_url.unwrap(),
+                    auth,
+                    tenant_shard_id,
+                    timeline_id,
+                    lsn,
+                )?,
+                PageserverProtocol::Libpq => acquire_lsn_lease_libpq(
+                    &pageserver.libpq_url.unwrap(),
+                    auth,
+                    tenant_shard_id,
+                    timeline_id,
+                    lsn,
+                )?,
+            };
+            leases.push(lease);
+        }
     }
 
     Ok(leases.into_iter().min().flatten())

compute_tools/src/spec_apply.rs

@@ -13,17 +13,19 @@ use tokio_postgres::Client;
 use tokio_postgres::error::SqlState;
 use tracing::{Instrument, debug, error, info, info_span, instrument, warn};
 
-use crate::compute::{ComputeNode, ComputeNodeParams, ComputeState};
+use crate::compute::{ComputeNode, ComputeNodeParams, ComputeState, create_databricks_roles};
+use crate::hadron_metrics::COMPUTE_CONFIGURE_STATEMENT_TIMEOUT_ERRORS;
 use crate::pg_helpers::{
     DatabaseExt, Escaping, GenericOptionsSearch, RoleExt, get_existing_dbs_async,
     get_existing_roles_async,
 };
 use crate::spec_apply::ApplySpecPhase::{
-    CreateAndAlterDatabases, CreateAndAlterRoles, CreateAvailabilityCheck, CreatePgauditExtension,
+    AddDatabricksGrants, AlterDatabricksRoles, CreateAndAlterDatabases, CreateAndAlterRoles,
+    CreateAvailabilityCheck, CreateDatabricksMisc, CreateDatabricksRoles, CreatePgauditExtension,
     CreatePgauditlogtofileExtension, CreatePrivilegedRole, CreateSchemaNeon,
     DisablePostgresDBPgAudit, DropInvalidDatabases, DropRoles, FinalizeDropLogicalSubscriptions,
-    HandleNeonExtension, HandleOtherExtensions, RenameAndDeleteDatabases, RenameRoles,
-    RunInEachDatabase,
+    HandleDatabricksAuthExtension, HandleNeonExtension, HandleOtherExtensions,
+    RenameAndDeleteDatabases, RenameRoles, RunInEachDatabase,
 };
 use crate::spec_apply::PerDatabasePhase::{
     ChangeSchemaPerms, DeleteDBRoleReferences, DropLogicalSubscriptions,
@@ -166,6 +168,7 @@ impl ComputeNode {
                         concurrency_token.clone(),
                         db,
                         [DropLogicalSubscriptions].to_vec(),
+                        self.params.lakebase_mode,
                     );
 
                     Ok(tokio::spawn(fut))
@@ -186,15 +189,33 @@ impl ComputeNode {
                 };
             }
 
-            for phase in [
-                CreatePrivilegedRole,
+            let phases = if self.params.lakebase_mode {
+                vec![
+                    CreatePrivilegedRole,
+                // BEGIN_HADRON
+                CreateDatabricksRoles,
+                AlterDatabricksRoles,
+                // END_HADRON
                 DropInvalidDatabases,
                 RenameRoles,
                 CreateAndAlterRoles,
                 RenameAndDeleteDatabases,
                 CreateAndAlterDatabases,
                 CreateSchemaNeon,
-            ] {
+            ]
+            } else {
+                vec![
+                    CreatePrivilegedRole,
+                DropInvalidDatabases,
+                RenameRoles,
+                CreateAndAlterRoles,
+                RenameAndDeleteDatabases,
+                CreateAndAlterDatabases,
+                CreateSchemaNeon,
+            ]
+            };
+
+            for phase in phases {
                 info!("Applying phase {:?}", &phase);
                 apply_operations(
                     params.clone(),
@@ -203,6 +224,7 @@ impl ComputeNode {
                     jwks_roles.clone(),
                     phase,
                     || async { Ok(&client) },
+                    self.params.lakebase_mode,
                 )
                 .await?;
             }
@@ -254,6 +276,7 @@ impl ComputeNode {
                         concurrency_token.clone(),
                         db,
                         phases,
+                        self.params.lakebase_mode,
                     );
 
                     Ok(tokio::spawn(fut))
@@ -265,12 +288,28 @@ impl ComputeNode {
                 handle.await??;
             }
 
-            let mut phases = vec![
+            let mut phases = if self.params.lakebase_mode {
+                vec![
+                HandleOtherExtensions,
+                HandleNeonExtension, // This step depends on CreateSchemaNeon
+                // BEGIN_HADRON
+                HandleDatabricksAuthExtension,
+                // END_HADRON
+                CreateAvailabilityCheck,
+                DropRoles,
+                // BEGIN_HADRON
+                AddDatabricksGrants,
+                CreateDatabricksMisc,
+                // END_HADRON
+            ]
+            } else {
+                vec![
                 HandleOtherExtensions,
                 HandleNeonExtension, // This step depends on CreateSchemaNeon
                 CreateAvailabilityCheck,
                 DropRoles,
-            ];
+            ]
+            };
 
             // This step depends on CreateSchemaNeon
             if spec.drop_subscriptions_before_start && !drop_subscriptions_done {
@@ -303,6 +342,7 @@ impl ComputeNode {
                     jwks_roles.clone(),
                     phase,
                     || async { Ok(&client) },
+                    self.params.lakebase_mode,
                 )
                 .await?;
             }
@@ -328,6 +368,7 @@ impl ComputeNode {
         concurrency_token: Arc<tokio::sync::Semaphore>,
         db: DB,
         subphases: Vec<PerDatabasePhase>,
+        lakebase_mode: bool,
     ) -> Result<()> {
         let _permit = concurrency_token.acquire().await?;
 
@@ -355,6 +396,7 @@ impl ComputeNode {
                     let client = client_conn.as_ref().unwrap();
                     Ok(client)
                 },
+                lakebase_mode,
             )
             .await?;
         }
@@ -477,6 +519,10 @@ pub enum PerDatabasePhase {
 #[derive(Clone, Debug)]
 pub enum ApplySpecPhase {
     CreatePrivilegedRole,
+    // BEGIN_HADRON
+    CreateDatabricksRoles,
+    AlterDatabricksRoles,
+    // END_HADRON
     DropInvalidDatabases,
     RenameRoles,
     CreateAndAlterRoles,
@@ -489,7 +535,14 @@ pub enum ApplySpecPhase {
     DisablePostgresDBPgAudit,
     HandleOtherExtensions,
     HandleNeonExtension,
+    // BEGIN_HADRON
+    HandleDatabricksAuthExtension,
+    // END_HADRON
     CreateAvailabilityCheck,
+    // BEGIN_HADRON
+    AddDatabricksGrants,
+    CreateDatabricksMisc,
+    // END_HADRON
     DropRoles,
     FinalizeDropLogicalSubscriptions,
 }
@@ -525,6 +578,7 @@ pub async fn apply_operations<'a, Fut, F>(
     jwks_roles: Arc<HashSet<String>>,
     apply_spec_phase: ApplySpecPhase,
     client: F,
+    lakebase_mode: bool,
 ) -> Result<()>
 where
     F: FnOnce() -> Fut,
@@ -571,6 +625,23 @@ where
                         },
                         query
                     );
+                    if !lakebase_mode {
+                        return res;
+                    }
+                    // BEGIN HADRON
+                    if let Err(e) = res.as_ref() {
+                        if let Some(sql_state) = e.code() {
+                            if sql_state.code() == "57014" {
+                                // SQL State 57014 (ERRCODE_QUERY_CANCELED) is used for statement timeouts.
+                                // Increment the counter whenever a statement timeout occurs. Timeouts on
+                                // this configuration path can only occur due to PS connectivity problems that
+                                // Postgres failed to recover from.
+                                COMPUTE_CONFIGURE_STATEMENT_TIMEOUT_ERRORS.inc();
+                            }
+                        }
+                    }
+                    // END HADRON
+
                     res
                 }
                 .instrument(inspan)
@@ -608,10 +679,44 @@ async fn get_operations<'a>(
         ApplySpecPhase::CreatePrivilegedRole => Ok(Box::new(once(Operation {
             query: format!(
                 include_str!("sql/create_privileged_role.sql"),
-                privileged_role_name = params.privileged_role_name
+                privileged_role_name = params.privileged_role_name,
+                privileges = if params.lakebase_mode {
+                    "CREATEDB CREATEROLE NOLOGIN BYPASSRLS"
+                } else {
+                    "CREATEDB CREATEROLE NOLOGIN REPLICATION BYPASSRLS"
+                }
             ),
             comment: None,
         }))),
+        // BEGIN_HADRON
+        // New Hadron phase
+        ApplySpecPhase::CreateDatabricksRoles => {
+            let queries = create_databricks_roles();
+            let operations = queries.into_iter().map(|query| Operation {
+                query,
+                comment: None,
+            });
+            Ok(Box::new(operations))
+        }
+
+        // Backfill existing databricks_reader_* roles with statement timeout from GUC
+        ApplySpecPhase::AlterDatabricksRoles => {
+            let query = String::from(include_str!(
+                "sql/alter_databricks_reader_roles_timeout.sql"
+            ));
+
+            let operations = once(Operation {
+                query,
+                comment: Some(
+                    "Backfill existing databricks_reader_* roles with statement timeout"
+                        .to_string(),
+                ),
+            });
+
+            Ok(Box::new(operations))
+        }
+        // End of new Hadron Phase
+        // END_HADRON
         ApplySpecPhase::DropInvalidDatabases => {
             let mut ctx = ctx.write().await;
             let databases = &mut ctx.dbs;
@@ -981,7 +1086,10 @@ async fn get_operations<'a>(
                                         // N.B. this has to be properly dollar-escaped with `pg_quote_dollar()`
                                         role_name = escaped_role,
                                         outer_tag = outer_tag,
-                                    ),
+                                    )
+                                    // HADRON change:
+                                    .replace("neon_superuser", &params.privileged_role_name),
+                                    // HADRON change end                                    ,
                                     comment: None,
                                 },
                                 // This now will only drop privileges of the role
@@ -1017,7 +1125,8 @@ async fn get_operations<'a>(
                             comment: None,
                         },
                         Operation {
-                            query: String::from(include_str!("sql/default_grants.sql")),
+                            query: String::from(include_str!("sql/default_grants.sql"))
+                                .replace("neon_superuser", &params.privileged_role_name),
                             comment: None,
                         },
                     ]
@@ -1086,6 +1195,28 @@ async fn get_operations<'a>(
 
             Ok(Box::new(operations))
         }
+        // BEGIN_HADRON
+        // Note: we may want to version the extension someday, but for now we just drop it and recreate it.
+        ApplySpecPhase::HandleDatabricksAuthExtension => {
+            let operations = vec![
+                Operation {
+                    query: String::from("DROP EXTENSION IF EXISTS databricks_auth"),
+                    comment: Some(String::from("dropping existing databricks_auth extension")),
+                },
+                Operation {
+                    query: String::from("CREATE EXTENSION databricks_auth"),
+                    comment: Some(String::from("creating databricks_auth extension")),
+                },
+                Operation {
+                    query: String::from("GRANT SELECT ON databricks_auth_metrics TO pg_monitor"),
+                    comment: Some(String::from("grant select on databricks auth counters")),
+                },
+            ]
+            .into_iter();
+
+            Ok(Box::new(operations))
+        }
+        // END_HADRON
         ApplySpecPhase::CreateAvailabilityCheck => Ok(Box::new(once(Operation {
             query: String::from(include_str!("sql/add_availabilitycheck_tables.sql")),
             comment: None,
@@ -1103,6 +1234,63 @@ async fn get_operations<'a>(
 
             Ok(Box::new(operations))
         }
+
+        // BEGIN_HADRON
+        // New Hadron phases
+        //
+        // Grants permissions to roles that are used by Databricks.
+        ApplySpecPhase::AddDatabricksGrants => {
+            let operations = vec![
+                Operation {
+                    query: String::from("GRANT USAGE ON SCHEMA neon TO databricks_monitor"),
+                    comment: Some(String::from(
+                        "Permissions needed to execute neon.* functions (in the postgres database)",
+                    )),
+                },
+                Operation {
+                    query: String::from(
+                        "GRANT SELECT, INSERT, UPDATE ON health_check TO databricks_monitor",
+                    ),
+                    comment: Some(String::from("Permissions needed for read and write probes")),
+                },
+                Operation {
+                    query: String::from(
+                        "GRANT EXECUTE ON FUNCTION pg_ls_dir(text) TO databricks_monitor",
+                    ),
+                    comment: Some(String::from(
+                        "Permissions needed to monitor .snap file counts",
+                    )),
+                },
+                Operation {
+                    query: String::from(
+                        "GRANT SELECT ON neon.neon_perf_counters TO databricks_monitor",
+                    ),
+                    comment: Some(String::from(
+                        "Permissions needed to access neon performance counters view",
+                    )),
+                },
+                Operation {
+                    query: String::from(
+                        "GRANT EXECUTE ON FUNCTION neon.get_perf_counters() TO databricks_monitor",
+                    ),
+                    comment: Some(String::from(
+                        "Permissions needed to execute the underlying performance counters function",
+                    )),
+                },
+            ]
+            .into_iter();
+
+            Ok(Box::new(operations))
+        }
+        // Creates minor objects that are used by Databricks.
+        ApplySpecPhase::CreateDatabricksMisc => Ok(Box::new(once(Operation {
+            query: String::from(include_str!("sql/create_databricks_misc.sql")),
+            comment: Some(String::from(
+                "The function databricks_monitor uses to convert exception to 0 or 1",
+            )),
+        }))),
+        // End of new Hadron phases
+        // END_HADRON
         ApplySpecPhase::FinalizeDropLogicalSubscriptions => Ok(Box::new(once(Operation {
             query: String::from(include_str!("sql/finalize_drop_subscriptions.sql")),
             comment: None,

compute_tools/src/sql/alter_databricks_reader_roles_timeout.sql

@@ -0,0 +1,25 @@
+DO $$
+DECLARE
+    reader_role RECORD;
+    timeout_value TEXT;
+BEGIN
+    -- Get the current GUC setting for reader statement timeout
+    SELECT current_setting('databricks.reader_statement_timeout', true) INTO timeout_value;
+    
+    -- Only proceed if timeout_value is not null/empty and not '0' (disabled)
+    IF timeout_value IS NOT NULL AND timeout_value != '' AND timeout_value != '0' THEN
+        -- Find all databricks_reader_* roles and update their statement_timeout
+        FOR reader_role IN 
+            SELECT r.rolname
+            FROM pg_roles r
+            WHERE r.rolname ~ '^databricks_reader_\d+$'
+        LOOP
+            -- Apply the timeout setting to the role (will overwrite existing setting)
+            EXECUTE format('ALTER ROLE %I SET statement_timeout = %L', 
+                         reader_role.rolname, timeout_value);
+            
+            RAISE LOG 'Updated statement_timeout = % for role %', timeout_value, reader_role.rolname;
+        END LOOP;
+    END IF;
+END
+$$;

compute_tools/src/sql/create_databricks_misc.sql

@@ -0,0 +1,15 @@
+ALTER ROLE databricks_monitor SET statement_timeout = '60s';
+
+CREATE OR REPLACE FUNCTION health_check_write_succeeds()
+RETURNS INTEGER AS $$
+BEGIN
+INSERT INTO health_check VALUES (1, now())
+ON CONFLICT (id) DO UPDATE
+    SET updated_at = now();
+
+RETURN 1;
+EXCEPTION WHEN OTHERS THEN
+RAISE EXCEPTION '[DATABRICKS_SMGR] health_check failed: [%] %', SQLSTATE, SQLERRM;
+RETURN 0;
+END;
+$$ LANGUAGE plpgsql;

compute_tools/src/sql/create_privileged_role.sql

@@ -2,7 +2,7 @@ DO $$
     BEGIN
         IF NOT EXISTS (SELECT FROM pg_catalog.pg_roles WHERE rolname = '{privileged_role_name}')
         THEN
-            CREATE ROLE {privileged_role_name} CREATEDB CREATEROLE NOLOGIN REPLICATION BYPASSRLS IN ROLE pg_read_all_data, pg_write_all_data;
+            CREATE ROLE {privileged_role_name} {privileges} IN ROLE pg_read_all_data, pg_write_all_data;
         END IF;
     END
 $$;

control_plane/src/bin/neon_local.rs

@@ -19,6 +19,9 @@ use compute_api::requests::ComputeClaimsScope;
 use compute_api::spec::{ComputeMode, PageserverProtocol};
 use control_plane::broker::StorageBroker;
 use control_plane::endpoint::{ComputeControlPlane, EndpointTerminateMode};
+use control_plane::endpoint::{
+    local_pageserver_conf_to_conn_info, tenant_locate_response_to_conn_info,
+};
 use control_plane::endpoint_storage::{ENDPOINT_STORAGE_DEFAULT_ADDR, EndpointStorage};
 use control_plane::local_env;
 use control_plane::local_env::{
@@ -44,7 +47,6 @@ 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, SafekeeperId};
 use safekeeper_api::{
     DEFAULT_HTTP_LISTEN_PORT as DEFAULT_SAFEKEEPER_HTTP_PORT,
@@ -52,7 +54,6 @@ use safekeeper_api::{
 };
 use storage_broker::DEFAULT_LISTEN_ADDR as DEFAULT_BROKER_ADDR;
 use tokio::task::JoinSet;
-use url::Host;
 use utils::auth::{Claims, Scope};
 use utils::id::{NodeId, TenantId, TenantTimelineId, TimelineId};
 use utils::lsn::Lsn;
@@ -1546,62 +1547,41 @@ async fn handle_endpoint(subcmd: &EndpointCmd, env: &local_env::LocalEnv) -> Res
                 )?;
             }
 
-            let (pageservers, stripe_size) = if let Some(pageserver_id) = pageserver_id {
-                let conf = env.get_pageserver_conf(pageserver_id).unwrap();
-                // Use gRPC if requested.
-                let pageserver = if endpoint.grpc {
-                    let grpc_addr = conf.listen_grpc_addr.as_ref().expect("bad config");
-                    let (host, port) = parse_host_port(grpc_addr)?;
-                    let port = port.unwrap_or(DEFAULT_PAGESERVER_GRPC_PORT);
-                    (PageserverProtocol::Grpc, host, port)
-                } else {
-                    let (host, port) = parse_host_port(&conf.listen_pg_addr)?;
-                    let port = port.unwrap_or(5432);
-                    (PageserverProtocol::Libpq, host, port)
-                };
-                // If caller is telling us what pageserver to use, this is not a tenant which is
-                // fully managed by storage controller, therefore not sharded.
-                (vec![pageserver], DEFAULT_STRIPE_SIZE)
+            let prefer_protocol = if endpoint.grpc {
+                PageserverProtocol::Grpc
+            } else {
+                PageserverProtocol::Libpq
+            };
+
+            let mut pageserver_conninfo = if let Some(ps_id) = pageserver_id {
+                let conf = env.get_pageserver_conf(ps_id).unwrap();
+                local_pageserver_conf_to_conn_info(conf)?
             } else {
                 // Look up the currently attached location of the tenant, and its striping metadata,
                 // to pass these on to postgres.
                 let storage_controller = StorageController::from_env(env);
                 let locate_result = storage_controller.tenant_locate(endpoint.tenant_id).await?;
-                let pageservers = futures::future::try_join_all(
-                    locate_result.shards.into_iter().map(|shard| async move {
-                        if let ComputeMode::Static(lsn) = endpoint.mode {
-                            // Initialize LSN leases for static computes.
+                assert!(!locate_result.shards.is_empty());
+
+                // Initialize LSN leases for static computes.
+                if let ComputeMode::Static(lsn) = endpoint.mode {
+                    futures::future::try_join_all(locate_result.shards.iter().map(
+                        |shard| async move {
                             let conf = env.get_pageserver_conf(shard.node_id).unwrap();
                             let pageserver = PageServerNode::from_env(env, conf);
 
                             pageserver
                                 .http_client
                                 .timeline_init_lsn_lease(shard.shard_id, endpoint.timeline_id, lsn)
-                                .await?;
-                        }
+                                .await
+                        },
+                    ))
+                    .await?;
+                }
 
-                        let pageserver = if endpoint.grpc {
-                            (
-                                PageserverProtocol::Grpc,
-                                Host::parse(&shard.listen_grpc_addr.expect("no gRPC address"))?,
-                                shard.listen_grpc_port.expect("no gRPC port"),
-                            )
-                        } else {
-                            (
-                                PageserverProtocol::Libpq,
-                                Host::parse(&shard.listen_pg_addr)?,
-                                shard.listen_pg_port,
-                            )
-                        };
-                        anyhow::Ok(pageserver)
-                    }),
-                )
-                .await?;
-                let stripe_size = locate_result.shard_params.stripe_size;
-
-                (pageservers, stripe_size)
+                tenant_locate_response_to_conn_info(&locate_result)?
             };
-            assert!(!pageservers.is_empty());
+            pageserver_conninfo.prefer_protocol = prefer_protocol;
 
             let ps_conf = env.get_pageserver_conf(DEFAULT_PAGESERVER_ID)?;
             let auth_token = if matches!(ps_conf.pg_auth_type, AuthType::NeonJWT) {
@@ -1631,9 +1611,8 @@ async fn handle_endpoint(subcmd: &EndpointCmd, env: &local_env::LocalEnv) -> Res
                 endpoint_storage_addr,
                 safekeepers_generation,
                 safekeepers,
-                pageservers,
+                pageserver_conninfo,
                 remote_ext_base_url: remote_ext_base_url.clone(),
-                shard_stripe_size: stripe_size.0 as usize,
                 create_test_user: args.create_test_user,
                 start_timeout: args.start_timeout,
                 autoprewarm: args.autoprewarm,
@@ -1650,37 +1629,29 @@ async fn handle_endpoint(subcmd: &EndpointCmd, env: &local_env::LocalEnv) -> Res
                 .endpoints
                 .get(endpoint_id.as_str())
                 .with_context(|| format!("postgres endpoint {endpoint_id} is not found"))?;
-            let pageservers = match args.pageserver_id {
+            let prefer_protocol = if endpoint.grpc {
+                PageserverProtocol::Grpc
+            } else {
+                PageserverProtocol::Libpq
+            };
+            let mut pageserver_conninfo = match args.pageserver_id {
                 Some(pageserver_id) => {
-                    let pageserver =
-                        PageServerNode::from_env(env, env.get_pageserver_conf(pageserver_id)?);
-
-                    vec![(
-                        PageserverProtocol::Libpq,
-                        pageserver.pg_connection_config.host().clone(),
-                        pageserver.pg_connection_config.port(),
-                    )]
+                    let conf = env.get_pageserver_conf(pageserver_id)?;
+                    local_pageserver_conf_to_conn_info(conf)?
                 }
                 None => {
                     let storage_controller = StorageController::from_env(env);
-                    storage_controller
-                        .tenant_locate(endpoint.tenant_id)
-                        .await?
-                        .shards
-                        .into_iter()
-                        .map(|shard| {
-                            (
-                                PageserverProtocol::Libpq,
-                                Host::parse(&shard.listen_pg_addr)
-                                    .expect("Storage controller reported malformed host"),
-                                shard.listen_pg_port,
-                            )
-                        })
-                        .collect::<Vec<_>>()
+                    let locate_result =
+                        storage_controller.tenant_locate(endpoint.tenant_id).await?;
+
+                    tenant_locate_response_to_conn_info(&locate_result)?
                 }
             };
+            pageserver_conninfo.prefer_protocol = prefer_protocol;
 
-            endpoint.update_pageservers_in_config(pageservers).await?;
+            endpoint
+                .update_pageservers_in_config(&pageserver_conninfo)
+                .await?;
         }
         EndpointCmd::Reconfigure(args) => {
             let endpoint_id = &args.endpoint_id;
@@ -1688,51 +1659,30 @@ async fn handle_endpoint(subcmd: &EndpointCmd, env: &local_env::LocalEnv) -> Res
                 .endpoints
                 .get(endpoint_id.as_str())
                 .with_context(|| format!("postgres endpoint {endpoint_id} is not found"))?;
-            let pageservers = if let Some(ps_id) = args.endpoint_pageserver_id {
-                let conf = env.get_pageserver_conf(ps_id)?;
-                // Use gRPC if requested.
-                let pageserver = if endpoint.grpc {
-                    let grpc_addr = conf.listen_grpc_addr.as_ref().expect("bad config");
-                    let (host, port) = parse_host_port(grpc_addr)?;
-                    let port = port.unwrap_or(DEFAULT_PAGESERVER_GRPC_PORT);
-                    (PageserverProtocol::Grpc, host, port)
-                } else {
-                    let (host, port) = parse_host_port(&conf.listen_pg_addr)?;
-                    let port = port.unwrap_or(5432);
-                    (PageserverProtocol::Libpq, host, port)
-                };
-                vec![pageserver]
+
+            let prefer_protocol = if endpoint.grpc {
+                PageserverProtocol::Grpc
             } else {
-                let storage_controller = StorageController::from_env(env);
-                storage_controller
-                    .tenant_locate(endpoint.tenant_id)
-                    .await?
-                    .shards
-                    .into_iter()
-                    .map(|shard| {
-                        // Use gRPC if requested.
-                        if endpoint.grpc {
-                            (
-                                PageserverProtocol::Grpc,
-                                Host::parse(&shard.listen_grpc_addr.expect("no gRPC address"))
-                                    .expect("bad hostname"),
-                                shard.listen_grpc_port.expect("no gRPC port"),
-                            )
-                        } else {
-                            (
-                                PageserverProtocol::Libpq,
-                                Host::parse(&shard.listen_pg_addr).expect("bad hostname"),
-                                shard.listen_pg_port,
-                            )
-                        }
-                    })
-                    .collect::<Vec<_>>()
+                PageserverProtocol::Libpq
             };
+            let mut pageserver_conninfo = if let Some(ps_id) = args.endpoint_pageserver_id {
+                let conf = env.get_pageserver_conf(ps_id)?;
+                local_pageserver_conf_to_conn_info(conf)?
+            } else {
+                // Look up the currently attached location of the tenant, and its striping metadata,
+                // to pass these on to postgres.
+                let storage_controller = StorageController::from_env(env);
+                let locate_result = storage_controller.tenant_locate(endpoint.tenant_id).await?;
+
+                tenant_locate_response_to_conn_info(&locate_result)?
+            };
+            pageserver_conninfo.prefer_protocol = prefer_protocol;
+
             // If --safekeepers argument is given, use only the listed
             // safekeeper nodes; otherwise all from the env.
             let safekeepers = parse_safekeepers(&args.safekeepers)?;
             endpoint
-                .reconfigure(Some(pageservers), None, safekeepers, None)
+                .reconfigure(Some(&pageserver_conninfo), safekeepers, None)
                 .await?;
         }
         EndpointCmd::RefreshConfiguration(args) => {

control_plane/src/endpoint.rs

@@ -37,7 +37,7 @@
 //!         <other PostgreSQL files>
 //! ```
 //!
-use std::collections::BTreeMap;
+use std::collections::{BTreeMap, HashMap};
 use std::fmt::Display;
 use std::net::{IpAddr, Ipv4Addr, SocketAddr, TcpStream};
 use std::path::PathBuf;
@@ -58,8 +58,12 @@ use compute_api::responses::{
 };
 use compute_api::spec::{
     Cluster, ComputeAudit, ComputeFeature, ComputeMode, ComputeSpec, Database, PageserverProtocol,
-    PgIdent, RemoteExtSpec, Role,
+    PageserverShardInfo, PgIdent, RemoteExtSpec, Role,
 };
+
+// re-export these, because they're used in the reconfigure() function
+pub use compute_api::spec::{PageserverConnectionInfo, PageserverShardConnectionInfo};
+
 use jsonwebtoken::jwk::{
     AlgorithmParameters, CommonParameters, EllipticCurve, Jwk, JwkSet, KeyAlgorithm, KeyOperations,
     OctetKeyPairParameters, OctetKeyPairType, PublicKeyUse,
@@ -74,9 +78,11 @@ use sha2::{Digest, Sha256};
 use spki::der::Decode;
 use spki::{SubjectPublicKeyInfo, SubjectPublicKeyInfoRef};
 use tracing::debug;
-use url::Host;
 use utils::id::{NodeId, TenantId, TimelineId};
-use utils::shard::ShardStripeSize;
+use utils::shard::{ShardCount, ShardIndex, ShardNumber};
+
+use pageserver_api::config::DEFAULT_GRPC_LISTEN_PORT as DEFAULT_PAGESERVER_GRPC_PORT;
+use postgres_connection::parse_host_port;
 
 use crate::local_env::LocalEnv;
 use crate::postgresql_conf::PostgresConf;
@@ -387,9 +393,8 @@ pub struct EndpointStartArgs {
     pub endpoint_storage_addr: String,
     pub safekeepers_generation: Option<SafekeeperGeneration>,
     pub safekeepers: Vec<NodeId>,
-    pub pageservers: Vec<(PageserverProtocol, Host, u16)>,
+    pub pageserver_conninfo: PageserverConnectionInfo,
     pub remote_ext_base_url: Option<String>,
-    pub shard_stripe_size: usize,
     pub create_test_user: bool,
     pub start_timeout: Duration,
     pub autoprewarm: bool,
@@ -662,14 +667,6 @@ impl Endpoint {
         }
     }
 
-    fn build_pageserver_connstr(pageservers: &[(PageserverProtocol, Host, u16)]) -> String {
-        pageservers
-            .iter()
-            .map(|(scheme, host, port)| format!("{scheme}://no_user@{host}:{port}"))
-            .collect::<Vec<_>>()
-            .join(",")
-    }
-
     /// Map safekeepers ids to the actual connection strings.
     fn build_safekeepers_connstrs(&self, sk_ids: Vec<NodeId>) -> Result<Vec<String>> {
         let mut safekeeper_connstrings = Vec::new();
@@ -715,9 +712,6 @@ impl Endpoint {
             std::fs::remove_dir_all(self.pgdata())?;
         }
 
-        let pageserver_connstring = Self::build_pageserver_connstr(&args.pageservers);
-        assert!(!pageserver_connstring.is_empty());
-
         let safekeeper_connstrings = self.build_safekeepers_connstrs(args.safekeepers)?;
 
         // check for file remote_extensions_spec.json
@@ -732,6 +726,44 @@ impl Endpoint {
             remote_extensions = None;
         };
 
+        // For the sake of backwards-compatibility, also fill in 'pageserver_connstring'
+        //
+        // XXX: I believe this is not really needed, except to make
+        // test_forward_compatibility happy.
+        //
+        // Use a closure so that we can conviniently return None in the middle of the
+        // loop.
+        let pageserver_connstring: Option<String> = (|| {
+            let num_shards = args.pageserver_conninfo.shard_count.count();
+            let mut connstrings = Vec::new();
+            for shard_no in 0..num_shards {
+                let shard_index = ShardIndex {
+                    shard_count: args.pageserver_conninfo.shard_count,
+                    shard_number: ShardNumber(shard_no),
+                };
+                let shard = args
+                    .pageserver_conninfo
+                    .shards
+                    .get(&shard_index)
+                    .ok_or_else(|| {
+                        anyhow!(
+                            "shard {} not found in pageserver_connection_info",
+                            shard_index
+                        )
+                    })?;
+                let pageserver = shard
+                    .pageservers
+                    .first()
+                    .ok_or(anyhow!("must have at least one pageserver"))?;
+                if let Some(libpq_url) = &pageserver.libpq_url {
+                    connstrings.push(libpq_url.clone());
+                } else {
+                    return Ok::<_, anyhow::Error>(None);
+                }
+            }
+            Ok(Some(connstrings.join(",")))
+        })()?;
+
         // Create config file
         let config = {
             let mut spec = ComputeSpec {
@@ -776,13 +808,14 @@ impl Endpoint {
                 branch_id: None,
                 endpoint_id: Some(self.endpoint_id.clone()),
                 mode: self.mode,
-                pageserver_connstring: Some(pageserver_connstring),
+                pageserver_connection_info: Some(args.pageserver_conninfo.clone()),
+                pageserver_connstring,
                 safekeepers_generation: args.safekeepers_generation.map(|g| g.into_inner()),
                 safekeeper_connstrings,
                 storage_auth_token: args.auth_token.clone(),
                 remote_extensions,
                 pgbouncer_settings: None,
-                shard_stripe_size: Some(args.shard_stripe_size),
+                shard_stripe_size: args.pageserver_conninfo.stripe_size, // redundant with pageserver_connection_info.stripe_size
                 local_proxy_config: None,
                 reconfigure_concurrency: self.reconfigure_concurrency,
                 drop_subscriptions_before_start: self.drop_subscriptions_before_start,
@@ -966,7 +999,7 @@ impl Endpoint {
     // Update the pageservers in the spec file of the endpoint. This is useful to test the spec refresh scenario.
     pub async fn update_pageservers_in_config(
         &self,
-        pageservers: Vec<(PageserverProtocol, Host, u16)>,
+        pageserver_conninfo: &PageserverConnectionInfo,
     ) -> Result<()> {
         let config_path = self.endpoint_path().join("config.json");
         let mut config: ComputeConfig = {
@@ -974,10 +1007,8 @@ impl Endpoint {
             serde_json::from_reader(file)?
         };
 
-        let pageserver_connstring = Self::build_pageserver_connstr(&pageservers);
-        assert!(!pageserver_connstring.is_empty());
         let mut spec = config.spec.unwrap();
-        spec.pageserver_connstring = Some(pageserver_connstring);
+        spec.pageserver_connection_info = Some(pageserver_conninfo.clone());
         config.spec = Some(spec);
 
         let file = std::fs::File::create(&config_path)?;
@@ -1020,8 +1051,7 @@ impl Endpoint {
 
     pub async fn reconfigure(
         &self,
-        pageservers: Option<Vec<(PageserverProtocol, Host, u16)>>,
-        stripe_size: Option<ShardStripeSize>,
+        pageserver_conninfo: Option<&PageserverConnectionInfo>,
         safekeepers: Option<Vec<NodeId>>,
         safekeeper_generation: Option<SafekeeperGeneration>,
     ) -> Result<()> {
@@ -1036,15 +1066,15 @@ impl Endpoint {
         let postgresql_conf = self.read_postgresql_conf()?;
         spec.cluster.postgresql_conf = Some(postgresql_conf);
 
-        // If pageservers are not specified, don't change them.
-        if let Some(pageservers) = pageservers {
-            anyhow::ensure!(!pageservers.is_empty(), "no pageservers provided");
-
-            let pageserver_connstr = Self::build_pageserver_connstr(&pageservers);
-            spec.pageserver_connstring = Some(pageserver_connstr);
-            if stripe_size.is_some() {
-                spec.shard_stripe_size = stripe_size.map(|s| s.0 as usize);
-            }
+        if let Some(pageserver_conninfo) = pageserver_conninfo {
+            // If pageservers are provided, we need to ensure that they are not empty.
+            // This is a requirement for the compute_ctl configuration.
+            anyhow::ensure!(
+                !pageserver_conninfo.shards.is_empty(),
+                "no pageservers provided"
+            );
+            spec.pageserver_connection_info = Some(pageserver_conninfo.clone());
+            spec.shard_stripe_size = pageserver_conninfo.stripe_size;
         }
 
         // If safekeepers are not specified, don't change them.
@@ -1093,11 +1123,9 @@ impl Endpoint {
 
     pub async fn reconfigure_pageservers(
         &self,
-        pageservers: Vec<(PageserverProtocol, Host, u16)>,
-        stripe_size: Option<ShardStripeSize>,
+        pageservers: &PageserverConnectionInfo,
     ) -> Result<()> {
-        self.reconfigure(Some(pageservers), stripe_size, None, None)
-            .await
+        self.reconfigure(Some(pageservers), None, None).await
     }
 
     pub async fn reconfigure_safekeepers(
@@ -1105,7 +1133,7 @@ impl Endpoint {
         safekeepers: Vec<NodeId>,
         generation: SafekeeperGeneration,
     ) -> Result<()> {
-        self.reconfigure(None, None, Some(safekeepers), Some(generation))
+        self.reconfigure(None, Some(safekeepers), Some(generation))
             .await
     }
 
@@ -1188,3 +1216,84 @@ impl Endpoint {
         )
     }
 }
+
+/// If caller is telling us what pageserver to use, this is not a tenant which is
+/// fully managed by storage controller, therefore not sharded.
+pub fn local_pageserver_conf_to_conn_info(
+    conf: &crate::local_env::PageServerConf,
+) -> Result<PageserverConnectionInfo> {
+    let libpq_url = {
+        let (host, port) = parse_host_port(&conf.listen_pg_addr)?;
+        let port = port.unwrap_or(5432);
+        Some(format!("postgres://no_user@{host}:{port}"))
+    };
+    let grpc_url = if let Some(grpc_addr) = &conf.listen_grpc_addr {
+        let (host, port) = parse_host_port(grpc_addr)?;
+        let port = port.unwrap_or(DEFAULT_PAGESERVER_GRPC_PORT);
+        Some(format!("grpc://no_user@{host}:{port}"))
+    } else {
+        None
+    };
+    let ps_conninfo = PageserverShardConnectionInfo {
+        id: Some(conf.id),
+        libpq_url,
+        grpc_url,
+    };
+
+    let shard_info = PageserverShardInfo {
+        pageservers: vec![ps_conninfo],
+    };
+
+    let shards: HashMap<_, _> = vec![(ShardIndex::unsharded(), shard_info)]
+        .into_iter()
+        .collect();
+    Ok(PageserverConnectionInfo {
+        shard_count: ShardCount::unsharded(),
+        stripe_size: None,
+        shards,
+        prefer_protocol: PageserverProtocol::default(),
+    })
+}
+
+pub fn tenant_locate_response_to_conn_info(
+    response: &pageserver_api::controller_api::TenantLocateResponse,
+) -> Result<PageserverConnectionInfo> {
+    let mut shards = HashMap::new();
+    for shard in response.shards.iter() {
+        tracing::info!("parsing {}", shard.listen_pg_addr);
+        let libpq_url = {
+            let host = &shard.listen_pg_addr;
+            let port = shard.listen_pg_port;
+            Some(format!("postgres://no_user@{host}:{port}"))
+        };
+        let grpc_url = if let Some(grpc_addr) = &shard.listen_grpc_addr {
+            let host = grpc_addr;
+            let port = shard.listen_grpc_port.expect("no gRPC port");
+            Some(format!("grpc://no_user@{host}:{port}"))
+        } else {
+            None
+        };
+
+        let shard_info = PageserverShardInfo {
+            pageservers: vec![PageserverShardConnectionInfo {
+                id: Some(shard.node_id),
+                libpq_url,
+                grpc_url,
+            }],
+        };
+
+        shards.insert(shard.shard_id.to_index(), shard_info);
+    }
+
+    let stripe_size = if response.shard_params.count.is_unsharded() {
+        None
+    } else {
+        Some(response.shard_params.stripe_size)
+    };
+    Ok(PageserverConnectionInfo {
+        shard_count: response.shard_params.count,
+        stripe_size,
+        shards,
+        prefer_protocol: PageserverProtocol::default(),
+    })
+}

libs/compute_api/src/spec.rs

@@ -12,8 +12,9 @@ use regex::Regex;
 use remote_storage::RemotePath;
 use serde::{Deserialize, Serialize};
 use url::Url;
-use utils::id::{TenantId, TimelineId};
+use utils::id::{NodeId, TenantId, TimelineId};
 use utils::lsn::Lsn;
+use utils::shard::{ShardCount, ShardIndex, ShardNumber, ShardStripeSize};
 
 use crate::responses::TlsConfig;
 
@@ -105,8 +106,27 @@ pub struct ComputeSpec {
     // updated to fill these fields, we can make these non optional.
     pub tenant_id: Option<TenantId>,
     pub timeline_id: Option<TimelineId>,
+
+    /// Pageserver information can be passed in three different ways:
+    /// 1. Here in `pageserver_connection_info`
+    /// 2. In the `pageserver_connstring` field.
+    /// 3. in `cluster.settings`.
+    ///
+    /// The goal is to use method 1. everywhere. But for backwards-compatibility with old
+    /// versions of the control plane, `compute_ctl` will check 2. and 3. if the
+    /// `pageserver_connection_info` field is missing.
+    ///
+    /// If both `pageserver_connection_info` and `pageserver_connstring`+`shard_stripe_size` are
+    /// given, they must contain the same information.
+    pub pageserver_connection_info: Option<PageserverConnectionInfo>,
+
     pub pageserver_connstring: Option<String>,
 
+    /// Stripe size for pageserver sharding, in pages. This is set together with the legacy
+    /// `pageserver_connstring` field. When the modern `pageserver_connection_info` field is used,
+    /// the stripe size is stored in `pageserver_connection_info.stripe_size` instead.
+    pub shard_stripe_size: Option<ShardStripeSize>,
+
     // More neon ids that we expose to the compute_ctl
     // and to postgres as neon extension GUCs.
     pub project_id: Option<String>,
@@ -139,10 +159,6 @@ pub struct ComputeSpec {
 
     pub pgbouncer_settings: Option<IndexMap<String, String>>,
 
-    // Stripe size for pageserver sharding, in pages
-    #[serde(default)]
-    pub shard_stripe_size: Option<usize>,
-
     /// Local Proxy configuration used for JWT authentication
     #[serde(default)]
     pub local_proxy_config: Option<LocalProxySpec>,
@@ -217,6 +233,140 @@ pub enum ComputeFeature {
     UnknownFeature,
 }
 
+#[derive(Clone, Debug, Deserialize, Serialize, Eq, PartialEq)]
+pub struct PageserverConnectionInfo {
+    /// NB: 0 for unsharded tenants, 1 for sharded tenants with 1 shard, following storage
+    pub shard_count: ShardCount,
+
+    /// INVARIANT: null if shard_count is 0, otherwise non-null and immutable
+    pub stripe_size: Option<ShardStripeSize>,
+
+    pub shards: HashMap<ShardIndex, PageserverShardInfo>,
+
+    /// If the compute supports both protocols, this indicates which one it should use.  The compute
+    /// may use other available protocols too, if it doesn't support the preferred one. The URL's
+    /// for the protocol specified here must be present for all shards, i.e. do not mark a protocol
+    /// as preferred if it cannot actually be used with all the pageservers.
+    #[serde(default)]
+    pub prefer_protocol: PageserverProtocol,
+}
+
+/// Extract PageserverConnectionInfo from a comma-separated list of libpq connection strings.
+///
+/// This is used for backwards-compatibility, to parse the legacy
+/// [ComputeSpec::pageserver_connstring] field, or the 'neon.pageserver_connstring' GUC. Nowadays,
+/// the 'pageserver_connection_info' field should be used instead.
+impl PageserverConnectionInfo {
+    pub fn from_connstr(
+        connstr: &str,
+        stripe_size: Option<ShardStripeSize>,
+    ) -> Result<PageserverConnectionInfo, anyhow::Error> {
+        let shard_infos: Vec<_> = connstr
+            .split(',')
+            .map(|connstr| PageserverShardInfo {
+                pageservers: vec![PageserverShardConnectionInfo {
+                    id: None,
+                    libpq_url: Some(connstr.to_string()),
+                    grpc_url: None,
+                }],
+            })
+            .collect();
+
+        match shard_infos.len() {
+            0 => anyhow::bail!("empty connection string"),
+            1 => {
+                // We assume that if there's only connection string, it means "unsharded",
+                // rather than a sharded system with just a single shard. The latter is
+                // possible in principle, but we never do it.
+                let shard_count = ShardCount::unsharded();
+                let only_shard = shard_infos.first().unwrap().clone();
+                let shards = vec![(ShardIndex::unsharded(), only_shard)];
+                Ok(PageserverConnectionInfo {
+                    shard_count,
+                    stripe_size: None,
+                    shards: shards.into_iter().collect(),
+                    prefer_protocol: PageserverProtocol::Libpq,
+                })
+            }
+            n => {
+                if stripe_size.is_none() {
+                    anyhow::bail!("{n} shards but no stripe_size");
+                }
+                let shard_count = ShardCount(n.try_into()?);
+                let shards = shard_infos
+                    .into_iter()
+                    .enumerate()
+                    .map(|(idx, shard_info)| {
+                        (
+                            ShardIndex {
+                                shard_count,
+                                shard_number: ShardNumber(
+                                    idx.try_into().expect("shard number fits in u8"),
+                                ),
+                            },
+                            shard_info,
+                        )
+                    })
+                    .collect();
+                Ok(PageserverConnectionInfo {
+                    shard_count,
+                    stripe_size,
+                    shards,
+                    prefer_protocol: PageserverProtocol::Libpq,
+                })
+            }
+        }
+    }
+
+    /// Convenience routine to get the connection string for a shard.
+    pub fn shard_url(
+        &self,
+        shard_number: ShardNumber,
+        protocol: PageserverProtocol,
+    ) -> anyhow::Result<&str> {
+        let shard_index = ShardIndex {
+            shard_number,
+            shard_count: self.shard_count,
+        };
+        let shard = self.shards.get(&shard_index).ok_or_else(|| {
+            anyhow::anyhow!("shard connection info missing for shard {}", shard_index)
+        })?;
+
+        // Just use the first pageserver in the list. That's good enough for this
+        // convenience routine; if you need more control, like round robin policy or
+        // failover support, roll your own. (As of this writing, we never have more than
+        // one pageserver per shard anyway, but that will change in the future.)
+        let pageserver = shard
+            .pageservers
+            .first()
+            .ok_or(anyhow::anyhow!("must have at least one pageserver"))?;
+
+        let result = match protocol {
+            PageserverProtocol::Grpc => pageserver
+                .grpc_url
+                .as_ref()
+                .ok_or(anyhow::anyhow!("no grpc_url for shard {shard_index}"))?,
+            PageserverProtocol::Libpq => pageserver
+                .libpq_url
+                .as_ref()
+                .ok_or(anyhow::anyhow!("no libpq_url for shard {shard_index}"))?,
+        };
+        Ok(result)
+    }
+}
+
+#[derive(Clone, Debug, Deserialize, Serialize, Eq, PartialEq)]
+pub struct PageserverShardInfo {
+    pub pageservers: Vec<PageserverShardConnectionInfo>,
+}
+
+#[derive(Clone, Debug, Deserialize, Serialize, Eq, PartialEq)]
+pub struct PageserverShardConnectionInfo {
+    pub id: Option<NodeId>,
+    pub libpq_url: Option<String>,
+    pub grpc_url: Option<String>,
+}
+
 #[derive(Clone, Debug, Default, Deserialize, Serialize)]
 pub struct RemoteExtSpec {
     pub public_extensions: Option<Vec<String>>,
@@ -334,6 +484,12 @@ impl ComputeMode {
     }
 }
 
+impl Display for ComputeMode {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        f.write_str(self.to_type_str())
+    }
+}
+
 /// Log level for audit logging
 #[derive(Clone, Debug, Default, Eq, PartialEq, Deserialize, Serialize)]
 pub enum ComputeAudit {
@@ -470,13 +626,15 @@ pub struct JwksSettings {
     pub jwt_audience: Option<String>,
 }
 
-/// Protocol used to connect to a Pageserver. Parsed from the connstring scheme.
-#[derive(Clone, Copy, Debug, Default, PartialEq, Eq)]
+/// Protocol used to connect to a Pageserver.
+#[derive(Clone, Copy, Debug, Default, Deserialize, Serialize, PartialEq, Eq)]
 pub enum PageserverProtocol {
     /// The original protocol based on libpq and COPY. Uses postgresql:// or postgres:// scheme.
     #[default]
+    #[serde(rename = "libpq")]
     Libpq,
     /// A newer, gRPC-based protocol. Uses grpc:// scheme.
+    #[serde(rename = "grpc")]
     Grpc,
 }
 

libs/postgres_ffi/Cargo.toml

@@ -9,10 +9,7 @@ regex.workspace = true
 bytes.workspace = true
 anyhow.workspace = true
 crc32c.workspace = true
-criterion.workspace = true
 once_cell.workspace = true
-log.workspace = true
-memoffset.workspace = true
 pprof.workspace = true
 thiserror.workspace = true
 serde.workspace = true
@@ -22,6 +19,7 @@ tracing.workspace = true
 postgres_versioninfo.workspace = true
 
 [dev-dependencies]
+criterion.workspace = true
 env_logger.workspace = true
 postgres.workspace = true
 

libs/postgres_ffi/src/controlfile_utils.rs

@@ -34,9 +34,8 @@ const SIZEOF_CONTROLDATA: usize = size_of::<ControlFileData>();
 impl ControlFileData {
     /// Compute the offset of the `crc` field within the `ControlFileData` struct.
     /// Equivalent to offsetof(ControlFileData, crc) in C.
-    // Someday this can be const when the right compiler features land.
-    fn pg_control_crc_offset() -> usize {
-        memoffset::offset_of!(ControlFileData, crc)
+    const fn pg_control_crc_offset() -> usize {
+        std::mem::offset_of!(ControlFileData, crc)
     }
 
     ///

libs/postgres_ffi/src/nonrelfile_utils.rs

@@ -4,12 +4,11 @@
 use crate::pg_constants;
 use crate::transaction_id_precedes;
 use bytes::BytesMut;
-use log::*;
 
 use super::bindings::MultiXactId;
 
 pub fn transaction_id_set_status(xid: u32, status: u8, page: &mut BytesMut) {
-    trace!(
+    tracing::trace!(
         "handle_apply_request for RM_XACT_ID-{} (1-commit, 2-abort, 3-sub_commit)",
         status
     );

libs/postgres_ffi/src/waldecoder_handler.rs

@@ -14,7 +14,6 @@ use super::xlog_utils::*;
 use crate::WAL_SEGMENT_SIZE;
 use bytes::{Buf, BufMut, Bytes, BytesMut};
 use crc32c::*;
-use log::*;
 use std::cmp::min;
 use std::num::NonZeroU32;
 use utils::lsn::Lsn;
@@ -236,7 +235,7 @@ impl WalStreamDecoderHandler for WalStreamDecoder {
         // XLOG_SWITCH records are special. If we see one, we need to skip
         // to the next WAL segment.
         let next_lsn = if xlogrec.is_xlog_switch_record() {
-            trace!("saw xlog switch record at {}", self.lsn);
+            tracing::trace!("saw xlog switch record at {}", self.lsn);
             self.lsn + self.lsn.calc_padding(WAL_SEGMENT_SIZE as u64)
         } else {
             // Pad to an 8-byte boundary

libs/postgres_ffi/src/xlog_utils.rs

@@ -23,8 +23,6 @@ use crate::{WAL_SEGMENT_SIZE, XLOG_BLCKSZ};
 use bytes::BytesMut;
 use bytes::{Buf, Bytes};
 
-use log::*;
-
 use serde::Serialize;
 use std::ffi::{CString, OsStr};
 use std::fs::File;
@@ -235,7 +233,7 @@ pub fn find_end_of_wal(
     let mut curr_lsn = start_lsn;
     let mut buf = [0u8; XLOG_BLCKSZ];
     let pg_version = MY_PGVERSION;
-    debug!("find_end_of_wal PG_VERSION: {}", pg_version);
+    tracing::debug!("find_end_of_wal PG_VERSION: {}", pg_version);
 
     let mut decoder = WalStreamDecoder::new(start_lsn, pg_version);
 
@@ -247,7 +245,7 @@ pub fn find_end_of_wal(
         match open_wal_segment(&seg_file_path)? {
             None => {
                 // no more segments
-                debug!(
+                tracing::debug!(
                     "find_end_of_wal reached end at {:?}, segment {:?} doesn't exist",
                     result, seg_file_path
                 );
@@ -260,7 +258,7 @@ pub fn find_end_of_wal(
                 while curr_lsn.segment_number(wal_seg_size) == segno {
                     let bytes_read = segment.read(&mut buf)?;
                     if bytes_read == 0 {
-                        debug!(
+                        tracing::debug!(
                             "find_end_of_wal reached end at {:?}, EOF in segment {:?} at offset {}",
                             result,
                             seg_file_path,
@@ -276,7 +274,7 @@ pub fn find_end_of_wal(
                         match decoder.poll_decode() {
                             Ok(Some(record)) => result = record.0,
                             Err(e) => {
-                                debug!(
+                                tracing::debug!(
                                     "find_end_of_wal reached end at {:?}, decode error: {:?}",
                                     result, e
                                 );

libs/proxy/tokio-postgres2/src/error/mod.rs

@@ -9,7 +9,7 @@ use postgres_protocol2::message::backend::{ErrorFields, ErrorResponseBody};
 pub use self::sqlstate::*;
 
 #[allow(clippy::unreadable_literal)]
-mod sqlstate;
+pub mod sqlstate;
 
 /// The severity of a Postgres error or notice.
 #[derive(Debug, Copy, Clone, PartialEq, Eq)]

libs/utils/src/logging.rs

@@ -34,13 +34,16 @@ macro_rules! critical {
 
 #[macro_export]
 macro_rules! critical_timeline {
-    ($tenant_shard_id:expr, $timeline_id:expr, $($arg:tt)*) => {{
+    ($tenant_shard_id:expr, $timeline_id:expr, $corruption_detected:expr, $($arg:tt)*) => {{
         if cfg!(debug_assertions) {
             panic!($($arg)*);
         }
         // Increment both metrics
         $crate::logging::TRACING_EVENT_COUNT_METRIC.inc_critical();
         $crate::logging::HADRON_CRITICAL_STORAGE_EVENT_COUNT_METRIC.inc(&$tenant_shard_id.to_string(), &$timeline_id.to_string());
+        if let Some(c) = $corruption_detected.as_ref() {
+            c.store(true, std::sync::atomic::Ordering::Relaxed);
+        }
         let backtrace = std::backtrace::Backtrace::capture();
         tracing::error!("CRITICAL: [tenant_shard_id: {}, timeline_id: {}] {}\n{backtrace}",
                        $tenant_shard_id, $timeline_id, format!($($arg)*));

libs/utils/src/pageserver_feedback.rs

@@ -32,6 +32,9 @@ pub struct PageserverFeedback {
     pub replytime: SystemTime,
     /// Used to track feedbacks from different shards. Always zero for unsharded tenants.
     pub shard_number: u32,
+    /// If true, the pageserver has detected corruption and the safekeeper and postgres
+    /// should stop sending WAL.
+    pub corruption_detected: bool,
 }
 
 impl PageserverFeedback {
@@ -43,6 +46,7 @@ impl PageserverFeedback {
             disk_consistent_lsn: Lsn::INVALID,
             replytime: *PG_EPOCH,
             shard_number: 0,
+            corruption_detected: false,
         }
     }
 
@@ -101,6 +105,13 @@ impl PageserverFeedback {
             buf.put_u32(self.shard_number);
         }
 
+        if self.corruption_detected {
+            nkeys += 1;
+            buf.put_slice(b"corruption_detected\0");
+            buf.put_i32(1);
+            buf.put_u8(1);
+        }
+
         buf[buf_ptr] = nkeys;
     }
 
@@ -147,6 +158,11 @@ impl PageserverFeedback {
                     assert_eq!(len, 4);
                     rf.shard_number = buf.get_u32();
                 }
+                b"corruption_detected" => {
+                    let len = buf.get_i32();
+                    assert_eq!(len, 1);
+                    rf.corruption_detected = buf.get_u8() != 0;
+                }
                 _ => {
                     let len = buf.get_i32();
                     warn!(
@@ -206,6 +222,26 @@ mod tests {
         assert_eq!(rf, rf_parsed);
     }
 
+    // Test that databricks-specific fields added to the PageserverFeedback message are serialized
+    // and deserialized correctly, in addition to the existing fields from upstream.
+    #[test]
+    fn test_replication_feedback_databricks_fields() {
+        let mut rf = PageserverFeedback::empty();
+        rf.current_timeline_size = 12345678;
+        rf.last_received_lsn = Lsn(23456789);
+        rf.disk_consistent_lsn = Lsn(34567890);
+        rf.remote_consistent_lsn = Lsn(45678901);
+        rf.replytime = *PG_EPOCH + Duration::from_secs(100_000_000);
+        rf.shard_number = 1;
+        rf.corruption_detected = true;
+
+        let mut data = BytesMut::new();
+        rf.serialize(&mut data);
+
+        let rf_parsed = PageserverFeedback::parse(data.freeze());
+        assert_eq!(rf, rf_parsed);
+    }
+
     #[test]
     fn test_replication_feedback_unknown_key() {
         let mut rf = PageserverFeedback::empty();

libs/utils/src/shard.rs

@@ -59,6 +59,10 @@ impl ShardCount {
     pub const MAX: Self = Self(u8::MAX);
     pub const MIN: Self = Self(0);
 
+    pub fn unsharded() -> Self {
+        ShardCount(0)
+    }
+
     /// The internal value of a ShardCount may be zero, which means "1 shard, but use
     /// legacy format for TenantShardId that excludes the shard suffix", also known
     /// as [`TenantShardId::unsharded`].

libs/walproposer/src/api_bindings.rs

@@ -426,6 +426,7 @@ pub fn empty_shmem() -> crate::bindings::WalproposerShmemState {
         remote_consistent_lsn: 0,
         replytime: 0,
         shard_number: 0,
+        corruption_detected: false,
     };
 
     let empty_wal_rate_limiter = crate::bindings::WalRateLimiter {

pageserver/client_grpc/src/client.rs

@@ -14,9 +14,9 @@ use utils::logging::warn_slow;
 
 use crate::pool::{ChannelPool, ClientGuard, ClientPool, StreamGuard, StreamPool};
 use crate::retry::Retry;
-use crate::split::GetPageSplitter;
 use compute_api::spec::PageserverProtocol;
 use pageserver_page_api as page_api;
+use pageserver_page_api::GetPageSplitter;
 use utils::id::{TenantId, TimelineId};
 use utils::shard::{ShardCount, ShardIndex, ShardNumber, ShardStripeSize};
 
@@ -230,16 +230,14 @@ impl PageserverClient {
     ) -> tonic::Result<page_api::GetPageResponse> {
         // Fast path: request is for a single shard.
         if let Some(shard_id) =
-            GetPageSplitter::for_single_shard(&req, shards.count, shards.stripe_size)
-                .map_err(|err| tonic::Status::internal(err.to_string()))?
+            GetPageSplitter::for_single_shard(&req, shards.count, shards.stripe_size)?
         {
             return Self::get_page_with_shard(req, shards.get(shard_id)?).await;
         }
 
         // Request spans multiple shards. Split it, dispatch concurrent per-shard requests, and
         // reassemble the responses.
-        let mut splitter = GetPageSplitter::split(req, shards.count, shards.stripe_size)
-            .map_err(|err| tonic::Status::internal(err.to_string()))?;
+        let mut splitter = GetPageSplitter::split(req, shards.count, shards.stripe_size)?;
 
         let mut shard_requests = FuturesUnordered::new();
         for (shard_id, shard_req) in splitter.drain_requests() {
@@ -249,14 +247,10 @@ impl PageserverClient {
         }
 
         while let Some((shard_id, shard_response)) = shard_requests.next().await.transpose()? {
-            splitter
-                .add_response(shard_id, shard_response)
-                .map_err(|err| tonic::Status::internal(err.to_string()))?;
+            splitter.add_response(shard_id, shard_response)?;
         }
 
-        splitter
-            .get_response()
-            .map_err(|err| tonic::Status::internal(err.to_string()))
+        Ok(splitter.collect_response()?)
     }
 
     /// Fetches pages on the given shard. Does not retry internally.

pageserver/client_grpc/src/lib.rs

@@ -1,6 +1,5 @@
 mod client;
 mod pool;
 mod retry;
-mod split;
 
 pub use client::{PageserverClient, ShardSpec};

pageserver/page_api/src/lib.rs

@@ -19,7 +19,9 @@ pub mod proto {
 }
 
 mod client;
-pub use client::Client;
 mod model;
+mod split;
 
+pub use client::Client;
 pub use model::*;
+pub use split::{GetPageSplitter, SplitError};

pageserver/page_api/src/split.rs

@@ -1,20 +1,19 @@
 use std::collections::HashMap;
 
-use anyhow::anyhow;
 use bytes::Bytes;
 
+use crate::model::*;
 use pageserver_api::key::rel_block_to_key;
 use pageserver_api::shard::key_to_shard_number;
-use pageserver_page_api as page_api;
 use utils::shard::{ShardCount, ShardIndex, ShardStripeSize};
 
 /// Splits GetPageRequests that straddle shard boundaries and assembles the responses.
 /// TODO: add tests for this.
 pub struct GetPageSplitter {
     /// Split requests by shard index.
-    requests: HashMap<ShardIndex, page_api::GetPageRequest>,
+    requests: HashMap<ShardIndex, GetPageRequest>,
     /// The response being assembled. Preallocated with empty pages, to be filled in.
-    response: page_api::GetPageResponse,
+    response: GetPageResponse,
     /// Maps the offset in `request.block_numbers` and `response.pages` to the owning shard. Used
     /// to assemble the response pages in the same order as the original request.
     block_shards: Vec<ShardIndex>,
@@ -24,22 +23,22 @@ impl GetPageSplitter {
     /// Checks if the given request only touches a single shard, and returns the shard ID. This is
     /// the common case, so we check first in order to avoid unnecessary allocations and overhead.
     pub fn for_single_shard(
-        req: &page_api::GetPageRequest,
+        req: &GetPageRequest,
         count: ShardCount,
         stripe_size: Option<ShardStripeSize>,
-    ) -> anyhow::Result<Option<ShardIndex>> {
+    ) -> Result<Option<ShardIndex>, SplitError> {
         // Fast path: unsharded tenant.
         if count.is_unsharded() {
             return Ok(Some(ShardIndex::unsharded()));
         }
 
         let Some(stripe_size) = stripe_size else {
-            return Err(anyhow!("stripe size must be given for sharded tenants"));
+            return Err("stripe size must be given for sharded tenants".into());
         };
 
         // Find the first page's shard, for comparison.
         let Some(&first_page) = req.block_numbers.first() else {
-            return Err(anyhow!("no block numbers in request"));
+            return Err("no block numbers in request".into());
         };
         let key = rel_block_to_key(req.rel, first_page);
         let shard_number = key_to_shard_number(count, stripe_size, &key);
@@ -57,10 +56,10 @@ impl GetPageSplitter {
 
     /// Splits the given request.
     pub fn split(
-        req: page_api::GetPageRequest,
+        req: GetPageRequest,
         count: ShardCount,
         stripe_size: Option<ShardStripeSize>,
-    ) -> anyhow::Result<Self> {
+    ) -> Result<Self, SplitError> {
         // The caller should make sure we don't split requests unnecessarily.
         debug_assert!(
             Self::for_single_shard(&req, count, stripe_size)?.is_none(),
@@ -68,10 +67,10 @@ impl GetPageSplitter {
         );
 
         if count.is_unsharded() {
-            return Err(anyhow!("unsharded tenant, no point in splitting request"));
+            return Err("unsharded tenant, no point in splitting request".into());
         }
         let Some(stripe_size) = stripe_size else {
-            return Err(anyhow!("stripe size must be given for sharded tenants"));
+            return Err("stripe size must be given for sharded tenants".into());
         };
 
         // Split the requests by shard index.
@@ -84,7 +83,7 @@ impl GetPageSplitter {
 
             requests
                 .entry(shard_id)
-                .or_insert_with(|| page_api::GetPageRequest {
+                .or_insert_with(|| GetPageRequest {
                     request_id: req.request_id,
                     request_class: req.request_class,
                     rel: req.rel,
@@ -98,16 +97,16 @@ impl GetPageSplitter {
 
         // Construct a response to be populated by shard responses. Preallocate empty page slots
         // with the expected block numbers.
-        let response = page_api::GetPageResponse {
+        let response = GetPageResponse {
             request_id: req.request_id,
-            status_code: page_api::GetPageStatusCode::Ok,
+            status_code: GetPageStatusCode::Ok,
             reason: None,
             rel: req.rel,
             pages: req
                 .block_numbers
                 .into_iter()
                 .map(|block_number| {
-                    page_api::Page {
+                    Page {
                         block_number,
                         image: Bytes::new(), // empty page slot to be filled in
                     }
@@ -123,43 +122,38 @@ impl GetPageSplitter {
     }
 
     /// Drains the per-shard requests, moving them out of the splitter to avoid extra allocations.
-    pub fn drain_requests(
-        &mut self,
-    ) -> impl Iterator<Item = (ShardIndex, page_api::GetPageRequest)> {
+    pub fn drain_requests(&mut self) -> impl Iterator<Item = (ShardIndex, GetPageRequest)> {
         self.requests.drain()
     }
 
     /// Adds a response from the given shard. The response must match the request ID and have an OK
     /// status code. A response must not already exist for the given shard ID.
-    #[allow(clippy::result_large_err)]
     pub fn add_response(
         &mut self,
         shard_id: ShardIndex,
-        response: page_api::GetPageResponse,
-    ) -> anyhow::Result<()> {
+        response: GetPageResponse,
+    ) -> Result<(), SplitError> {
         // The caller should already have converted status codes into tonic::Status.
-        if response.status_code != page_api::GetPageStatusCode::Ok {
-            return Err(anyhow!(
+        if response.status_code != GetPageStatusCode::Ok {
+            return Err(SplitError(format!(
                 "unexpected non-OK response for shard {shard_id}: {} {}",
                 response.status_code,
                 response.reason.unwrap_or_default()
-            ));
+            )));
         }
 
         if response.request_id != self.response.request_id {
-            return Err(anyhow!(
+            return Err(SplitError(format!(
                 "response ID mismatch for shard {shard_id}: expected {}, got {}",
-                self.response.request_id,
-                response.request_id
-            ));
+                self.response.request_id, response.request_id
+            )));
         }
 
         if response.request_id != self.response.request_id {
-            return Err(anyhow!(
+            return Err(SplitError(format!(
                 "response ID mismatch for shard {shard_id}: expected {}, got {}",
-                self.response.request_id,
-                response.request_id
-            ));
+                self.response.request_id, response.request_id
+            )));
         }
 
         // Place the shard response pages into the assembled response, in request order.
@@ -171,26 +165,27 @@ impl GetPageSplitter {
             }
 
             let Some(slot) = self.response.pages.get_mut(i) else {
-                return Err(anyhow!("no block_shards slot {i} for shard {shard_id}"));
+                return Err(SplitError(format!(
+                    "no block_shards slot {i} for shard {shard_id}"
+                )));
             };
             let Some(page) = pages.next() else {
-                return Err(anyhow!(
+                return Err(SplitError(format!(
                     "missing page {} in shard {shard_id} response",
                     slot.block_number
-                ));
+                )));
             };
             if page.block_number != slot.block_number {
-                return Err(anyhow!(
+                return Err(SplitError(format!(
                     "shard {shard_id} returned wrong page at index {i}, expected {} got {}",
-                    slot.block_number,
-                    page.block_number
-                ));
+                    slot.block_number, page.block_number
+                )));
             }
             if !slot.image.is_empty() {
-                return Err(anyhow!(
+                return Err(SplitError(format!(
                     "shard {shard_id} returned duplicate page {} at index {i}",
                     slot.block_number
-                ));
+                )));
             }
 
             *slot = page;
@@ -198,32 +193,54 @@ impl GetPageSplitter {
 
         // Make sure we've consumed all pages from the shard response.
         if let Some(extra_page) = pages.next() {
-            return Err(anyhow!(
+            return Err(SplitError(format!(
                 "shard {shard_id} returned extra page: {}",
                 extra_page.block_number
-            ));
+            )));
         }
 
         Ok(())
     }
 
-    /// Fetches the final, assembled response.
-    #[allow(clippy::result_large_err)]
-    pub fn get_response(self) -> anyhow::Result<page_api::GetPageResponse> {
+    /// Collects the final, assembled response.
+    pub fn collect_response(self) -> Result<GetPageResponse, SplitError> {
         // Check that the response is complete.
         for (i, page) in self.response.pages.iter().enumerate() {
             if page.image.is_empty() {
-                return Err(anyhow!(
+                return Err(SplitError(format!(
                     "missing page {} for shard {}",
                     page.block_number,
                     self.block_shards
                         .get(i)
                         .map(|s| s.to_string())
                         .unwrap_or_else(|| "?".to_string())
-                ));
+                )));
             }
         }
 
         Ok(self.response)
     }
 }
+
+/// A GetPageSplitter error.
+#[derive(Debug, thiserror::Error)]
+#[error("{0}")]
+pub struct SplitError(String);
+
+impl From<&str> for SplitError {
+    fn from(err: &str) -> Self {
+        SplitError(err.to_string())
+    }
+}
+
+impl From<String> for SplitError {
+    fn from(err: String) -> Self {
+        SplitError(err)
+    }
+}
+
+impl From<SplitError> for tonic::Status {
+    fn from(err: SplitError) -> Self {
+        tonic::Status::internal(err.0)
+    }
+}

pageserver/src/page_service.rs

@@ -16,7 +16,8 @@ use anyhow::{Context as _, bail};
 use bytes::{Buf as _, BufMut as _, BytesMut};
 use chrono::Utc;
 use futures::future::BoxFuture;
-use futures::{FutureExt, Stream};
+use futures::stream::FuturesUnordered;
+use futures::{FutureExt, Stream, StreamExt as _};
 use itertools::Itertools;
 use jsonwebtoken::TokenData;
 use once_cell::sync::OnceCell;
@@ -35,8 +36,8 @@ use pageserver_api::pagestream_api::{
 };
 use pageserver_api::reltag::SlruKind;
 use pageserver_api::shard::TenantShardId;
-use pageserver_page_api as page_api;
 use pageserver_page_api::proto;
+use pageserver_page_api::{self as page_api, GetPageSplitter};
 use postgres_backend::{
     AuthType, PostgresBackend, PostgresBackendReader, QueryError, is_expected_io_error,
 };
@@ -466,13 +467,6 @@ impl TimelineHandles {
         self.handles
             .get(timeline_id, shard_selector, &self.wrapper)
             .await
-            .map_err(|e| match e {
-                timeline::handle::GetError::TenantManager(e) => e,
-                timeline::handle::GetError::PerTimelineStateShutDown => {
-                    trace!("per-timeline state shut down");
-                    GetActiveTimelineError::Timeline(GetTimelineError::ShuttingDown)
-                }
-            })
     }
 
     fn tenant_id(&self) -> Option<TenantId> {
@@ -488,11 +482,9 @@ pub(crate) struct TenantManagerWrapper {
     tenant_id: once_cell::sync::OnceCell<TenantId>,
 }
 
-#[derive(Debug)]
 pub(crate) struct TenantManagerTypes;
 
 impl timeline::handle::Types for TenantManagerTypes {
-    type TenantManagerError = GetActiveTimelineError;
     type TenantManager = TenantManagerWrapper;
     type Timeline = TenantManagerCacheItem;
 }
@@ -3432,18 +3424,6 @@ impl GrpcPageServiceHandler {
         Ok(CancellableTask { task, cancel })
     }
 
-    /// Errors if the request is executed on a non-zero shard. Only shard 0 has a complete view of
-    /// relations and their sizes, as well as SLRU segments and similar data.
-    #[allow(clippy::result_large_err)]
-    fn ensure_shard_zero(timeline: &Handle<TenantManagerTypes>) -> Result<(), tonic::Status> {
-        match timeline.get_shard_index().shard_number.0 {
-            0 => Ok(()),
-            shard => Err(tonic::Status::invalid_argument(format!(
-                "request must execute on shard zero (is shard {shard})",
-            ))),
-        }
-    }
-
     /// Generates a PagestreamRequest header from a ReadLsn and request ID.
     fn make_hdr(
         read_lsn: page_api::ReadLsn,
@@ -3458,30 +3438,72 @@ impl GrpcPageServiceHandler {
         }
     }
 
-    /// Acquires a timeline handle for the given request.
+    /// Acquires a timeline handle for the given request. The shard index must match a local shard.
     ///
-    /// TODO: during shard splits, the compute may still be sending requests to the parent shard
-    /// until the entire split is committed and the compute is notified. Consider installing a
-    /// temporary shard router from the parent to the children while the split is in progress.
-    ///
-    /// TODO: consider moving this to a middleware layer; all requests need it. Needs to manage
-    /// the TimelineHandles lifecycle.
-    ///
-    /// TODO: untangle acquisition from TenantManagerWrapper::resolve() and Cache::get(), to avoid
-    /// the unnecessary overhead.
+    /// NB: this will fail during shard splits, see comment on [`Self::maybe_split_get_page`].
     async fn get_request_timeline(
         &self,
         req: &tonic::Request<impl Any>,
     ) -> Result<Handle<TenantManagerTypes>, GetActiveTimelineError> {
-        let ttid = *extract::<TenantTimelineId>(req);
+        let TenantTimelineId {
+            tenant_id,
+            timeline_id,
+        } = *extract::<TenantTimelineId>(req);
         let shard_index = *extract::<ShardIndex>(req);
-        let shard_selector = ShardSelector::Known(shard_index);
 
+        // TODO: untangle acquisition from TenantManagerWrapper::resolve() and Cache::get(), to
+        // avoid the unnecessary overhead.
         TimelineHandles::new(self.tenant_manager.clone())
-            .get(ttid.tenant_id, ttid.timeline_id, shard_selector)
+            .get(tenant_id, timeline_id, ShardSelector::Known(shard_index))
             .await
     }
 
+    /// Acquires a timeline handle for the given request, which must be for shard zero. Most
+    /// metadata requests are only valid on shard zero.
+    ///
+    /// NB: during an ongoing shard split, the compute will keep talking to the parent shard until
+    /// the split is committed, but the parent shard may have been removed in the meanwhile. In that
+    /// case, we reroute the request to the new child shard. See [`Self::maybe_split_get_page`].
+    ///
+    /// TODO: revamp the split protocol to avoid this child routing.
+    async fn get_request_timeline_shard_zero(
+        &self,
+        req: &tonic::Request<impl Any>,
+    ) -> Result<Handle<TenantManagerTypes>, tonic::Status> {
+        let TenantTimelineId {
+            tenant_id,
+            timeline_id,
+        } = *extract::<TenantTimelineId>(req);
+        let shard_index = *extract::<ShardIndex>(req);
+
+        if shard_index.shard_number.0 != 0 {
+            return Err(tonic::Status::invalid_argument(format!(
+                "request only valid on shard zero (requested shard {shard_index})",
+            )));
+        }
+
+        // TODO: untangle acquisition from TenantManagerWrapper::resolve() and Cache::get(), to
+        // avoid the unnecessary overhead.
+        let mut handles = TimelineHandles::new(self.tenant_manager.clone());
+        match handles
+            .get(tenant_id, timeline_id, ShardSelector::Known(shard_index))
+            .await
+        {
+            Ok(timeline) => Ok(timeline),
+            Err(err) => {
+                // We may be in the middle of a shard split. Try to find a child shard 0.
+                if let Ok(timeline) = handles
+                    .get(tenant_id, timeline_id, ShardSelector::Zero)
+                    .await
+                    && timeline.get_shard_index().shard_count > shard_index.shard_count
+                {
+                    return Ok(timeline);
+                }
+                Err(err.into())
+            }
+        }
+    }
+
     /// Starts a SmgrOpTimer at received_at, throttles the request, and records execution start.
     /// Only errors if the timeline is shutting down.
     ///
@@ -3511,28 +3533,22 @@ impl GrpcPageServiceHandler {
     /// TODO: get_vectored() currently enforces a batch limit of 32. Postgres will typically send
     /// batches up to effective_io_concurrency = 100. Either we have to accept large batches, or
     /// split them up in the client or server.
-    #[instrument(skip_all, fields(req_id, rel, blkno, blks, req_lsn, mod_lsn))]
+    #[instrument(skip_all, fields(
+        req_id = %req.request_id,
+        rel = %req.rel,
+        blkno = %req.block_numbers[0],
+        blks = %req.block_numbers.len(),
+        lsn = %req.read_lsn,
+    ))]
     async fn get_page(
         ctx: &RequestContext,
-        timeline: &WeakHandle<TenantManagerTypes>,
-        req: proto::GetPageRequest,
+        timeline: Handle<TenantManagerTypes>,
+        req: page_api::GetPageRequest,
         io_concurrency: IoConcurrency,
-    ) -> Result<proto::GetPageResponse, tonic::Status> {
-        let received_at = Instant::now();
-        let timeline = timeline.upgrade()?;
+        received_at: Instant,
+    ) -> Result<page_api::GetPageResponse, tonic::Status> {
         let ctx = ctx.with_scope_page_service_pagestream(&timeline);
 
-        // Validate the request, decorate the span, and convert it to a Pagestream request.
-        let req = page_api::GetPageRequest::try_from(req)?;
-
-        span_record!(
-            req_id = %req.request_id,
-            rel = %req.rel,
-            blkno = %req.block_numbers[0],
-            blks = %req.block_numbers.len(),
-            lsn = %req.read_lsn,
-        );
-
         for &blkno in &req.block_numbers {
             let shard = timeline.get_shard_identity();
             let key = rel_block_to_key(req.rel, blkno);
@@ -3620,7 +3636,89 @@ impl GrpcPageServiceHandler {
             };
         }
 
-        Ok(resp.into())
+        Ok(resp)
+    }
+
+    /// Processes a GetPage request when there is a potential shard split in progress. We have to
+    /// reroute the request to any local child shards, and split batch requests that straddle
+    /// multiple child shards.
+    ///
+    /// Parent shards are split and removed incrementally (there may be many parent shards when
+    /// splitting an already-sharded tenant), but the compute is only notified once the overall
+    /// split commits, which can take several minutes. In the meanwhile, the compute will be sending
+    /// requests to the parent shards.
+    ///
+    /// TODO: add test infrastructure to provoke this situation frequently and for long periods of
+    /// time, to properly exercise it.
+    ///
+    /// TODO: revamp the split protocol to avoid this, e.g.:
+    /// * Keep the parent shard until the split commits and the compute is notified.
+    /// * Notify the compute about each subsplit.
+    /// * Return an error that updates the compute's shard map.
+    #[instrument(skip_all)]
+    #[allow(clippy::too_many_arguments)]
+    async fn maybe_split_get_page(
+        ctx: &RequestContext,
+        handles: &mut TimelineHandles,
+        tenant_id: TenantId,
+        timeline_id: TimelineId,
+        parent: ShardIndex,
+        req: page_api::GetPageRequest,
+        io_concurrency: IoConcurrency,
+        received_at: Instant,
+    ) -> Result<page_api::GetPageResponse, tonic::Status> {
+        // Check the first page to see if we have any child shards at all. Otherwise, the compute is
+        // just talking to the wrong Pageserver. If the parent has been split, the shard now owning
+        // the page must have a higher shard count.
+        let timeline = handles
+            .get(
+                tenant_id,
+                timeline_id,
+                ShardSelector::Page(rel_block_to_key(req.rel, req.block_numbers[0])),
+            )
+            .await?;
+
+        let shard_id = timeline.get_shard_identity();
+        if shard_id.count <= parent.shard_count {
+            return Err(HandleUpgradeError::ShutDown.into()); // emulate original error
+        }
+
+        // Fast path: the request fits in a single shard.
+        if let Some(shard_index) =
+            GetPageSplitter::for_single_shard(&req, shard_id.count, Some(shard_id.stripe_size))?
+        {
+            // We got the shard ID from the first page, so these must be equal.
+            assert_eq!(shard_index.shard_number, shard_id.number);
+            assert_eq!(shard_index.shard_count, shard_id.count);
+            return Self::get_page(ctx, timeline, req, io_concurrency, received_at).await;
+        }
+
+        // The request spans multiple shards; split it and dispatch parallel requests. All pages
+        // were originally in the parent shard, and during a split all children are local, so we
+        // expect to find local shards for all pages.
+        let mut splitter = GetPageSplitter::split(req, shard_id.count, Some(shard_id.stripe_size))?;
+
+        let mut shard_requests = FuturesUnordered::new();
+        for (shard_index, shard_req) in splitter.drain_requests() {
+            let timeline = handles
+                .get(tenant_id, timeline_id, ShardSelector::Known(shard_index))
+                .await?;
+            let future = Self::get_page(
+                ctx,
+                timeline,
+                shard_req,
+                io_concurrency.clone(),
+                received_at,
+            )
+            .map(move |result| result.map(|resp| (shard_index, resp)));
+            shard_requests.push(future);
+        }
+
+        while let Some((shard_index, shard_response)) = shard_requests.next().await.transpose()? {
+            splitter.add_response(shard_index, shard_response)?;
+        }
+
+        Ok(splitter.collect_response()?)
     }
 }
 
@@ -3649,11 +3747,10 @@ impl proto::PageService for GrpcPageServiceHandler {
         // to be the sweet spot where throughput is saturated.
         const CHUNK_SIZE: usize = 256 * 1024;
 
-        let timeline = self.get_request_timeline(&req).await?;
+        let timeline = self.get_request_timeline_shard_zero(&req).await?;
         let ctx = self.ctx.with_scope_timeline(&timeline);
 
         // Validate the request and decorate the span.
-        Self::ensure_shard_zero(&timeline)?;
         if timeline.is_archived() == Some(true) {
             return Err(tonic::Status::failed_precondition("timeline is archived"));
         }
@@ -3769,11 +3866,10 @@ impl proto::PageService for GrpcPageServiceHandler {
         req: tonic::Request<proto::GetDbSizeRequest>,
     ) -> Result<tonic::Response<proto::GetDbSizeResponse>, tonic::Status> {
         let received_at = extract::<ReceivedAt>(&req).0;
-        let timeline = self.get_request_timeline(&req).await?;
+        let timeline = self.get_request_timeline_shard_zero(&req).await?;
         let ctx = self.ctx.with_scope_page_service_pagestream(&timeline);
 
         // Validate the request, decorate the span, and convert it to a Pagestream request.
-        Self::ensure_shard_zero(&timeline)?;
         let req: page_api::GetDbSizeRequest = req.into_inner().try_into()?;
 
         span_record!(db_oid=%req.db_oid, lsn=%req.read_lsn);
@@ -3802,14 +3898,29 @@ impl proto::PageService for GrpcPageServiceHandler {
         req: tonic::Request<tonic::Streaming<proto::GetPageRequest>>,
     ) -> Result<tonic::Response<Self::GetPagesStream>, tonic::Status> {
         // Extract the timeline from the request and check that it exists.
-        let ttid = *extract::<TenantTimelineId>(&req);
+        //
+        // NB: during shard splits, the compute may still send requests to the parent shard. We'll
+        // reroute requests to the child shards below, but we also detect the common cases here
+        // where either the shard exists or no shards exist at all. If we have a child shard, we
+        // can't acquire a weak handle because we don't know which child shard to use yet.
+        let TenantTimelineId {
+            tenant_id,
+            timeline_id,
+        } = *extract::<TenantTimelineId>(&req);
         let shard_index = *extract::<ShardIndex>(&req);
-        let shard_selector = ShardSelector::Known(shard_index);
 
         let mut handles = TimelineHandles::new(self.tenant_manager.clone());
-        handles
-            .get(ttid.tenant_id, ttid.timeline_id, shard_selector)
-            .await?;
+        let timeline = match handles
+            .get(tenant_id, timeline_id, ShardSelector::Known(shard_index))
+            .await
+        {
+            // The timeline shard exists. Keep a weak handle to reuse for each request.
+            Ok(timeline) => Some(timeline.downgrade()),
+            // The shard doesn't exist, but a child shard does. We'll reroute requests later.
+            Err(_) if self.tenant_manager.has_child_shard(tenant_id, shard_index) => None,
+            // Failed to fetch the timeline, and no child shard exists. Error out.
+            Err(err) => return Err(err.into()),
+        };
 
         // Spawn an IoConcurrency sidecar, if enabled.
         let gate_guard = self
@@ -3826,11 +3937,9 @@ impl proto::PageService for GrpcPageServiceHandler {
         let mut reqs = req.into_inner();
 
         let resps = async_stream::try_stream! {
-            let timeline = handles
-                .get(ttid.tenant_id, ttid.timeline_id, shard_selector)
-                .await?
-                .downgrade();
             loop {
+                // Wait for the next client request.
+                //
                 // NB: Tonic considers the entire stream to be an in-flight request and will wait
                 // for it to complete before shutting down. React to cancellation between requests.
                 let req = tokio::select! {
@@ -3843,16 +3952,44 @@ impl proto::PageService for GrpcPageServiceHandler {
                         Err(err) => Err(err),
                     },
                 }?;
+
+                let received_at = Instant::now();
                 let req_id = req.request_id.map(page_api::RequestID::from).unwrap_or_default();
-                let result = Self::get_page(&ctx, &timeline, req, io_concurrency.clone())
+
+                // Process the request, using a closure to capture errors.
+                let process_request = async || {
+                    let req = page_api::GetPageRequest::try_from(req)?;
+
+                    // Fast path: use the pre-acquired timeline handle.
+                    if let Some(Ok(timeline)) = timeline.as_ref().map(|t| t.upgrade()) {
+                        return Self::get_page(&ctx, timeline, req, io_concurrency.clone(), received_at)
+                            .instrument(span.clone()) // propagate request span
+                            .await
+                    }
+
+                    // The timeline handle is stale. During shard splits, the compute may still be
+                    // sending requests to the parent shard. Try to re-route requests to the child
+                    // shards, and split any batch requests that straddle multiple child shards.
+                    Self::maybe_split_get_page(
+                        &ctx,
+                        &mut handles,
+                        tenant_id,
+                        timeline_id,
+                        shard_index,
+                        req,
+                        io_concurrency.clone(),
+                        received_at,
+                    )
                     .instrument(span.clone()) // propagate request span
-                    .await;
-                yield match result {
-                    Ok(resp) => resp,
-                    // Convert per-request errors to GetPageResponses as appropriate, or terminate
-                    // the stream with a tonic::Status. Log the error regardless, since
-                    // ObservabilityLayer can't automatically log stream errors.
+                    .await
+                };
+
+                // Return the response. Convert per-request errors to GetPageResponses if
+                // appropriate, or terminate the stream with a tonic::Status.
+                yield match process_request().await {
+                    Ok(resp) => resp.into(),
                     Err(status) => {
+                        // Log the error, since ObservabilityLayer won't see stream errors.
                         // TODO: it would be nice if we could propagate the get_page() fields here.
                         span.in_scope(|| {
                             warn!("request failed with {:?}: {}", status.code(), status.message());
@@ -3872,11 +4009,10 @@ impl proto::PageService for GrpcPageServiceHandler {
         req: tonic::Request<proto::GetRelSizeRequest>,
     ) -> Result<tonic::Response<proto::GetRelSizeResponse>, tonic::Status> {
         let received_at = extract::<ReceivedAt>(&req).0;
-        let timeline = self.get_request_timeline(&req).await?;
+        let timeline = self.get_request_timeline_shard_zero(&req).await?;
         let ctx = self.ctx.with_scope_page_service_pagestream(&timeline);
 
         // Validate the request, decorate the span, and convert it to a Pagestream request.
-        Self::ensure_shard_zero(&timeline)?;
         let req: page_api::GetRelSizeRequest = req.into_inner().try_into()?;
         let allow_missing = req.allow_missing;
 
@@ -3909,11 +4045,10 @@ impl proto::PageService for GrpcPageServiceHandler {
         req: tonic::Request<proto::GetSlruSegmentRequest>,
     ) -> Result<tonic::Response<proto::GetSlruSegmentResponse>, tonic::Status> {
         let received_at = extract::<ReceivedAt>(&req).0;
-        let timeline = self.get_request_timeline(&req).await?;
+        let timeline = self.get_request_timeline_shard_zero(&req).await?;
         let ctx = self.ctx.with_scope_page_service_pagestream(&timeline);
 
         // Validate the request, decorate the span, and convert it to a Pagestream request.
-        Self::ensure_shard_zero(&timeline)?;
         let req: page_api::GetSlruSegmentRequest = req.into_inner().try_into()?;
 
         span_record!(kind=%req.kind, segno=%req.segno, lsn=%req.read_lsn);
@@ -3943,6 +4078,10 @@ impl proto::PageService for GrpcPageServiceHandler {
         &self,
         req: tonic::Request<proto::LeaseLsnRequest>,
     ) -> Result<tonic::Response<proto::LeaseLsnResponse>, tonic::Status> {
+        // TODO: this won't work during shard splits, as the request is directed at a specific shard
+        // but the parent shard is removed before the split commits and the compute is notified
+        // (which can take several minutes for large tenants). That's also the case for the libpq
+        // implementation, so we keep the behavior for now.
         let timeline = self.get_request_timeline(&req).await?;
         let ctx = self.ctx.with_scope_timeline(&timeline);
 

pageserver/src/tenant/mgr.rs

@@ -826,6 +826,18 @@ impl TenantManager {
         peek_slot.is_some()
     }
 
+    /// Returns whether a local shard exists that's a child of the given tenant shard. Note that
+    /// this just checks for any shard with a larger shard count, and it may not be a direct child
+    /// of the given shard (their keyspace may not overlap).
+    pub(crate) fn has_child_shard(&self, tenant_id: TenantId, shard_index: ShardIndex) -> bool {
+        match &*self.tenants.read().unwrap() {
+            TenantsMap::Initializing => false,
+            TenantsMap::Open(slots) | TenantsMap::ShuttingDown(slots) => slots
+                .range(TenantShardId::tenant_range(tenant_id))
+                .any(|(tsid, _)| tsid.shard_count > shard_index.shard_count),
+        }
+    }
+
     #[instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug()))]
     pub(crate) async fn upsert_location(
         &self,
@@ -1522,6 +1534,13 @@ impl TenantManager {
         self.resources.deletion_queue_client.flush_advisory();
 
         // Phase 2: Put the parent shard to InProgress and grab a reference to the parent Tenant
+        //
+        // TODO: keeping the parent as InProgress while spawning the children causes read
+        // unavailability, as we can't acquire a new timeline handle for it (existing handles appear
+        // to still work though, even downgraded ones). The parent should be available for reads
+        // until the children are ready -- potentially until *all* subsplits across all parent
+        // shards are complete and the compute has been notified. See:
+        // <https://databricks.atlassian.net/browse/LKB-672>.
         drop(tenant);
         let mut parent_slot_guard =
             self.tenant_map_acquire_slot(&tenant_shard_id, TenantSlotAcquireMode::Any)?;

pageserver/src/tenant/timeline.rs

@@ -397,6 +397,11 @@ pub struct Timeline {
     /// If true, the last compaction failed.
     compaction_failed: AtomicBool,
 
+    /// Begin Hadron: If true, the pageserver has likely detected data corruption in the timeline.
+    /// We need to feed this information back to the Safekeeper and postgres for them to take the
+    /// appropriate action.
+    corruption_detected: AtomicBool,
+
     /// Notifies the tenant compaction loop that there is pending L0 compaction work.
     l0_compaction_trigger: Arc<Notify>,
 
@@ -3310,6 +3315,7 @@ impl Timeline {
 
                 compaction_lock: tokio::sync::Mutex::default(),
                 compaction_failed: AtomicBool::default(),
+                corruption_detected: AtomicBool::default(),
                 l0_compaction_trigger: resources.l0_compaction_trigger,
                 gc_lock: tokio::sync::Mutex::default(),
 
@@ -6004,6 +6010,17 @@ impl Timeline {
                 )))
             });
 
+            // Begin Hadron
+            //
+            fail_point!("create-image-layer-fail-simulated-corruption", |_| {
+                self.corruption_detected
+                    .store(true, std::sync::atomic::Ordering::Relaxed);
+                Err(CreateImageLayersError::Other(anyhow::anyhow!(
+                    "failpoint create-image-layer-fail-simulated-corruption"
+                )))
+            });
+            // End Hadron
+
             let io_concurrency = IoConcurrency::spawn_from_conf(
                 self.conf.get_vectored_concurrent_io,
                 self.gate
@@ -7149,6 +7166,7 @@ impl Timeline {
                             critical_timeline!(
                                 self.tenant_shard_id,
                                 self.timeline_id,
+                                Some(&self.corruption_detected),
                                 "walredo failure during page reconstruction: {err:?}"
                             );
                         }

pageserver/src/tenant/timeline/compaction.rs

@@ -1397,6 +1397,7 @@ impl Timeline {
                             critical_timeline!(
                                 self.tenant_shard_id,
                                 self.timeline_id,
+                                Some(&self.corruption_detected),
                                 "missing key during compaction: {err:?}"
                             );
                         }
@@ -1441,6 +1442,7 @@ impl Timeline {
                 critical_timeline!(
                     self.tenant_shard_id,
                     self.timeline_id,
+                    Some(&self.corruption_detected),
                     "could not compact, repartitioning keyspace failed: {e:?}"
                 );
             }

pageserver/src/tenant/timeline/handle.rs

@@ -224,11 +224,11 @@ use tracing::{instrument, trace};
 use utils::id::TimelineId;
 use utils::shard::{ShardIndex, ShardNumber};
 
-use crate::tenant::mgr::ShardSelector;
+use crate::page_service::GetActiveTimelineError;
+use crate::tenant::GetTimelineError;
+use crate::tenant::mgr::{GetActiveTenantError, ShardSelector};
 
-/// The requirement for Debug is so that #[derive(Debug)] works in some places.
-pub(crate) trait Types: Sized + std::fmt::Debug {
-    type TenantManagerError: Sized + std::fmt::Debug;
+pub(crate) trait Types: Sized {
     type TenantManager: TenantManager<Self> + Sized;
     type Timeline: Timeline<Self> + Sized;
 }
@@ -307,12 +307,11 @@ impl<T: Types> Default for PerTimelineState<T> {
 /// Abstract view of [`crate::tenant::mgr`], for testability.
 pub(crate) trait TenantManager<T: Types> {
     /// Invoked by [`Cache::get`] to resolve a [`ShardTimelineId`] to a [`Types::Timeline`].
-    /// Errors are returned as [`GetError::TenantManager`].
     async fn resolve(
         &self,
         timeline_id: TimelineId,
         shard_selector: ShardSelector,
-    ) -> Result<T::Timeline, T::TenantManagerError>;
+    ) -> Result<T::Timeline, GetActiveTimelineError>;
 }
 
 /// Abstract view of an [`Arc<Timeline>`], for testability.
@@ -322,13 +321,6 @@ pub(crate) trait Timeline<T: Types> {
     fn per_timeline_state(&self) -> &PerTimelineState<T>;
 }
 
-/// Errors returned by [`Cache::get`].
-#[derive(Debug)]
-pub(crate) enum GetError<T: Types> {
-    TenantManager(T::TenantManagerError),
-    PerTimelineStateShutDown,
-}
-
 /// Internal type used in [`Cache::get`].
 enum RoutingResult<T: Types> {
     FastPath(Handle<T>),
@@ -345,7 +337,7 @@ impl<T: Types> Cache<T> {
         timeline_id: TimelineId,
         shard_selector: ShardSelector,
         tenant_manager: &T::TenantManager,
-    ) -> Result<Handle<T>, GetError<T>> {
+    ) -> Result<Handle<T>, GetActiveTimelineError> {
         const GET_MAX_RETRIES: usize = 10;
         const RETRY_BACKOFF: Duration = Duration::from_millis(100);
         let mut attempt = 0;
@@ -356,7 +348,11 @@ impl<T: Types> Cache<T> {
                 .await
             {
                 Ok(handle) => return Ok(handle),
-                Err(e) => {
+                Err(
+                    e @ GetActiveTimelineError::Tenant(GetActiveTenantError::WaitForActiveTimeout {
+                        ..
+                    }),
+                ) => {
                     // Retry on tenant manager error to handle tenant split more gracefully
                     if attempt < GET_MAX_RETRIES {
                         tokio::time::sleep(RETRY_BACKOFF).await;
@@ -370,6 +366,7 @@ impl<T: Types> Cache<T> {
                         return Err(e);
                     }
                 }
+                Err(err) => return Err(err),
             }
         }
     }
@@ -388,7 +385,7 @@ impl<T: Types> Cache<T> {
         timeline_id: TimelineId,
         shard_selector: ShardSelector,
         tenant_manager: &T::TenantManager,
-    ) -> Result<Handle<T>, GetError<T>> {
+    ) -> Result<Handle<T>, GetActiveTimelineError> {
         // terminates because when every iteration we remove an element from the map
         let miss: ShardSelector = loop {
             let routing_state = self.shard_routing(timeline_id, shard_selector);
@@ -468,60 +465,50 @@ impl<T: Types> Cache<T> {
         timeline_id: TimelineId,
         shard_selector: ShardSelector,
         tenant_manager: &T::TenantManager,
-    ) -> Result<Handle<T>, GetError<T>> {
-        match tenant_manager.resolve(timeline_id, shard_selector).await {
-            Ok(timeline) => {
-                let key = timeline.shard_timeline_id();
-                match &shard_selector {
-                    ShardSelector::Zero => assert_eq!(key.shard_index.shard_number, ShardNumber(0)),
-                    ShardSelector::Page(_) => (), // gotta trust tenant_manager
-                    ShardSelector::Known(idx) => assert_eq!(idx, &key.shard_index),
-                }
-
-                trace!("creating new HandleInner");
-                let timeline = Arc::new(timeline);
-                let handle_inner_arc =
-                    Arc::new(Mutex::new(HandleInner::Open(Arc::clone(&timeline))));
-                let handle_weak = WeakHandle {
-                    inner: Arc::downgrade(&handle_inner_arc),
-                };
-                let handle = handle_weak
-                    .upgrade()
-                    .ok()
-                    .expect("we just created it and it's not linked anywhere yet");
-                {
-                    let mut lock_guard = timeline
-                        .per_timeline_state()
-                        .handles
-                        .lock()
-                        .expect("mutex poisoned");
-                    match &mut *lock_guard {
-                        Some(per_timeline_state) => {
-                            let replaced =
-                                per_timeline_state.insert(self.id, Arc::clone(&handle_inner_arc));
-                            assert!(replaced.is_none(), "some earlier code left a stale handle");
-                            match self.map.entry(key) {
-                                hash_map::Entry::Occupied(_o) => {
-                                    // This cannot not happen because
-                                    // 1. we're the _miss_ handle, i.e., `self.map` didn't contain an entry and
-                                    // 2. we were holding &mut self during .resolve().await above, so, no other thread can have inserted a handle
-                                    //    while we were waiting for the tenant manager.
-                                    unreachable!()
-                                }
-                                hash_map::Entry::Vacant(v) => {
-                                    v.insert(handle_weak);
-                                }
-                            }
-                        }
-                        None => {
-                            return Err(GetError::PerTimelineStateShutDown);
-                        }
-                    }
-                }
-                Ok(handle)
-            }
-            Err(e) => Err(GetError::TenantManager(e)),
+    ) -> Result<Handle<T>, GetActiveTimelineError> {
+        let timeline = tenant_manager.resolve(timeline_id, shard_selector).await?;
+        let key = timeline.shard_timeline_id();
+        match &shard_selector {
+            ShardSelector::Zero => assert_eq!(key.shard_index.shard_number, ShardNumber(0)),
+            ShardSelector::Page(_) => (), // gotta trust tenant_manager
+            ShardSelector::Known(idx) => assert_eq!(idx, &key.shard_index),
         }
+
+        trace!("creating new HandleInner");
+        let timeline = Arc::new(timeline);
+        let handle_inner_arc = Arc::new(Mutex::new(HandleInner::Open(Arc::clone(&timeline))));
+        let handle_weak = WeakHandle {
+            inner: Arc::downgrade(&handle_inner_arc),
+        };
+        let handle = handle_weak
+            .upgrade()
+            .ok()
+            .expect("we just created it and it's not linked anywhere yet");
+        let mut lock_guard = timeline
+            .per_timeline_state()
+            .handles
+            .lock()
+            .expect("mutex poisoned");
+        let Some(per_timeline_state) = &mut *lock_guard else {
+            return Err(GetActiveTimelineError::Timeline(
+                GetTimelineError::ShuttingDown,
+            ));
+        };
+        let replaced = per_timeline_state.insert(self.id, Arc::clone(&handle_inner_arc));
+        assert!(replaced.is_none(), "some earlier code left a stale handle");
+        match self.map.entry(key) {
+            hash_map::Entry::Occupied(_o) => {
+                // This cannot not happen because
+                // 1. we're the _miss_ handle, i.e., `self.map` didn't contain an entry and
+                // 2. we were holding &mut self during .resolve().await above, so, no other thread can have inserted a handle
+                //    while we were waiting for the tenant manager.
+                unreachable!()
+            }
+            hash_map::Entry::Vacant(v) => {
+                v.insert(handle_weak);
+            }
+        }
+        Ok(handle)
     }
 }
 
@@ -655,7 +642,8 @@ mod tests {
     use pageserver_api::models::ShardParameters;
     use pageserver_api::reltag::RelTag;
     use pageserver_api::shard::DEFAULT_STRIPE_SIZE;
-    use utils::shard::ShardCount;
+    use utils::id::TenantId;
+    use utils::shard::{ShardCount, TenantShardId};
     use utils::sync::gate::GateGuard;
 
     use super::*;
@@ -665,7 +653,6 @@ mod tests {
     #[derive(Debug)]
     struct TestTypes;
     impl Types for TestTypes {
-        type TenantManagerError = anyhow::Error;
         type TenantManager = StubManager;
         type Timeline = Entered;
     }
@@ -716,40 +703,48 @@ mod tests {
             &self,
             timeline_id: TimelineId,
             shard_selector: ShardSelector,
-        ) -> anyhow::Result<Entered> {
+        ) -> Result<Entered, GetActiveTimelineError> {
+            fn enter_gate(
+                timeline: &StubTimeline,
+            ) -> Result<Arc<GateGuard>, GetActiveTimelineError> {
+                Ok(Arc::new(timeline.gate.enter().map_err(|_| {
+                    GetActiveTimelineError::Timeline(GetTimelineError::ShuttingDown)
+                })?))
+            }
+
             for timeline in &self.shards {
                 if timeline.id == timeline_id {
-                    let enter_gate = || {
-                        let gate_guard = timeline.gate.enter()?;
-                        let gate_guard = Arc::new(gate_guard);
-                        anyhow::Ok(gate_guard)
-                    };
                     match &shard_selector {
                         ShardSelector::Zero if timeline.shard.is_shard_zero() => {
                             return Ok(Entered {
                                 timeline: Arc::clone(timeline),
-                                gate_guard: enter_gate()?,
+                                gate_guard: enter_gate(timeline)?,
                             });
                         }
                         ShardSelector::Zero => continue,
                         ShardSelector::Page(key) if timeline.shard.is_key_local(key) => {
                             return Ok(Entered {
                                 timeline: Arc::clone(timeline),
-                                gate_guard: enter_gate()?,
+                                gate_guard: enter_gate(timeline)?,
                             });
                         }
                         ShardSelector::Page(_) => continue,
                         ShardSelector::Known(idx) if idx == &timeline.shard.shard_index() => {
                             return Ok(Entered {
                                 timeline: Arc::clone(timeline),
-                                gate_guard: enter_gate()?,
+                                gate_guard: enter_gate(timeline)?,
                             });
                         }
                         ShardSelector::Known(_) => continue,
                     }
                 }
             }
-            anyhow::bail!("not found")
+            Err(GetActiveTimelineError::Timeline(
+                GetTimelineError::NotFound {
+                    tenant_id: TenantShardId::unsharded(TenantId::from([0; 16])),
+                    timeline_id,
+                },
+            ))
         }
     }
 

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

@@ -365,6 +365,7 @@ pub(super) async fn handle_walreceiver_connection(
                                 critical_timeline!(
                                     timeline.tenant_shard_id,
                                     timeline.timeline_id,
+                                    Some(&timeline.corruption_detected),
                                     "{msg}"
                                 );
                                 return Err(WalReceiverError::Other(anyhow!(msg)));
@@ -382,6 +383,7 @@ pub(super) async fn handle_walreceiver_connection(
                                         critical_timeline!(
                                             timeline.tenant_shard_id,
                                             timeline.timeline_id,
+                                            Some(&timeline.corruption_detected),
                                             "{msg}"
                                         );
                                         return Err(WalReceiverError::Other(anyhow!(msg)));
@@ -455,6 +457,7 @@ pub(super) async fn handle_walreceiver_connection(
                                 critical_timeline!(
                                     timeline.tenant_shard_id,
                                     timeline.timeline_id,
+                                    Some(&timeline.corruption_detected),
                                     "{err:?}"
                                 );
                             }
@@ -586,6 +589,9 @@ pub(super) async fn handle_walreceiver_connection(
                 remote_consistent_lsn,
                 replytime: ts,
                 shard_number: timeline.tenant_shard_id.shard_number.0 as u32,
+                corruption_detected: timeline
+                    .corruption_detected
+                    .load(std::sync::atomic::Ordering::Relaxed),
             };
 
             debug!("neon_status_update {status_update:?}");

pageserver/src/walingest.rs

@@ -23,6 +23,7 @@
 
 use std::backtrace::Backtrace;
 use std::collections::HashMap;
+use std::sync::atomic::AtomicBool;
 use std::sync::{Arc, OnceLock};
 use std::time::{Duration, Instant, SystemTime};
 
@@ -422,6 +423,8 @@ impl WalIngest {
             critical_timeline!(
                 modification.tline.tenant_shard_id,
                 modification.tline.timeline_id,
+                // Hadron: No need to raise the corruption flag here; the caller of `ingest_record()` will do it.
+                None::<&AtomicBool>,
                 "clear_vm_bits for unknown VM relation {vm_rel}"
             );
             return Ok(());
@@ -431,6 +434,8 @@ impl WalIngest {
                 critical_timeline!(
                     modification.tline.tenant_shard_id,
                     modification.tline.timeline_id,
+                    // Hadron: No need to raise the corruption flag here; the caller of `ingest_record()` will do it.
+                    None::<&AtomicBool>,
                     "new_vm_blk {blknum} not in {vm_rel} of size {vm_size}"
                 );
                 new_vm_blk = None;
@@ -441,6 +446,8 @@ impl WalIngest {
                 critical_timeline!(
                     modification.tline.tenant_shard_id,
                     modification.tline.timeline_id,
+                    // Hadron: No need to raise the corruption flag here; the caller of `ingest_record()` will do it.
+                    None::<&AtomicBool>,
                     "old_vm_blk {blknum} not in {vm_rel} of size {vm_size}"
                 );
                 old_vm_blk = None;

pgxn/neon/file_cache.c

@@ -49,6 +49,7 @@
 #include "neon.h"
 #include "neon_lwlsncache.h"
 #include "neon_perf_counters.h"
+#include "neon_utils.h"
 #include "pagestore_client.h"
 #include "communicator.h"
 
@@ -673,8 +674,19 @@ lfc_get_state(size_t max_entries)
 			{
 				if (GET_STATE(entry, j) != UNAVAILABLE)
 				{
-					BITMAP_SET(bitmap, i*lfc_blocks_per_chunk + j);
-					n_pages += 1;
+					/* Validate the buffer tag before including it */
+					BufferTag test_tag = entry->key;
+					test_tag.blockNum += j;
+
+					if (BufferTagIsValid(&test_tag))
+					{
+						BITMAP_SET(bitmap, i*lfc_blocks_per_chunk + j);
+						n_pages += 1;
+					}
+					else
+					{
+						elog(ERROR, "LFC: Skipping invalid buffer tag during cache state capture: blockNum=%u", test_tag.blockNum);
+					}
 				}
 			}
 			if (++i == n_entries)
@@ -683,7 +695,7 @@ lfc_get_state(size_t max_entries)
 		Assert(i == n_entries);
 		fcs->n_pages = n_pages;
 		Assert(pg_popcount((char*)bitmap, ((n_entries << lfc_chunk_size_log) + 7)/8) == n_pages);
-		elog(LOG, "LFC: save state of %d chunks %d pages", (int)n_entries, (int)n_pages);
+		elog(LOG, "LFC: save state of %d chunks %d pages (validated)", (int)n_entries, (int)n_pages);
 	}
 
 	LWLockRelease(lfc_lock);
@@ -702,6 +714,7 @@ lfc_prewarm(FileCacheState* fcs, uint32 n_workers)
 	size_t n_entries;
 	size_t prewarm_batch = Min(lfc_prewarm_batch, readahead_buffer_size);
 	size_t fcs_size;
+	uint32_t max_prefetch_pages;
 	dsm_segment *seg;
 	BackgroundWorkerHandle* bgw_handle[MAX_PREWARM_WORKERS];
 
@@ -746,6 +759,11 @@ lfc_prewarm(FileCacheState* fcs, uint32 n_workers)
 	n_entries = Min(fcs->n_chunks, lfc_prewarm_limit);
 	Assert(n_entries != 0);
 
+	max_prefetch_pages = n_entries << fcs_chunk_size_log;
+	if (fcs->n_pages > max_prefetch_pages) {
+		elog(ERROR, "LFC: Number of pages in file cache state (%d) is more than the limit (%d)", fcs->n_pages, max_prefetch_pages);
+	}
+
 	LWLockAcquire(lfc_lock, LW_EXCLUSIVE);
 
 	/* Do not prewarm more entries than LFC limit */
@@ -898,6 +916,11 @@ lfc_prewarm_main(Datum main_arg)
 				{
 					tag = fcs->chunks[snd_idx >> fcs_chunk_size_log];
 					tag.blockNum += snd_idx & ((1 << fcs_chunk_size_log) - 1);
+
+					if (!BufferTagIsValid(&tag)) {
+						elog(ERROR, "LFC: Invalid buffer tag: %u", tag.blockNum);
+					}
+
 					if (!lfc_cache_contains(BufTagGetNRelFileInfo(tag), tag.forkNum, tag.blockNum))
 					{
 						(void)communicator_prefetch_register_bufferv(tag, NULL, 1, NULL);
@@ -1832,125 +1855,46 @@ lfc_writev(NRelFileInfo rinfo, ForkNumber forkNum, BlockNumber blkno,
 	LWLockRelease(lfc_lock);
 }
 
-typedef struct
+/*
+ * Return metrics about the LFC.
+ *
+ * The return format is a palloc'd array of LfcStatsEntrys. The size
+ * of the returned array is returned in *num_entries.
+ */
+LfcStatsEntry *
+lfc_get_stats(size_t *num_entries)
 {
-	TupleDesc	tupdesc;
-} NeonGetStatsCtx;
+	LfcStatsEntry *entries;
+	size_t		n = 0;
 
-#define NUM_NEON_GET_STATS_COLS	2
+#define MAX_ENTRIES 10
+	entries = palloc(sizeof(LfcStatsEntry) * MAX_ENTRIES);
 
-PG_FUNCTION_INFO_V1(neon_get_lfc_stats);
-Datum
-neon_get_lfc_stats(PG_FUNCTION_ARGS)
-{
-	FuncCallContext *funcctx;
-	NeonGetStatsCtx *fctx;
-	MemoryContext oldcontext;
-	TupleDesc	tupledesc;
-	Datum		result;
-	HeapTuple	tuple;
-	char const *key;
-	uint64		value = 0;
-	Datum		values[NUM_NEON_GET_STATS_COLS];
-	bool		nulls[NUM_NEON_GET_STATS_COLS];
+	entries[n++] = (LfcStatsEntry) {"file_cache_chunk_size_pages", lfc_ctl == NULL,
+									lfc_ctl ? lfc_blocks_per_chunk : 0 };
+	entries[n++] = (LfcStatsEntry) {"file_cache_misses", lfc_ctl == NULL,
+									lfc_ctl ? lfc_ctl->misses : 0};
+	entries[n++] = (LfcStatsEntry) {"file_cache_hits", lfc_ctl == NULL,
+									lfc_ctl ? lfc_ctl->hits : 0 };
+	entries[n++] = (LfcStatsEntry) {"file_cache_used", lfc_ctl == NULL,
+									lfc_ctl ? lfc_ctl->used : 0 };
+	entries[n++] = (LfcStatsEntry) {"file_cache_writes", lfc_ctl == NULL,
+									lfc_ctl ? lfc_ctl->writes : 0 };
+	entries[n++] = (LfcStatsEntry) {"file_cache_size", lfc_ctl == NULL,
+									lfc_ctl ? lfc_ctl->size : 0 };
+	entries[n++] = (LfcStatsEntry) {"file_cache_used_pages", lfc_ctl == NULL,
+									lfc_ctl ? lfc_ctl->used_pages : 0 };
+	entries[n++] = (LfcStatsEntry) {"file_cache_evicted_pages", lfc_ctl == NULL,
+									lfc_ctl ? lfc_ctl->evicted_pages : 0 };
+	entries[n++] = (LfcStatsEntry) {"file_cache_limit", lfc_ctl == NULL,
+									lfc_ctl ? lfc_ctl->limit : 0 };
+	entries[n++] = (LfcStatsEntry) {"file_cache_chunks_pinned", lfc_ctl == NULL,
+									lfc_ctl ? lfc_ctl->pinned : 0 };
+	Assert(n <= MAX_ENTRIES);
+#undef MAX_ENTRIES
 
-	if (SRF_IS_FIRSTCALL())
-	{
-		funcctx = SRF_FIRSTCALL_INIT();
-
-		/* Switch context when allocating stuff to be used in later calls */
-		oldcontext = MemoryContextSwitchTo(funcctx->multi_call_memory_ctx);
-
-		/* Create a user function context for cross-call persistence */
-		fctx = (NeonGetStatsCtx *) palloc(sizeof(NeonGetStatsCtx));
-
-		/* Construct a tuple descriptor for the result rows. */
-		tupledesc = CreateTemplateTupleDesc(NUM_NEON_GET_STATS_COLS);
-
-		TupleDescInitEntry(tupledesc, (AttrNumber) 1, "lfc_key",
-						   TEXTOID, -1, 0);
-		TupleDescInitEntry(tupledesc, (AttrNumber) 2, "lfc_value",
-						   INT8OID, -1, 0);
-
-		fctx->tupdesc = BlessTupleDesc(tupledesc);
-		funcctx->user_fctx = fctx;
-
-		/* Return to original context when allocating transient memory */
-		MemoryContextSwitchTo(oldcontext);
-	}
-
-	funcctx = SRF_PERCALL_SETUP();
-
-	/* Get the saved state */
-	fctx = (NeonGetStatsCtx *) funcctx->user_fctx;
-
-	switch (funcctx->call_cntr)
-	{
-		case 0:
-			key = "file_cache_misses";
-			if (lfc_ctl)
-				value = lfc_ctl->misses;
-			break;
-		case 1:
-			key = "file_cache_hits";
-			if (lfc_ctl)
-				value = lfc_ctl->hits;
-			break;
-		case 2:
-			key = "file_cache_used";
-			if (lfc_ctl)
-				value = lfc_ctl->used;
-			break;
-		case 3:
-			key = "file_cache_writes";
-			if (lfc_ctl)
-				value = lfc_ctl->writes;
-			break;
-		case 4:
-			key = "file_cache_size";
-			if (lfc_ctl)
-				value = lfc_ctl->size;
-			break;
-		case 5:
-			key = "file_cache_used_pages";
-			if (lfc_ctl)
-				value = lfc_ctl->used_pages;
-			break;
-		case 6:
-			key = "file_cache_evicted_pages";
-			if (lfc_ctl)
-				value = lfc_ctl->evicted_pages;
-			break;
-		case 7:
-			key = "file_cache_limit";
-			if (lfc_ctl)
-				value = lfc_ctl->limit;
-			break;
-		case 8:
-			key = "file_cache_chunk_size_pages";
-			value = lfc_blocks_per_chunk;
-			break;
-		case 9:
-			key = "file_cache_chunks_pinned";
-			if (lfc_ctl)
-				value = lfc_ctl->pinned;
-			break;
-		default:
-			SRF_RETURN_DONE(funcctx);
-	}
-	values[0] = PointerGetDatum(cstring_to_text(key));
-	nulls[0] = false;
-	if (lfc_ctl)
-	{
-		nulls[1] = false;
-		values[1] = Int64GetDatum(value);
-	}
-	else
-		nulls[1] = true;
-
-	tuple = heap_form_tuple(fctx->tupdesc, values, nulls);
-	result = HeapTupleGetDatum(tuple);
-	SRF_RETURN_NEXT(funcctx, result);
+	*num_entries = n;
+	return entries;
 }
 
 
@@ -1958,193 +1902,86 @@ neon_get_lfc_stats(PG_FUNCTION_ARGS)
  * Function returning data from the local file cache
  * relation node/tablespace/database/blocknum and access_counter
  */
-PG_FUNCTION_INFO_V1(local_cache_pages);
-
-/*
- * Record structure holding the to be exposed cache data.
- */
-typedef struct
+LocalCachePagesRec *
+lfc_local_cache_pages(size_t *num_entries)
 {
-	uint32		pageoffs;
-	Oid			relfilenode;
-	Oid			reltablespace;
-	Oid			reldatabase;
-	ForkNumber	forknum;
-	BlockNumber blocknum;
-	uint16		accesscount;
-} LocalCachePagesRec;
+	HASH_SEQ_STATUS status;
+	FileCacheEntry *entry;
+	size_t		n_pages;
+	size_t		n;
+	LocalCachePagesRec *result;
 
-/*
- * Function context for data persisting over repeated calls.
- */
-typedef struct
-{
-	TupleDesc	tupdesc;
-	LocalCachePagesRec *record;
-} LocalCachePagesContext;
-
-
-#define NUM_LOCALCACHE_PAGES_ELEM	7
-
-Datum
-local_cache_pages(PG_FUNCTION_ARGS)
-{
-	FuncCallContext *funcctx;
-	Datum		result;
-	MemoryContext oldcontext;
-	LocalCachePagesContext *fctx;	/* User function context. */
-	TupleDesc	tupledesc;
-	TupleDesc	expected_tupledesc;
-	HeapTuple	tuple;
-
-	if (SRF_IS_FIRSTCALL())
+	if (!lfc_ctl)
 	{
-		HASH_SEQ_STATUS status;
-		FileCacheEntry *entry;
-		uint32		n_pages = 0;
+		*num_entries = 0;
+		return NULL;
+	}
 
-		funcctx = SRF_FIRSTCALL_INIT();
+	LWLockAcquire(lfc_lock, LW_SHARED);
+	if (!LFC_ENABLED())
+	{
+		LWLockRelease(lfc_lock);
+		*num_entries = 0;
+		return NULL;
+	}
 
-		/* Switch context when allocating stuff to be used in later calls */
-		oldcontext = MemoryContextSwitchTo(funcctx->multi_call_memory_ctx);
-
-		/* Create a user function context for cross-call persistence */
-		fctx = (LocalCachePagesContext *) palloc(sizeof(LocalCachePagesContext));
-
-		/*
-		 * To smoothly support upgrades from version 1.0 of this extension
-		 * transparently handle the (non-)existence of the pinning_backends
-		 * column. We unfortunately have to get the result type for that... -
-		 * we can't use the result type determined by the function definition
-		 * without potentially crashing when somebody uses the old (or even
-		 * wrong) function definition though.
-		 */
-		if (get_call_result_type(fcinfo, NULL, &expected_tupledesc) != TYPEFUNC_COMPOSITE)
-			neon_log(ERROR, "return type must be a row type");
-
-		if (expected_tupledesc->natts != NUM_LOCALCACHE_PAGES_ELEM)
-			neon_log(ERROR, "incorrect number of output arguments");
-
-		/* Construct a tuple descriptor for the result rows. */
-		tupledesc = CreateTemplateTupleDesc(expected_tupledesc->natts);
-		TupleDescInitEntry(tupledesc, (AttrNumber) 1, "pageoffs",
-						   INT8OID, -1, 0);
-#if PG_MAJORVERSION_NUM < 16
-		TupleDescInitEntry(tupledesc, (AttrNumber) 2, "relfilenode",
-						   OIDOID, -1, 0);
-#else
-		TupleDescInitEntry(tupledesc, (AttrNumber) 2, "relfilenumber",
-						   OIDOID, -1, 0);
-#endif
-		TupleDescInitEntry(tupledesc, (AttrNumber) 3, "reltablespace",
-						   OIDOID, -1, 0);
-		TupleDescInitEntry(tupledesc, (AttrNumber) 4, "reldatabase",
-						   OIDOID, -1, 0);
-		TupleDescInitEntry(tupledesc, (AttrNumber) 5, "relforknumber",
-						   INT2OID, -1, 0);
-		TupleDescInitEntry(tupledesc, (AttrNumber) 6, "relblocknumber",
-						   INT8OID, -1, 0);
-		TupleDescInitEntry(tupledesc, (AttrNumber) 7, "accesscount",
-						   INT4OID, -1, 0);
-
-		fctx->tupdesc = BlessTupleDesc(tupledesc);
-
-		if (lfc_ctl)
+	/* Count the pages first */
+	n_pages = 0;
+	hash_seq_init(&status, lfc_hash);
+	while ((entry = hash_seq_search(&status)) != NULL)
+	{
+		/* Skip hole tags */
+		if (NInfoGetRelNumber(BufTagGetNRelFileInfo(entry->key)) != 0)
 		{
-			LWLockAcquire(lfc_lock, LW_SHARED);
+			for (int i = 0; i < lfc_blocks_per_chunk; i++)
+				n_pages += GET_STATE(entry, i) == AVAILABLE;
+		}
+	}
 
-			if (LFC_ENABLED())
+	if (n_pages == 0)
+	{
+		LWLockRelease(lfc_lock);
+		*num_entries = 0;
+		return NULL;
+	}
+
+	result = (LocalCachePagesRec *)
+		MemoryContextAllocHuge(CurrentMemoryContext,
+							   sizeof(LocalCachePagesRec) * n_pages);
+
+	/*
+	 * Scan through all the cache entries, saving the relevant fields
+	 * in the result structure.
+	 */
+	n = 0;
+	hash_seq_init(&status, lfc_hash);
+	while ((entry = hash_seq_search(&status)) != NULL)
+	{
+		for (int i = 0; i < lfc_blocks_per_chunk; i++)
+		{
+			if (NInfoGetRelNumber(BufTagGetNRelFileInfo(entry->key)) != 0)
 			{
-				hash_seq_init(&status, lfc_hash);
-				while ((entry = hash_seq_search(&status)) != NULL)
+				if (GET_STATE(entry, i) == AVAILABLE)
 				{
-					/* Skip hole tags */
-					if (NInfoGetRelNumber(BufTagGetNRelFileInfo(entry->key)) != 0)
-					{
-						for (int i = 0; i < lfc_blocks_per_chunk; i++)
-							n_pages += GET_STATE(entry, i) == AVAILABLE;
-					}
+					result[n].pageoffs = entry->offset * lfc_blocks_per_chunk + i;
+					result[n].relfilenode = NInfoGetRelNumber(BufTagGetNRelFileInfo(entry->key));
+					result[n].reltablespace = NInfoGetSpcOid(BufTagGetNRelFileInfo(entry->key));
+					result[n].reldatabase = NInfoGetDbOid(BufTagGetNRelFileInfo(entry->key));
+					result[n].forknum = entry->key.forkNum;
+					result[n].blocknum = entry->key.blockNum + i;
+					result[n].accesscount = entry->access_count;
+					n += 1;
 				}
 			}
 		}
-		fctx->record = (LocalCachePagesRec *)
-			MemoryContextAllocHuge(CurrentMemoryContext,
-								   sizeof(LocalCachePagesRec) * n_pages);
-
-		/* Set max calls and remember the user function context. */
-		funcctx->max_calls = n_pages;
-		funcctx->user_fctx = fctx;
-
-		/* Return to original context when allocating transient memory */
-		MemoryContextSwitchTo(oldcontext);
-
-		if (n_pages != 0)
-		{
-			/*
-			 * Scan through all the cache entries, saving the relevant fields
-			 * in the fctx->record structure.
-			 */
-			uint32		n = 0;
-
-			hash_seq_init(&status, lfc_hash);
-			while ((entry = hash_seq_search(&status)) != NULL)
-			{
-				for (int i = 0; i < lfc_blocks_per_chunk; i++)
-				{
-					if (NInfoGetRelNumber(BufTagGetNRelFileInfo(entry->key)) != 0)
-					{
-						if (GET_STATE(entry, i) == AVAILABLE)
-						{
-							fctx->record[n].pageoffs = entry->offset * lfc_blocks_per_chunk + i;
-							fctx->record[n].relfilenode = NInfoGetRelNumber(BufTagGetNRelFileInfo(entry->key));
-							fctx->record[n].reltablespace = NInfoGetSpcOid(BufTagGetNRelFileInfo(entry->key));
-							fctx->record[n].reldatabase = NInfoGetDbOid(BufTagGetNRelFileInfo(entry->key));
-							fctx->record[n].forknum = entry->key.forkNum;
-							fctx->record[n].blocknum = entry->key.blockNum + i;
-							fctx->record[n].accesscount = entry->access_count;
-							n += 1;
-						}
-					}
-				}
-			}
-			Assert(n_pages == n);
-		}
-		if (lfc_ctl)
-			LWLockRelease(lfc_lock);
 	}
+	Assert(n_pages == n);
+	LWLockRelease(lfc_lock);
 
-	funcctx = SRF_PERCALL_SETUP();
-
-	/* Get the saved state */
-	fctx = funcctx->user_fctx;
-
-	if (funcctx->call_cntr < funcctx->max_calls)
-	{
-		uint32		i = funcctx->call_cntr;
-		Datum		values[NUM_LOCALCACHE_PAGES_ELEM];
-		bool		nulls[NUM_LOCALCACHE_PAGES_ELEM] = {
-			false, false, false, false, false, false, false
-		};
-
-		values[0] = Int64GetDatum((int64) fctx->record[i].pageoffs);
-		values[1] = ObjectIdGetDatum(fctx->record[i].relfilenode);
-		values[2] = ObjectIdGetDatum(fctx->record[i].reltablespace);
-		values[3] = ObjectIdGetDatum(fctx->record[i].reldatabase);
-		values[4] = ObjectIdGetDatum(fctx->record[i].forknum);
-		values[5] = Int64GetDatum((int64) fctx->record[i].blocknum);
-		values[6] = Int32GetDatum(fctx->record[i].accesscount);
-
-		/* Build and return the tuple. */
-		tuple = heap_form_tuple(fctx->tupdesc, values, nulls);
-		result = HeapTupleGetDatum(tuple);
-
-		SRF_RETURN_NEXT(funcctx, result);
-	}
-	else
-		SRF_RETURN_DONE(funcctx);
+	*num_entries = n_pages;
+	return result;
 }
 
-
 /*
  * Internal implementation of the approximate_working_set_size_seconds()
  * function.

pgxn/neon/file_cache.h

@@ -47,6 +47,26 @@ extern bool lfc_prefetch(NRelFileInfo rinfo, ForkNumber forknum, BlockNumber blk
 extern FileCacheState* lfc_get_state(size_t max_entries);
 extern void lfc_prewarm(FileCacheState* fcs, uint32 n_workers);
 
+typedef struct LfcStatsEntry
+{
+	const char *metric_name;
+	bool		isnull;
+	uint64		value;
+} LfcStatsEntry;
+extern LfcStatsEntry *lfc_get_stats(size_t *num_entries);
+
+typedef struct
+{
+	uint32		pageoffs;
+	Oid			relfilenode;
+	Oid			reltablespace;
+	Oid			reldatabase;
+	ForkNumber	forknum;
+	BlockNumber blocknum;
+	uint16		accesscount;
+} LocalCachePagesRec;
+extern LocalCachePagesRec *lfc_local_cache_pages(size_t *num_entries);
+
 extern int32 lfc_approximate_working_set_size_seconds(time_t duration, bool reset);
 
 

pgxn/neon/libpagestore.c

@@ -71,7 +71,7 @@ char	   *neon_project_id;
 char	   *neon_branch_id;
 char	   *neon_endpoint_id;
 int32		max_cluster_size;
-char	   *page_server_connstring;
+char	   *pageserver_connstring;
 char	   *neon_auth_token;
 
 int			readahead_buffer_size = 128;
@@ -1440,7 +1440,6 @@ check_neon_id(char **newval, void **extra, GucSource source)
 	return **newval == '\0' || HexDecodeString(id, *newval, 16);
 }
 
-
 void
 PagestoreShmemInit(void)
 {
@@ -1454,7 +1453,7 @@ PagestoreShmemInit(void)
 		pg_atomic_init_u64(&pagestore_shared->begin_update_counter, 0);
 		pg_atomic_init_u64(&pagestore_shared->end_update_counter, 0);
 		memset(&pagestore_shared->shard_map, 0, sizeof(ShardMap));
-		AssignPageserverConnstring(page_server_connstring, NULL);
+		AssignPageserverConnstring(pageserver_connstring, NULL);
 	}
 }
 
@@ -1473,7 +1472,7 @@ pg_init_libpagestore(void)
 	DefineCustomStringVariable("neon.pageserver_connstring",
 							   "connection string to the page server",
 							   NULL,
-							   &page_server_connstring,
+							   &pageserver_connstring,
 							   "",
 							   PGC_SIGHUP,
 							   0,	/* no flags required */
@@ -1644,7 +1643,7 @@ pg_init_libpagestore(void)
 	if (neon_auth_token)
 		neon_log(LOG, "using storage auth token from NEON_AUTH_TOKEN environment variable");
 
-	if (page_server_connstring && page_server_connstring[0])
+	if (pageserver_connstring[0])
 	{
 		neon_log(PageStoreTrace, "set neon_smgr hook");
 		smgr_hook = smgr_neon;

pgxn/neon/neon.c

@@ -51,6 +51,7 @@ void		_PG_init(void);
 bool lakebase_mode = false;
 
 static int  running_xacts_overflow_policy;
+static emit_log_hook_type prev_emit_log_hook;
 static bool monitor_query_exec_time = false;
 
 static ExecutorStart_hook_type prev_ExecutorStart = NULL;
@@ -81,6 +82,8 @@ uint32		WAIT_EVENT_NEON_PS_READ;
 uint32		WAIT_EVENT_NEON_WAL_DL;
 #endif
 
+int databricks_test_hook = 0;
+
 enum RunningXactsOverflowPolicies {
 	OP_IGNORE,
 	OP_SKIP,
@@ -445,6 +448,20 @@ ReportSearchPath(void)
 static int neon_pgstat_file_size_limit;
 #endif
 
+static void DatabricksSqlErrorHookImpl(ErrorData *edata) {
+	if (prev_emit_log_hook != NULL) {
+		prev_emit_log_hook(edata);
+	}
+
+	if (edata->sqlerrcode == ERRCODE_DATA_CORRUPTED) {
+		pg_atomic_fetch_add_u32(&databricks_metrics_shared->data_corruption_count, 1);
+	} else if (edata->sqlerrcode == ERRCODE_INDEX_CORRUPTED) {
+		pg_atomic_fetch_add_u32(&databricks_metrics_shared->index_corruption_count, 1);
+	} else if (edata->sqlerrcode == ERRCODE_INTERNAL_ERROR) {
+		pg_atomic_fetch_add_u32(&databricks_metrics_shared->internal_error_count, 1);
+	}
+}
+
 void
 _PG_init(void)
 {
@@ -456,6 +473,11 @@ _PG_init(void)
 	load_file("$libdir/neon_rmgr", false);
 #endif
 
+	if (lakebase_mode) {
+		prev_emit_log_hook = emit_log_hook;
+		emit_log_hook = DatabricksSqlErrorHookImpl;
+	}
+
 	/*
 	 * Initializing a pre-loaded Postgres extension happens in three stages:
 	 *
@@ -594,6 +616,19 @@ _PG_init(void)
 							0,
 							NULL, NULL, NULL);
 
+	// A test hook used in sql regress to trigger specific behaviors
+	// to test features easily.
+	DefineCustomIntVariable(
+							"databricks.test_hook",
+							"The test hook used in sql regress tests only",
+							NULL,
+							&databricks_test_hook,
+							0,
+							0, INT32_MAX,
+							PGC_SUSET,
+							0,
+							NULL, NULL, NULL);
+
 	/*
 	 * Important: This must happen after other parts of the extension are
 	 * loaded, otherwise any settings to GUCs that were set before the
@@ -625,11 +660,15 @@ _PG_init(void)
 	ExecutorEnd_hook = neon_ExecutorEnd;
 }
 
+/* Various functions exposed at SQL level */
+
 PG_FUNCTION_INFO_V1(pg_cluster_size);
 PG_FUNCTION_INFO_V1(backpressure_lsns);
 PG_FUNCTION_INFO_V1(backpressure_throttling_time);
 PG_FUNCTION_INFO_V1(approximate_working_set_size_seconds);
 PG_FUNCTION_INFO_V1(approximate_working_set_size);
+PG_FUNCTION_INFO_V1(neon_get_lfc_stats);
+PG_FUNCTION_INFO_V1(local_cache_pages);
 
 Datum
 pg_cluster_size(PG_FUNCTION_ARGS)
@@ -704,6 +743,76 @@ approximate_working_set_size(PG_FUNCTION_ARGS)
 		PG_RETURN_INT32(dc);
 }
 
+Datum
+neon_get_lfc_stats(PG_FUNCTION_ARGS)
+{
+#define NUM_NEON_GET_STATS_COLS        2
+	ReturnSetInfo *rsinfo = (ReturnSetInfo *) fcinfo->resultinfo;
+	LfcStatsEntry *entries;
+	size_t		num_entries;
+
+	InitMaterializedSRF(fcinfo, 0);
+
+	/* lfc_get_stats() does all the heavy lifting */
+	entries = lfc_get_stats(&num_entries);
+
+	/* Convert the LfcStatsEntrys to a result set */
+	for (size_t i = 0; i < num_entries; i++)
+	{
+		LfcStatsEntry *entry = &entries[i];
+		Datum		values[NUM_NEON_GET_STATS_COLS];
+		bool		nulls[NUM_NEON_GET_STATS_COLS];
+
+		values[0] = CStringGetTextDatum(entry->metric_name);
+		nulls[0] = false;
+		values[1] = Int64GetDatum(entry->isnull ? 0 : entry->value);
+		nulls[1] = entry->isnull;
+		tuplestore_putvalues(rsinfo->setResult, rsinfo->setDesc, values, nulls);
+	}
+	PG_RETURN_VOID();
+
+#undef NUM_NEON_GET_STATS_COLS
+}
+
+Datum
+local_cache_pages(PG_FUNCTION_ARGS)
+{
+#define NUM_LOCALCACHE_PAGES_COLS	7
+	ReturnSetInfo *rsinfo = (ReturnSetInfo *) fcinfo->resultinfo;
+	LocalCachePagesRec *entries;
+	size_t		num_entries;
+
+	InitMaterializedSRF(fcinfo, 0);
+
+	/* lfc_local_cache_pages() does all the heavy lifting */
+	entries = lfc_local_cache_pages(&num_entries);
+
+	/* Convert the LocalCachePagesRec structs to a result set */
+	for (size_t i = 0; i < num_entries; i++)
+	{
+		LocalCachePagesRec *entry = &entries[i];
+		Datum		values[NUM_LOCALCACHE_PAGES_COLS];
+		bool		nulls[NUM_LOCALCACHE_PAGES_COLS] = {
+			false, false, false, false, false, false, false
+		};
+
+		values[0] = Int64GetDatum((int64) entry->pageoffs);
+		values[1] = ObjectIdGetDatum(entry->relfilenode);
+		values[2] = ObjectIdGetDatum(entry->reltablespace);
+		values[3] = ObjectIdGetDatum(entry->reldatabase);
+		values[4] = ObjectIdGetDatum(entry->forknum);
+		values[5] = Int64GetDatum((int64) entry->blocknum);
+		values[6] = Int32GetDatum(entry->accesscount);
+
+		/* Build and return the tuple. */
+		tuplestore_putvalues(rsinfo->setResult, rsinfo->setDesc, values, nulls);
+	}
+
+	PG_RETURN_VOID();
+
+#undef NUM_LOCALCACHE_PAGES_COLS
+}
+
 /*
  * Initialization stage 2: make requests for the amount of shared memory we
  * will need.
@@ -742,6 +851,9 @@ neon_shmem_startup_hook(void)
 
 	LfcShmemInit();
 	NeonPerfCountersShmemInit();
+	if (lakebase_mode) {
+		DatabricksMetricsShmemInit();
+	}
 	PagestoreShmemInit();
 	RelsizeCacheShmemInit();
 	WalproposerShmemInit();

pgxn/neon/neon_perf_counters.c

@@ -19,7 +19,36 @@
 
 #include "neon.h"
 #include "neon_perf_counters.h"
-#include "neon_pgversioncompat.h"
+#include "walproposer.h"
+
+/* BEGIN_HADRON */
+databricks_metrics *databricks_metrics_shared;
+
+Size
+DatabricksMetricsShmemSize(void)
+{
+	return sizeof(databricks_metrics);
+}
+
+void
+DatabricksMetricsShmemInit(void)
+{
+	bool		found;
+
+	databricks_metrics_shared =
+		ShmemInitStruct("Databricks counters",
+						DatabricksMetricsShmemSize(),
+						&found);
+	Assert(found == IsUnderPostmaster);
+	if (!found)
+	{
+		pg_atomic_init_u32(&databricks_metrics_shared->index_corruption_count, 0);
+		pg_atomic_init_u32(&databricks_metrics_shared->data_corruption_count, 0);
+		pg_atomic_init_u32(&databricks_metrics_shared->internal_error_count, 0);
+		pg_atomic_init_u32(&databricks_metrics_shared->ps_corruption_detected, 0);
+	}
+}
+/* END_HADRON */
 
 neon_per_backend_counters *neon_per_backend_counters_shared;
 
@@ -38,11 +67,12 @@ NeonPerfCountersShmemRequest(void)
 #else
 	size = mul_size(NUM_NEON_PERF_COUNTER_SLOTS, sizeof(neon_per_backend_counters));
 #endif
+	if (lakebase_mode) {
+		size = add_size(size, DatabricksMetricsShmemSize());
+	}
 	RequestAddinShmemSpace(size);
 }
 
-
-
 void
 NeonPerfCountersShmemInit(void)
 {
@@ -395,6 +425,34 @@ neon_get_perf_counters(PG_FUNCTION_ARGS)
 		metric_to_datums(&metrics[i], &values[0], &nulls[0]);
 		tuplestore_putvalues(rsinfo->setResult, rsinfo->setDesc, values, nulls);
 	}
+
+	if (lakebase_mode) {
+
+		if (databricks_test_hook == TestHookCorruption) {
+			ereport(ERROR,
+						(errcode(ERRCODE_DATA_CORRUPTED),
+						errmsg("test corruption")));
+		}
+
+		// Not ideal but piggyback our databricks counters into the neon perf counters view
+		// so that we don't need to introduce neon--1.x+1.sql to add a new view.
+		{
+			metric_t databricks_metrics[] = {
+				{"sql_index_corruption_count", false, 0, (double) pg_atomic_read_u32(&databricks_metrics_shared->index_corruption_count)},
+				{"sql_data_corruption_count", false, 0, (double) pg_atomic_read_u32(&databricks_metrics_shared->data_corruption_count)},
+				{"sql_internal_error_count", false, 0, (double) pg_atomic_read_u32(&databricks_metrics_shared->internal_error_count)},
+				{"ps_corruption_detected", false, 0, (double) pg_atomic_read_u32(&databricks_metrics_shared->ps_corruption_detected)},
+				{NULL, false, 0, 0},
+			};
+			for (int i = 0; databricks_metrics[i].name != NULL; i++)
+			{
+				metric_to_datums(&databricks_metrics[i], &values[0], &nulls[0]);
+				tuplestore_putvalues(rsinfo->setResult, rsinfo->setDesc, values, nulls);
+			}
+		}
+		/* END_HADRON */
+	}
+
 	pfree(metrics);
 
 	return (Datum) 0;

pgxn/neon/neon_perf_counters.h

@@ -177,5 +177,24 @@ extern void inc_query_time(uint64 elapsed);
 extern Size NeonPerfCountersShmemSize(void);
 extern void NeonPerfCountersShmemInit(void);
 
+/* BEGIN_HADRON */
+typedef struct
+{
+	pg_atomic_uint32 index_corruption_count;
+	pg_atomic_uint32 data_corruption_count;
+	pg_atomic_uint32 internal_error_count;
+	pg_atomic_uint32 ps_corruption_detected;
+} databricks_metrics;
+
+extern databricks_metrics *databricks_metrics_shared;
+
+extern Size DatabricksMetricsShmemSize(void);
+extern void DatabricksMetricsShmemInit(void);
+
+extern int databricks_test_hook;
+
+static const int TestHookCorruption = 1;
+/* END_HADRON */
+
 
 #endif							/* NEON_PERF_COUNTERS_H */

pgxn/neon/neon_utils.c

@@ -183,3 +183,22 @@ alloc_curl_handle(void)
 }
 
 #endif
+
+/*
+ * Check if a BufferTag is valid by verifying all its fields are not invalid.
+ */
+bool
+BufferTagIsValid(const BufferTag *tag)
+{
+	#if PG_MAJORVERSION_NUM >= 16
+	return (tag->spcOid != InvalidOid) &&
+		(tag->relNumber != InvalidRelFileNumber) &&
+		(tag->forkNum != InvalidForkNumber) &&
+		(tag->blockNum != InvalidBlockNumber);
+	#else
+	return (tag->rnode.spcNode != InvalidOid) &&
+		(tag->rnode.relNode != InvalidOid) &&
+		(tag->forkNum != InvalidForkNumber) &&
+		(tag->blockNum != InvalidBlockNumber);
+	#endif
+}

pgxn/neon/neon_utils.h

@@ -2,6 +2,7 @@
 #define __NEON_UTILS_H__
 
 #include "lib/stringinfo.h"
+#include "storage/buf_internals.h"
 
 #ifndef WALPROPOSER_LIB
 #include <curl/curl.h>
@@ -16,6 +17,9 @@ void		pq_sendint32_le(StringInfo buf, uint32 i);
 void		pq_sendint64_le(StringInfo buf, uint64 i);
 void        disable_core_dump(void);
 
+/* Buffer tag validation function */
+bool		BufferTagIsValid(const BufferTag *tag);
+
 #ifndef WALPROPOSER_LIB
 
 CURL *		alloc_curl_handle(void);

pgxn/neon/pagestore_client.h

@@ -236,7 +236,7 @@ extern void prefetch_on_ps_disconnect(void);
 
 extern page_server_api *page_server;
 
-extern char *page_server_connstring;
+extern char *pageserver_connstring;
 extern int	flush_every_n_requests;
 extern int	readahead_buffer_size;
 extern char *neon_timeline;

pgxn/neon/walproposer.c

@@ -1887,6 +1887,12 @@ ParsePageserverFeedbackMessage(WalProposer *wp, StringInfo reply_message, Pagese
 			ps_feedback->shard_number = pq_getmsgint(reply_message, sizeof(uint32));
 			psfeedback_log("%u", key, ps_feedback->shard_number);
 		}
+		else if (strcmp(key, "corruption_detected") == 0)
+		{
+			Assert(value_len == 1);
+			ps_feedback->corruption_detected = pq_getmsgbyte(reply_message) != 0;
+			psfeedback_log("%s", key, ps_feedback->corruption_detected ? "true" : "false");
+		}
 		else
 		{
 			/*

pgxn/neon/walproposer.h

@@ -374,6 +374,8 @@ typedef struct PageserverFeedback
 	XLogRecPtr	remote_consistent_lsn;
 	TimestampTz replytime;
 	uint32		shard_number;
+	/* true if the pageserver has detected data corruption in the timeline */
+	bool		corruption_detected;
 } PageserverFeedback;
 
 /* BEGIN_HADRON */

pgxn/neon/walproposer_pg.c

@@ -49,6 +49,7 @@
 
 #include "libpqwalproposer.h"
 #include "neon.h"
+#include "neon_perf_counters.h"
 #include "neon_walreader.h"
 #include "walproposer.h"
 
@@ -741,6 +742,11 @@ record_pageserver_feedback(PageserverFeedback *ps_feedback, shardno_t num_shards
 	Assert(ps_feedback->shard_number < MAX_SHARDS);
 	Assert(ps_feedback->shard_number < num_shards);
 
+	// Begin Hadron: Record any corruption signal from the pageserver first.
+	if (ps_feedback->corruption_detected) {
+		pg_atomic_write_u32(&databricks_metrics_shared->ps_corruption_detected, 1);
+	}
+
 	SpinLockAcquire(&walprop_shared->mutex);
 
 	// Hadron: Update the num_shards from the source-of-truth (shard map) lazily when we receive

proxy/Cargo.toml

@@ -107,6 +107,7 @@ uuid.workspace = true
 x509-cert.workspace = true
 redis.workspace = true
 zerocopy.workspace = true
+zeroize.workspace = true
 # uncomment this to use the real subzero-core crate
 # subzero-core = { git = "https://github.com/neondatabase/subzero", rev = "396264617e78e8be428682f87469bb25429af88a", features = ["postgresql"], optional = true }
 # this is a stub for the subzero-core crate

proxy/src/auth/backend/hacks.rs

@@ -6,7 +6,7 @@ use crate::auth::{self, AuthFlow};
 use crate::config::AuthenticationConfig;
 use crate::context::RequestContext;
 use crate::control_plane::AuthSecret;
-use crate::intern::EndpointIdInt;
+use crate::intern::{EndpointIdInt, RoleNameInt};
 use crate::sasl;
 use crate::stream::{self, Stream};
 
@@ -25,13 +25,15 @@ pub(crate) async fn authenticate_cleartext(
     ctx.set_auth_method(crate::context::AuthMethod::Cleartext);
 
     let ep = EndpointIdInt::from(&info.endpoint);
+    let role = RoleNameInt::from(&info.user);
 
     let auth_flow = AuthFlow::new(
         client,
         auth::CleartextPassword {
             secret,
             endpoint: ep,
-            pool: config.thread_pool.clone(),
+            role,
+            pool: config.scram_thread_pool.clone(),
         },
     );
     let auth_outcome = {

proxy/src/auth/backend/mod.rs

@@ -25,7 +25,7 @@ use crate::control_plane::messages::EndpointRateLimitConfig;
 use crate::control_plane::{
     self, AccessBlockerFlags, AuthSecret, ControlPlaneApi, EndpointAccessControl, RoleAccessControl,
 };
-use crate::intern::EndpointIdInt;
+use crate::intern::{EndpointIdInt, RoleNameInt};
 use crate::pqproto::BeMessage;
 use crate::proxy::NeonOptions;
 use crate::proxy::wake_compute::WakeComputeBackend;
@@ -273,9 +273,11 @@ async fn authenticate_with_secret(
 ) -> auth::Result<ComputeCredentials> {
     if let Some(password) = unauthenticated_password {
         let ep = EndpointIdInt::from(&info.endpoint);
+        let role = RoleNameInt::from(&info.user);
 
         let auth_outcome =
-            validate_password_and_exchange(&config.thread_pool, ep, &password, secret).await?;
+            validate_password_and_exchange(&config.scram_thread_pool, ep, role, &password, secret)
+                .await?;
         let keys = match auth_outcome {
             crate::sasl::Outcome::Success(key) => key,
             crate::sasl::Outcome::Failure(reason) => {
@@ -499,7 +501,7 @@ mod tests {
 
     static CONFIG: Lazy<AuthenticationConfig> = Lazy::new(|| AuthenticationConfig {
         jwks_cache: JwkCache::default(),
-        thread_pool: ThreadPool::new(1),
+        scram_thread_pool: ThreadPool::new(1),
         scram_protocol_timeout: std::time::Duration::from_secs(5),
         ip_allowlist_check_enabled: true,
         is_vpc_acccess_proxy: false,

proxy/src/auth/flow.rs

@@ -10,7 +10,7 @@ use super::backend::ComputeCredentialKeys;
 use super::{AuthError, PasswordHackPayload};
 use crate::context::RequestContext;
 use crate::control_plane::AuthSecret;
-use crate::intern::EndpointIdInt;
+use crate::intern::{EndpointIdInt, RoleNameInt};
 use crate::pqproto::{BeAuthenticationSaslMessage, BeMessage};
 use crate::sasl;
 use crate::scram::threadpool::ThreadPool;
@@ -46,6 +46,7 @@ pub(crate) struct PasswordHack;
 pub(crate) struct CleartextPassword {
     pub(crate) pool: Arc<ThreadPool>,
     pub(crate) endpoint: EndpointIdInt,
+    pub(crate) role: RoleNameInt,
     pub(crate) secret: AuthSecret,
 }
 
@@ -111,6 +112,7 @@ impl<S: AsyncRead + AsyncWrite + Unpin> AuthFlow<'_, S, CleartextPassword> {
         let outcome = validate_password_and_exchange(
             &self.state.pool,
             self.state.endpoint,
+            self.state.role,
             password,
             self.state.secret,
         )
@@ -165,13 +167,15 @@ impl<S: AsyncRead + AsyncWrite + Unpin> AuthFlow<'_, S, Scram<'_>> {
 pub(crate) async fn validate_password_and_exchange(
     pool: &ThreadPool,
     endpoint: EndpointIdInt,
+    role: RoleNameInt,
     password: &[u8],
     secret: AuthSecret,
 ) -> super::Result<sasl::Outcome<ComputeCredentialKeys>> {
     match secret {
         // 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?;
+            let outcome =
+                crate::scram::exchange(pool, endpoint, role, &scram_secret, password).await?;
 
             let client_key = match outcome {
                 sasl::Outcome::Success(client_key) => client_key,

proxy/src/binary/local_proxy.rs

@@ -29,7 +29,7 @@ use crate::config::{
 };
 use crate::control_plane::locks::ApiLocks;
 use crate::http::health_server::AppMetrics;
-use crate::metrics::{Metrics, ServiceInfo, ThreadPoolMetrics};
+use crate::metrics::{Metrics, ServiceInfo};
 use crate::rate_limiter::{EndpointRateLimiter, LeakyBucketConfig, RateBucketInfo};
 use crate::scram::threadpool::ThreadPool;
 use crate::serverless::cancel_set::CancelSet;
@@ -114,8 +114,6 @@ pub async fn run() -> anyhow::Result<()> {
     let _panic_hook_guard = utils::logging::replace_panic_hook_with_tracing_panic_hook();
     let _sentry_guard = init_sentry(Some(GIT_VERSION.into()), &[]);
 
-    Metrics::install(Arc::new(ThreadPoolMetrics::new(0)));
-
     // TODO: refactor these to use labels
     debug!("Version: {GIT_VERSION}");
     debug!("Build_tag: {BUILD_TAG}");
@@ -284,7 +282,7 @@ fn build_config(args: &LocalProxyCliArgs) -> anyhow::Result<&'static ProxyConfig
         http_config,
         authentication_config: AuthenticationConfig {
             jwks_cache: JwkCache::default(),
-            thread_pool: ThreadPool::new(0),
+            scram_thread_pool: ThreadPool::new(0),
             scram_protocol_timeout: Duration::from_secs(10),
             ip_allowlist_check_enabled: true,
             is_vpc_acccess_proxy: false,

proxy/src/binary/pg_sni_router.rs

@@ -26,7 +26,7 @@ use utils::project_git_version;
 use utils::sentry_init::init_sentry;
 
 use crate::context::RequestContext;
-use crate::metrics::{Metrics, ServiceInfo, ThreadPoolMetrics};
+use crate::metrics::{Metrics, ServiceInfo};
 use crate::pglb::TlsRequired;
 use crate::pqproto::FeStartupPacket;
 use crate::protocol2::ConnectionInfo;
@@ -80,8 +80,6 @@ pub async fn run() -> anyhow::Result<()> {
     let _panic_hook_guard = utils::logging::replace_panic_hook_with_tracing_panic_hook();
     let _sentry_guard = init_sentry(Some(GIT_VERSION.into()), &[]);
 
-    Metrics::install(Arc::new(ThreadPoolMetrics::new(0)));
-
     let args = cli().get_matches();
     let destination: String = args
         .get_one::<String>("dest")

proxy/src/binary/proxy.rs

@@ -617,7 +617,12 @@ pub async fn run() -> anyhow::Result<()> {
 /// ProxyConfig is created at proxy startup, and lives forever.
 fn build_config(args: &ProxyCliArgs) -> anyhow::Result<&'static ProxyConfig> {
     let thread_pool = ThreadPool::new(args.scram_thread_pool_size);
-    Metrics::install(thread_pool.metrics.clone());
+    Metrics::get()
+        .proxy
+        .scram_pool
+        .0
+        .set(thread_pool.metrics.clone())
+        .ok();
 
     let tls_config = match (&args.tls_key, &args.tls_cert) {
         (Some(key_path), Some(cert_path)) => Some(config::configure_tls(
@@ -690,12 +695,15 @@ fn build_config(args: &ProxyCliArgs) -> anyhow::Result<&'static ProxyConfig> {
     };
     let authentication_config = AuthenticationConfig {
         jwks_cache: JwkCache::default(),
-        thread_pool,
+        scram_thread_pool: thread_pool,
         scram_protocol_timeout: args.scram_protocol_timeout,
         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,
+        #[cfg(not(feature = "rest_broker"))]
         accept_jwts: args.is_auth_broker,
+        #[cfg(feature = "rest_broker")]
+        accept_jwts: args.is_auth_broker || args.is_rest_broker,
         console_redirect_confirmation_timeout: args.webauth_confirmation_timeout,
     };
 

proxy/src/compute/mod.rs

@@ -8,6 +8,7 @@ use futures::{FutureExt, TryFutureExt};
 use itertools::Itertools;
 use postgres_client::config::{AuthKeys, ChannelBinding, SslMode};
 use postgres_client::connect_raw::StartupStream;
+use postgres_client::error::SqlState;
 use postgres_client::maybe_tls_stream::MaybeTlsStream;
 use postgres_client::tls::MakeTlsConnect;
 use thiserror::Error;
@@ -22,7 +23,7 @@ use crate::context::RequestContext;
 use crate::control_plane::client::ApiLockError;
 use crate::control_plane::errors::WakeComputeError;
 use crate::control_plane::messages::MetricsAuxInfo;
-use crate::error::{ReportableError, UserFacingError};
+use crate::error::{ErrorKind, ReportableError, UserFacingError};
 use crate::metrics::{Metrics, NumDbConnectionsGuard};
 use crate::pqproto::StartupMessageParams;
 use crate::proxy::connect_compute::TlsNegotiation;
@@ -65,12 +66,13 @@ impl UserFacingError for PostgresError {
 }
 
 impl ReportableError for PostgresError {
-    fn get_error_kind(&self) -> crate::error::ErrorKind {
+    fn get_error_kind(&self) -> ErrorKind {
         match self {
-            PostgresError::Postgres(e) if e.as_db_error().is_some() => {
-                crate::error::ErrorKind::Postgres
-            }
-            PostgresError::Postgres(_) => crate::error::ErrorKind::Compute,
+            PostgresError::Postgres(err) => match err.as_db_error() {
+                Some(err) if err.code() == &SqlState::INVALID_CATALOG_NAME => ErrorKind::User,
+                Some(_) => ErrorKind::Postgres,
+                None => ErrorKind::Compute,
+            },
         }
     }
 }
@@ -110,9 +112,9 @@ impl UserFacingError for ConnectionError {
 }
 
 impl ReportableError for ConnectionError {
-    fn get_error_kind(&self) -> crate::error::ErrorKind {
+    fn get_error_kind(&self) -> ErrorKind {
         match self {
-            ConnectionError::TlsError(_) => crate::error::ErrorKind::Compute,
+            ConnectionError::TlsError(_) => ErrorKind::Compute,
             ConnectionError::WakeComputeError(e) => e.get_error_kind(),
             ConnectionError::TooManyConnectionAttempts(e) => e.get_error_kind(),
             #[cfg(test)]

proxy/src/config.rs

@@ -19,7 +19,7 @@ use crate::control_plane::messages::{EndpointJwksResponse, JwksSettings};
 use crate::ext::TaskExt;
 use crate::intern::RoleNameInt;
 use crate::rate_limiter::{RateLimitAlgorithm, RateLimiterConfig};
-use crate::scram::threadpool::ThreadPool;
+use crate::scram;
 use crate::serverless::GlobalConnPoolOptions;
 use crate::serverless::cancel_set::CancelSet;
 #[cfg(feature = "rest_broker")]
@@ -75,7 +75,7 @@ pub struct HttpConfig {
 }
 
 pub struct AuthenticationConfig {
-    pub thread_pool: Arc<ThreadPool>,
+    pub scram_thread_pool: Arc<scram::threadpool::ThreadPool>,
     pub scram_protocol_timeout: tokio::time::Duration,
     pub ip_allowlist_check_enabled: bool,
     pub is_vpc_acccess_proxy: bool,

proxy/src/metrics.rs

@@ -5,6 +5,7 @@ use measured::label::{
     FixedCardinalitySet, LabelGroupSet, LabelGroupVisitor, LabelName, LabelSet, LabelValue,
     StaticLabelSet,
 };
+use measured::metric::group::Encoding;
 use measured::metric::histogram::Thresholds;
 use measured::metric::name::MetricName;
 use measured::{
@@ -18,10 +19,10 @@ use crate::control_plane::messages::ColdStartInfo;
 use crate::error::ErrorKind;
 
 #[derive(MetricGroup)]
-#[metric(new(thread_pool: Arc<ThreadPoolMetrics>))]
+#[metric(new())]
 pub struct Metrics {
     #[metric(namespace = "proxy")]
-    #[metric(init = ProxyMetrics::new(thread_pool))]
+    #[metric(init = ProxyMetrics::new())]
     pub proxy: ProxyMetrics,
 
     #[metric(namespace = "wake_compute_lock")]
@@ -34,34 +35,27 @@ pub struct Metrics {
     pub cache: CacheMetrics,
 }
 
-static SELF: OnceLock<Metrics> = OnceLock::new();
 impl Metrics {
-    pub fn install(thread_pool: Arc<ThreadPoolMetrics>) {
-        let mut metrics = Metrics::new(thread_pool);
-
-        metrics.proxy.errors_total.init_all_dense();
-        metrics.proxy.redis_errors_total.init_all_dense();
-        metrics.proxy.redis_events_count.init_all_dense();
-        metrics.proxy.retries_metric.init_all_dense();
-        metrics.proxy.connection_failures_total.init_all_dense();
-
-        SELF.set(metrics)
-            .ok()
-            .expect("proxy metrics must not be installed more than once");
-    }
-
+    #[track_caller]
     pub fn get() -> &'static Self {
-        #[cfg(test)]
-        return SELF.get_or_init(|| Metrics::new(Arc::new(ThreadPoolMetrics::new(0))));
+        static SELF: OnceLock<Metrics> = OnceLock::new();
 
-        #[cfg(not(test))]
-        SELF.get()
-            .expect("proxy metrics must be installed by the main() function")
+        SELF.get_or_init(|| {
+            let mut metrics = Metrics::new();
+
+            metrics.proxy.errors_total.init_all_dense();
+            metrics.proxy.redis_errors_total.init_all_dense();
+            metrics.proxy.redis_events_count.init_all_dense();
+            metrics.proxy.retries_metric.init_all_dense();
+            metrics.proxy.connection_failures_total.init_all_dense();
+
+            metrics
+        })
     }
 }
 
 #[derive(MetricGroup)]
-#[metric(new(thread_pool: Arc<ThreadPoolMetrics>))]
+#[metric(new())]
 pub struct ProxyMetrics {
     #[metric(flatten)]
     pub db_connections: CounterPairVec<NumDbConnectionsGauge>,
@@ -134,6 +128,9 @@ pub struct ProxyMetrics {
     /// Number of TLS handshake failures
     pub tls_handshake_failures: Counter,
 
+    /// Number of SHA 256 rounds executed.
+    pub sha_rounds: Counter,
+
     /// HLL approximate cardinality of endpoints that are connecting
     pub connecting_endpoints: HyperLogLogVec<StaticLabelSet<Protocol>, 32>,
 
@@ -151,8 +148,25 @@ pub struct ProxyMetrics {
     pub connect_compute_lock: ApiLockMetrics,
 
     #[metric(namespace = "scram_pool")]
-    #[metric(init = thread_pool)]
-    pub scram_pool: Arc<ThreadPoolMetrics>,
+    pub scram_pool: OnceLockWrapper<Arc<ThreadPoolMetrics>>,
+}
+
+/// A Wrapper over [`OnceLock`] to implement [`MetricGroup`].
+pub struct OnceLockWrapper<T>(pub OnceLock<T>);
+
+impl<T> Default for OnceLockWrapper<T> {
+    fn default() -> Self {
+        Self(OnceLock::new())
+    }
+}
+
+impl<Enc: Encoding, T: MetricGroup<Enc>> MetricGroup<Enc> for OnceLockWrapper<T> {
+    fn collect_group_into(&self, enc: &mut Enc) -> Result<(), Enc::Err> {
+        if let Some(inner) = self.0.get() {
+            inner.collect_group_into(enc)?;
+        }
+        Ok(())
+    }
 }
 
 #[derive(MetricGroup)]
@@ -719,6 +733,7 @@ pub enum CacheKind {
     ProjectInfoEndpoints,
     ProjectInfoRoles,
     Schema,
+    Pbkdf2,
 }
 
 #[derive(FixedCardinalityLabel, Clone, Copy, Debug)]

proxy/src/scram/cache.rs

@@ -0,0 +1,84 @@
+use tokio::time::Instant;
+use zeroize::Zeroize as _;
+
+use super::pbkdf2;
+use crate::cache::Cached;
+use crate::cache::common::{Cache, count_cache_insert, count_cache_outcome, eviction_listener};
+use crate::intern::{EndpointIdInt, RoleNameInt};
+use crate::metrics::{CacheKind, Metrics};
+
+pub(crate) struct Pbkdf2Cache(moka::sync::Cache<(EndpointIdInt, RoleNameInt), Pbkdf2CacheEntry>);
+pub(crate) type CachedPbkdf2<'a> = Cached<&'a Pbkdf2Cache>;
+
+impl Cache for Pbkdf2Cache {
+    type Key = (EndpointIdInt, RoleNameInt);
+    type Value = Pbkdf2CacheEntry;
+
+    fn invalidate(&self, info: &(EndpointIdInt, RoleNameInt)) {
+        self.0.invalidate(info);
+    }
+}
+
+/// To speed up password hashing for more active customers, we store the tail results of the
+/// PBKDF2 algorithm. If the output of PBKDF2 is U1 ^ U2 ^ ⋯ ^ Uc, then we store
+/// suffix = U17 ^ U18 ^ ⋯ ^ Uc. We only need to calculate U1 ^ U2 ^ ⋯ ^ U15 ^ U16
+/// to determine the final result.
+///
+/// The suffix alone isn't enough to crack the password. The stored_key is still required.
+/// While both are cached in memory, given they're in different locations is makes it much
+/// harder to exploit, even if any such memory exploit exists in proxy.
+#[derive(Clone)]
+pub struct Pbkdf2CacheEntry {
+    /// corresponds to [`super::ServerSecret::cached_at`]
+    pub(super) cached_from: Instant,
+    pub(super) suffix: pbkdf2::Block,
+}
+
+impl Drop for Pbkdf2CacheEntry {
+    fn drop(&mut self) {
+        self.suffix.zeroize();
+    }
+}
+
+impl Pbkdf2Cache {
+    pub fn new() -> Self {
+        const SIZE: u64 = 100;
+        const TTL: std::time::Duration = std::time::Duration::from_secs(60);
+
+        let builder = moka::sync::Cache::builder()
+            .name("pbkdf2")
+            .max_capacity(SIZE)
+            // We use time_to_live so we don't refresh the lifetime for an invalid password attempt.
+            .time_to_live(TTL);
+
+        Metrics::get()
+            .cache
+            .capacity
+            .set(CacheKind::Pbkdf2, SIZE as i64);
+
+        let builder =
+            builder.eviction_listener(|_k, _v, cause| eviction_listener(CacheKind::Pbkdf2, cause));
+
+        Self(builder.build())
+    }
+
+    pub fn insert(&self, endpoint: EndpointIdInt, role: RoleNameInt, value: Pbkdf2CacheEntry) {
+        count_cache_insert(CacheKind::Pbkdf2);
+        self.0.insert((endpoint, role), value);
+    }
+
+    fn get(&self, endpoint: EndpointIdInt, role: RoleNameInt) -> Option<Pbkdf2CacheEntry> {
+        count_cache_outcome(CacheKind::Pbkdf2, self.0.get(&(endpoint, role)))
+    }
+
+    pub fn get_entry(
+        &self,
+        endpoint: EndpointIdInt,
+        role: RoleNameInt,
+    ) -> Option<CachedPbkdf2<'_>> {
+        self.get(endpoint, role).map(|value| Cached {
+            token: Some((self, (endpoint, role))),
+            value,
+        })
+    }
+}

proxy/src/scram/exchange.rs

@@ -4,10 +4,8 @@ use std::convert::Infallible;
 
 use base64::Engine as _;
 use base64::prelude::BASE64_STANDARD;
-use hmac::{Hmac, Mac};
-use sha2::Sha256;
+use tracing::{debug, trace};
 
-use super::ScramKey;
 use super::messages::{
     ClientFinalMessage, ClientFirstMessage, OwnedServerFirstMessage, SCRAM_RAW_NONCE_LEN,
 };
@@ -15,8 +13,10 @@ use super::pbkdf2::Pbkdf2;
 use super::secret::ServerSecret;
 use super::signature::SignatureBuilder;
 use super::threadpool::ThreadPool;
-use crate::intern::EndpointIdInt;
+use super::{ScramKey, pbkdf2};
+use crate::intern::{EndpointIdInt, RoleNameInt};
 use crate::sasl::{self, ChannelBinding, Error as SaslError};
+use crate::scram::cache::Pbkdf2CacheEntry;
 
 /// The only channel binding mode we currently support.
 #[derive(Debug)]
@@ -77,46 +77,113 @@ impl<'a> Exchange<'a> {
     }
 }
 
-// copied from <https://github.com/neondatabase/rust-postgres/blob/20031d7a9ee1addeae6e0968e3899ae6bf01cee2/postgres-protocol/src/authentication/sasl.rs#L236-L248>
 async fn derive_client_key(
     pool: &ThreadPool,
     endpoint: EndpointIdInt,
     password: &[u8],
     salt: &[u8],
     iterations: u32,
-) -> ScramKey {
-    let salted_password = pool
-        .spawn_job(endpoint, Pbkdf2::start(password, salt, iterations))
-        .await;
-
-    let make_key = |name| {
-        let key = Hmac::<Sha256>::new_from_slice(&salted_password)
-            .expect("HMAC is able to accept all key sizes")
-            .chain_update(name)
-            .finalize();
-
-        <[u8; 32]>::from(key.into_bytes())
-    };
-
-    make_key(b"Client Key").into()
+) -> pbkdf2::Block {
+    pool.spawn_job(endpoint, Pbkdf2::start(password, salt, iterations))
+        .await
 }
 
+/// For cleartext flow, we need to derive the client key to
+/// 1. authenticate the client.
+/// 2. authenticate with compute.
 pub(crate) async fn exchange(
     pool: &ThreadPool,
     endpoint: EndpointIdInt,
+    role: RoleNameInt,
+    secret: &ServerSecret,
+    password: &[u8],
+) -> sasl::Result<sasl::Outcome<super::ScramKey>> {
+    if secret.iterations > CACHED_ROUNDS {
+        exchange_with_cache(pool, endpoint, role, secret, password).await
+    } else {
+        let salt = BASE64_STANDARD.decode(&*secret.salt_base64)?;
+        let hash = derive_client_key(pool, endpoint, password, &salt, secret.iterations).await;
+        Ok(validate_pbkdf2(secret, &hash))
+    }
+}
+
+/// Compute the client key using a cache. We cache the suffix of the pbkdf2 result only,
+/// which is not enough by itself to perform an offline brute force.
+async fn exchange_with_cache(
+    pool: &ThreadPool,
+    endpoint: EndpointIdInt,
+    role: RoleNameInt,
     secret: &ServerSecret,
     password: &[u8],
 ) -> sasl::Result<sasl::Outcome<super::ScramKey>> {
     let salt = BASE64_STANDARD.decode(&*secret.salt_base64)?;
-    let client_key = derive_client_key(pool, endpoint, password, &salt, secret.iterations).await;
 
+    debug_assert!(
+        secret.iterations > CACHED_ROUNDS,
+        "we should not cache password data if there isn't enough rounds needed"
+    );
+
+    // compute the prefix of the pbkdf2 output.
+    let prefix = derive_client_key(pool, endpoint, password, &salt, CACHED_ROUNDS).await;
+
+    if let Some(entry) = pool.cache.get_entry(endpoint, role) {
+        // hot path: let's check the threadpool cache
+        if secret.cached_at == entry.cached_from {
+            // cache is valid. compute the full hash by adding the prefix to the suffix.
+            let mut hash = prefix;
+            pbkdf2::xor_assign(&mut hash, &entry.suffix);
+            let outcome = validate_pbkdf2(secret, &hash);
+
+            if matches!(outcome, sasl::Outcome::Success(_)) {
+                trace!("password validated from cache");
+            }
+
+            return Ok(outcome);
+        }
+
+        // cached key is no longer valid.
+        debug!("invalidating cached password");
+        entry.invalidate();
+    }
+
+    // slow path: full password hash.
+    let hash = derive_client_key(pool, endpoint, password, &salt, secret.iterations).await;
+    let outcome = validate_pbkdf2(secret, &hash);
+
+    let client_key = match outcome {
+        sasl::Outcome::Success(client_key) => client_key,
+        sasl::Outcome::Failure(_) => return Ok(outcome),
+    };
+
+    trace!("storing cached password");
+
+    // time to cache, compute the suffix by subtracting the prefix from the hash.
+    let mut suffix = hash;
+    pbkdf2::xor_assign(&mut suffix, &prefix);
+
+    pool.cache.insert(
+        endpoint,
+        role,
+        Pbkdf2CacheEntry {
+            cached_from: secret.cached_at,
+            suffix,
+        },
+    );
+
+    Ok(sasl::Outcome::Success(client_key))
+}
+
+fn validate_pbkdf2(secret: &ServerSecret, hash: &pbkdf2::Block) -> sasl::Outcome<ScramKey> {
+    let client_key = super::ScramKey::client_key(&(*hash).into());
     if secret.is_password_invalid(&client_key).into() {
-        Ok(sasl::Outcome::Failure("password doesn't match"))
+        sasl::Outcome::Failure("password doesn't match")
     } else {
-        Ok(sasl::Outcome::Success(client_key))
+        sasl::Outcome::Success(client_key)
     }
 }
 
+const CACHED_ROUNDS: u32 = 16;
+
 impl SaslInitial {
     fn transition(
         &self,

proxy/src/scram/key.rs

@@ -1,6 +1,12 @@
 //! Tools for client/server/stored key management.
 
+use hmac::Mac as _;
+use sha2::Digest as _;
 use subtle::ConstantTimeEq;
+use zeroize::Zeroize as _;
+
+use crate::metrics::Metrics;
+use crate::scram::pbkdf2::Prf;
 
 /// Faithfully taken from PostgreSQL.
 pub(crate) const SCRAM_KEY_LEN: usize = 32;
@@ -14,6 +20,12 @@ pub(crate) struct ScramKey {
     bytes: [u8; SCRAM_KEY_LEN],
 }
 
+impl Drop for ScramKey {
+    fn drop(&mut self) {
+        self.bytes.zeroize();
+    }
+}
+
 impl PartialEq for ScramKey {
     fn eq(&self, other: &Self) -> bool {
         self.ct_eq(other).into()
@@ -28,12 +40,26 @@ impl ConstantTimeEq for ScramKey {
 
 impl ScramKey {
     pub(crate) fn sha256(&self) -> Self {
-        super::sha256([self.as_ref()]).into()
+        Metrics::get().proxy.sha_rounds.inc_by(1);
+        Self {
+            bytes: sha2::Sha256::digest(self.as_bytes()).into(),
+        }
     }
 
     pub(crate) fn as_bytes(&self) -> [u8; SCRAM_KEY_LEN] {
         self.bytes
     }
+
+    pub(crate) fn client_key(b: &[u8; 32]) -> Self {
+        // Prf::new_from_slice will run 2 sha256 rounds.
+        // Update + Finalize run 2 sha256 rounds.
+        Metrics::get().proxy.sha_rounds.inc_by(4);
+
+        let mut prf = Prf::new_from_slice(b).expect("HMAC is able to accept all key sizes");
+        prf.update(b"Client Key");
+        let client_key: [u8; 32] = prf.finalize().into_bytes().into();
+        client_key.into()
+    }
 }
 
 impl From<[u8; SCRAM_KEY_LEN]> for ScramKey {

proxy/src/scram/mod.rs

@@ -6,6 +6,7 @@
 //! * <https://github.com/postgres/postgres/blob/94226d4506e66d6e7cbf4b391f1e7393c1962841/src/backend/libpq/auth-scram.c>
 //! * <https://github.com/postgres/postgres/blob/94226d4506e66d6e7cbf4b391f1e7393c1962841/src/interfaces/libpq/fe-auth-scram.c>
 
+mod cache;
 mod countmin;
 mod exchange;
 mod key;
@@ -18,10 +19,8 @@ pub mod threadpool;
 use base64::Engine as _;
 use base64::prelude::BASE64_STANDARD;
 pub(crate) use exchange::{Exchange, exchange};
-use hmac::{Hmac, Mac};
 pub(crate) use key::ScramKey;
 pub(crate) use secret::ServerSecret;
-use sha2::{Digest, Sha256};
 
 const SCRAM_SHA_256: &str = "SCRAM-SHA-256";
 const SCRAM_SHA_256_PLUS: &str = "SCRAM-SHA-256-PLUS";
@@ -42,29 +41,13 @@ fn base64_decode_array<const N: usize>(input: impl AsRef<[u8]>) -> Option<[u8; N
     Some(bytes)
 }
 
-/// This function essentially is `Hmac(sha256, key, input)`.
-/// Further reading: <https://datatracker.ietf.org/doc/html/rfc2104>.
-fn hmac_sha256<'a>(key: &[u8], parts: impl IntoIterator<Item = &'a [u8]>) -> [u8; 32] {
-    let mut mac = Hmac::<Sha256>::new_from_slice(key).expect("bad key size");
-    parts.into_iter().for_each(|s| mac.update(s));
-
-    mac.finalize().into_bytes().into()
-}
-
-fn sha256<'a>(parts: impl IntoIterator<Item = &'a [u8]>) -> [u8; 32] {
-    let mut hasher = Sha256::new();
-    parts.into_iter().for_each(|s| hasher.update(s));
-
-    hasher.finalize().into()
-}
-
 #[cfg(test)]
 mod tests {
     use super::threadpool::ThreadPool;
     use super::{Exchange, ServerSecret};
-    use crate::intern::EndpointIdInt;
+    use crate::intern::{EndpointIdInt, RoleNameInt};
     use crate::sasl::{Mechanism, Step};
-    use crate::types::EndpointId;
+    use crate::types::{EndpointId, RoleName};
 
     #[test]
     fn snapshot() {
@@ -114,23 +97,34 @@ mod tests {
         );
     }
 
-    async fn run_round_trip_test(server_password: &str, client_password: &str) {
-        let pool = ThreadPool::new(1);
-
+    async fn check(
+        pool: &ThreadPool,
+        scram_secret: &ServerSecret,
+        password: &[u8],
+    ) -> Result<(), &'static str> {
         let ep = EndpointId::from("foo");
         let ep = EndpointIdInt::from(ep);
+        let role = RoleName::from("user");
+        let role = RoleNameInt::from(&role);
 
-        let scram_secret = ServerSecret::build(server_password).await.unwrap();
-        let outcome = super::exchange(&pool, ep, &scram_secret, client_password.as_bytes())
+        let outcome = super::exchange(pool, ep, role, scram_secret, password)
             .await
             .unwrap();
 
         match outcome {
-            crate::sasl::Outcome::Success(_) => {}
-            crate::sasl::Outcome::Failure(r) => panic!("{r}"),
+            crate::sasl::Outcome::Success(_) => Ok(()),
+            crate::sasl::Outcome::Failure(r) => Err(r),
         }
     }
 
+    async fn run_round_trip_test(server_password: &str, client_password: &str) {
+        let pool = ThreadPool::new(1);
+        let scram_secret = ServerSecret::build(server_password).await.unwrap();
+        check(&pool, &scram_secret, client_password.as_bytes())
+            .await
+            .unwrap();
+    }
+
     #[tokio::test]
     async fn round_trip() {
         run_round_trip_test("pencil", "pencil").await;
@@ -141,4 +135,27 @@ mod tests {
     async fn failure() {
         run_round_trip_test("pencil", "eraser").await;
     }
+
+    #[tokio::test]
+    #[tracing_test::traced_test]
+    async fn password_cache() {
+        let pool = ThreadPool::new(1);
+        let scram_secret = ServerSecret::build("password").await.unwrap();
+
+        // wrong passwords are not added to cache
+        check(&pool, &scram_secret, b"wrong").await.unwrap_err();
+        assert!(!logs_contain("storing cached password"));
+
+        // correct passwords get cached
+        check(&pool, &scram_secret, b"password").await.unwrap();
+        assert!(logs_contain("storing cached password"));
+
+        // wrong passwords do not match the cache
+        check(&pool, &scram_secret, b"wrong").await.unwrap_err();
+        assert!(!logs_contain("password validated from cache"));
+
+        // correct passwords match the cache
+        check(&pool, &scram_secret, b"password").await.unwrap();
+        assert!(logs_contain("password validated from cache"));
+    }
 }

proxy/src/scram/pbkdf2.rs

@@ -1,25 +1,50 @@
+//! For postgres password authentication, we need to perform a PBKDF2 using
+//! PRF=HMAC-SHA2-256, producing only 1 block (32 bytes) of output key.
+
+use hmac::Mac as _;
 use hmac::digest::consts::U32;
 use hmac::digest::generic_array::GenericArray;
-use hmac::{Hmac, Mac};
-use sha2::Sha256;
+use zeroize::Zeroize as _;
+
+use crate::metrics::Metrics;
+
+/// The Psuedo-random function used during PBKDF2 and the SCRAM-SHA-256 handshake.
+pub type Prf = hmac::Hmac<sha2::Sha256>;
+pub(crate) type Block = GenericArray<u8, U32>;
 
 pub(crate) struct Pbkdf2 {
-    hmac: Hmac<Sha256>,
-    prev: GenericArray<u8, U32>,
-    hi: GenericArray<u8, U32>,
+    hmac: Prf,
+    /// U{r-1} for whatever iteration r we are currently on.
+    prev: Block,
+    /// the output of `fold(xor, U{1}..U{r})` for whatever iteration r we are currently on.
+    hi: Block,
+    /// number of iterations left
     iterations: u32,
 }
 
+impl Drop for Pbkdf2 {
+    fn drop(&mut self) {
+        self.prev.zeroize();
+        self.hi.zeroize();
+    }
+}
+
 // inspired from <https://github.com/neondatabase/rust-postgres/blob/20031d7a9ee1addeae6e0968e3899ae6bf01cee2/postgres-protocol/src/authentication/sasl.rs#L36-L61>
 impl Pbkdf2 {
-    pub(crate) fn start(str: &[u8], salt: &[u8], iterations: u32) -> Self {
+    pub(crate) fn start(pw: &[u8], salt: &[u8], iterations: u32) -> Self {
         // key the HMAC and derive the first block in-place
-        let mut hmac =
-            Hmac::<Sha256>::new_from_slice(str).expect("HMAC is able to accept all key sizes");
+        let mut hmac = Prf::new_from_slice(pw).expect("HMAC is able to accept all key sizes");
+
+        // U1 = PRF(Password, Salt + INT_32_BE(i))
+        // i = 1 since we only need 1 block of output.
         hmac.update(salt);
         hmac.update(&1u32.to_be_bytes());
         let init_block = hmac.finalize_reset().into_bytes();
 
+        // Prf::new_from_slice will run 2 sha256 rounds.
+        // Our update + finalize run 2 sha256 rounds for each pbkdf2 round.
+        Metrics::get().proxy.sha_rounds.inc_by(4);
+
         Self {
             hmac,
             // one iteration spent above
@@ -33,7 +58,11 @@ impl Pbkdf2 {
         (self.iterations).clamp(0, 4096)
     }
 
-    pub(crate) fn turn(&mut self) -> std::task::Poll<[u8; 32]> {
+    /// For "fairness", we implement PBKDF2 with cooperative yielding, which is why we use this `turn`
+    /// function that only executes a fixed number of iterations before continuing.
+    ///
+    /// Task must be rescheuled if this returns [`std::task::Poll::Pending`].
+    pub(crate) fn turn(&mut self) -> std::task::Poll<Block> {
         let Self {
             hmac,
             prev,
@@ -44,25 +73,37 @@ impl Pbkdf2 {
         // only do up to 4096 iterations per turn for fairness
         let n = (*iterations).clamp(0, 4096);
         for _ in 0..n {
-            hmac.update(prev);
-            let block = hmac.finalize_reset().into_bytes();
-
-            for (hi_byte, &b) in hi.iter_mut().zip(block.iter()) {
-                *hi_byte ^= b;
-            }
-
-            *prev = block;
+            let next = single_round(hmac, prev);
+            xor_assign(hi, &next);
+            *prev = next;
         }
 
+        // Our update + finalize run 2 sha256 rounds for each pbkdf2 round.
+        Metrics::get().proxy.sha_rounds.inc_by(2 * n as u64);
+
         *iterations -= n;
         if *iterations == 0 {
-            std::task::Poll::Ready((*hi).into())
+            std::task::Poll::Ready(*hi)
         } else {
             std::task::Poll::Pending
         }
     }
 }
 
+#[inline(always)]
+pub fn xor_assign(x: &mut Block, y: &Block) {
+    for (x, &y) in std::iter::zip(x, y) {
+        *x ^= y;
+    }
+}
+
+#[inline(always)]
+fn single_round(prf: &mut Prf, ui: &Block) -> Block {
+    // Ui = PRF(Password, Ui-1)
+    prf.update(ui);
+    prf.finalize_reset().into_bytes()
+}
+
 #[cfg(test)]
 mod tests {
     use pbkdf2::pbkdf2_hmac_array;
@@ -76,11 +117,11 @@ mod tests {
         let pass = b"Ne0n_!5_50_C007";
 
         let mut job = Pbkdf2::start(pass, salt, 60000);
-        let hash = loop {
+        let hash: [u8; 32] = loop {
             let std::task::Poll::Ready(hash) = job.turn() else {
                 continue;
             };
-            break hash;
+            break hash.into();
         };
 
         let expected = pbkdf2_hmac_array::<Sha256, 32>(pass, salt, 60000);

proxy/src/scram/secret.rs

@@ -3,6 +3,7 @@
 use base64::Engine as _;
 use base64::prelude::BASE64_STANDARD;
 use subtle::{Choice, ConstantTimeEq};
+use tokio::time::Instant;
 
 use super::base64_decode_array;
 use super::key::ScramKey;
@@ -11,6 +12,9 @@ use super::key::ScramKey;
 /// and is used throughout the authentication process.
 #[derive(Clone, Eq, PartialEq, Debug)]
 pub(crate) struct ServerSecret {
+    /// When this secret was cached.
+    pub(crate) cached_at: Instant,
+
     /// Number of iterations for `PBKDF2` function.
     pub(crate) iterations: u32,
     /// Salt used to hash user's password.
@@ -34,6 +38,7 @@ impl ServerSecret {
             params.split_once(':').zip(keys.split_once(':'))?;
 
         let secret = ServerSecret {
+            cached_at: Instant::now(),
             iterations: iterations.parse().ok()?,
             salt_base64: salt.into(),
             stored_key: base64_decode_array(stored_key)?.into(),
@@ -54,6 +59,7 @@ impl ServerSecret {
     /// See `auth-scram.c : mock_scram_secret` for details.
     pub(crate) fn mock(nonce: [u8; 32]) -> Self {
         Self {
+            cached_at: Instant::now(),
             // this doesn't reveal much information as we're going to use
             // iteration count 1 for our generated passwords going forward.
             // PG16 users can set iteration count=1 already today.

proxy/src/scram/signature.rs

@@ -1,6 +1,10 @@
 //! Tools for client/server signature management.
 
+use hmac::Mac as _;
+
 use super::key::{SCRAM_KEY_LEN, ScramKey};
+use crate::metrics::Metrics;
+use crate::scram::pbkdf2::Prf;
 
 /// A collection of message parts needed to derive the client's signature.
 #[derive(Debug)]
@@ -12,15 +16,18 @@ pub(crate) struct SignatureBuilder<'a> {
 
 impl SignatureBuilder<'_> {
     pub(crate) fn build(&self, key: &ScramKey) -> Signature {
-        let parts = [
-            self.client_first_message_bare.as_bytes(),
-            b",",
-            self.server_first_message.as_bytes(),
-            b",",
-            self.client_final_message_without_proof.as_bytes(),
-        ];
+        // don't know exactly. this is a rough approx
+        Metrics::get().proxy.sha_rounds.inc_by(8);
 
-        super::hmac_sha256(key.as_ref(), parts).into()
+        let mut mac = Prf::new_from_slice(key.as_ref()).expect("HMAC accepts all key sizes");
+        mac.update(self.client_first_message_bare.as_bytes());
+        mac.update(b",");
+        mac.update(self.server_first_message.as_bytes());
+        mac.update(b",");
+        mac.update(self.client_final_message_without_proof.as_bytes());
+        Signature {
+            bytes: mac.finalize().into_bytes().into(),
+        }
     }
 }
 

proxy/src/scram/threadpool.rs

@@ -15,6 +15,8 @@ use futures::FutureExt;
 use rand::rngs::SmallRng;
 use rand::{Rng, SeedableRng};
 
+use super::cache::Pbkdf2Cache;
+use super::pbkdf2;
 use super::pbkdf2::Pbkdf2;
 use crate::intern::EndpointIdInt;
 use crate::metrics::{ThreadPoolMetrics, ThreadPoolWorkerId};
@@ -23,6 +25,10 @@ use crate::scram::countmin::CountMinSketch;
 pub struct ThreadPool {
     runtime: Option<tokio::runtime::Runtime>,
     pub metrics: Arc<ThreadPoolMetrics>,
+
+    // we hash a lot of passwords.
+    // we keep a cache of partial hashes for faster validation.
+    pub(super) cache: Pbkdf2Cache,
 }
 
 /// How often to reset the sketch values
@@ -68,6 +74,7 @@ impl ThreadPool {
             Self {
                 runtime: Some(runtime),
                 metrics: Arc::new(ThreadPoolMetrics::new(n_workers as usize)),
+                cache: Pbkdf2Cache::new(),
             }
         })
     }
@@ -130,7 +137,7 @@ struct JobSpec {
 }
 
 impl Future for JobSpec {
-    type Output = [u8; 32];
+    type Output = pbkdf2::Block;
 
     fn poll(mut self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<Self::Output> {
         STATE.with_borrow_mut(|state| {
@@ -166,10 +173,10 @@ impl Future for JobSpec {
     }
 }
 
-pub(crate) struct JobHandle(tokio::task::JoinHandle<[u8; 32]>);
+pub(crate) struct JobHandle(tokio::task::JoinHandle<pbkdf2::Block>);
 
 impl Future for JobHandle {
-    type Output = [u8; 32];
+    type Output = pbkdf2::Block;
 
     fn poll(mut self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<Self::Output> {
         match self.0.poll_unpin(cx) {
@@ -203,10 +210,10 @@ mod tests {
             .spawn_job(ep, Pbkdf2::start(b"password", &salt, 4096))
             .await;
 
-        let expected = [
+        let expected = &[
             10, 114, 73, 188, 140, 222, 196, 156, 214, 184, 79, 157, 119, 242, 16, 31, 53, 242,
             178, 43, 95, 8, 225, 182, 122, 40, 219, 21, 89, 147, 64, 140,
         ];
-        assert_eq!(actual, expected);
+        assert_eq!(actual.as_slice(), expected);
     }
 }

proxy/src/serverless/backend.rs

@@ -4,6 +4,7 @@ use std::time::Duration;
 use ed25519_dalek::SigningKey;
 use hyper_util::rt::{TokioExecutor, TokioIo, TokioTimer};
 use jose_jwk::jose_b64;
+use postgres_client::error::SqlState;
 use postgres_client::maybe_tls_stream::MaybeTlsStream;
 use rand_core::OsRng;
 use tracing::field::display;
@@ -26,7 +27,7 @@ use crate::context::RequestContext;
 use crate::control_plane::client::ApiLockError;
 use crate::control_plane::errors::{GetAuthInfoError, WakeComputeError};
 use crate::error::{ErrorKind, ReportableError, UserFacingError};
-use crate::intern::EndpointIdInt;
+use crate::intern::{EndpointIdInt, RoleNameInt};
 use crate::pqproto::StartupMessageParams;
 use crate::proxy::{connect_auth, connect_compute};
 use crate::rate_limiter::EndpointRateLimiter;
@@ -76,9 +77,11 @@ impl PoolingBackend {
         };
 
         let ep = EndpointIdInt::from(&user_info.endpoint);
+        let role = RoleNameInt::from(&user_info.user);
         let auth_outcome = crate::auth::validate_password_and_exchange(
-            &self.config.authentication_config.thread_pool,
+            &self.config.authentication_config.scram_thread_pool,
             ep,
+            role,
             password,
             secret,
         )
@@ -457,15 +460,14 @@ impl ReportableError for HttpConnError {
         match self {
             HttpConnError::ConnectError(_) => ErrorKind::Compute,
             HttpConnError::ConnectionClosedAbruptly(_) => ErrorKind::Compute,
-            HttpConnError::PostgresConnectionError(p) => {
-                if p.as_db_error().is_some() {
-                    // postgres rejected the connection
-                    ErrorKind::Postgres
-                } else {
-                    // couldn't even reach postgres
-                    ErrorKind::Compute
-                }
-            }
+            HttpConnError::PostgresConnectionError(p) => match p.as_db_error() {
+                // user provided a wrong database name
+                Some(err) if err.code() == &SqlState::INVALID_CATALOG_NAME => ErrorKind::User,
+                // postgres rejected the connection
+                Some(_) => ErrorKind::Postgres,
+                // couldn't even reach postgres
+                None => ErrorKind::Compute,
+            },
             HttpConnError::LocalProxyConnectionError(_) => ErrorKind::Compute,
             HttpConnError::ComputeCtl(_) => ErrorKind::Service,
             HttpConnError::JwtPayloadError(_) => ErrorKind::User,

proxy/src/serverless/sql_over_http.rs

@@ -192,34 +192,29 @@ pub(crate) async fn handle(
             let line = get(db_error, |db| db.line().map(|l| l.to_string()));
             let routine = get(db_error, |db| db.routine());
 
-            match &e {
-                SqlOverHttpError::Postgres(e)
-                    if e.as_db_error().is_some() && error_kind == ErrorKind::User =>
-                {
-                    // this error contains too much info, and it's not an error we care about.
-                    if tracing::enabled!(Level::DEBUG) {
-                        tracing::debug!(
-                            kind=error_kind.to_metric_label(),
-                            error=%e,
-                            msg=message,
-                            "forwarding error to user"
-                        );
-                    } else {
-                        tracing::info!(
-                            kind = error_kind.to_metric_label(),
-                            error = "bad query",
-                            "forwarding error to user"
-                        );
-                    }
-                }
-                _ => {
-                    tracing::info!(
+            if db_error.is_some() && error_kind == ErrorKind::User {
+                // this error contains too much info, and it's not an error we care about.
+                if tracing::enabled!(Level::DEBUG) {
+                    debug!(
                         kind=error_kind.to_metric_label(),
                         error=%e,
                         msg=message,
                         "forwarding error to user"
                     );
+                } else {
+                    info!(
+                        kind = error_kind.to_metric_label(),
+                        error = "bad query",
+                        "forwarding error to user"
+                    );
                 }
+            } else {
+                info!(
+                    kind=error_kind.to_metric_label(),
+                    error=%e,
+                    msg=message,
+                    "forwarding error to user"
+                );
             }
 
             json_response(

safekeeper/src/hadron.rs

@@ -387,6 +387,7 @@ pub fn get_filesystem_usage(path: &std::path::Path) -> u64 {
             critical_timeline!(
                 placeholder_ttid.tenant_id,
                 placeholder_ttid.timeline_id,
+                None::<&AtomicBool>,
                 "Global disk usage watcher failed to read filesystem usage: {:?}",
                 e
             );

safekeeper/src/metrics.rs

@@ -518,6 +518,7 @@ pub async fn time_io_closure<E: Into<anyhow::Error>>(
 pub struct FullTimelineInfo {
     pub ttid: TenantTimelineId,
     pub ps_feedback_count: u64,
+    pub ps_corruption_detected: bool,
     pub last_ps_feedback: PageserverFeedback,
     pub wal_backup_active: bool,
     pub timeline_is_active: bool,
@@ -547,6 +548,7 @@ pub struct TimelineCollector {
     ps_last_received_lsn: GenericGaugeVec<AtomicU64>,
     feedback_last_time_seconds: GenericGaugeVec<AtomicU64>,
     ps_feedback_count: GenericGaugeVec<AtomicU64>,
+    ps_corruption_detected: IntGaugeVec,
     timeline_active: GenericGaugeVec<AtomicU64>,
     wal_backup_active: GenericGaugeVec<AtomicU64>,
     connected_computes: IntGaugeVec,
@@ -654,6 +656,15 @@ impl TimelineCollector {
         )
         .unwrap();
 
+        let ps_corruption_detected = IntGaugeVec::new(
+            Opts::new(
+                "safekeeper_ps_corruption_detected",
+                "1 if corruption was detected in the timeline according to feedback from the pageserver, 0 otherwise",
+            ),
+            &["tenant_id", "timeline_id"],
+        )
+        .unwrap();
+
         let timeline_active = GenericGaugeVec::new(
             Opts::new(
                 "safekeeper_timeline_active",
@@ -774,6 +785,7 @@ impl TimelineCollector {
             ps_last_received_lsn,
             feedback_last_time_seconds,
             ps_feedback_count,
+            ps_corruption_detected,
             timeline_active,
             wal_backup_active,
             connected_computes,
@@ -892,6 +904,9 @@ impl Collector for TimelineCollector {
             self.ps_feedback_count
                 .with_label_values(labels)
                 .set(tli.ps_feedback_count);
+            self.ps_corruption_detected
+                .with_label_values(labels)
+                .set(tli.ps_corruption_detected as i64);
             if let Ok(unix_time) = tli
                 .last_ps_feedback
                 .replytime
@@ -925,6 +940,7 @@ impl Collector for TimelineCollector {
         mfs.extend(self.ps_last_received_lsn.collect());
         mfs.extend(self.feedback_last_time_seconds.collect());
         mfs.extend(self.ps_feedback_count.collect());
+        mfs.extend(self.ps_corruption_detected.collect());
         mfs.extend(self.timeline_active.collect());
         mfs.extend(self.wal_backup_active.collect());
         mfs.extend(self.connected_computes.collect());

safekeeper/src/send_interpreted_wal.rs

@@ -1,6 +1,7 @@
 use std::collections::HashMap;
 use std::fmt::Display;
 use std::sync::Arc;
+use std::sync::atomic::AtomicBool;
 use std::time::Duration;
 
 use anyhow::{Context, anyhow};
@@ -305,6 +306,9 @@ impl InterpretedWalReader {
                             critical_timeline!(
                                 ttid.tenant_id,
                                 ttid.timeline_id,
+                                // Hadron: The corruption flag is only used in PS so that it can feed this information back to SKs.
+                                // We do not use these flags in SKs.
+                                None::<&AtomicBool>,
                                 "failed to read WAL record: {err:?}"
                             );
                         }
@@ -375,6 +379,9 @@ impl InterpretedWalReader {
                 critical_timeline!(
                     ttid.tenant_id,
                     ttid.timeline_id,
+                    // Hadron: The corruption flag is only used in PS so that it can feed this information back to SKs.
+                    // We do not use these flags in SKs.
+                    None::<&AtomicBool>,
                     "failed to decode WAL record: {err:?}"
                 );
             }

safekeeper/src/send_wal.rs

@@ -55,6 +55,7 @@ pub struct WalSenders {
 
 pub struct WalSendersTimelineMetricValues {
     pub ps_feedback_counter: u64,
+    pub ps_corruption_detected: bool,
     pub last_ps_feedback: PageserverFeedback,
     pub interpreted_wal_reader_tasks: usize,
 }
@@ -193,6 +194,7 @@ impl WalSenders {
 
         WalSendersTimelineMetricValues {
             ps_feedback_counter: shared.ps_feedback_counter,
+            ps_corruption_detected: shared.ps_corruption_detected,
             last_ps_feedback: shared.last_ps_feedback,
             interpreted_wal_reader_tasks,
         }
@@ -209,6 +211,9 @@ impl WalSenders {
         *shared.get_slot_mut(id).get_mut_feedback() = ReplicationFeedback::Pageserver(*feedback);
         shared.last_ps_feedback = *feedback;
         shared.ps_feedback_counter += 1;
+        if feedback.corruption_detected {
+            shared.ps_corruption_detected = true;
+        }
         drop(shared);
 
         RECEIVED_PS_FEEDBACKS.inc();
@@ -278,6 +283,9 @@ struct WalSendersShared {
     last_ps_feedback: PageserverFeedback,
     // total counter of pageserver feedbacks received
     ps_feedback_counter: u64,
+    // Hadron: true iff we received a pageserver feedback that incidated
+    // data corruption in the timeline
+    ps_corruption_detected: bool,
     slots: Vec<Option<WalSenderState>>,
 }
 
@@ -328,6 +336,7 @@ impl WalSendersShared {
             agg_standby_feedback: StandbyFeedback::empty(),
             last_ps_feedback: PageserverFeedback::empty(),
             ps_feedback_counter: 0,
+            ps_corruption_detected: false,
             slots: Vec::new(),
         }
     }

safekeeper/src/timeline.rs

@@ -839,6 +839,7 @@ impl Timeline {
 
         let WalSendersTimelineMetricValues {
             ps_feedback_counter,
+            ps_corruption_detected,
             last_ps_feedback,
             interpreted_wal_reader_tasks,
         } = self.walsenders.info_for_metrics();
@@ -847,6 +848,7 @@ impl Timeline {
         Some(FullTimelineInfo {
             ttid: self.ttid,
             ps_feedback_count: ps_feedback_counter,
+            ps_corruption_detected,
             last_ps_feedback,
             wal_backup_active: self.wal_backup_active.load(Ordering::Relaxed),
             timeline_is_active: self.broker_active.load(Ordering::Relaxed),

safekeeper/src/wal_backup.rs

@@ -12,7 +12,7 @@ use futures::stream::{self, FuturesOrdered};
 use postgres_ffi::v14::xlog_utils::XLogSegNoOffsetToRecPtr;
 use postgres_ffi::{PG_TLI, XLogFileName, XLogSegNo};
 use remote_storage::{
-    DownloadOpts, GenericRemoteStorage, ListingMode, RemotePath, StorageMetadata,
+    DownloadError, DownloadOpts, GenericRemoteStorage, ListingMode, RemotePath, StorageMetadata,
 };
 use safekeeper_api::models::PeerInfo;
 use tokio::fs::File;
@@ -607,6 +607,9 @@ pub(crate) async fn copy_partial_segment(
     storage.copy_object(source, destination, &cancel).await
 }
 
+const WAL_READ_WARN_THRESHOLD: u32 = 2;
+const WAL_READ_MAX_RETRIES: u32 = 3;
+
 pub async fn read_object(
     storage: &GenericRemoteStorage,
     file_path: &RemotePath,
@@ -620,12 +623,23 @@ pub async fn read_object(
         byte_start: std::ops::Bound::Included(offset),
         ..Default::default()
     };
-    let download = storage
-        .download(file_path, &opts, &cancel)
-        .await
-        .with_context(|| {
-            format!("Failed to open WAL segment download stream for remote path {file_path:?}")
-        })?;
+
+    // This retry only solves the connect errors: subsequent reads can still fail as this function returns
+    // a stream.
+    let download = backoff::retry(
+        || async { storage.download(file_path, &opts, &cancel).await },
+        DownloadError::is_permanent,
+        WAL_READ_WARN_THRESHOLD,
+        WAL_READ_MAX_RETRIES,
+        "download WAL segment",
+        &cancel,
+    )
+    .await
+    .ok_or_else(|| DownloadError::Cancelled)
+    .and_then(|x| x)
+    .with_context(|| {
+        format!("Failed to open WAL segment download stream for remote path {file_path:?}")
+    })?;
 
     let reader = tokio_util::io::StreamReader::new(download.download_stream);
 

storage_controller/src/compute_hook.rs

@@ -6,13 +6,16 @@ use std::time::Duration;
 
 use anyhow::Context;
 use compute_api::spec::PageserverProtocol;
-use control_plane::endpoint::{ComputeControlPlane, EndpointStatus};
+use compute_api::spec::PageserverShardInfo;
+use control_plane::endpoint::{
+    ComputeControlPlane, EndpointStatus, PageserverConnectionInfo, PageserverShardConnectionInfo,
+};
 use control_plane::local_env::LocalEnv;
 use futures::StreamExt;
 use hyper::StatusCode;
 use pageserver_api::config::DEFAULT_GRPC_LISTEN_PORT;
 use pageserver_api::controller_api::AvailabilityZone;
-use pageserver_api::shard::{ShardCount, ShardNumber, ShardStripeSize, TenantShardId};
+use pageserver_api::shard::{ShardCount, ShardIndex, ShardNumber, ShardStripeSize, TenantShardId};
 use postgres_connection::parse_host_port;
 use safekeeper_api::membership::SafekeeperGeneration;
 use serde::{Deserialize, Serialize};
@@ -506,27 +509,64 @@ impl ApiMethod for ComputeHookTenant {
             if endpoint.tenant_id == *tenant_id && endpoint.status() == EndpointStatus::Running {
                 tracing::info!("Reconfiguring pageservers for endpoint {endpoint_name}");
 
-                let pageservers = shards
-                    .iter()
-                    .map(|shard| {
-                        let ps_conf = env
-                            .get_pageserver_conf(shard.node_id)
-                            .expect("Unknown pageserver");
-                        if endpoint.grpc {
-                            let addr = ps_conf.listen_grpc_addr.as_ref().expect("no gRPC address");
-                            let (host, port) = parse_host_port(addr).expect("invalid gRPC address");
-                            let port = port.unwrap_or(DEFAULT_GRPC_LISTEN_PORT);
-                            (PageserverProtocol::Grpc, host, port)
-                        } else {
-                            let (host, port) = parse_host_port(&ps_conf.listen_pg_addr)
-                                .expect("Unable to parse listen_pg_addr");
-                            (PageserverProtocol::Libpq, host, port.unwrap_or(5432))
-                        }
-                    })
-                    .collect::<Vec<_>>();
+                let shard_count = match shards.len() {
+                    1 => ShardCount::unsharded(),
+                    n => ShardCount(n.try_into().expect("too many shards")),
+                };
+
+                let mut shard_infos: HashMap<ShardIndex, PageserverShardInfo> = HashMap::new();
+
+                let prefer_protocol = if endpoint.grpc {
+                    PageserverProtocol::Grpc
+                } else {
+                    PageserverProtocol::Libpq
+                };
+
+                for shard in shards.iter() {
+                    let ps_conf = env
+                        .get_pageserver_conf(shard.node_id)
+                        .expect("Unknown pageserver");
+
+                    let libpq_url = Some({
+                        let (host, port) = parse_host_port(&ps_conf.listen_pg_addr)
+                            .expect("Unable to parse listen_pg_addr");
+                        let port = port.unwrap_or(5432);
+                        format!("postgres://no_user@{host}:{port}")
+                    });
+                    let grpc_url = if let Some(grpc_addr) = &ps_conf.listen_grpc_addr {
+                        let (host, port) =
+                            parse_host_port(grpc_addr).expect("invalid gRPC address");
+                        let port = port.unwrap_or(DEFAULT_GRPC_LISTEN_PORT);
+                        Some(format!("grpc://no_user@{host}:{port}"))
+                    } else {
+                        None
+                    };
+                    let pageserver = PageserverShardConnectionInfo {
+                        id: Some(shard.node_id),
+                        libpq_url,
+                        grpc_url,
+                    };
+                    let shard_info = PageserverShardInfo {
+                        pageservers: vec![pageserver],
+                    };
+                    shard_infos.insert(
+                        ShardIndex {
+                            shard_number: shard.shard_number,
+                            shard_count,
+                        },
+                        shard_info,
+                    );
+                }
+
+                let pageserver_conninfo = PageserverConnectionInfo {
+                    shard_count,
+                    stripe_size: stripe_size.map(|val| ShardStripeSize(val.0)),
+                    shards: shard_infos,
+                    prefer_protocol,
+                };
 
                 endpoint
-                    .reconfigure_pageservers(pageservers, *stripe_size)
+                    .reconfigure_pageservers(&pageserver_conninfo)
                     .await
                     .map_err(NotifyError::NeonLocal)?;
             }

storage_controller/src/operation_utils.rs

@@ -46,11 +46,31 @@ impl TenantShardDrain {
         &self,
         tenants: &BTreeMap<TenantShardId, TenantShard>,
         scheduler: &Scheduler,
-    ) -> Option<NodeId> {
-        let tenant_shard = tenants.get(&self.tenant_shard_id)?;
+    ) -> TenantShardDrainAction {
+        let Some(tenant_shard) = tenants.get(&self.tenant_shard_id) else {
+            return TenantShardDrainAction::Skip;
+        };
 
         if *tenant_shard.intent.get_attached() != Some(self.drained_node) {
-            return None;
+            // If the intent attached node is not the drained node, check the observed state
+            // of the shard on the drained node. If it is Attached*, it means the shard is
+            // beeing migrated from the drained node. The drain loop needs to wait for the
+            // reconciliation to complete for a smooth draining.
+
+            use pageserver_api::models::LocationConfigMode::*;
+
+            let attach_mode = tenant_shard
+                .observed
+                .locations
+                .get(&self.drained_node)
+                .and_then(|observed| observed.conf.as_ref().map(|conf| conf.mode));
+
+            return match (attach_mode, tenant_shard.intent.get_attached()) {
+                (Some(AttachedSingle | AttachedMulti | AttachedStale), Some(intent_node_id)) => {
+                    TenantShardDrainAction::Reconcile(*intent_node_id)
+                }
+                _ => TenantShardDrainAction::Skip,
+            };
         }
 
         // Only tenants with a normal (Active) scheduling policy are proactively moved
@@ -63,19 +83,19 @@ impl TenantShardDrain {
             }
             ShardSchedulingPolicy::Pause | ShardSchedulingPolicy::Stop => {
                 // If we have been asked to avoid rescheduling this shard, then do not migrate it during a drain
-                return None;
+                return TenantShardDrainAction::Skip;
             }
         }
 
         match tenant_shard.preferred_secondary(scheduler) {
-            Some(node) => Some(node),
+            Some(node) => TenantShardDrainAction::RescheduleToSecondary(node),
             None => {
                 tracing::warn!(
                     tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug(),
                     "No eligible secondary while draining {}", self.drained_node
                 );
 
-                None
+                TenantShardDrainAction::Skip
             }
         }
     }
@@ -138,3 +158,17 @@ impl TenantShardDrain {
         }
     }
 }
+
+/// Action to take when draining a tenant shard.
+pub(crate) enum TenantShardDrainAction {
+    /// The tenant shard is on the draining node.
+    /// Reschedule the tenant shard to a secondary location.
+    /// Holds a destination node id to reschedule to.
+    RescheduleToSecondary(NodeId),
+    /// The tenant shard is beeing migrated from the draining node.
+    /// Wait for the reconciliation to complete.
+    /// Holds the intent attached node id.
+    Reconcile(NodeId),
+    /// The tenant shard is not eligible for drainining, skip it.
+    Skip,
+}

storage_controller/src/service.rs

@@ -79,7 +79,7 @@ use crate::id_lock_map::{
 use crate::leadership::Leadership;
 use crate::metrics;
 use crate::node::{AvailabilityTransition, Node};
-use crate::operation_utils::{self, TenantShardDrain};
+use crate::operation_utils::{self, TenantShardDrain, TenantShardDrainAction};
 use crate::pageserver_client::PageserverClient;
 use crate::peer_client::GlobalObservedState;
 use crate::persistence::split_state::SplitState;
@@ -1274,7 +1274,7 @@ impl Service {
                 // Always attempt autosplits. Sharding is crucial for bulk ingest performance, so we
                 // must be responsive when new projects begin ingesting and reach the threshold.
                 self.autosplit_tenants().await;
-            }
+              },
               _ = self.reconcilers_cancel.cancelled() => return
             }
         }
@@ -8876,6 +8876,9 @@ impl Service {
         for (_tenant_id, schedule_context, shards) in
             TenantShardExclusiveIterator::new(tenants, ScheduleMode::Speculative)
         {
+            if work.len() >= MAX_OPTIMIZATIONS_PLAN_PER_PASS {
+                break;
+            }
             for shard in shards {
                 if work.len() >= MAX_OPTIMIZATIONS_PLAN_PER_PASS {
                     break;
@@ -9640,16 +9643,16 @@ impl Service {
                     tenant_shard_id: tid,
                 };
 
-                let dest_node_id = {
+                let drain_action = {
                     let locked = self.inner.read().unwrap();
+                    tid_drain.tenant_shard_eligible_for_drain(&locked.tenants, &locked.scheduler)
+                };
 
-                    match tid_drain
-                        .tenant_shard_eligible_for_drain(&locked.tenants, &locked.scheduler)
-                    {
-                        Some(node_id) => node_id,
-                        None => {
-                            continue;
-                        }
+                let dest_node_id = match drain_action {
+                    TenantShardDrainAction::RescheduleToSecondary(dest_node_id) => dest_node_id,
+                    TenantShardDrainAction::Reconcile(intent_node_id) => intent_node_id,
+                    TenantShardDrainAction::Skip => {
+                        continue;
                     }
                 };
 
@@ -9684,14 +9687,16 @@ impl Service {
                 {
                     let mut locked = self.inner.write().unwrap();
                     let (nodes, tenants, scheduler) = locked.parts_mut();
-                    let rescheduled = tid_drain.reschedule_to_secondary(
-                        dest_node_id,
-                        tenants,
-                        scheduler,
-                        nodes,
-                    )?;
 
-                    if let Some(tenant_shard) = rescheduled {
+                    let tenant_shard = match drain_action {
+                        TenantShardDrainAction::RescheduleToSecondary(dest_node_id) => tid_drain
+                            .reschedule_to_secondary(dest_node_id, tenants, scheduler, nodes)?,
+                        TenantShardDrainAction::Reconcile(_) => tenants.get_mut(&tid),
+                        // Note: Unreachable, handled above.
+                        TenantShardDrainAction::Skip => None,
+                    };
+
+                    if let Some(tenant_shard) = tenant_shard {
                         let waiter = self.maybe_configured_reconcile_shard(
                             tenant_shard,
                             nodes,

storage_controller/src/tenant_shard.rs

@@ -812,8 +812,6 @@ impl TenantShard {
     /// if the swap is not possible and leaves the intent state in its original state.
     ///
     /// Arguments:
-    /// `attached_to`: the currently attached location matching the intent state (may be None if the
-    /// shard is not attached)
     /// `promote_to`: an optional secondary location of this tenant shard. If set to None, we ask
     /// the scheduler to recommend a node
     pub(crate) fn reschedule_to_secondary(

test_runner/fixtures/neon_fixtures.py

@@ -5280,16 +5280,32 @@ class EndpointFactory:
         )
 
     def stop_all(self, fail_on_error=True) -> Self:
-        exception = None
-        for ep in self.endpoints:
+        """
+        Stop all the endpoints in parallel.
+        """
+
+        # Note: raising an exception from a task in a task group cancels
+        # all the other tasks. We don't want that, hence the 'stop_one'
+        # function catches exceptions and puts them on the 'exceptions'
+        # list for later processing.
+        exceptions = []
+
+        async def stop_one(ep):
             try:
-                ep.stop()
+                await asyncio.to_thread(ep.stop)
             except Exception as e:
                 log.error(f"Failed to stop endpoint {ep.endpoint_id}: {e}")
-                exception = e
+                exceptions.append(e)
 
-        if fail_on_error and exception is not None:
-            raise exception
+        async def async_stop_all():
+            async with asyncio.TaskGroup() as tg:
+                for ep in self.endpoints:
+                    tg.create_task(stop_one(ep))
+
+        asyncio.run(async_stop_all())
+
+        if fail_on_error and exceptions:
+            raise ExceptionGroup("stopping an endpoint failed", exceptions)
 
         return self
 

test_runner/regress/test_basebackup.py

@@ -2,13 +2,15 @@ from __future__ import annotations
 
 from typing import TYPE_CHECKING
 
+import pytest
 from fixtures.utils import wait_until
 
 if TYPE_CHECKING:
     from fixtures.neon_fixtures import NeonEnvBuilder
 
 
-def test_basebackup_cache(neon_env_builder: NeonEnvBuilder):
+@pytest.mark.parametrize("grpc", [True, False])
+def test_basebackup_cache(neon_env_builder: NeonEnvBuilder, grpc: bool):
     """
     Simple test for basebackup cache.
     1. Check that we always hit the cache after compute restart.
@@ -22,7 +24,7 @@ def test_basebackup_cache(neon_env_builder: NeonEnvBuilder):
     """
 
     env = neon_env_builder.init_start()
-    ep = env.endpoints.create("main")
+    ep = env.endpoints.create("main", grpc=grpc)
     ps = env.pageserver
     ps_http = ps.http_client()
 

test_runner/regress/test_compaction.py

@@ -863,6 +863,88 @@ def test_pageserver_compaction_circuit_breaker(neon_env_builder: NeonEnvBuilder)
     assert not env.pageserver.log_contains(".*Circuit breaker failure ended.*")
 
 
+@pytest.mark.skip(reason="Lakebase mode")
+def test_ps_corruption_detection_feedback(neon_env_builder: NeonEnvBuilder):
+    """
+    Test that when the pageserver detects corruption during image layer creation,
+    it sends corruption feedback to the safekeeper which gets recorded in its
+    safekeeper_ps_corruption_detected metric.
+    """
+    # Configure tenant with aggressive compaction settings to easily trigger compaction
+    TENANT_CONF = {
+        # Small checkpoint distance to create many layers
+        "checkpoint_distance": 1024 * 128,
+        # Compact small layers
+        "compaction_target_size": 1024 * 128,
+        # Create image layers eagerly
+        "image_creation_threshold": 1,
+        "image_layer_creation_check_threshold": 0,
+        # Force frequent compaction
+        "compaction_period": "1s",
+    }
+
+    env = neon_env_builder.init_start(initial_tenant_conf=TENANT_CONF)
+    # We are simulating compaction failures so we should allow these error messages.
+    env.pageserver.allowed_errors.append(".*Compaction failed.*")
+    tenant_id = env.initial_tenant
+    timeline_id = env.initial_timeline
+
+    pageserver_http = env.pageserver.http_client()
+    workload = Workload(env, tenant_id, timeline_id)
+    workload.init()
+
+    # Enable the failpoint that will cause image layer creation to fail due to a (simulated) detected
+    # corruption.
+    pageserver_http.configure_failpoints(("create-image-layer-fail-simulated-corruption", "return"))
+
+    # Write some data to trigger compaction and image layer creation
+    log.info("Writing data to trigger compaction...")
+    workload.write_rows(1024 * 64, upload=False)
+    workload.write_rows(1024 * 64, upload=False)
+
+    # Returns True if the corruption signal from PS is propagated to the SK according to the "safekeeper_ps_corruption_detected" metric.
+    # Raises an exception otherwise.
+    def check_corruption_signal_propagated_to_sk():
+        # Get metrics from all safekeepers
+        for sk in env.safekeepers:
+            sk_metrics = sk.http_client().get_metrics()
+            # Look for our corruption detected metric with the right tenant and timeline
+            corruption_metrics = sk_metrics.query_all("safekeeper_ps_corruption_detected")
+
+            for metric in corruption_metrics:
+                # Check if there's a metric for our tenant and timeline that has value 1
+                if (
+                    metric.labels.get("tenant_id") == str(tenant_id)
+                    and metric.labels.get("timeline_id") == str(timeline_id)
+                    and metric.value == 1
+                ):
+                    log.info(f"Corruption detected by safekeeper {sk.id}: {metric}")
+                    return True
+        raise Exception("Corruption detection feedback not found in any safekeeper metrics")
+
+    # Returns True if the corruption signal from PS is propagated to the PG according to the "ps_corruption_detected" metric
+    # in "neon_perf_counters".
+    # Raises an exception otherwise.
+    def check_corruption_signal_propagated_to_pg():
+        endpoint = workload.endpoint()
+        results = endpoint.safe_psql("CREATE EXTENSION IF NOT EXISTS neon")
+        results = endpoint.safe_psql(
+            "SELECT value FROM neon_perf_counters WHERE metric = 'ps_corruption_detected'"
+        )
+        log.info("Query corruption detection metric, results: %s", results)
+        if results[0][0] == 1:
+            log.info("Corruption detection signal is raised on Postgres")
+            return True
+        raise Exception("Corruption detection signal is not raise on Postgres")
+
+    # Confirm that the corruption signal propagates to both the safekeeper and Postgres
+    wait_until(check_corruption_signal_propagated_to_sk, timeout=10, interval=0.1)
+    wait_until(check_corruption_signal_propagated_to_pg, timeout=10, interval=0.1)
+
+    # Cleanup the failpoint
+    pageserver_http.configure_failpoints(("create-image-layer-fail-simulated-corruption", "off"))
+
+
 @pytest.mark.parametrize("enabled", [True, False])
 def test_image_layer_compression(neon_env_builder: NeonEnvBuilder, enabled: bool):
     tenant_conf = {

test_runner/regress/test_readonly_node.py

@@ -129,7 +129,10 @@ def test_readonly_node_gc(neon_env_builder: NeonEnvBuilder):
     Test static endpoint is protected from GC by acquiring and renewing lsn leases.
     """
 
-    LSN_LEASE_LENGTH = 8
+    LSN_LEASE_LENGTH = (
+        14  # This value needs to be large enough for compute_ctl to send two lease requests.
+    )
+
     neon_env_builder.num_pageservers = 2
     # GC is manual triggered.
     env = neon_env_builder.init_start(
@@ -230,6 +233,15 @@ def test_readonly_node_gc(neon_env_builder: NeonEnvBuilder):
         log.info(f"`SELECT` query succeed after GC, {ctx=}")
         return offset
 
+    # It's not reliable to let the compute renew the lease in this test case as we have a very tight
+    # lease timeout. Therefore, the test case itself will renew the lease.
+    #
+    # This is a workaround to make the test case more deterministic.
+    def renew_lease(env: NeonEnv, lease_lsn: Lsn):
+        env.storage_controller.pageserver_api().timeline_lsn_lease(
+            env.initial_tenant, env.initial_timeline, lease_lsn
+        )
+
     # Insert some records on main branch
     with env.endpoints.create_start("main", config_lines=["shared_buffers=1MB"]) as ep_main:
         with ep_main.cursor() as cur:
@@ -242,6 +254,9 @@ def test_readonly_node_gc(neon_env_builder: NeonEnvBuilder):
         XLOG_BLCKSZ = 8192
         lsn = Lsn((int(lsn) // XLOG_BLCKSZ) * XLOG_BLCKSZ)
 
+        # We need to mock the way cplane works: it gets a lease for a branch before starting the compute.
+        renew_lease(env, lsn)
+
         with env.endpoints.create_start(
             branch_name="main",
             endpoint_id="static",
@@ -251,9 +266,6 @@ def test_readonly_node_gc(neon_env_builder: NeonEnvBuilder):
                 cur.execute("SELECT count(*) FROM t0")
                 assert cur.fetchone() == (ROW_COUNT,)
 
-            # Wait for static compute to renew lease at least once.
-            time.sleep(LSN_LEASE_LENGTH / 2)
-
             generate_updates_on_main(env, ep_main, 3, end=100)
 
             offset = trigger_gc_and_select(
@@ -263,10 +275,10 @@ def test_readonly_node_gc(neon_env_builder: NeonEnvBuilder):
             # Trigger Pageserver restarts
             for ps in env.pageservers:
                 ps.stop()
-                # Static compute should have at least one lease request failure due to connection.
-                time.sleep(LSN_LEASE_LENGTH / 2)
                 ps.start()
 
+            renew_lease(env, lsn)
+
             trigger_gc_and_select(
                 env,
                 ep_static,
@@ -282,6 +294,9 @@ def test_readonly_node_gc(neon_env_builder: NeonEnvBuilder):
             )
             env.storage_controller.reconcile_until_idle()
 
+            # Wait for static compute to renew lease on the new pageserver.
+            time.sleep(LSN_LEASE_LENGTH + 3)
+
             trigger_gc_and_select(
                 env,
                 ep_static,
@@ -292,7 +307,6 @@ def test_readonly_node_gc(neon_env_builder: NeonEnvBuilder):
 
         # Do some update so we can increment gc_cutoff
         generate_updates_on_main(env, ep_main, i, end=100)
-
     # Wait for the existing lease to expire.
     time.sleep(LSN_LEASE_LENGTH + 1)
     # Now trigger GC again, layers should be removed.

test_runner/regress/test_tenant_size.py

@@ -1,5 +1,6 @@
 from __future__ import annotations
 
+import os
 from concurrent.futures import ThreadPoolExecutor
 from typing import TYPE_CHECKING
 
@@ -768,6 +769,14 @@ def test_lsn_lease_storcon(neon_env_builder: NeonEnvBuilder):
         "compaction_period": "0s",
     }
     env = neon_env_builder.init_start(initial_tenant_conf=conf)
+    # ShardSplit is slow in debug builds, so ignore the warning
+    if os.getenv("BUILD_TYPE", "debug") == "debug":
+        env.storage_controller.allowed_errors.extend(
+            [
+                ".*Exclusive lock by ShardSplit was held.*",
+            ]
+        )
+
     with env.endpoints.create_start(
         "main",
     ) as ep: