Proxy release 2025-06-02 06:12 UTC

libs/pqproto is designed for safekeeper/pageserver with maximum
throughput.

proxy only needs it for handshakes/authentication where throughput is
not a concern but memory efficiency is. For this reason, we switch to
using read_exact and only allocating as much memory as we need to.

All reads return a `&'a [u8]` instead of a `Bytes` because accidental
sharing of bytes can cause fragmentation. Returning the reference
enforces all callers only hold onto the bytes they absolutely need. For
example, before this change, `pqproto` was allocating 8KiB for the
initial read `BytesMut`, and proxy was holding the `Bytes` in the
`StartupMessageParams` for the entire connection through to passthrough.

Cargo.lock

@@ -4330,6 +4330,7 @@ dependencies = [
  "postgres_connection",
  "postgres_ffi",
  "postgres_initdb",
+ "posthog_client_lite",
  "pprof",
  "pq_proto",
  "procfs",
@@ -4458,9 +4459,15 @@ dependencies = [
 name = "pageserver_page_api"
 version = "0.1.0"
 dependencies = [
+ "bytes",
+ "pageserver_api",
+ "postgres_ffi",
  "prost 0.13.5",
+ "smallvec",
+ "thiserror 1.0.69",
  "tonic 0.13.1",
  "tonic-build",
+ "utils",
  "workspace_hack",
 ]
 
@@ -4901,11 +4908,16 @@ name = "posthog_client_lite"
 version = "0.1.0"
 dependencies = [
  "anyhow",
+ "arc-swap",
  "reqwest",
  "serde",
  "serde_json",
  "sha2",
  "thiserror 1.0.69",
+ "tokio",
+ "tokio-util",
+ "tracing",
+ "tracing-utils",
  "workspace_hack",
 ]
 
@@ -8569,10 +8581,8 @@ dependencies = [
  "fail",
  "form_urlencoded",
  "futures-channel",
- "futures-core",
  "futures-executor",
  "futures-io",
- "futures-task",
  "futures-util",
  "generic-array",
  "getrandom 0.2.11",
@@ -8602,7 +8612,6 @@ dependencies = [
  "once_cell",
  "p256 0.13.2",
  "parquet",
- "percent-encoding",
  "prettyplease",
  "proc-macro2",
  "prost 0.13.5",

Cargo.toml

@@ -247,6 +247,7 @@ azure_storage_blobs = { git = "https://github.com/neondatabase/azure-sdk-for-rus
 ## Local libraries
 compute_api = { version = "0.1", path = "./libs/compute_api/" }
 consumption_metrics = { version = "0.1", path = "./libs/consumption_metrics/" }
+desim = { version = "0.1", path = "./libs/desim" }
 endpoint_storage = { version = "0.0.1", path = "./endpoint_storage/" }
 http-utils = { version = "0.1", path = "./libs/http-utils/" }
 metrics = { version = "0.1", path = "./libs/metrics/" }
@@ -259,19 +260,19 @@ postgres_backend = { version = "0.1", path = "./libs/postgres_backend/" }
 postgres_connection = { version = "0.1", path = "./libs/postgres_connection/" }
 postgres_ffi = { version = "0.1", path = "./libs/postgres_ffi/" }
 postgres_initdb = { path = "./libs/postgres_initdb" }
+posthog_client_lite = { version = "0.1", path = "./libs/posthog_client_lite" }
 pq_proto = { version = "0.1", path = "./libs/pq_proto/" }
 remote_storage = { version = "0.1", path = "./libs/remote_storage/" }
 safekeeper_api = { version = "0.1", path = "./libs/safekeeper_api" }
 safekeeper_client = { path = "./safekeeper/client" }
-desim = { version = "0.1", path = "./libs/desim" }
 storage_broker = { version = "0.1", path = "./storage_broker/" } # Note: main broker code is inside the binary crate, so linking with the library shouldn't be heavy.
 storage_controller_client = { path = "./storage_controller/client" }
 tenant_size_model = { version = "0.1", path = "./libs/tenant_size_model/" }
 tracing-utils = { version = "0.1", path = "./libs/tracing-utils/" }
 utils = { version = "0.1", path = "./libs/utils/" }
 vm_monitor = { version = "0.1", path = "./libs/vm_monitor/" }
-walproposer = { version = "0.1", path = "./libs/walproposer/" }
 wal_decoder = { version = "0.1", path = "./libs/wal_decoder" }
+walproposer = { version = "0.1", path = "./libs/walproposer/" }
 
 ## Common library dependency
 workspace_hack = { version = "0.1", path = "./workspace_hack/" }

build-tools.Dockerfile

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

compute/compute-node.Dockerfile

@@ -1847,7 +1847,7 @@ COPY docker-compose/ext-src/ /ext-src/
 COPY --from=pg-build /postgres /postgres
 #COPY --from=postgis-src /ext-src/ /ext-src/
 COPY --from=plv8-src /ext-src/ /ext-src/
-#COPY --from=h3-pg-src /ext-src/ /ext-src/
+COPY --from=h3-pg-src /ext-src/h3-pg-src /ext-src/h3-pg-src
 COPY --from=postgresql-unit-src /ext-src/ /ext-src/
 COPY --from=pgvector-src /ext-src/ /ext-src/
 COPY --from=pgjwt-src /ext-src/ /ext-src/

compute_tools/src/bin/compute_ctl.rs

@@ -57,21 +57,6 @@ use tracing::{error, info};
 use url::Url;
 use utils::failpoint_support;
 
-// Compatibility hack: if the control plane specified any remote-ext-config
-// use the default value for extension storage proxy gateway.
-// Remove this once the control plane is updated to pass the gateway URL
-fn parse_remote_ext_base_url(arg: &str) -> Result<String> {
-    const FALLBACK_PG_EXT_GATEWAY_BASE_URL: &str =
-        "http://pg-ext-s3-gateway.pg-ext-s3-gateway.svc.cluster.local";
-
-    Ok(if arg.starts_with("http") {
-        arg
-    } else {
-        FALLBACK_PG_EXT_GATEWAY_BASE_URL
-    }
-    .to_owned())
-}
-
 #[derive(Parser)]
 #[command(rename_all = "kebab-case")]
 struct Cli {
@@ -79,9 +64,8 @@ struct Cli {
     pub pgbin: String,
 
     /// The base URL for the remote extension storage proxy gateway.
-    /// Should be in the form of `http(s)://<gateway-hostname>[:<port>]`.
-    #[arg(short = 'r', long, value_parser = parse_remote_ext_base_url, alias = "remote-ext-config")]
-    pub remote_ext_base_url: Option<String>,
+    #[arg(short = 'r', long)]
+    pub remote_ext_base_url: Option<Url>,
 
     /// The port to bind the external listening HTTP server to. Clients running
     /// outside the compute will talk to the compute through this port. Keep
@@ -136,6 +120,10 @@ struct Cli {
         requires = "compute-id"
     )]
     pub control_plane_uri: Option<String>,
+
+    /// Interval in seconds for collecting installed extensions statistics
+    #[arg(long, default_value = "3600")]
+    pub installed_extensions_collection_interval: u64,
 }
 
 fn main() -> Result<()> {
@@ -179,6 +167,7 @@ fn main() -> Result<()> {
             cgroup: cli.cgroup,
             #[cfg(target_os = "linux")]
             vm_monitor_addr: cli.vm_monitor_addr,
+            installed_extensions_collection_interval: cli.installed_extensions_collection_interval,
         },
         config,
     )?;
@@ -271,18 +260,4 @@ mod test {
     fn verify_cli() {
         Cli::command().debug_assert()
     }
-
-    #[test]
-    fn parse_pg_ext_gateway_base_url() {
-        let arg = "http://pg-ext-s3-gateway2";
-        let result = super::parse_remote_ext_base_url(arg).unwrap();
-        assert_eq!(result, arg);
-
-        let arg = "pg-ext-s3-gateway";
-        let result = super::parse_remote_ext_base_url(arg).unwrap();
-        assert_eq!(
-            result,
-            "http://pg-ext-s3-gateway.pg-ext-s3-gateway.svc.cluster.local"
-        );
-    }
 }

compute_tools/src/bin/fast_import.rs

@@ -339,6 +339,8 @@ async fn run_dump_restore(
     destination_connstring: String,
 ) -> Result<(), anyhow::Error> {
     let dumpdir = workdir.join("dumpdir");
+    let num_jobs = num_cpus::get().to_string();
+    info!("using {num_jobs} jobs for dump/restore");
 
     let common_args = [
         // schema mapping (prob suffices to specify them on one side)
@@ -354,7 +356,7 @@ async fn run_dump_restore(
         "directory".to_string(),
         // concurrency
         "--jobs".to_string(),
-        num_cpus::get().to_string(),
+        num_jobs,
         // progress updates
         "--verbose".to_string(),
     ];

compute_tools/src/compute.rs

@@ -31,6 +31,7 @@ use std::time::{Duration, Instant};
 use std::{env, fs};
 use tokio::spawn;
 use tracing::{Instrument, debug, error, info, instrument, warn};
+use url::Url;
 use utils::id::{TenantId, TimelineId};
 use utils::lsn::Lsn;
 use utils::measured_stream::MeasuredReader;
@@ -96,7 +97,10 @@ pub struct ComputeNodeParams {
     pub internal_http_port: u16,
 
     /// the address of extension storage proxy gateway
-    pub remote_ext_base_url: Option<String>,
+    pub remote_ext_base_url: Option<Url>,
+
+    /// Interval for installed extensions collection
+    pub installed_extensions_collection_interval: u64,
 }
 
 /// Compute node info shared across several `compute_ctl` threads.
@@ -695,25 +699,18 @@ impl ComputeNode {
                 let log_directory_path = Path::new(&self.params.pgdata).join("log");
                 let log_directory_path = log_directory_path.to_string_lossy().to_string();
 
-                // Add project_id,endpoint_id tag to identify the logs.
+                // Add project_id,endpoint_id to identify the logs.
                 //
                 // These ids are passed from cplane,
-                // for backwards compatibility (old computes that don't have them),
-                // we set them to None.
-                // TODO: Clean up this code when all computes have them.
-                let tag: Option<String> = match (
-                    pspec.spec.project_id.as_deref(),
-                    pspec.spec.endpoint_id.as_deref(),
-                ) {
-                    (Some(project_id), Some(endpoint_id)) => {
-                        Some(format!("{project_id}/{endpoint_id}"))
-                    }
-                    (Some(project_id), None) => Some(format!("{project_id}/None")),
-                    (None, Some(endpoint_id)) => Some(format!("None,{endpoint_id}")),
-                    (None, None) => None,
-                };
+                let endpoint_id = pspec.spec.endpoint_id.as_deref().unwrap_or("");
+                let project_id = pspec.spec.project_id.as_deref().unwrap_or("");
 
-                configure_audit_rsyslog(log_directory_path.clone(), tag, &remote_endpoint)?;
+                configure_audit_rsyslog(
+                    log_directory_path.clone(),
+                    endpoint_id,
+                    project_id,
+                    &remote_endpoint,
+                )?;
 
                 // Launch a background task to clean up the audit logs
                 launch_pgaudit_gc(log_directory_path);
@@ -749,17 +746,7 @@ impl ComputeNode {
 
             let conf = self.get_tokio_conn_conf(None);
             tokio::task::spawn(async {
-                let res = get_installed_extensions(conf).await;
-                match res {
-                    Ok(extensions) => {
-                        info!(
-                            "[NEON_EXT_STAT] {}",
-                            serde_json::to_string(&extensions)
-                                .expect("failed to serialize extensions list")
-                        );
-                    }
-                    Err(err) => error!("could not get installed extensions: {err:?}"),
-                }
+                let _ = installed_extensions(conf).await;
             });
         }
 
@@ -789,6 +776,9 @@ impl ComputeNode {
         // Log metrics so that we can search for slow operations in logs
         info!(?metrics, postmaster_pid = %postmaster_pid, "compute start finished");
 
+        // Spawn the extension stats background task
+        self.spawn_extension_stats_task();
+
         if pspec.spec.prewarm_lfc_on_startup {
             self.prewarm_lfc();
         }
@@ -2199,6 +2189,41 @@ LIMIT 100",
             info!("Pageserver config changed");
         }
     }
+
+    pub fn spawn_extension_stats_task(&self) {
+        let conf = self.tokio_conn_conf.clone();
+        let installed_extensions_collection_interval =
+            self.params.installed_extensions_collection_interval;
+        tokio::spawn(async move {
+            // An initial sleep is added to ensure that two collections don't happen at the same time.
+            // The first collection happens during compute startup.
+            tokio::time::sleep(tokio::time::Duration::from_secs(
+                installed_extensions_collection_interval,
+            ))
+            .await;
+            let mut interval = tokio::time::interval(tokio::time::Duration::from_secs(
+                installed_extensions_collection_interval,
+            ));
+            loop {
+                interval.tick().await;
+                let _ = installed_extensions(conf.clone()).await;
+            }
+        });
+    }
+}
+
+pub async fn installed_extensions(conf: tokio_postgres::Config) -> Result<()> {
+    let res = get_installed_extensions(conf).await;
+    match res {
+        Ok(extensions) => {
+            info!(
+                "[NEON_EXT_STAT] {}",
+                serde_json::to_string(&extensions).expect("failed to serialize extensions list")
+            );
+        }
+        Err(err) => error!("could not get installed extensions: {err:?}"),
+    }
+    Ok(())
 }
 
 pub fn forward_termination_signal() {

compute_tools/src/config_template/compute_audit_rsyslog_template.conf

@@ -2,10 +2,24 @@
 module(load="imfile")
 
 # Input configuration for log files in the specified directory
-# Replace {log_directory} with the directory containing the log files
-input(type="imfile" File="{log_directory}/*.log" Tag="{tag}" Severity="info" Facility="local0")
+# The messages can be multiline. The start of the message is a timestamp
+# in "%Y-%m-%d %H:%M:%S.%3N GMT" (so timezone hardcoded).
+# Replace log_directory with the directory containing the log files
+input(type="imfile" File="{log_directory}/*.log"
+  Tag="pgaudit_log" Severity="info" Facility="local5"
+  startmsg.regex="^[[:digit:]]{{4}}-[[:digit:]]{{2}}-[[:digit:]]{{2}} [[:digit:]]{{2}}:[[:digit:]]{{2}}:[[:digit:]]{{2}}.[[:digit:]]{{3}} GMT,")
+
 # the directory to store rsyslog state files
 global(workDirectory="/var/log/rsyslog")
 
-# Forward logs to remote syslog server
-*.* @@{remote_endpoint}
+# Construct json, endpoint_id and project_id as additional metadata
+set $.json_log!endpoint_id = "{endpoint_id}";
+set $.json_log!project_id = "{project_id}";
+set $.json_log!msg = $msg;
+
+# Template suitable for rfc5424 syslog format
+template(name="PgAuditLog" type="string"
+    string="<%PRI%>1 %TIMESTAMP:::date-rfc3339% %HOSTNAME% - - - - %$.json_log%")
+
+# Forward to remote syslog receiver (@@<hostname>:<port>;format
+local5.info @@{remote_endpoint};PgAuditLog

compute_tools/src/extension_server.rs

@@ -83,6 +83,7 @@ use reqwest::StatusCode;
 use tar::Archive;
 use tracing::info;
 use tracing::log::warn;
+use url::Url;
 use zstd::stream::read::Decoder;
 
 use crate::metrics::{REMOTE_EXT_REQUESTS_TOTAL, UNKNOWN_HTTP_STATUS};
@@ -158,14 +159,14 @@ fn parse_pg_version(human_version: &str) -> PostgresMajorVersion {
 pub async fn download_extension(
     ext_name: &str,
     ext_path: &RemotePath,
-    remote_ext_base_url: &str,
+    remote_ext_base_url: &Url,
     pgbin: &str,
 ) -> Result<u64> {
     info!("Download extension {:?} from {:?}", ext_name, ext_path);
 
     // TODO add retry logic
     let download_buffer =
-        match download_extension_tar(remote_ext_base_url, &ext_path.to_string()).await {
+        match download_extension_tar(remote_ext_base_url.as_str(), &ext_path.to_string()).await {
             Ok(buffer) => buffer,
             Err(error_message) => {
                 return Err(anyhow::anyhow!(

compute_tools/src/rsyslog.rs

@@ -84,13 +84,15 @@ fn restart_rsyslog() -> Result<()> {
 
 pub fn configure_audit_rsyslog(
     log_directory: String,
-    tag: Option<String>,
+    endpoint_id: &str,
+    project_id: &str,
     remote_endpoint: &str,
 ) -> Result<()> {
     let config_content: String = format!(
         include_str!("config_template/compute_audit_rsyslog_template.conf"),
         log_directory = log_directory,
-        tag = tag.unwrap_or("".to_string()),
+        endpoint_id = endpoint_id,
+        project_id = project_id,
         remote_endpoint = remote_endpoint
     );
 

control_plane/src/bin/neon_local.rs

@@ -1279,6 +1279,7 @@ async fn handle_timeline(cmd: &TimelineCmd, env: &mut local_env::LocalEnv) -> Re
                 mode: pageserver_api::models::TimelineCreateRequestMode::Branch {
                     ancestor_timeline_id,
                     ancestor_start_lsn: start_lsn,
+                    read_only: false,
                     pg_version: None,
                 },
             };

docker-compose/compute_wrapper/shell/compute.sh

@@ -20,7 +20,7 @@ first_path="$(ldconfig --verbose 2>/dev/null \
     | grep --invert-match ^$'\t' \
     | cut --delimiter=: --fields=1 \
     | head --lines=1)"
-test "$first_path" == '/usr/local/lib' || true # Remove the || true in a follow-up PR. Needed for backwards compat.
+test "$first_path" == '/usr/local/lib'
 
 echo "Waiting pageserver become ready."
 while ! nc -z pageserver 6400; do

docker-compose/ext-src/h3-pg-src/neon-test.sh

@@ -0,0 +1,16 @@
+#!/usr/bin/env bash
+set -ex
+cd "$(dirname "${0}")"
+PG_REGRESS=$(dirname "$(pg_config --pgxs)")/../test/regress/pg_regress
+dropdb --if-exists contrib_regression
+createdb contrib_regression
+cd h3_postgis/test
+psql -d contrib_regression -c "CREATE EXTENSION postgis" -c "CREATE EXTENSION postgis_raster" -c "CREATE EXTENSION h3" -c "CREATE EXTENSION h3_postgis"
+TESTS=$(echo sql/* | sed 's|sql/||g; s|\.sql||g')
+${PG_REGRESS} --use-existing --dbname contrib_regression ${TESTS}
+cd ../../h3/test
+TESTS=$(echo sql/* | sed 's|sql/||g; s|\.sql||g')
+dropdb --if-exists contrib_regression
+createdb contrib_regression
+psql -d contrib_regression -c "CREATE EXTENSION h3"
+${PG_REGRESS} --use-existing --dbname contrib_regression ${TESTS}

docker-compose/ext-src/h3-pg-src/test-upgrade.sh

@@ -0,0 +1,7 @@
+#!/bin/sh
+set -ex
+cd "$(dirname ${0})"
+PG_REGRESS=$(dirname "$(pg_config --pgxs)")/../test/regress/pg_regress
+cd h3/test
+TESTS=$(echo sql/* | sed 's|sql/||g; s|\.sql||g')
+${PG_REGRESS} --use-existing --inputdir=./ --bindir='/usr/local/pgsql/bin'  --dbname=contrib_regression  ${TESTS}

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

@@ -0,0 +1,6 @@
+#!/bin/sh
+set -ex
+cd "$(dirname "${0}")"
+if [ -f Makefile ]; then
+  make installcheck
+fi

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

@@ -0,0 +1,9 @@
+#!/bin/sh
+set -ex
+cd "$(dirname ${0})"
+[ -f Makefile ] || exit 0
+dropdb --if-exist contrib_regression
+createdb contrib_regression
+PG_REGRESS=$(dirname "$(pg_config --pgxs)")/../test/regress/pg_regress
+TESTS=$(echo sql/* | sed 's|sql/||g; s|\.sql||g')
+${PG_REGRESS} --use-existing --inputdir=./ --bindir='/usr/local/pgsql/bin' --dbname=contrib_regression ${TESTS}

docker-compose/test_extensions_upgrade.sh

@@ -82,7 +82,8 @@ EXTENSIONS='[
 {"extname": "pg_ivm", "extdir": "pg_ivm-src"},
 {"extname": "pgjwt", "extdir": "pgjwt-src"},
 {"extname": "pgtap", "extdir": "pgtap-src"},
-{"extname": "pg_repack", "extdir": "pg_repack-src"}
+{"extname": "pg_repack", "extdir": "pg_repack-src"},
+{"extname": "h3", "extdir": "h3-pg-src"}
 ]'
 EXTNAMES=$(echo ${EXTENSIONS} | jq -r '.[].extname' | paste -sd ' ' -)
 COMPUTE_TAG=${NEW_COMPUTE_TAG} docker compose --profile test-extensions up --quiet-pull --build -d

libs/metrics/src/lib.rs

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

libs/pageserver_api/src/config.rs

@@ -45,6 +45,21 @@ pub struct NodeMetadata {
     pub other: HashMap<String, serde_json::Value>,
 }
 
+/// PostHog integration config.
+#[derive(Debug, Clone, PartialEq, Eq, serde::Serialize, serde::Deserialize)]
+pub struct PostHogConfig {
+    /// PostHog project ID
+    pub project_id: String,
+    /// Server-side (private) API key
+    pub server_api_key: String,
+    /// Client-side (public) API key
+    pub client_api_key: String,
+    /// Private API URL
+    pub private_api_url: String,
+    /// Public API URL
+    pub public_api_url: String,
+}
+
 /// `pageserver.toml`
 ///
 /// We use serde derive with `#[serde(default)]` to generate a deserializer
@@ -186,6 +201,8 @@ pub struct ConfigToml {
     pub tracing: Option<Tracing>,
     pub enable_tls_page_service_api: bool,
     pub dev_mode: bool,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub posthog_config: Option<PostHogConfig>,
     pub timeline_import_config: TimelineImportConfig,
     #[serde(skip_serializing_if = "Option::is_none")]
     pub basebackup_cache_config: Option<BasebackupCacheConfig>,
@@ -701,6 +718,7 @@ impl Default for ConfigToml {
                 import_job_checkpoint_threshold: NonZeroUsize::new(128).unwrap(),
             },
             basebackup_cache_config: None,
+            posthog_config: None,
         }
     }
 }

libs/pageserver_api/src/models.rs

@@ -354,6 +354,9 @@ pub struct ShardImportProgressV1 {
     pub completed: usize,
     /// Hash of the plan
     pub import_plan_hash: u64,
+    /// Soft limit for the job size
+    /// This needs to remain constant throughout the import
+    pub job_soft_size_limit: usize,
 }
 
 impl ShardImportStatus {
@@ -402,6 +405,8 @@ pub enum TimelineCreateRequestMode {
         // using a flattened enum, so, it was an accepted field, and
         // we continue to accept it by having it here.
         pg_version: Option<u32>,
+        #[serde(default, skip_serializing_if = "std::ops::Not::not")]
+        read_only: bool,
     },
     ImportPgdata {
         import_pgdata: TimelineCreateRequestModeImportPgdata,

libs/posthog_client_lite/Cargo.toml

@@ -6,9 +6,14 @@ license.workspace = true
 
 [dependencies]
 anyhow.workspace = true
+arc-swap.workspace = true
 reqwest.workspace = true
-serde.workspace = true
 serde_json.workspace = true
+serde.workspace = true
 sha2.workspace = true
-workspace_hack.workspace = true
 thiserror.workspace = true
+tokio = { workspace = true, features = ["process", "sync", "fs", "rt", "io-util", "time"] }
+tokio-util.workspace = true
+tracing-utils.workspace = true
+tracing.workspace = true
+workspace_hack.workspace = true

libs/posthog_client_lite/src/background_loop.rs

@@ -0,0 +1,64 @@
+//! A background loop that fetches feature flags from PostHog and updates the feature store.
+
+use std::{sync::Arc, time::Duration};
+
+use arc_swap::ArcSwap;
+use tokio_util::sync::CancellationToken;
+use tracing::{Instrument, info_span};
+
+use crate::{FeatureStore, PostHogClient, PostHogClientConfig};
+
+/// A background loop that fetches feature flags from PostHog and updates the feature store.
+pub struct FeatureResolverBackgroundLoop {
+    posthog_client: PostHogClient,
+    feature_store: ArcSwap<FeatureStore>,
+    cancel: CancellationToken,
+}
+
+impl FeatureResolverBackgroundLoop {
+    pub fn new(config: PostHogClientConfig, shutdown_pageserver: CancellationToken) -> Self {
+        Self {
+            posthog_client: PostHogClient::new(config),
+            feature_store: ArcSwap::new(Arc::new(FeatureStore::new())),
+            cancel: shutdown_pageserver,
+        }
+    }
+
+    pub fn spawn(self: Arc<Self>, handle: &tokio::runtime::Handle, refresh_period: Duration) {
+        let this = self.clone();
+        let cancel = self.cancel.clone();
+        handle.spawn(
+            async move {
+                tracing::info!("Starting PostHog feature resolver");
+                let mut ticker = tokio::time::interval(refresh_period);
+                ticker.set_missed_tick_behavior(tokio::time::MissedTickBehavior::Skip);
+                loop {
+                    tokio::select! {
+                        _ = ticker.tick() => {}
+                        _ = cancel.cancelled() => break
+                    }
+                    let resp = match this
+                        .posthog_client
+                        .get_feature_flags_local_evaluation()
+                        .await
+                    {
+                        Ok(resp) => resp,
+                        Err(e) => {
+                            tracing::warn!("Cannot get feature flags: {}", e);
+                            continue;
+                        }
+                    };
+                    let feature_store = FeatureStore::new_with_flags(resp.flags);
+                    this.feature_store.store(Arc::new(feature_store));
+                    tracing::info!("Feature flag updated");
+                }
+                tracing::info!("PostHog feature resolver stopped");
+            }
+            .instrument(info_span!("posthog_feature_resolver")),
+        );
+    }
+
+    pub fn feature_store(&self) -> Arc<FeatureStore> {
+        self.feature_store.load_full()
+    }
+}

libs/posthog_client_lite/src/lib.rs

@@ -1,5 +1,9 @@
 //! A lite version of the PostHog client that only supports local evaluation of feature flags.
 
+mod background_loop;
+
+pub use background_loop::FeatureResolverBackgroundLoop;
+
 use std::collections::HashMap;
 
 use serde::{Deserialize, Serialize};
@@ -20,8 +24,7 @@ pub enum PostHogEvaluationError {
 
 #[derive(Deserialize)]
 pub struct LocalEvaluationResponse {
-    #[allow(dead_code)]
-    flags: Vec<LocalEvaluationFlag>,
+    pub flags: Vec<LocalEvaluationFlag>,
 }
 
 #[derive(Deserialize)]
@@ -34,7 +37,7 @@ pub struct LocalEvaluationFlag {
 #[derive(Deserialize)]
 pub struct LocalEvaluationFlagFilters {
     groups: Vec<LocalEvaluationFlagFilterGroup>,
-    multivariate: LocalEvaluationFlagMultivariate,
+    multivariate: Option<LocalEvaluationFlagMultivariate>,
 }
 
 #[derive(Deserialize)]
@@ -94,6 +97,12 @@ impl FeatureStore {
         }
     }
 
+    pub fn new_with_flags(flags: Vec<LocalEvaluationFlag>) -> Self {
+        let mut store = Self::new();
+        store.set_flags(flags);
+        store
+    }
+
     pub fn set_flags(&mut self, flags: Vec<LocalEvaluationFlag>) {
         self.flags.clear();
         for flag in flags {
@@ -245,7 +254,7 @@ impl FeatureStore {
         }
     }
 
-    /// Evaluate a multivariate feature flag. Returns `None` if the flag is not available or if there are errors
+    /// Evaluate a multivariate feature flag. Returns an error if the flag is not available or if there are errors
     /// during the evaluation.
     ///
     /// The parsing logic is as follows:
@@ -263,10 +272,15 @@ impl FeatureStore {
     /// Example: we have a multivariate flag with 3 groups of the configured global rollout percentage: A (10%), B (20%), C (70%).
     /// There is a single group with a condition that has a rollout percentage of 10% and it does not have a variant override.
     /// Then, we will have 1% of the users evaluated to A, 2% to B, and 7% to C.
+    ///
+    /// Error handling: the caller should inspect the error and decide the behavior when a feature flag
+    /// cannot be evaluated (i.e., default to false if it cannot be resolved). The error should *not* be
+    /// propagated beyond where the feature flag gets resolved.
     pub fn evaluate_multivariate(
         &self,
         flag_key: &str,
         user_id: &str,
+        properties: &HashMap<String, PostHogFlagFilterPropertyValue>,
     ) -> Result<String, PostHogEvaluationError> {
         let hash_on_global_rollout_percentage =
             Self::consistent_hash(user_id, flag_key, "multivariate");
@@ -276,10 +290,39 @@ impl FeatureStore {
             flag_key,
             hash_on_global_rollout_percentage,
             hash_on_group_rollout_percentage,
-            &HashMap::new(),
+            properties,
         )
     }
 
+    /// Evaluate a boolean feature flag. Returns  an error if the flag is not available or if there are errors
+    /// during the evaluation.
+    ///
+    /// The parsing logic is as follows:
+    ///
+    /// * Generate a consistent hash for the tenant-feature.
+    /// * Match each filter group.
+    ///   - If a group is matched, it will first determine whether the user is in the range of the rollout
+    ///     percentage.
+    ///   - If the hash falls within the group's rollout percentage, return true.
+    /// * Otherwise, continue with the next group until all groups are evaluated and no group is within the
+    ///   rollout percentage.
+    /// * If there are no matching groups, return an error.
+    ///
+    /// Returns `Ok(())` if the feature flag evaluates to true. In the future, it will return a payload.
+    ///
+    /// Error handling: the caller should inspect the error and decide the behavior when a feature flag
+    /// cannot be evaluated (i.e., default to false if it cannot be resolved). The error should *not* be
+    /// propagated beyond where the feature flag gets resolved.
+    pub fn evaluate_boolean(
+        &self,
+        flag_key: &str,
+        user_id: &str,
+        properties: &HashMap<String, PostHogFlagFilterPropertyValue>,
+    ) -> Result<(), PostHogEvaluationError> {
+        let hash_on_global_rollout_percentage = Self::consistent_hash(user_id, flag_key, "boolean");
+        self.evaluate_boolean_inner(flag_key, hash_on_global_rollout_percentage, properties)
+    }
+
     /// Evaluate a multivariate feature flag. Note that we directly take the mapped user ID
     /// (a consistent hash ranging from 0 to 1) so that it is easier to use it in the tests
     /// and avoid duplicate computations.
@@ -306,6 +349,11 @@ impl FeatureStore {
                     flag_key
                 )));
             }
+            let Some(ref multivariate) = flag_config.filters.multivariate else {
+                return Err(PostHogEvaluationError::Internal(format!(
+                    "No multivariate available, should use evaluate_boolean?: {flag_key}"
+                )));
+            };
             // TODO: sort the groups so that variant overrides always get evaluated first and it follows the PostHog
             // Python SDK behavior; for now we do not configure conditions without variant overrides in Neon so it
             // does not matter.
@@ -314,7 +362,7 @@ impl FeatureStore {
                     GroupEvaluationResult::MatchedAndOverride(variant) => return Ok(variant),
                     GroupEvaluationResult::MatchedAndEvaluate => {
                         let mut percentage = 0;
-                        for variant in &flag_config.filters.multivariate.variants {
+                        for variant in &multivariate.variants {
                             percentage += variant.rollout_percentage;
                             if self
                                 .evaluate_percentage(hash_on_global_rollout_percentage, percentage)
@@ -342,6 +390,89 @@ impl FeatureStore {
             )))
         }
     }
+
+    /// Evaluate a multivariate feature flag. Note that we directly take the mapped user ID
+    /// (a consistent hash ranging from 0 to 1) so that it is easier to use it in the tests
+    /// and avoid duplicate computations.
+    ///
+    /// Use a different consistent hash for evaluating the group rollout percentage.
+    /// The behavior: if the condition is set to rolling out to 10% of the users, and
+    /// we set the variant A to 20% in the global config, then 2% of the total users will
+    /// be evaluated to variant A.
+    ///
+    /// Note that the hash to determine group rollout percentage is shared across all groups. So if we have two
+    /// exactly-the-same conditions with 10% and 20% rollout percentage respectively, a total of 20% of the users
+    /// will be evaluated (versus 30% if group evaluation is done independently).
+    pub(crate) fn evaluate_boolean_inner(
+        &self,
+        flag_key: &str,
+        hash_on_global_rollout_percentage: f64,
+        properties: &HashMap<String, PostHogFlagFilterPropertyValue>,
+    ) -> Result<(), PostHogEvaluationError> {
+        if let Some(flag_config) = self.flags.get(flag_key) {
+            if !flag_config.active {
+                return Err(PostHogEvaluationError::NotAvailable(format!(
+                    "The feature flag is not active: {}",
+                    flag_key
+                )));
+            }
+            if flag_config.filters.multivariate.is_some() {
+                return Err(PostHogEvaluationError::Internal(format!(
+                    "This looks like a multivariate flag, should use evaluate_multivariate?: {flag_key}"
+                )));
+            };
+            // TODO: sort the groups so that variant overrides always get evaluated first and it follows the PostHog
+            // Python SDK behavior; for now we do not configure conditions without variant overrides in Neon so it
+            // does not matter.
+            for group in &flag_config.filters.groups {
+                match self.evaluate_group(group, hash_on_global_rollout_percentage, properties)? {
+                    GroupEvaluationResult::MatchedAndOverride(_) => {
+                        return Err(PostHogEvaluationError::Internal(format!(
+                            "Boolean flag cannot have overrides: {}",
+                            flag_key
+                        )));
+                    }
+                    GroupEvaluationResult::MatchedAndEvaluate => {
+                        return Ok(());
+                    }
+                    GroupEvaluationResult::Unmatched => continue,
+                }
+            }
+            // If no group is matched, the feature is not available, and up to the caller to decide what to do.
+            Err(PostHogEvaluationError::NoConditionGroupMatched)
+        } else {
+            // The feature flag is not available yet
+            Err(PostHogEvaluationError::NotAvailable(format!(
+                "Not found in the local evaluation spec: {}",
+                flag_key
+            )))
+        }
+    }
+
+    /// Infer whether a feature flag is a boolean flag by checking if it has a multivariate filter.
+    pub fn is_feature_flag_boolean(&self, flag_key: &str) -> Result<bool, PostHogEvaluationError> {
+        if let Some(flag_config) = self.flags.get(flag_key) {
+            Ok(flag_config.filters.multivariate.is_none())
+        } else {
+            Err(PostHogEvaluationError::NotAvailable(format!(
+                "Not found in the local evaluation spec: {}",
+                flag_key
+            )))
+        }
+    }
+}
+
+pub struct PostHogClientConfig {
+    /// The server API key.
+    pub server_api_key: String,
+    /// The client API key.
+    pub client_api_key: String,
+    /// The project ID.
+    pub project_id: String,
+    /// The private API URL.
+    pub private_api_url: String,
+    /// The public API URL.
+    pub public_api_url: String,
 }
 
 /// A lite PostHog client.
@@ -360,37 +491,16 @@ impl FeatureStore {
 /// want to report the feature flag usage back to PostHog. The current plan is to use PostHog only as an UI to
 /// configure feature flags so it is very likely that the client API will not be used.
 pub struct PostHogClient {
-    /// The server API key.
-    server_api_key: String,
-    /// The client API key.
-    client_api_key: String,
-    /// The project ID.
-    project_id: String,
-    /// The private API URL.
-    private_api_url: String,
-    /// The public API URL.
-    public_api_url: String,
+    /// The config.
+    config: PostHogClientConfig,
     /// The HTTP client.
     client: reqwest::Client,
 }
 
 impl PostHogClient {
-    pub fn new(
-        server_api_key: String,
-        client_api_key: String,
-        project_id: String,
-        private_api_url: String,
-        public_api_url: String,
-    ) -> Self {
+    pub fn new(config: PostHogClientConfig) -> Self {
         let client = reqwest::Client::new();
-        Self {
-            server_api_key,
-            client_api_key,
-            project_id,
-            private_api_url,
-            public_api_url,
-            client,
-        }
+        Self { config, client }
     }
 
     pub fn new_with_us_region(
@@ -398,13 +508,13 @@ impl PostHogClient {
         client_api_key: String,
         project_id: String,
     ) -> Self {
-        Self::new(
+        Self::new(PostHogClientConfig {
             server_api_key,
             client_api_key,
             project_id,
-            "https://us.posthog.com".to_string(),
-            "https://us.i.posthog.com".to_string(),
-        )
+            private_api_url: "https://us.posthog.com".to_string(),
+            public_api_url: "https://us.i.posthog.com".to_string(),
+        })
     }
 
     /// Fetch the feature flag specs from the server.
@@ -422,15 +532,23 @@ impl PostHogClient {
         // with bearer token of self.server_api_key
         let url = format!(
             "{}/api/projects/{}/feature_flags/local_evaluation",
-            self.private_api_url, self.project_id
+            self.config.private_api_url, self.config.project_id
         );
         let response = self
             .client
             .get(url)
-            .bearer_auth(&self.server_api_key)
+            .bearer_auth(&self.config.server_api_key)
             .send()
             .await?;
+        let status = response.status();
         let body = response.text().await?;
+        if !status.is_success() {
+            return Err(anyhow::anyhow!(
+                "Failed to get feature flags: {}, {}",
+                status,
+                body
+            ));
+        }
         Ok(serde_json::from_str(&body)?)
     }
 
@@ -446,11 +564,11 @@ impl PostHogClient {
     ) -> anyhow::Result<()> {
         // PUBLIC_URL/capture/
         // with bearer token of self.client_api_key
-        let url = format!("{}/capture/", self.public_api_url);
+        let url = format!("{}/capture/", self.config.public_api_url);
         self.client
             .post(url)
             .body(serde_json::to_string(&json!({
-                "api_key": self.client_api_key,
+                "api_key": self.config.client_api_key,
                 "distinct_id": distinct_id,
                 "event": event,
                 "properties": properties,
@@ -467,95 +585,162 @@ mod tests {
 
     fn data() -> &'static str {
         r#"{
-            "flags": [
-                {
-                    "id": 132794,
-                    "team_id": 152860,
-                    "name": "",
-                    "key": "gc-compaction",
-                    "filters": {
-                        "groups": [
-                            {
-                                "variant": "enabled-stage-2",
-                                "properties": [
-                                    {
-                                        "key": "plan_type",
-                                        "type": "person",
-                                        "value": [
-                                            "free"
-                                        ],
-                                        "operator": "exact"
-                                    },
-                                    {
-                                        "key": "pageserver_remote_size",
-                                        "type": "person",
-                                        "value": "10000000",
-                                        "operator": "lt"
-                                    }
-                                ],
-                                "rollout_percentage": 50
-                            },
-                            {
-                                "properties": [
-                                    {
-                                        "key": "plan_type",
-                                        "type": "person",
-                                        "value": [
-                                            "free"
-                                        ],
-                                        "operator": "exact"
-                                    },
-                                    {
-                                        "key": "pageserver_remote_size",
-                                        "type": "person",
-                                        "value": "10000000",
-                                        "operator": "lt"
-                                    }
-                                ],
-                                "rollout_percentage": 80
-                            }
-                        ],
-                        "payloads": {},
-                        "multivariate": {
-                            "variants": [
-                                {
-                                    "key": "disabled",
-                                    "name": "",
-                                    "rollout_percentage": 90
-                                },
-                                {
-                                    "key": "enabled-stage-1",
-                                    "name": "",
-                                    "rollout_percentage": 10
-                                },
-                                {
-                                    "key": "enabled-stage-2",
-                                    "name": "",
-                                    "rollout_percentage": 0
-                                },
-                                {
-                                    "key": "enabled-stage-3",
-                                    "name": "",
-                                    "rollout_percentage": 0
-                                },
-                                {
-                                    "key": "enabled",
-                                    "name": "",
-                                    "rollout_percentage": 0
-                                }
-                            ]
-                        }
-                    },
-                    "deleted": false,
-                    "active": true,
-                    "ensure_experience_continuity": false,
-                    "has_encrypted_payloads": false,
-                    "version": 6
-                }
+  "flags": [
+    {
+      "id": 141807,
+      "team_id": 152860,
+      "name": "",
+      "key": "image-compaction-boundary",
+      "filters": {
+        "groups": [
+          {
+            "variant": null,
+            "properties": [
+              {
+                "key": "plan_type",
+                "type": "person",
+                "value": [
+                  "free"
+                ],
+                "operator": "exact"
+              }
             ],
-            "group_type_mapping": {},
-            "cohorts": {}
-        }"#
+            "rollout_percentage": 40
+          },
+          {
+            "variant": null,
+            "properties": [],
+            "rollout_percentage": 10
+          }
+        ],
+        "payloads": {},
+        "multivariate": null
+      },
+      "deleted": false,
+      "active": true,
+      "ensure_experience_continuity": false,
+      "has_encrypted_payloads": false,
+      "version": 1
+    },
+    {
+      "id": 135586,
+      "team_id": 152860,
+      "name": "",
+      "key": "boolean-flag",
+      "filters": {
+        "groups": [
+          {
+            "variant": null,
+            "properties": [
+              {
+                "key": "plan_type",
+                "type": "person",
+                "value": [
+                  "free"
+                ],
+                "operator": "exact"
+              }
+            ],
+            "rollout_percentage": 47
+          }
+        ],
+        "payloads": {},
+        "multivariate": null
+      },
+      "deleted": false,
+      "active": true,
+      "ensure_experience_continuity": false,
+      "has_encrypted_payloads": false,
+      "version": 1
+    },
+    {
+      "id": 132794,
+      "team_id": 152860,
+      "name": "",
+      "key": "gc-compaction",
+      "filters": {
+        "groups": [
+          {
+            "variant": "enabled-stage-2",
+            "properties": [
+              {
+                "key": "plan_type",
+                "type": "person",
+                "value": [
+                  "free"
+                ],
+                "operator": "exact"
+              },
+              {
+                "key": "pageserver_remote_size",
+                "type": "person",
+                "value": "10000000",
+                "operator": "lt"
+              }
+            ],
+             "rollout_percentage": 50
+          },
+          {
+            "properties": [
+              {
+                "key": "plan_type",
+                "type": "person",
+                "value": [
+                  "free"
+                ],
+                "operator": "exact"
+              },
+              {
+                "key": "pageserver_remote_size",
+                "type": "person",
+                "value": "10000000",
+                "operator": "lt"
+              }
+            ],
+            "rollout_percentage": 80
+          }
+        ],
+        "payloads": {},
+        "multivariate": {
+          "variants": [
+            {
+              "key": "disabled",
+              "name": "",
+              "rollout_percentage": 90
+            },
+            {
+              "key": "enabled-stage-1",
+              "name": "",
+              "rollout_percentage": 10
+            },
+            {
+              "key": "enabled-stage-2",
+              "name": "",
+              "rollout_percentage": 0
+            },
+            {
+              "key": "enabled-stage-3",
+              "name": "",
+              "rollout_percentage": 0
+            },
+            {
+              "key": "enabled",
+              "name": "",
+              "rollout_percentage": 0
+            }
+          ]
+        }
+      },
+      "deleted": false,
+      "active": true,
+      "ensure_experience_continuity": false,
+      "has_encrypted_payloads": false,
+      "version": 7
+    }
+  ],
+  "group_type_mapping": {},
+  "cohorts": {}
+}"#
     }
 
     #[test]
@@ -631,4 +816,125 @@ mod tests {
             Err(PostHogEvaluationError::NoConditionGroupMatched)
         ),);
     }
+
+    #[test]
+    fn evaluate_boolean_1() {
+        // The `boolean-flag` feature flag only has one group that matches on the free user.
+
+        let mut store = FeatureStore::new();
+        let response: LocalEvaluationResponse = serde_json::from_str(data()).unwrap();
+        store.set_flags(response.flags);
+
+        // This lacks the required properties and cannot be evaluated.
+        let variant = store.evaluate_boolean_inner("boolean-flag", 1.00, &HashMap::new());
+        assert!(matches!(
+            variant,
+            Err(PostHogEvaluationError::NotAvailable(_))
+        ),);
+
+        let properties_unmatched = HashMap::from([
+            (
+                "plan_type".to_string(),
+                PostHogFlagFilterPropertyValue::String("paid".to_string()),
+            ),
+            (
+                "pageserver_remote_size".to_string(),
+                PostHogFlagFilterPropertyValue::Number(1000.0),
+            ),
+        ]);
+
+        // This does not match any group so there will be an error.
+        let variant = store.evaluate_boolean_inner("boolean-flag", 1.00, &properties_unmatched);
+        assert!(matches!(
+            variant,
+            Err(PostHogEvaluationError::NoConditionGroupMatched)
+        ),);
+
+        let properties = HashMap::from([
+            (
+                "plan_type".to_string(),
+                PostHogFlagFilterPropertyValue::String("free".to_string()),
+            ),
+            (
+                "pageserver_remote_size".to_string(),
+                PostHogFlagFilterPropertyValue::Number(1000.0),
+            ),
+        ]);
+
+        // It matches the first group as 0.10 <= 0.50 and the properties are matched. Then it gets evaluated to the variant override.
+        let variant = store.evaluate_boolean_inner("boolean-flag", 0.10, &properties);
+        assert!(variant.is_ok());
+
+        // It matches the group conditions but not the group rollout percentage.
+        let variant = store.evaluate_boolean_inner("boolean-flag", 1.00, &properties);
+        assert!(matches!(
+            variant,
+            Err(PostHogEvaluationError::NoConditionGroupMatched)
+        ),);
+    }
+
+    #[test]
+    fn evaluate_boolean_2() {
+        // The `image-compaction-boundary` feature flag has one group that matches on the free user and a group that matches on all users.
+
+        let mut store = FeatureStore::new();
+        let response: LocalEvaluationResponse = serde_json::from_str(data()).unwrap();
+        store.set_flags(response.flags);
+
+        // This lacks the required properties and cannot be evaluated.
+        let variant =
+            store.evaluate_boolean_inner("image-compaction-boundary", 1.00, &HashMap::new());
+        assert!(matches!(
+            variant,
+            Err(PostHogEvaluationError::NotAvailable(_))
+        ),);
+
+        let properties_unmatched = HashMap::from([
+            (
+                "plan_type".to_string(),
+                PostHogFlagFilterPropertyValue::String("paid".to_string()),
+            ),
+            (
+                "pageserver_remote_size".to_string(),
+                PostHogFlagFilterPropertyValue::Number(1000.0),
+            ),
+        ]);
+
+        // This does not match the filtered group but the all user group.
+        let variant =
+            store.evaluate_boolean_inner("image-compaction-boundary", 1.00, &properties_unmatched);
+        assert!(matches!(
+            variant,
+            Err(PostHogEvaluationError::NoConditionGroupMatched)
+        ),);
+        let variant =
+            store.evaluate_boolean_inner("image-compaction-boundary", 0.05, &properties_unmatched);
+        assert!(variant.is_ok());
+
+        let properties = HashMap::from([
+            (
+                "plan_type".to_string(),
+                PostHogFlagFilterPropertyValue::String("free".to_string()),
+            ),
+            (
+                "pageserver_remote_size".to_string(),
+                PostHogFlagFilterPropertyValue::Number(1000.0),
+            ),
+        ]);
+
+        // It matches the first group as 0.30 <= 0.40 and the properties are matched. Then it gets evaluated to the variant override.
+        let variant = store.evaluate_boolean_inner("image-compaction-boundary", 0.30, &properties);
+        assert!(variant.is_ok());
+
+        // It matches the group conditions but not the group rollout percentage.
+        let variant = store.evaluate_boolean_inner("image-compaction-boundary", 1.00, &properties);
+        assert!(matches!(
+            variant,
+            Err(PostHogEvaluationError::NoConditionGroupMatched)
+        ),);
+
+        // It matches the second "all" group conditions.
+        let variant = store.evaluate_boolean_inner("image-compaction-boundary", 0.09, &properties);
+        assert!(variant.is_ok());
+    }
 }

libs/walproposer/src/walproposer.rs

@@ -1,6 +1,7 @@
 #![allow(clippy::todo)]
 
 use std::ffi::CString;
+use std::str::FromStr;
 
 use postgres_ffi::WAL_SEGMENT_SIZE;
 use utils::id::TenantTimelineId;
@@ -173,6 +174,8 @@ pub struct Config {
     pub ttid: TenantTimelineId,
     /// List of safekeepers in format `host:port`
     pub safekeepers_list: Vec<String>,
+    /// libpq connection info options
+    pub safekeeper_conninfo_options: String,
     /// Safekeeper reconnect timeout in milliseconds
     pub safekeeper_reconnect_timeout: i32,
     /// Safekeeper connection timeout in milliseconds
@@ -202,6 +205,9 @@ impl Wrapper {
             .into_bytes_with_nul();
         assert!(safekeepers_list_vec.len() == safekeepers_list_vec.capacity());
         let safekeepers_list = safekeepers_list_vec.as_mut_ptr() as *mut std::ffi::c_char;
+        let safekeeper_conninfo_options = CString::from_str(&config.safekeeper_conninfo_options)
+            .unwrap()
+            .into_raw();
 
         let callback_data = Box::into_raw(Box::new(api)) as *mut ::std::os::raw::c_void;
 
@@ -209,6 +215,7 @@ impl Wrapper {
             neon_tenant,
             neon_timeline,
             safekeepers_list,
+            safekeeper_conninfo_options,
             safekeeper_reconnect_timeout: config.safekeeper_reconnect_timeout,
             safekeeper_connection_timeout: config.safekeeper_connection_timeout,
             wal_segment_size: WAL_SEGMENT_SIZE as i32, // default 16MB
@@ -576,6 +583,7 @@ mod tests {
         let config = crate::walproposer::Config {
             ttid,
             safekeepers_list: vec!["localhost:5000".to_string()],
+            safekeeper_conninfo_options: String::new(),
             safekeeper_reconnect_timeout: 1000,
             safekeeper_connection_timeout: 10000,
             sync_safekeepers: true,

pageserver/Cargo.toml

@@ -17,51 +17,69 @@ anyhow.workspace = true
 arc-swap.workspace = true
 async-compression.workspace = true
 async-stream.workspace = true
-bit_field.workspace = true
 bincode.workspace = true
+bit_field.workspace = true
 byteorder.workspace = true
 bytes.workspace = true
-camino.workspace = true
 camino-tempfile.workspace = true
+camino.workspace = true
 chrono = { workspace = true, features = ["serde"] }
 clap = { workspace = true, features = ["string"] }
 consumption_metrics.workspace = true
 crc32c.workspace = true
 either.workspace = true
+enum-map.workspace = true
+enumset = { workspace = true, features = ["serde"]}
 fail.workspace = true
 futures.workspace = true
 hashlink.workspace = true
 hex.workspace = true
-humantime.workspace = true
+http-utils.workspace = true
 humantime-serde.workspace = true
+humantime.workspace = true
 hyper0.workspace = true
 itertools.workspace = true
 jsonwebtoken.workspace = true
 md5.workspace = true
+metrics.workspace = true
 nix.workspace = true
-# hack to get the number of worker threads tokio uses
-num_cpus.workspace = true
+num_cpus.workspace = true # hack to get the number of worker threads tokio uses
 num-traits.workspace = true
 once_cell.workspace = true
+pageserver_api.workspace = true
+pageserver_client.workspace = true # for ResponseErrorMessageExt TOOD refactor that
+pageserver_compaction.workspace = true
 pageserver_page_api.workspace = true
+pem.workspace = true
 pin-project-lite.workspace = true
 postgres_backend.workspace = true
+postgres_connection.workspace = true
+postgres_ffi.workspace = true
+postgres_initdb.workspace = true
 postgres-protocol.workspace = true
 postgres-types.workspace = true
-postgres_initdb.workspace = true
+posthog_client_lite.workspace = true
 pprof.workspace = true
+pq_proto.workspace = true
 rand.workspace = true
 range-set-blaze = { version = "0.1.16", features = ["alloc"] }
 regex.workspace = true
+remote_storage.workspace = true
+reqwest.workspace = true
+rpds.workspace = true
 rustls.workspace = true
 scopeguard.workspace = true
 send-future.workspace = true
-serde.workspace = true
 serde_json = { workspace = true, features = ["raw_value"] }
 serde_path_to_error.workspace = true
 serde_with.workspace = true
+serde.workspace = true
+smallvec.workspace = true
+storage_broker.workspace = true
+strum_macros.workspace = true
+strum.workspace = true
 sysinfo.workspace = true
-tokio-tar.workspace = true
+tenant_size_model.workspace = true
 thiserror.workspace = true
 tikv-jemallocator.workspace = true
 tokio = { workspace = true, features = ["process", "sync", "fs", "rt", "io-util", "time"] }
@@ -70,6 +88,7 @@ tokio-io-timeout.workspace = true
 tokio-postgres.workspace = true
 tokio-rustls.workspace = true
 tokio-stream.workspace = true
+tokio-tar.workspace = true
 tokio-util.workspace = true
 toml_edit = { workspace = true, features = [ "serde" ] }
 tonic.workspace = true
@@ -77,29 +96,10 @@ tonic-reflection.workspace = true
 tracing.workspace = true
 tracing-utils.workspace = true
 url.workspace = true
-walkdir.workspace = true
-metrics.workspace = true
-pageserver_api.workspace = true
-pageserver_client.workspace = true # for ResponseErrorMessageExt TOOD refactor that
-pageserver_compaction.workspace = true
-pem.workspace = true
-postgres_connection.workspace = true
-postgres_ffi.workspace = true
-pq_proto.workspace = true
-remote_storage.workspace = true
-storage_broker.workspace = true
-tenant_size_model.workspace = true
-http-utils.workspace = true
 utils.workspace = true
-workspace_hack.workspace = true
-reqwest.workspace = true
-rpds.workspace = true
-enum-map.workspace = true
-enumset = { workspace = true, features = ["serde"]}
-strum.workspace = true
-strum_macros.workspace = true
 wal_decoder.workspace = true
-smallvec.workspace = true
+walkdir.workspace = true
+workspace_hack.workspace = true
 twox-hash.workspace = true
 
 [target.'cfg(target_os = "linux")'.dependencies]

pageserver/benches/bench_metrics.rs

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

pageserver/page_api/Cargo.toml

@@ -5,8 +5,14 @@ edition.workspace = true
 license.workspace = true
 
 [dependencies]
+bytes.workspace = true
+pageserver_api.workspace = true
+postgres_ffi.workspace = true
 prost.workspace = true
+smallvec.workspace = true
+thiserror.workspace = true
 tonic.workspace = true
+utils.workspace = true
 workspace_hack.workspace = true
 
 [build-dependencies]

pageserver/page_api/proto/page_service.proto

@@ -54,9 +54,9 @@ service PageService {
   // RPCs use regular unary requests, since they are not as frequent and
   // performance-critical, and this simplifies implementation.
   //
-  // NB: a status response (e.g. errors) will terminate the stream. The stream
-  // may be shared by e.g. multiple Postgres backends, so we should avoid this.
-  // Most errors are therefore sent as GetPageResponse.status instead.
+  // NB: a gRPC status response (e.g. errors) will terminate the stream. The
+  // stream may be shared by multiple Postgres backends, so we avoid this by
+  // sending them as GetPageResponse.status_code instead.
   rpc GetPages (stream GetPageRequest) returns (stream GetPageResponse);
 
   // Returns the size of a relation, as # of blocks.
@@ -159,8 +159,8 @@ message GetPageRequest {
 // A GetPageRequest class. Primarily intended for observability, but may also be
 // used for prioritization in the future.
 enum GetPageClass {
-  // Unknown class. For forwards compatibility: used when the client sends a
-  // class that the server doesn't know about.
+  // Unknown class. For backwards compatibility: used when an older client version sends a class
+  // that a newer server version has removed.
   GET_PAGE_CLASS_UNKNOWN = 0;
   // A normal request. This is the default.
   GET_PAGE_CLASS_NORMAL = 1;
@@ -180,31 +180,37 @@ message GetPageResponse {
   // The original request's ID.
   uint64 request_id = 1;
   // The response status code.
-  GetPageStatus status = 2;
+  GetPageStatusCode status_code = 2;
   // A string describing the status, if any.
   string reason = 3;
-  // The 8KB page images, in the same order as the request. Empty if status != OK.
+  // The 8KB page images, in the same order as the request. Empty if status_code != OK.
   repeated bytes page_image = 4;
 }
 
-// A GetPageResponse status code. Since we use a bidirectional stream, we don't
-// want to send errors as gRPC statuses, since this would terminate the stream.
-enum GetPageStatus {
-  // Unknown status. For forwards compatibility: used when the server sends a
-  // status code that the client doesn't know about.
-  GET_PAGE_STATUS_UNKNOWN = 0;
+// A GetPageResponse status code.
+//
+// These are effectively equivalent to gRPC statuses. However, we use a bidirectional stream
+// (potentially shared by many backends), and a gRPC status response would terminate the stream so
+// we send GetPageResponse messages with these codes instead.
+enum GetPageStatusCode {
+  // Unknown status. For forwards compatibility: used when an older client version receives a new
+  // status code from a newer server version.
+  GET_PAGE_STATUS_CODE_UNKNOWN = 0;
   // The request was successful.
-  GET_PAGE_STATUS_OK = 1;
+  GET_PAGE_STATUS_CODE_OK = 1;
   // The page did not exist. The tenant/timeline/shard has already been
   // validated during stream setup.
-  GET_PAGE_STATUS_NOT_FOUND = 2;
+  GET_PAGE_STATUS_CODE_NOT_FOUND = 2;
   // The request was invalid.
-  GET_PAGE_STATUS_INVALID = 3;
+  GET_PAGE_STATUS_CODE_INVALID_REQUEST = 3;
+  // The request failed due to an internal server error.
+  GET_PAGE_STATUS_CODE_INTERNAL_ERROR = 4;
   // The tenant is rate limited. Slow down and retry later.
-  GET_PAGE_STATUS_SLOW_DOWN = 4;
-  // TODO: consider adding a GET_PAGE_STATUS_LAYER_DOWNLOAD in the case of a
-  // layer download. This could free up the server task to process other
-  // requests while the layer download is in progress.
+  GET_PAGE_STATUS_CODE_SLOW_DOWN = 5;
+  // NB: shutdown errors are emitted as a gRPC Unavailable status.
+  //
+  // TODO: consider adding a GET_PAGE_STATUS_CODE_LAYER_DOWNLOAD in the case of a layer download.
+  // This could free up the server task to process other requests while the download is in progress.
 }
 
 // Fetches the size of a relation at a given LSN, as # of blocks. Only valid on

pageserver/page_api/src/lib.rs

@@ -17,3 +17,7 @@ pub mod proto {
     pub use page_service_client::PageServiceClient;
     pub use page_service_server::{PageService, PageServiceServer};
 }
+
+mod model;
+
+pub use model::*;

pageserver/page_api/src/model.rs

@@ -0,0 +1,595 @@
+//! Structs representing the canonical page service API.
+//!
+//! These mirror the autogenerated Protobuf types. The differences are:
+//!
+//! - Types that are in fact required by the API are not Options. The protobuf "required"
+//!   attribute is deprecated and 'prost' marks a lot of members as optional because of that.
+//!   (See <https://github.com/tokio-rs/prost/issues/800> for a gripe on this)
+//!
+//! - Use more precise datatypes, e.g. Lsn and uints shorter than 32 bits.
+//!
+//! - Validate protocol invariants, via try_from() and try_into().
+
+use bytes::Bytes;
+use postgres_ffi::Oid;
+use smallvec::SmallVec;
+// TODO: split out Lsn, RelTag, SlruKind, Oid and other basic types to a separate crate, to avoid
+// pulling in all of their other crate dependencies when building the client.
+use utils::lsn::Lsn;
+
+use crate::proto;
+
+/// A protocol error. Typically returned via try_from() or try_into().
+#[derive(thiserror::Error, Debug)]
+pub enum ProtocolError {
+    #[error("field '{0}' has invalid value '{1}'")]
+    Invalid(&'static str, String),
+    #[error("required field '{0}' is missing")]
+    Missing(&'static str),
+}
+
+impl ProtocolError {
+    /// Helper to generate a new ProtocolError::Invalid for the given field and value.
+    pub fn invalid(field: &'static str, value: impl std::fmt::Debug) -> Self {
+        Self::Invalid(field, format!("{value:?}"))
+    }
+}
+
+impl From<ProtocolError> for tonic::Status {
+    fn from(err: ProtocolError) -> Self {
+        tonic::Status::invalid_argument(format!("{err}"))
+    }
+}
+
+/// The LSN a request should read at.
+#[derive(Clone, Copy, Debug)]
+pub struct ReadLsn {
+    /// The request's read LSN.
+    pub request_lsn: Lsn,
+    /// If given, the caller guarantees that the page has not been modified since this LSN. Must be
+    /// smaller than or equal to request_lsn. This allows the Pageserver to serve an old page
+    /// without waiting for the request LSN to arrive. Valid for all request types.
+    ///
+    /// It is undefined behaviour to make a request such that the page was, in fact, modified
+    /// between request_lsn and not_modified_since_lsn. The Pageserver might detect it and return an
+    /// error, or it might return the old page version or the new page version. Setting
+    /// not_modified_since_lsn equal to request_lsn is always safe, but can lead to unnecessary
+    /// waiting.
+    pub not_modified_since_lsn: Option<Lsn>,
+}
+
+impl ReadLsn {
+    /// Validates the ReadLsn.
+    pub fn validate(&self) -> Result<(), ProtocolError> {
+        if self.request_lsn == Lsn::INVALID {
+            return Err(ProtocolError::invalid("request_lsn", self.request_lsn));
+        }
+        if self.not_modified_since_lsn > Some(self.request_lsn) {
+            return Err(ProtocolError::invalid(
+                "not_modified_since_lsn",
+                self.not_modified_since_lsn,
+            ));
+        }
+        Ok(())
+    }
+}
+
+impl TryFrom<proto::ReadLsn> for ReadLsn {
+    type Error = ProtocolError;
+
+    fn try_from(pb: proto::ReadLsn) -> Result<Self, Self::Error> {
+        let read_lsn = Self {
+            request_lsn: Lsn(pb.request_lsn),
+            not_modified_since_lsn: match pb.not_modified_since_lsn {
+                0 => None,
+                lsn => Some(Lsn(lsn)),
+            },
+        };
+        read_lsn.validate()?;
+        Ok(read_lsn)
+    }
+}
+
+impl TryFrom<ReadLsn> for proto::ReadLsn {
+    type Error = ProtocolError;
+
+    fn try_from(read_lsn: ReadLsn) -> Result<Self, Self::Error> {
+        read_lsn.validate()?;
+        Ok(Self {
+            request_lsn: read_lsn.request_lsn.0,
+            not_modified_since_lsn: read_lsn.not_modified_since_lsn.unwrap_or_default().0,
+        })
+    }
+}
+
+// RelTag is defined in pageserver_api::reltag.
+pub type RelTag = pageserver_api::reltag::RelTag;
+
+impl TryFrom<proto::RelTag> for RelTag {
+    type Error = ProtocolError;
+
+    fn try_from(pb: proto::RelTag) -> Result<Self, Self::Error> {
+        Ok(Self {
+            spcnode: pb.spc_oid,
+            dbnode: pb.db_oid,
+            relnode: pb.rel_number,
+            forknum: pb
+                .fork_number
+                .try_into()
+                .map_err(|_| ProtocolError::invalid("fork_number", pb.fork_number))?,
+        })
+    }
+}
+
+impl From<RelTag> for proto::RelTag {
+    fn from(rel_tag: RelTag) -> Self {
+        Self {
+            spc_oid: rel_tag.spcnode,
+            db_oid: rel_tag.dbnode,
+            rel_number: rel_tag.relnode,
+            fork_number: rel_tag.forknum as u32,
+        }
+    }
+}
+
+/// Checks whether a relation exists, at the given LSN. Only valid on shard 0, other shards error.
+#[derive(Clone, Copy, Debug)]
+pub struct CheckRelExistsRequest {
+    pub read_lsn: ReadLsn,
+    pub rel: RelTag,
+}
+
+impl TryFrom<proto::CheckRelExistsRequest> for CheckRelExistsRequest {
+    type Error = ProtocolError;
+
+    fn try_from(pb: proto::CheckRelExistsRequest) -> Result<Self, Self::Error> {
+        Ok(Self {
+            read_lsn: pb
+                .read_lsn
+                .ok_or(ProtocolError::Missing("read_lsn"))?
+                .try_into()?,
+            rel: pb.rel.ok_or(ProtocolError::Missing("rel"))?.try_into()?,
+        })
+    }
+}
+
+pub type CheckRelExistsResponse = bool;
+
+impl From<proto::CheckRelExistsResponse> for CheckRelExistsResponse {
+    fn from(pb: proto::CheckRelExistsResponse) -> Self {
+        pb.exists
+    }
+}
+
+impl From<CheckRelExistsResponse> for proto::CheckRelExistsResponse {
+    fn from(exists: CheckRelExistsResponse) -> Self {
+        Self { exists }
+    }
+}
+
+/// Requests a base backup at a given LSN.
+#[derive(Clone, Copy, Debug)]
+pub struct GetBaseBackupRequest {
+    /// The LSN to fetch a base backup at.
+    pub read_lsn: ReadLsn,
+    /// If true, logical replication slots will not be created.
+    pub replica: bool,
+}
+
+impl TryFrom<proto::GetBaseBackupRequest> for GetBaseBackupRequest {
+    type Error = ProtocolError;
+
+    fn try_from(pb: proto::GetBaseBackupRequest) -> Result<Self, Self::Error> {
+        Ok(Self {
+            read_lsn: pb
+                .read_lsn
+                .ok_or(ProtocolError::Missing("read_lsn"))?
+                .try_into()?,
+            replica: pb.replica,
+        })
+    }
+}
+
+impl TryFrom<GetBaseBackupRequest> for proto::GetBaseBackupRequest {
+    type Error = ProtocolError;
+
+    fn try_from(request: GetBaseBackupRequest) -> Result<Self, Self::Error> {
+        Ok(Self {
+            read_lsn: Some(request.read_lsn.try_into()?),
+            replica: request.replica,
+        })
+    }
+}
+
+pub type GetBaseBackupResponseChunk = Bytes;
+
+impl TryFrom<proto::GetBaseBackupResponseChunk> for GetBaseBackupResponseChunk {
+    type Error = ProtocolError;
+
+    fn try_from(pb: proto::GetBaseBackupResponseChunk) -> Result<Self, Self::Error> {
+        if pb.chunk.is_empty() {
+            return Err(ProtocolError::Missing("chunk"));
+        }
+        Ok(pb.chunk)
+    }
+}
+
+impl TryFrom<GetBaseBackupResponseChunk> for proto::GetBaseBackupResponseChunk {
+    type Error = ProtocolError;
+
+    fn try_from(chunk: GetBaseBackupResponseChunk) -> Result<Self, Self::Error> {
+        if chunk.is_empty() {
+            return Err(ProtocolError::Missing("chunk"));
+        }
+        Ok(Self { chunk })
+    }
+}
+
+/// Requests the size of a database, as # of bytes. Only valid on shard 0, other shards will error.
+#[derive(Clone, Copy, Debug)]
+pub struct GetDbSizeRequest {
+    pub read_lsn: ReadLsn,
+    pub db_oid: Oid,
+}
+
+impl TryFrom<proto::GetDbSizeRequest> for GetDbSizeRequest {
+    type Error = ProtocolError;
+
+    fn try_from(pb: proto::GetDbSizeRequest) -> Result<Self, Self::Error> {
+        Ok(Self {
+            read_lsn: pb
+                .read_lsn
+                .ok_or(ProtocolError::Missing("read_lsn"))?
+                .try_into()?,
+            db_oid: pb.db_oid,
+        })
+    }
+}
+
+impl TryFrom<GetDbSizeRequest> for proto::GetDbSizeRequest {
+    type Error = ProtocolError;
+
+    fn try_from(request: GetDbSizeRequest) -> Result<Self, Self::Error> {
+        Ok(Self {
+            read_lsn: Some(request.read_lsn.try_into()?),
+            db_oid: request.db_oid,
+        })
+    }
+}
+
+pub type GetDbSizeResponse = u64;
+
+impl From<proto::GetDbSizeResponse> for GetDbSizeResponse {
+    fn from(pb: proto::GetDbSizeResponse) -> Self {
+        pb.num_bytes
+    }
+}
+
+impl From<GetDbSizeResponse> for proto::GetDbSizeResponse {
+    fn from(num_bytes: GetDbSizeResponse) -> Self {
+        Self { num_bytes }
+    }
+}
+
+/// Requests one or more pages.
+#[derive(Clone, Debug)]
+pub struct GetPageRequest {
+    /// A request ID. Will be included in the response. Should be unique for in-flight requests on
+    /// the stream.
+    pub request_id: RequestID,
+    /// The request class.
+    pub request_class: GetPageClass,
+    /// The LSN to read at.
+    pub read_lsn: ReadLsn,
+    /// The relation to read from.
+    pub rel: RelTag,
+    /// Page numbers to read. Must belong to the remote shard.
+    ///
+    /// Multiple pages will be executed as a single batch by the Pageserver, amortizing layer access
+    /// costs and parallelizing them. This may increase the latency of any individual request, but
+    /// improves the overall latency and throughput of the batch as a whole.
+    pub block_numbers: SmallVec<[u32; 1]>,
+}
+
+impl TryFrom<proto::GetPageRequest> for GetPageRequest {
+    type Error = ProtocolError;
+
+    fn try_from(pb: proto::GetPageRequest) -> Result<Self, Self::Error> {
+        if pb.block_number.is_empty() {
+            return Err(ProtocolError::Missing("block_number"));
+        }
+        Ok(Self {
+            request_id: pb.request_id,
+            request_class: pb.request_class.into(),
+            read_lsn: pb
+                .read_lsn
+                .ok_or(ProtocolError::Missing("read_lsn"))?
+                .try_into()?,
+            rel: pb.rel.ok_or(ProtocolError::Missing("rel"))?.try_into()?,
+            block_numbers: pb.block_number.into(),
+        })
+    }
+}
+
+impl TryFrom<GetPageRequest> for proto::GetPageRequest {
+    type Error = ProtocolError;
+
+    fn try_from(request: GetPageRequest) -> Result<Self, Self::Error> {
+        if request.block_numbers.is_empty() {
+            return Err(ProtocolError::Missing("block_number"));
+        }
+        Ok(Self {
+            request_id: request.request_id,
+            request_class: request.request_class.into(),
+            read_lsn: Some(request.read_lsn.try_into()?),
+            rel: Some(request.rel.into()),
+            block_number: request.block_numbers.into_vec(),
+        })
+    }
+}
+
+/// A GetPage request ID.
+pub type RequestID = u64;
+
+/// A GetPage request class.
+#[derive(Clone, Copy, Debug)]
+pub enum GetPageClass {
+    /// Unknown class. For backwards compatibility: used when an older client version sends a class
+    /// that a newer server version has removed.
+    Unknown,
+    /// A normal request. This is the default.
+    Normal,
+    /// A prefetch request. NB: can only be classified on pg < 18.
+    Prefetch,
+    /// A background request (e.g. vacuum).
+    Background,
+}
+
+impl From<proto::GetPageClass> for GetPageClass {
+    fn from(pb: proto::GetPageClass) -> Self {
+        match pb {
+            proto::GetPageClass::Unknown => Self::Unknown,
+            proto::GetPageClass::Normal => Self::Normal,
+            proto::GetPageClass::Prefetch => Self::Prefetch,
+            proto::GetPageClass::Background => Self::Background,
+        }
+    }
+}
+
+impl From<i32> for GetPageClass {
+    fn from(class: i32) -> Self {
+        proto::GetPageClass::try_from(class)
+            .unwrap_or(proto::GetPageClass::Unknown)
+            .into()
+    }
+}
+
+impl From<GetPageClass> for proto::GetPageClass {
+    fn from(class: GetPageClass) -> Self {
+        match class {
+            GetPageClass::Unknown => Self::Unknown,
+            GetPageClass::Normal => Self::Normal,
+            GetPageClass::Prefetch => Self::Prefetch,
+            GetPageClass::Background => Self::Background,
+        }
+    }
+}
+
+impl From<GetPageClass> for i32 {
+    fn from(class: GetPageClass) -> Self {
+        proto::GetPageClass::from(class).into()
+    }
+}
+
+/// A GetPage response.
+///
+/// A batch response will contain all of the requested pages. We could eagerly emit individual pages
+/// as soon as they are ready, but on a readv() Postgres holds buffer pool locks on all pages in the
+/// batch and we'll only return once the entire batch is ready, so no one can make use of the
+/// individual pages.
+#[derive(Clone, Debug)]
+pub struct GetPageResponse {
+    /// The original request's ID.
+    pub request_id: RequestID,
+    /// The response status code.
+    pub status_code: GetPageStatusCode,
+    /// A string describing the status, if any.
+    pub reason: Option<String>,
+    /// The 8KB page images, in the same order as the request. Empty if status != OK.
+    pub page_images: SmallVec<[Bytes; 1]>,
+}
+
+impl From<proto::GetPageResponse> for GetPageResponse {
+    fn from(pb: proto::GetPageResponse) -> Self {
+        Self {
+            request_id: pb.request_id,
+            status_code: pb.status_code.into(),
+            reason: Some(pb.reason).filter(|r| !r.is_empty()),
+            page_images: pb.page_image.into(),
+        }
+    }
+}
+
+impl From<GetPageResponse> for proto::GetPageResponse {
+    fn from(response: GetPageResponse) -> Self {
+        Self {
+            request_id: response.request_id,
+            status_code: response.status_code.into(),
+            reason: response.reason.unwrap_or_default(),
+            page_image: response.page_images.into_vec(),
+        }
+    }
+}
+
+/// A GetPage response status code.
+///
+/// These are effectively equivalent to gRPC statuses. However, we use a bidirectional stream
+/// (potentially shared by many backends), and a gRPC status response would terminate the stream so
+/// we send GetPageResponse messages with these codes instead.
+#[derive(Clone, Copy, Debug)]
+pub enum GetPageStatusCode {
+    /// Unknown status. For forwards compatibility: used when an older client version receives a new
+    /// status code from a newer server version.
+    Unknown,
+    /// The request was successful.
+    Ok,
+    /// The page did not exist. The tenant/timeline/shard has already been validated during stream
+    /// setup.
+    NotFound,
+    /// The request was invalid.
+    InvalidRequest,
+    /// The request failed due to an internal server error.
+    InternalError,
+    /// The tenant is rate limited. Slow down and retry later.
+    SlowDown,
+}
+
+impl From<proto::GetPageStatusCode> for GetPageStatusCode {
+    fn from(pb: proto::GetPageStatusCode) -> Self {
+        match pb {
+            proto::GetPageStatusCode::Unknown => Self::Unknown,
+            proto::GetPageStatusCode::Ok => Self::Ok,
+            proto::GetPageStatusCode::NotFound => Self::NotFound,
+            proto::GetPageStatusCode::InvalidRequest => Self::InvalidRequest,
+            proto::GetPageStatusCode::InternalError => Self::InternalError,
+            proto::GetPageStatusCode::SlowDown => Self::SlowDown,
+        }
+    }
+}
+
+impl From<i32> for GetPageStatusCode {
+    fn from(status_code: i32) -> Self {
+        proto::GetPageStatusCode::try_from(status_code)
+            .unwrap_or(proto::GetPageStatusCode::Unknown)
+            .into()
+    }
+}
+
+impl From<GetPageStatusCode> for proto::GetPageStatusCode {
+    fn from(status_code: GetPageStatusCode) -> Self {
+        match status_code {
+            GetPageStatusCode::Unknown => Self::Unknown,
+            GetPageStatusCode::Ok => Self::Ok,
+            GetPageStatusCode::NotFound => Self::NotFound,
+            GetPageStatusCode::InvalidRequest => Self::InvalidRequest,
+            GetPageStatusCode::InternalError => Self::InternalError,
+            GetPageStatusCode::SlowDown => Self::SlowDown,
+        }
+    }
+}
+
+impl From<GetPageStatusCode> for i32 {
+    fn from(status_code: GetPageStatusCode) -> Self {
+        proto::GetPageStatusCode::from(status_code).into()
+    }
+}
+
+// Fetches the size of a relation at a given LSN, as # of blocks. Only valid on shard 0, other
+// shards will error.
+pub struct GetRelSizeRequest {
+    pub read_lsn: ReadLsn,
+    pub rel: RelTag,
+}
+
+impl TryFrom<proto::GetRelSizeRequest> for GetRelSizeRequest {
+    type Error = ProtocolError;
+
+    fn try_from(proto: proto::GetRelSizeRequest) -> Result<Self, Self::Error> {
+        Ok(Self {
+            read_lsn: proto
+                .read_lsn
+                .ok_or(ProtocolError::Missing("read_lsn"))?
+                .try_into()?,
+            rel: proto.rel.ok_or(ProtocolError::Missing("rel"))?.try_into()?,
+        })
+    }
+}
+
+impl TryFrom<GetRelSizeRequest> for proto::GetRelSizeRequest {
+    type Error = ProtocolError;
+
+    fn try_from(request: GetRelSizeRequest) -> Result<Self, Self::Error> {
+        Ok(Self {
+            read_lsn: Some(request.read_lsn.try_into()?),
+            rel: Some(request.rel.into()),
+        })
+    }
+}
+
+pub type GetRelSizeResponse = u32;
+
+impl From<proto::GetRelSizeResponse> for GetRelSizeResponse {
+    fn from(proto: proto::GetRelSizeResponse) -> Self {
+        proto.num_blocks
+    }
+}
+
+impl From<GetRelSizeResponse> for proto::GetRelSizeResponse {
+    fn from(num_blocks: GetRelSizeResponse) -> Self {
+        Self { num_blocks }
+    }
+}
+
+/// Requests an SLRU segment. Only valid on shard 0, other shards will error.
+pub struct GetSlruSegmentRequest {
+    pub read_lsn: ReadLsn,
+    pub kind: SlruKind,
+    pub segno: u32,
+}
+
+impl TryFrom<proto::GetSlruSegmentRequest> for GetSlruSegmentRequest {
+    type Error = ProtocolError;
+
+    fn try_from(pb: proto::GetSlruSegmentRequest) -> Result<Self, Self::Error> {
+        Ok(Self {
+            read_lsn: pb
+                .read_lsn
+                .ok_or(ProtocolError::Missing("read_lsn"))?
+                .try_into()?,
+            kind: u8::try_from(pb.kind)
+                .ok()
+                .and_then(SlruKind::from_repr)
+                .ok_or_else(|| ProtocolError::invalid("slru_kind", pb.kind))?,
+            segno: pb.segno,
+        })
+    }
+}
+
+impl TryFrom<GetSlruSegmentRequest> for proto::GetSlruSegmentRequest {
+    type Error = ProtocolError;
+
+    fn try_from(request: GetSlruSegmentRequest) -> Result<Self, Self::Error> {
+        Ok(Self {
+            read_lsn: Some(request.read_lsn.try_into()?),
+            kind: request.kind as u32,
+            segno: request.segno,
+        })
+    }
+}
+
+pub type GetSlruSegmentResponse = Bytes;
+
+impl TryFrom<proto::GetSlruSegmentResponse> for GetSlruSegmentResponse {
+    type Error = ProtocolError;
+
+    fn try_from(pb: proto::GetSlruSegmentResponse) -> Result<Self, Self::Error> {
+        if pb.segment.is_empty() {
+            return Err(ProtocolError::Missing("segment"));
+        }
+        Ok(pb.segment)
+    }
+}
+
+impl TryFrom<GetSlruSegmentResponse> for proto::GetSlruSegmentResponse {
+    type Error = ProtocolError;
+
+    fn try_from(segment: GetSlruSegmentResponse) -> Result<Self, Self::Error> {
+        if segment.is_empty() {
+            return Err(ProtocolError::Missing("segment"));
+        }
+        Ok(Self { segment })
+    }
+}
+
+// SlruKind is defined in pageserver_api::reltag.
+pub type SlruKind = pageserver_api::reltag::SlruKind;

pageserver/src/bin/pageserver.rs

@@ -21,6 +21,7 @@ use pageserver::config::{PageServerConf, PageserverIdentity, ignored_fields};
 use pageserver::controller_upcall_client::StorageControllerUpcallClient;
 use pageserver::deletion_queue::DeletionQueue;
 use pageserver::disk_usage_eviction_task::{self, launch_disk_usage_global_eviction_task};
+use pageserver::feature_resolver::FeatureResolver;
 use pageserver::metrics::{STARTUP_DURATION, STARTUP_IS_LOADING};
 use pageserver::task_mgr::{
     BACKGROUND_RUNTIME, COMPUTE_REQUEST_RUNTIME, MGMT_REQUEST_RUNTIME, WALRECEIVER_RUNTIME,
@@ -522,6 +523,12 @@ fn start_pageserver(
     // Set up remote storage client
     let remote_storage = BACKGROUND_RUNTIME.block_on(create_remote_storage_client(conf))?;
 
+    let feature_resolver = create_feature_resolver(
+        conf,
+        shutdown_pageserver.clone(),
+        BACKGROUND_RUNTIME.handle(),
+    )?;
+
     // Set up deletion queue
     let (deletion_queue, deletion_workers) = DeletionQueue::new(
         remote_storage.clone(),
@@ -575,6 +582,7 @@ fn start_pageserver(
             deletion_queue_client,
             l0_flush_global_state,
             basebackup_prepare_sender,
+            feature_resolver,
         },
         order,
         shutdown_pageserver.clone(),
@@ -849,6 +857,14 @@ fn start_pageserver(
     })
 }
 
+fn create_feature_resolver(
+    conf: &'static PageServerConf,
+    shutdown_pageserver: CancellationToken,
+    handle: &tokio::runtime::Handle,
+) -> anyhow::Result<FeatureResolver> {
+    FeatureResolver::spawn(conf, shutdown_pageserver, handle)
+}
+
 async fn create_remote_storage_client(
     conf: &'static PageServerConf,
 ) -> anyhow::Result<GenericRemoteStorage> {

pageserver/src/config.rs

@@ -14,7 +14,7 @@ use std::time::Duration;
 use anyhow::{Context, bail, ensure};
 use camino::{Utf8Path, Utf8PathBuf};
 use once_cell::sync::OnceCell;
-use pageserver_api::config::{DiskUsageEvictionTaskConfig, MaxVectoredReadBytes};
+use pageserver_api::config::{DiskUsageEvictionTaskConfig, MaxVectoredReadBytes, PostHogConfig};
 use pageserver_api::models::ImageCompressionAlgorithm;
 use pageserver_api::shard::TenantShardId;
 use pem::Pem;
@@ -238,6 +238,9 @@ pub struct PageServerConf {
     /// This is insecure and should only be used in development environments.
     pub dev_mode: bool,
 
+    /// PostHog integration config.
+    pub posthog_config: Option<PostHogConfig>,
+
     pub timeline_import_config: pageserver_api::config::TimelineImportConfig,
 
     pub basebackup_cache_config: Option<pageserver_api::config::BasebackupCacheConfig>,
@@ -421,6 +424,7 @@ impl PageServerConf {
             tracing,
             enable_tls_page_service_api,
             dev_mode,
+            posthog_config,
             timeline_import_config,
             basebackup_cache_config,
         } = config_toml;
@@ -536,6 +540,7 @@ impl PageServerConf {
                 }
                 None => Vec::new(),
             },
+            posthog_config,
         };
 
         // ------------------------------------------------------------

pageserver/src/disk_usage_eviction_task.rs

@@ -837,7 +837,30 @@ async fn collect_eviction_candidates(
                 continue;
             }
             let info = tl.get_local_layers_for_disk_usage_eviction().await;
-            debug!(tenant_id=%tl.tenant_shard_id.tenant_id, shard_id=%tl.tenant_shard_id.shard_slug(), timeline_id=%tl.timeline_id, "timeline resident layers count: {}", info.resident_layers.len());
+            debug!(
+                tenant_id=%tl.tenant_shard_id.tenant_id,
+                shard_id=%tl.tenant_shard_id.shard_slug(),
+                timeline_id=%tl.timeline_id,
+                "timeline resident layers count: {}", info.resident_layers.len()
+            );
+
+            tenant_candidates.extend(info.resident_layers.into_iter());
+            max_layer_size = max_layer_size.max(info.max_layer_size.unwrap_or(0));
+
+            if cancel.is_cancelled() {
+                return Ok(EvictionCandidates::Cancelled);
+            }
+        }
+
+        // Also consider layers of timelines being imported for eviction
+        for tl in tenant.list_importing_timelines() {
+            let info = tl.timeline.get_local_layers_for_disk_usage_eviction().await;
+            debug!(
+                tenant_id=%tl.timeline.tenant_shard_id.tenant_id,
+                shard_id=%tl.timeline.tenant_shard_id.shard_slug(),
+                timeline_id=%tl.timeline.timeline_id,
+                "timeline resident layers count: {}", info.resident_layers.len()
+            );
 
             tenant_candidates.extend(info.resident_layers.into_iter());
             max_layer_size = max_layer_size.max(info.max_layer_size.unwrap_or(0));

pageserver/src/feature_resolver.rs

@@ -0,0 +1,104 @@
+use std::{collections::HashMap, sync::Arc, time::Duration};
+
+use posthog_client_lite::{
+    FeatureResolverBackgroundLoop, PostHogClientConfig, PostHogEvaluationError,
+};
+use tokio_util::sync::CancellationToken;
+use utils::id::TenantId;
+
+use crate::config::PageServerConf;
+
+#[derive(Clone)]
+pub struct FeatureResolver {
+    inner: Option<Arc<FeatureResolverBackgroundLoop>>,
+}
+
+impl FeatureResolver {
+    pub fn new_disabled() -> Self {
+        Self { inner: None }
+    }
+
+    pub fn spawn(
+        conf: &PageServerConf,
+        shutdown_pageserver: CancellationToken,
+        handle: &tokio::runtime::Handle,
+    ) -> anyhow::Result<Self> {
+        // DO NOT block in this function: make it return as fast as possible to avoid startup delays.
+        if let Some(posthog_config) = &conf.posthog_config {
+            let inner = FeatureResolverBackgroundLoop::new(
+                PostHogClientConfig {
+                    server_api_key: posthog_config.server_api_key.clone(),
+                    client_api_key: posthog_config.client_api_key.clone(),
+                    project_id: posthog_config.project_id.clone(),
+                    private_api_url: posthog_config.private_api_url.clone(),
+                    public_api_url: posthog_config.public_api_url.clone(),
+                },
+                shutdown_pageserver,
+            );
+            let inner = Arc::new(inner);
+            // TODO: make this configurable
+            inner.clone().spawn(handle, Duration::from_secs(60));
+            Ok(FeatureResolver { inner: Some(inner) })
+        } else {
+            Ok(FeatureResolver { inner: None })
+        }
+    }
+
+    /// Evaluate a multivariate feature flag. Currently, we do not support any properties.
+    ///
+    /// Error handling: the caller should inspect the error and decide the behavior when a feature flag
+    /// cannot be evaluated (i.e., default to false if it cannot be resolved). The error should *not* be
+    /// propagated beyond where the feature flag gets resolved.
+    pub fn evaluate_multivariate(
+        &self,
+        flag_key: &str,
+        tenant_id: TenantId,
+    ) -> Result<String, PostHogEvaluationError> {
+        if let Some(inner) = &self.inner {
+            inner.feature_store().evaluate_multivariate(
+                flag_key,
+                &tenant_id.to_string(),
+                &HashMap::new(),
+            )
+        } else {
+            Err(PostHogEvaluationError::NotAvailable(
+                "PostHog integration is not enabled".to_string(),
+            ))
+        }
+    }
+
+    /// Evaluate a boolean feature flag. Currently, we do not support any properties.
+    ///
+    /// Returns `Ok(())` if the flag is evaluated to true, otherwise returns an error.
+    ///
+    /// Error handling: the caller should inspect the error and decide the behavior when a feature flag
+    /// cannot be evaluated (i.e., default to false if it cannot be resolved). The error should *not* be
+    /// propagated beyond where the feature flag gets resolved.
+    pub fn evaluate_boolean(
+        &self,
+        flag_key: &str,
+        tenant_id: TenantId,
+    ) -> Result<(), PostHogEvaluationError> {
+        if let Some(inner) = &self.inner {
+            inner.feature_store().evaluate_boolean(
+                flag_key,
+                &tenant_id.to_string(),
+                &HashMap::new(),
+            )
+        } else {
+            Err(PostHogEvaluationError::NotAvailable(
+                "PostHog integration is not enabled".to_string(),
+            ))
+        }
+    }
+
+    pub fn is_feature_flag_boolean(&self, flag_key: &str) -> Result<bool, PostHogEvaluationError> {
+        if let Some(inner) = &self.inner {
+            inner.feature_store().is_feature_flag_boolean(flag_key)
+        } else {
+            Err(PostHogEvaluationError::NotAvailable(
+                "PostHog integration is not enabled".to_string(),
+            ))
+        }
+    }
+}

pageserver/src/http/openapi_spec.yml

@@ -353,6 +353,33 @@ paths:
         "200":
           description: OK
 
+  /v1/tenant/{tenant_shard_id}/timeline/{timeline_id}/mark_invisible:
+    parameters:
+      - name: tenant_shard_id
+        in: path
+        required: true
+        schema:
+          type: string
+      - name: timeline_id
+        in: path
+        required: true
+        schema:
+          type: string
+          format: hex
+    put:
+      requestBody:
+        content:
+          application/json:
+            schema:
+              type: object
+              properties:
+                is_visible:
+                  type: boolean
+                  default: false
+      responses:
+        "200":
+          description: OK
+
   /v1/tenant/{tenant_shard_id}/location_config:
     parameters:
       - name: tenant_shard_id
@@ -626,6 +653,8 @@ paths:
                   format: hex
                 pg_version:
                   type: integer
+                read_only:
+                  type: boolean
                 existing_initdb_timeline_id:
                   type: string
                   format: hex

pageserver/src/http/routes.rs

@@ -370,6 +370,18 @@ impl From<crate::tenant::secondary::SecondaryTenantError> for ApiError {
     }
 }
 
+impl From<crate::tenant::FinalizeTimelineImportError> for ApiError {
+    fn from(err: crate::tenant::FinalizeTimelineImportError) -> ApiError {
+        use crate::tenant::FinalizeTimelineImportError::*;
+        match err {
+            ImportTaskStillRunning => {
+                ApiError::ResourceUnavailable("Import task still running".into())
+            }
+            ShuttingDown => ApiError::ShuttingDown,
+        }
+    }
+}
+
 // Helper function to construct a TimelineInfo struct for a timeline
 async fn build_timeline_info(
     timeline: &Arc<Timeline>,
@@ -572,6 +584,7 @@ async fn timeline_create_handler(
         TimelineCreateRequestMode::Branch {
             ancestor_timeline_id,
             ancestor_start_lsn,
+            read_only: _,
             pg_version: _,
         } => tenant::CreateTimelineParams::Branch(tenant::CreateTimelineParamsBranch {
             new_timeline_id,
@@ -3532,10 +3545,7 @@ async fn activate_post_import_handler(
 
         tenant.wait_to_become_active(ACTIVE_TENANT_TIMEOUT).await?;
 
-        tenant
-            .finalize_importing_timeline(timeline_id)
-            .await
-            .map_err(ApiError::InternalServerError)?;
+        tenant.finalize_importing_timeline(timeline_id).await?;
 
         match tenant.get_timeline(timeline_id, false) {
             Ok(_timeline) => {
@@ -3653,6 +3663,46 @@ async fn read_tar_eof(mut reader: (impl tokio::io::AsyncRead + Unpin)) -> anyhow
     Ok(())
 }
 
+async fn tenant_evaluate_feature_flag(
+    request: Request<Body>,
+    _cancel: CancellationToken,
+) -> Result<Response<Body>, ApiError> {
+    let tenant_shard_id: TenantShardId = parse_request_param(&request, "tenant_shard_id")?;
+    check_permission(&request, Some(tenant_shard_id.tenant_id))?;
+
+    let flag: String = must_parse_query_param(&request, "flag")?;
+    let as_type: String = must_parse_query_param(&request, "as")?;
+
+    let state = get_state(&request);
+
+    async {
+        let tenant = state
+            .tenant_manager
+            .get_attached_tenant_shard(tenant_shard_id)?;
+        if as_type == "boolean" {
+            let result = tenant.feature_resolver.evaluate_boolean(&flag, tenant_shard_id.tenant_id);
+            let result = result.map(|_| true).map_err(|e| e.to_string());
+            json_response(StatusCode::OK, result)
+        } else if as_type == "multivariate" {
+            let result = tenant.feature_resolver.evaluate_multivariate(&flag, tenant_shard_id.tenant_id).map_err(|e| e.to_string());
+            json_response(StatusCode::OK, result)
+        } else {
+            // Auto infer the type of the feature flag.
+            let is_boolean = tenant.feature_resolver.is_feature_flag_boolean(&flag).map_err(|e| ApiError::InternalServerError(anyhow::anyhow!("{e}")))?;
+            if is_boolean {
+                let result = tenant.feature_resolver.evaluate_boolean(&flag, tenant_shard_id.tenant_id);
+                let result = result.map(|_| true).map_err(|e| e.to_string());
+                json_response(StatusCode::OK, result)
+            } else {
+                let result = tenant.feature_resolver.evaluate_multivariate(&flag, tenant_shard_id.tenant_id).map_err(|e| e.to_string());
+                json_response(StatusCode::OK, result)
+            }
+        }
+    }
+    .instrument(info_span!("tenant_evaluate_feature_flag", tenant_id = %tenant_shard_id.tenant_id, shard_id = %tenant_shard_id.shard_slug()))
+    .await
+}
+
 /// Common functionality of all the HTTP API handlers.
 ///
 /// - Adds a tracing span to each request (by `request_span`)
@@ -4029,5 +4079,8 @@ pub fn make_router(
             "/v1/tenant/:tenant_shard_id/timeline/:timeline_id/activate_post_import",
             |r| api_handler(r, activate_post_import_handler),
         )
+        .get("/v1/tenant/:tenant_shard_id/feature_flag", |r| {
+            api_handler(r, tenant_evaluate_feature_flag)
+        })
         .any(handler_404))
 }

pageserver/src/lib.rs

@@ -10,6 +10,7 @@ pub mod context;
 pub mod controller_upcall_client;
 pub mod deletion_queue;
 pub mod disk_usage_eviction_task;
+pub mod feature_resolver;
 pub mod http;
 pub mod import_datadir;
 pub mod l0_flush;

pageserver/src/metrics.rs

@@ -1312,11 +1312,44 @@ impl EvictionsWithLowResidenceDuration {
 //
 // Roughly logarithmic scale.
 const STORAGE_IO_TIME_BUCKETS: &[f64] = &[
-    0.000030, // 30 usec
-    0.001000, // 1000 usec
-    0.030,    // 30 ms
-    1.000,    // 1000 ms
-    30.000,   // 30000 ms
+    0.00005,  // 50us
+    0.00006,  // 60us
+    0.00007,  // 70us
+    0.00008,  // 80us
+    0.00009,  // 90us
+    0.0001,   // 100us
+    0.000110, // 110us
+    0.000120, // 120us
+    0.000130, // 130us
+    0.000140, // 140us
+    0.000150, // 150us
+    0.000160, // 160us
+    0.000170, // 170us
+    0.000180, // 180us
+    0.000190, // 190us
+    0.000200, // 200us
+    0.000210, // 210us
+    0.000220, // 220us
+    0.000230, // 230us
+    0.000240, // 240us
+    0.000250, // 250us
+    0.000300, // 300us
+    0.000350, // 350us
+    0.000400, // 400us
+    0.000450, // 450us
+    0.000500, // 500us
+    0.000600, // 600us
+    0.000700, // 700us
+    0.000800, // 800us
+    0.000900, // 900us
+    0.001000, // 1ms
+    0.002000, // 2ms
+    0.003000, // 3ms
+    0.004000, // 4ms
+    0.005000, // 5ms
+    0.01000,  // 10ms
+    0.02000,  // 20ms
+    0.05000,  // 50ms
 ];
 
 /// VirtualFile fs operation variants.
@@ -2234,8 +2267,10 @@ impl BasebackupQueryTimeOngoingRecording<'_> {
         // If you want to change categorize of a specific error, also change it in `log_query_error`.
         let metric = match res {
             Ok(_) => &self.parent.ok,
-            Err(QueryError::Shutdown) => {
-                // Do not observe ok/err for shutdown
+            Err(QueryError::Shutdown) | Err(QueryError::Reconnect) => {
+                // Do not observe ok/err for shutdown/reconnect.
+                // Reconnect error might be raised when the operation is waiting for LSN and the tenant shutdown interrupts
+                // the operation. A reconnect error will be issued and the client will retry.
                 return;
             }
             Err(QueryError::Disconnected(ConnectionError::Io(io_error)))

pageserver/src/page_service.rs

@@ -43,12 +43,14 @@ use strum_macros::IntoStaticStr;
 use tokio::io::{AsyncRead, AsyncWrite, AsyncWriteExt, BufWriter};
 use tokio::task::JoinHandle;
 use tokio_util::sync::CancellationToken;
+use tonic::service::Interceptor as _;
 use tracing::*;
 use utils::auth::{Claims, Scope, SwappableJwtAuth};
 use utils::failpoint_support;
-use utils::id::{TenantId, TimelineId};
+use utils::id::{TenantId, TenantTimelineId, TimelineId};
 use utils::logging::log_slow;
 use utils::lsn::Lsn;
+use utils::shard::ShardIndex;
 use utils::simple_rcu::RcuReadGuard;
 use utils::sync::gate::{Gate, GateGuard};
 use utils::sync::spsc_fold;
@@ -200,9 +202,9 @@ pub fn spawn_grpc(
         .max_concurrent_streams(Some(GRPC_MAX_CONCURRENT_STREAMS));
 
     // Main page service.
-    let page_service = proto::PageServiceServer::new(PageServerHandler::new(
+    let page_service_handler = PageServerHandler::new(
         tenant_manager,
-        auth,
+        auth.clone(),
         PageServicePipeliningConfig::Serial, // TODO: unused with gRPC
         conf.get_vectored_concurrent_io,
         ConnectionPerfSpanFields::default(),
@@ -210,7 +212,18 @@ pub fn spawn_grpc(
         ctx,
         cancel.clone(),
         gate.enter().expect("just created"),
-    ));
+    );
+
+    let mut tenant_interceptor = TenantMetadataInterceptor;
+    let mut auth_interceptor = TenantAuthInterceptor::new(auth);
+    let interceptors = move |mut req: tonic::Request<()>| {
+        req = tenant_interceptor.call(req)?;
+        req = auth_interceptor.call(req)?;
+        Ok(req)
+    };
+
+    let page_service =
+        proto::PageServiceServer::with_interceptor(page_service_handler, interceptors);
     let server = server.add_service(page_service);
 
     // Reflection service for use with e.g. grpcurl.
@@ -756,6 +769,9 @@ struct BatchedGetPageRequest {
     timer: SmgrOpTimer,
     lsn_range: LsnRange,
     ctx: RequestContext,
+    // If the request is perf enabled, this contains a context
+    // with a perf span tracking the time spent waiting for the executor.
+    batch_wait_ctx: Option<RequestContext>,
 }
 
 #[cfg(feature = "testing")]
@@ -768,6 +784,7 @@ struct BatchedTestRequest {
 /// so that we don't keep the [`Timeline::gate`] open while the batch
 /// is being built up inside the [`spsc_fold`] (pagestream pipelining).
 #[derive(IntoStaticStr)]
+#[allow(clippy::large_enum_variant)]
 enum BatchedFeMessage {
     Exists {
         span: Span,
@@ -1285,6 +1302,22 @@ impl PageServerHandler {
                     }
                 };
 
+                let batch_wait_ctx = if ctx.has_perf_span() {
+                    Some(
+                        RequestContextBuilder::from(&ctx)
+                            .perf_span(|crnt_perf_span| {
+                                info_span!(
+                                    target: PERF_TRACE_TARGET,
+                                    parent: crnt_perf_span,
+                                    "WAIT_EXECUTOR",
+                                )
+                            })
+                            .attached_child(),
+                    )
+                } else {
+                    None
+                };
+
                 BatchedFeMessage::GetPage {
                     span,
                     shard: shard.downgrade(),
@@ -1296,6 +1329,7 @@ impl PageServerHandler {
                             request_lsn: req.hdr.request_lsn
                         },
                         ctx,
+                        batch_wait_ctx,
                     }],
                     // The executor grabs the batch when it becomes idle.
                     // Hence, [`GetPageBatchBreakReason::ExecutorSteal`] is the
@@ -1451,7 +1485,7 @@ impl PageServerHandler {
             let mut flush_timers = Vec::with_capacity(handler_results.len());
             for handler_result in &mut handler_results {
                 let flush_timer = match handler_result {
-                    Ok((_, timer)) => Some(
+                    Ok((_response, timer, _ctx)) => Some(
                         timer
                             .observe_execution_end(flushing_start_time)
                             .expect("we are the first caller"),
@@ -1471,7 +1505,7 @@ impl PageServerHandler {
         // Some handler errors cause exit from pagestream protocol.
         // Other handler errors are sent back as an error message and we stay in pagestream protocol.
         for (handler_result, flushing_timer) in handler_results.into_iter().zip(flush_timers) {
-            let response_msg = match handler_result {
+            let (response_msg, ctx) = match handler_result {
                 Err(e) => match &e.err {
                     PageStreamError::Shutdown => {
                         // If we fail to fulfil a request during shutdown, which may be _because_ of
@@ -1496,15 +1530,30 @@ impl PageServerHandler {
                             error!("error reading relation or page version: {full:#}")
                         });
 
-                        PagestreamBeMessage::Error(PagestreamErrorResponse {
-                            req: e.req,
-                            message: e.err.to_string(),
-                        })
+                        (
+                            PagestreamBeMessage::Error(PagestreamErrorResponse {
+                                req: e.req,
+                                message: e.err.to_string(),
+                            }),
+                            None,
+                        )
                     }
                 },
-                Ok((response_msg, _op_timer_already_observed)) => response_msg,
+                Ok((response_msg, _op_timer_already_observed, ctx)) => (response_msg, Some(ctx)),
             };
 
+            let ctx = ctx.map(|req_ctx| {
+                RequestContextBuilder::from(&req_ctx)
+                    .perf_span(|crnt_perf_span| {
+                        info_span!(
+                            target: PERF_TRACE_TARGET,
+                            parent: crnt_perf_span,
+                            "FLUSH_RESPONSE",
+                        )
+                    })
+                    .attached_child()
+            });
+
             //
             // marshal & transmit response message
             //
@@ -1527,6 +1576,17 @@ impl PageServerHandler {
                 )),
                 None => futures::future::Either::Right(flush_fut),
             };
+
+            let flush_fut = if let Some(req_ctx) = ctx.as_ref() {
+                futures::future::Either::Left(
+                    flush_fut.maybe_perf_instrument(req_ctx, |current_perf_span| {
+                        current_perf_span.clone()
+                    }),
+                )
+            } else {
+                futures::future::Either::Right(flush_fut)
+            };
+
             // do it while respecting cancellation
             let _: () = async move {
                 tokio::select! {
@@ -1556,7 +1616,7 @@ impl PageServerHandler {
         ctx: &RequestContext,
     ) -> Result<
         (
-            Vec<Result<(PagestreamBeMessage, SmgrOpTimer), BatchedPageStreamError>>,
+            Vec<Result<(PagestreamBeMessage, SmgrOpTimer, RequestContext), BatchedPageStreamError>>,
             Span,
         ),
         QueryError,
@@ -1583,7 +1643,7 @@ impl PageServerHandler {
                         self.handle_get_rel_exists_request(&shard, &req, &ctx)
                             .instrument(span.clone())
                             .await
-                            .map(|msg| (msg, timer))
+                            .map(|msg| (msg, timer, ctx))
                             .map_err(|err| BatchedPageStreamError { err, req: req.hdr }),
                     ],
                     span,
@@ -1602,7 +1662,7 @@ impl PageServerHandler {
                         self.handle_get_nblocks_request(&shard, &req, &ctx)
                             .instrument(span.clone())
                             .await
-                            .map(|msg| (msg, timer))
+                            .map(|msg| (msg, timer, ctx))
                             .map_err(|err| BatchedPageStreamError { err, req: req.hdr }),
                     ],
                     span,
@@ -1649,7 +1709,7 @@ impl PageServerHandler {
                         self.handle_db_size_request(&shard, &req, &ctx)
                             .instrument(span.clone())
                             .await
-                            .map(|msg| (msg, timer))
+                            .map(|msg| (msg, timer, ctx))
                             .map_err(|err| BatchedPageStreamError { err, req: req.hdr }),
                     ],
                     span,
@@ -1668,7 +1728,7 @@ impl PageServerHandler {
                         self.handle_get_slru_segment_request(&shard, &req, &ctx)
                             .instrument(span.clone())
                             .await
-                            .map(|msg| (msg, timer))
+                            .map(|msg| (msg, timer, ctx))
                             .map_err(|err| BatchedPageStreamError { err, req: req.hdr }),
                     ],
                     span,
@@ -2020,12 +2080,25 @@ impl PageServerHandler {
                             return Ok(());
                         }
                     };
-                    let batch = match batch {
+                    let mut batch = match batch {
                         Ok(batch) => batch,
                         Err(e) => {
                             return Err(e);
                         }
                     };
+
+                    if let BatchedFeMessage::GetPage {
+                        pages,
+                        span: _,
+                        shard: _,
+                        batch_break_reason: _,
+                    } = &mut batch
+                    {
+                        for req in pages {
+                            req.batch_wait_ctx.take();
+                        }
+                    }
+
                     self.pagestream_handle_batched_message(
                         pgb_writer,
                         batch,
@@ -2338,7 +2411,8 @@ impl PageServerHandler {
         io_concurrency: IoConcurrency,
         batch_break_reason: GetPageBatchBreakReason,
         ctx: &RequestContext,
-    ) -> Vec<Result<(PagestreamBeMessage, SmgrOpTimer), BatchedPageStreamError>> {
+    ) -> Vec<Result<(PagestreamBeMessage, SmgrOpTimer, RequestContext), BatchedPageStreamError>>
+    {
         debug_assert_current_span_has_tenant_and_timeline_id();
 
         timeline
@@ -2445,6 +2519,7 @@ impl PageServerHandler {
                                 page,
                             }),
                             req.timer,
+                            req.ctx,
                         )
                     })
                     .map_err(|e| BatchedPageStreamError {
@@ -2489,7 +2564,8 @@ impl PageServerHandler {
         timeline: &Timeline,
         requests: Vec<BatchedTestRequest>,
         _ctx: &RequestContext,
-    ) -> Vec<Result<(PagestreamBeMessage, SmgrOpTimer), BatchedPageStreamError>> {
+    ) -> Vec<Result<(PagestreamBeMessage, SmgrOpTimer, RequestContext), BatchedPageStreamError>>
+    {
         // real requests would do something with the timeline
         let mut results = Vec::with_capacity(requests.len());
         for _req in requests.iter() {
@@ -2516,6 +2592,10 @@ impl PageServerHandler {
                                 req: req.req.clone(),
                             }),
                             req.timer,
+                            RequestContext::new(
+                                TaskKind::PageRequestHandler,
+                                DownloadBehavior::Warn,
+                            ),
                         )
                     })
                     .map_err(|e| BatchedPageStreamError {
@@ -3290,6 +3370,104 @@ impl From<GetActiveTenantError> for QueryError {
     }
 }
 
+/// gRPC interceptor that decodes tenant metadata and stores it as request extensions of type
+/// TenantTimelineId and ShardIndex.
+///
+/// TODO: consider looking up the timeline handle here and storing it.
+#[derive(Clone)]
+struct TenantMetadataInterceptor;
+
+impl tonic::service::Interceptor for TenantMetadataInterceptor {
+    fn call(&mut self, mut req: tonic::Request<()>) -> Result<tonic::Request<()>, tonic::Status> {
+        // Decode the tenant ID.
+        let tenant_id = req
+            .metadata()
+            .get("neon-tenant-id")
+            .ok_or_else(|| tonic::Status::invalid_argument("missing neon-tenant-id"))?
+            .to_str()
+            .map_err(|_| tonic::Status::invalid_argument("invalid neon-tenant-id"))?;
+        let tenant_id = TenantId::from_str(tenant_id)
+            .map_err(|_| tonic::Status::invalid_argument("invalid neon-tenant-id"))?;
+
+        // Decode the timeline ID.
+        let timeline_id = req
+            .metadata()
+            .get("neon-timeline-id")
+            .ok_or_else(|| tonic::Status::invalid_argument("missing neon-timeline-id"))?
+            .to_str()
+            .map_err(|_| tonic::Status::invalid_argument("invalid neon-timeline-id"))?;
+        let timeline_id = TimelineId::from_str(timeline_id)
+            .map_err(|_| tonic::Status::invalid_argument("invalid neon-timeline-id"))?;
+
+        // Decode the shard ID.
+        let shard_index = req
+            .metadata()
+            .get("neon-shard-id")
+            .ok_or_else(|| tonic::Status::invalid_argument("missing neon-shard-id"))?
+            .to_str()
+            .map_err(|_| tonic::Status::invalid_argument("invalid neon-shard-id"))?;
+        let shard_index = ShardIndex::from_str(shard_index)
+            .map_err(|_| tonic::Status::invalid_argument("invalid neon-shard-id"))?;
+
+        // Stash them in the request.
+        let extensions = req.extensions_mut();
+        extensions.insert(TenantTimelineId::new(tenant_id, timeline_id));
+        extensions.insert(shard_index);
+
+        Ok(req)
+    }
+}
+
+/// Authenticates gRPC page service requests. Must run after TenantMetadataInterceptor.
+#[derive(Clone)]
+struct TenantAuthInterceptor {
+    auth: Option<Arc<SwappableJwtAuth>>,
+}
+
+impl TenantAuthInterceptor {
+    fn new(auth: Option<Arc<SwappableJwtAuth>>) -> Self {
+        Self { auth }
+    }
+}
+
+impl tonic::service::Interceptor for TenantAuthInterceptor {
+    fn call(&mut self, req: tonic::Request<()>) -> Result<tonic::Request<()>, tonic::Status> {
+        // Do nothing if auth is disabled.
+        let Some(auth) = self.auth.as_ref() else {
+            return Ok(req);
+        };
+
+        // Fetch the tenant ID that's been set by TenantMetadataInterceptor.
+        let ttid = req
+            .extensions()
+            .get::<TenantTimelineId>()
+            .expect("TenantMetadataInterceptor must run before TenantAuthInterceptor");
+
+        // Fetch and decode the JWT token.
+        let jwt = req
+            .metadata()
+            .get("authorization")
+            .ok_or_else(|| tonic::Status::unauthenticated("no authorization header"))?
+            .to_str()
+            .map_err(|_| tonic::Status::invalid_argument("invalid authorization header"))?
+            .strip_prefix("Bearer ")
+            .ok_or_else(|| tonic::Status::invalid_argument("invalid authorization header"))?
+            .trim();
+        let jwtdata: TokenData<Claims> = auth
+            .decode(jwt)
+            .map_err(|err| tonic::Status::invalid_argument(format!("invalid JWT token: {err}")))?;
+        let claims = jwtdata.claims;
+
+        // Check if the token is valid for this tenant.
+        check_permission(&claims, Some(ttid.tenant_id))
+            .map_err(|err| tonic::Status::permission_denied(err.to_string()))?;
+
+        // TODO: consider stashing the claims in the request extensions, if needed.
+
+        Ok(req)
+    }
+}
+
 #[derive(Debug, thiserror::Error)]
 pub(crate) enum GetActiveTimelineError {
     #[error(transparent)]

pageserver/src/tenant.rs

@@ -84,6 +84,7 @@ use crate::context;
 use crate::context::RequestContextBuilder;
 use crate::context::{DownloadBehavior, RequestContext};
 use crate::deletion_queue::{DeletionQueueClient, DeletionQueueError};
+use crate::feature_resolver::FeatureResolver;
 use crate::l0_flush::L0FlushGlobalState;
 use crate::metrics::{
     BROKEN_TENANTS_SET, CIRCUIT_BREAKERS_BROKEN, CIRCUIT_BREAKERS_UNBROKEN, CONCURRENT_INITDBS,
@@ -159,6 +160,7 @@ pub struct TenantSharedResources {
     pub deletion_queue_client: DeletionQueueClient,
     pub l0_flush_global_state: L0FlushGlobalState,
     pub basebackup_prepare_sender: BasebackupPrepareSender,
+    pub feature_resolver: FeatureResolver,
 }
 
 /// A [`TenantShard`] is really an _attached_ tenant.  The configuration
@@ -298,7 +300,7 @@ pub struct TenantShard {
     ///   as in progress.
     /// * Imported timelines are removed when the storage controller calls the post timeline
     ///   import activation endpoint.
-    timelines_importing: std::sync::Mutex<HashMap<TimelineId, ImportingTimeline>>,
+    timelines_importing: std::sync::Mutex<HashMap<TimelineId, Arc<ImportingTimeline>>>,
 
     /// The last tenant manifest known to be in remote storage. None if the manifest has not yet
     /// been either downloaded or uploaded. Always Some after tenant attach.
@@ -380,6 +382,8 @@ pub struct TenantShard {
     pub(crate) gc_block: gc_block::GcBlock,
 
     l0_flush_global_state: L0FlushGlobalState,
+
+    pub(crate) feature_resolver: FeatureResolver,
 }
 impl std::fmt::Debug for TenantShard {
     fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
@@ -668,6 +672,7 @@ pub enum MaybeOffloaded {
 pub enum TimelineOrOffloaded {
     Timeline(Arc<Timeline>),
     Offloaded(Arc<OffloadedTimeline>),
+    Importing(Arc<ImportingTimeline>),
 }
 
 impl TimelineOrOffloaded {
@@ -679,6 +684,9 @@ impl TimelineOrOffloaded {
             TimelineOrOffloaded::Offloaded(offloaded) => {
                 TimelineOrOffloadedArcRef::Offloaded(offloaded)
             }
+            TimelineOrOffloaded::Importing(importing) => {
+                TimelineOrOffloadedArcRef::Importing(importing)
+            }
         }
     }
     pub fn tenant_shard_id(&self) -> TenantShardId {
@@ -691,12 +699,16 @@ impl TimelineOrOffloaded {
         match self {
             TimelineOrOffloaded::Timeline(timeline) => &timeline.delete_progress,
             TimelineOrOffloaded::Offloaded(offloaded) => &offloaded.delete_progress,
+            TimelineOrOffloaded::Importing(importing) => &importing.delete_progress,
         }
     }
     fn maybe_remote_client(&self) -> Option<Arc<RemoteTimelineClient>> {
         match self {
             TimelineOrOffloaded::Timeline(timeline) => Some(timeline.remote_client.clone()),
             TimelineOrOffloaded::Offloaded(_offloaded) => None,
+            TimelineOrOffloaded::Importing(importing) => {
+                Some(importing.timeline.remote_client.clone())
+            }
         }
     }
 }
@@ -704,6 +716,7 @@ impl TimelineOrOffloaded {
 pub enum TimelineOrOffloadedArcRef<'a> {
     Timeline(&'a Arc<Timeline>),
     Offloaded(&'a Arc<OffloadedTimeline>),
+    Importing(&'a Arc<ImportingTimeline>),
 }
 
 impl TimelineOrOffloadedArcRef<'_> {
@@ -711,12 +724,14 @@ impl TimelineOrOffloadedArcRef<'_> {
         match self {
             TimelineOrOffloadedArcRef::Timeline(timeline) => timeline.tenant_shard_id,
             TimelineOrOffloadedArcRef::Offloaded(offloaded) => offloaded.tenant_shard_id,
+            TimelineOrOffloadedArcRef::Importing(importing) => importing.timeline.tenant_shard_id,
         }
     }
     pub fn timeline_id(&self) -> TimelineId {
         match self {
             TimelineOrOffloadedArcRef::Timeline(timeline) => timeline.timeline_id,
             TimelineOrOffloadedArcRef::Offloaded(offloaded) => offloaded.timeline_id,
+            TimelineOrOffloadedArcRef::Importing(importing) => importing.timeline.timeline_id,
         }
     }
 }
@@ -733,6 +748,12 @@ impl<'a> From<&'a Arc<OffloadedTimeline>> for TimelineOrOffloadedArcRef<'a> {
     }
 }
 
+impl<'a> From<&'a Arc<ImportingTimeline>> for TimelineOrOffloadedArcRef<'a> {
+    fn from(timeline: &'a Arc<ImportingTimeline>) -> Self {
+        Self::Importing(timeline)
+    }
+}
+
 #[derive(Debug, thiserror::Error, PartialEq, Eq)]
 pub enum GetTimelineError {
     #[error("Timeline is shutting down")]
@@ -860,6 +881,14 @@ impl Debug for SetStoppingError {
     }
 }
 
+#[derive(thiserror::Error, Debug)]
+pub(crate) enum FinalizeTimelineImportError {
+    #[error("Import task not done yet")]
+    ImportTaskStillRunning,
+    #[error("Shutting down")]
+    ShuttingDown,
+}
+
 /// Arguments to [`TenantShard::create_timeline`].
 ///
 /// Not usable as an idempotency key for timeline creation because if [`CreateTimelineParamsBranch::ancestor_start_lsn`]
@@ -1146,10 +1175,20 @@ impl TenantShard {
             ctx,
         )?;
         let disk_consistent_lsn = timeline.get_disk_consistent_lsn();
-        anyhow::ensure!(
-            disk_consistent_lsn.is_valid(),
-            "Timeline {tenant_id}/{timeline_id} has invalid disk_consistent_lsn"
-        );
+
+        if !disk_consistent_lsn.is_valid() {
+            // As opposed to normal timelines which get initialised with a disk consitent LSN
+            // via initdb, imported timelines start from 0. If the import task stops before
+            // it advances disk consitent LSN, allow it to resume.
+            let in_progress_import = import_pgdata
+                .as_ref()
+                .map(|import| !import.is_done())
+                .unwrap_or(false);
+            if !in_progress_import {
+                anyhow::bail!("Timeline {tenant_id}/{timeline_id} has invalid disk_consistent_lsn");
+            }
+        }
+
         assert_eq!(
             disk_consistent_lsn,
             metadata.disk_consistent_lsn(),
@@ -1243,20 +1282,25 @@ impl TenantShard {
                     }
                 }
 
-                // Sanity check: a timeline should have some content.
-                anyhow::ensure!(
-                    ancestor.is_some()
-                        || timeline
-                            .layers
-                            .read()
-                            .await
-                            .layer_map()
-                            .expect("currently loading, layer manager cannot be shutdown already")
-                            .iter_historic_layers()
-                            .next()
-                            .is_some(),
-                    "Timeline has no ancestor and no layer files"
-                );
+                if disk_consistent_lsn.is_valid() {
+                    // Sanity check: a timeline should have some content.
+                    // Exception: importing timelines might not yet have any
+                    anyhow::ensure!(
+                        ancestor.is_some()
+                            || timeline
+                                .layers
+                                .read()
+                                .await
+                                .layer_map()
+                                .expect(
+                                    "currently loading, layer manager cannot be shutdown already"
+                                )
+                                .iter_historic_layers()
+                                .next()
+                                .is_some(),
+                        "Timeline has no ancestor and no layer files"
+                    );
+                }
 
                 Ok(TimelineInitAndSyncResult::ReadyToActivate)
             }
@@ -1292,6 +1336,7 @@ impl TenantShard {
             deletion_queue_client,
             l0_flush_global_state,
             basebackup_prepare_sender,
+            feature_resolver,
         } = resources;
 
         let attach_mode = attached_conf.location.attach_mode;
@@ -1308,6 +1353,7 @@ impl TenantShard {
             deletion_queue_client,
             l0_flush_global_state,
             basebackup_prepare_sender,
+            feature_resolver,
         ));
 
         // The attach task will carry a GateGuard, so that shutdown() reliably waits for it to drop out if
@@ -1760,20 +1806,25 @@ impl TenantShard {
                     },
                 ) => {
                     let timeline_id = timeline.timeline_id;
+                    let import_task_gate = Gate::default();
+                    let import_task_guard = import_task_gate.enter().unwrap();
                     let import_task_handle =
                         tokio::task::spawn(self.clone().create_timeline_import_pgdata_task(
                             timeline.clone(),
                             import_pgdata,
                             guard,
+                            import_task_guard,
                             ctx.detached_child(TaskKind::ImportPgdata, DownloadBehavior::Warn),
                         ));
 
                     let prev = self.timelines_importing.lock().unwrap().insert(
                         timeline_id,
-                        ImportingTimeline {
+                        Arc::new(ImportingTimeline {
                             timeline: timeline.clone(),
                             import_task_handle,
-                        },
+                            import_task_gate,
+                            delete_progress: TimelineDeleteProgress::default(),
+                        }),
                     );
 
                     assert!(prev.is_none());
@@ -2391,6 +2442,17 @@ impl TenantShard {
             .collect()
     }
 
+    /// Lists timelines the tenant contains.
+    /// It's up to callers to omit certain timelines that are not considered ready for use.
+    pub fn list_importing_timelines(&self) -> Vec<Arc<ImportingTimeline>> {
+        self.timelines_importing
+            .lock()
+            .unwrap()
+            .values()
+            .map(Arc::clone)
+            .collect()
+    }
+
     /// Lists timelines the tenant manages, including offloaded ones.
     ///
     /// It's up to callers to omit certain timelines that are not considered ready for use.
@@ -2824,19 +2886,25 @@ impl TenantShard {
 
         let (timeline, timeline_create_guard) = uninit_timeline.finish_creation_myself();
 
+        let import_task_gate = Gate::default();
+        let import_task_guard = import_task_gate.enter().unwrap();
+
         let import_task_handle = tokio::spawn(self.clone().create_timeline_import_pgdata_task(
             timeline.clone(),
             index_part,
             timeline_create_guard,
+            import_task_guard,
             timeline_ctx.detached_child(TaskKind::ImportPgdata, DownloadBehavior::Warn),
         ));
 
         let prev = self.timelines_importing.lock().unwrap().insert(
             timeline.timeline_id,
-            ImportingTimeline {
+            Arc::new(ImportingTimeline {
                 timeline: timeline.clone(),
                 import_task_handle,
-            },
+                import_task_gate,
+                delete_progress: TimelineDeleteProgress::default(),
+            }),
         );
 
         // Idempotency is enforced higher up the stack
@@ -2854,13 +2922,13 @@ impl TenantShard {
     pub(crate) async fn finalize_importing_timeline(
         &self,
         timeline_id: TimelineId,
-    ) -> anyhow::Result<()> {
+    ) -> Result<(), FinalizeTimelineImportError> {
         let timeline = {
             let locked = self.timelines_importing.lock().unwrap();
             match locked.get(&timeline_id) {
                 Some(importing_timeline) => {
                     if !importing_timeline.import_task_handle.is_finished() {
-                        return Err(anyhow::anyhow!("Import task not done yet"));
+                        return Err(FinalizeTimelineImportError::ImportTaskStillRunning);
                     }
 
                     importing_timeline.timeline.clone()
@@ -2873,8 +2941,13 @@ impl TenantShard {
 
         timeline
             .remote_client
-            .schedule_index_upload_for_import_pgdata_finalize()?;
-        timeline.remote_client.wait_completion().await?;
+            .schedule_index_upload_for_import_pgdata_finalize()
+            .map_err(|_err| FinalizeTimelineImportError::ShuttingDown)?;
+        timeline
+            .remote_client
+            .wait_completion()
+            .await
+            .map_err(|_err| FinalizeTimelineImportError::ShuttingDown)?;
 
         self.timelines_importing
             .lock()
@@ -2890,6 +2963,7 @@ impl TenantShard {
         timeline: Arc<Timeline>,
         index_part: import_pgdata::index_part_format::Root,
         timeline_create_guard: TimelineCreateGuard,
+        _import_task_guard: GateGuard,
         ctx: RequestContext,
     ) {
         debug_assert_current_span_has_tenant_and_timeline_id();
@@ -3135,11 +3209,18 @@ impl TenantShard {
                         .or_insert_with(|| Arc::new(GcCompactionQueue::new()))
                         .clone()
                 };
+                let gc_compaction_strategy = self
+                    .feature_resolver
+                    .evaluate_multivariate("gc-comapction-strategy", self.tenant_shard_id.tenant_id)
+                    .ok();
+                let span = if let Some(gc_compaction_strategy) = gc_compaction_strategy {
+                    info_span!("gc_compact_timeline", timeline_id = %timeline.timeline_id, strategy = %gc_compaction_strategy)
+                } else {
+                    info_span!("gc_compact_timeline", timeline_id = %timeline.timeline_id)
+                };
                 outcome = queue
                     .iteration(cancel, ctx, &self.gc_block, &timeline)
-                    .instrument(
-                        info_span!("gc_compact_timeline", timeline_id = %timeline.timeline_id),
-                    )
+                    .instrument(span)
                     .await?;
             }
 
@@ -3471,8 +3552,9 @@ impl TenantShard {
             let mut timelines_importing = self.timelines_importing.lock().unwrap();
             timelines_importing
                 .drain()
-                .for_each(|(_timeline_id, importing_timeline)| {
-                    importing_timeline.shutdown();
+                .for_each(|(timeline_id, importing_timeline)| {
+                    let span = tracing::info_span!("importing_timeline_shutdown", %timeline_id);
+                    js.spawn(async move { importing_timeline.shutdown().instrument(span).await });
                 });
         }
         // test_long_timeline_create_then_tenant_delete is leaning on this message
@@ -3793,6 +3875,9 @@ impl TenantShard {
                         .build_timeline_client(offloaded.timeline_id, self.remote_storage.clone());
                     Arc::new(remote_client)
                 }
+                TimelineOrOffloadedArcRef::Importing(_) => {
+                    unreachable!("Importing timelines are not included in the iterator")
+                }
             };
 
             // Shut down the timeline's remote client: this means that the indices we write
@@ -4247,6 +4332,7 @@ impl TenantShard {
         deletion_queue_client: DeletionQueueClient,
         l0_flush_global_state: L0FlushGlobalState,
         basebackup_prepare_sender: BasebackupPrepareSender,
+        feature_resolver: FeatureResolver,
     ) -> TenantShard {
         assert!(!attached_conf.location.generation.is_none());
 
@@ -4351,6 +4437,7 @@ impl TenantShard {
             gc_block: Default::default(),
             l0_flush_global_state,
             basebackup_prepare_sender,
+            feature_resolver,
         }
     }
 
@@ -5000,6 +5087,14 @@ impl TenantShard {
                 info!("timeline already exists but is offloaded");
                 Err(CreateTimelineError::Conflict)
             }
+            Err(TimelineExclusionError::AlreadyExists {
+                existing: TimelineOrOffloaded::Importing(_existing),
+                ..
+            }) => {
+                // If there's a timeline already importing, then we would hit
+                // the [`TimelineExclusionError::AlreadyCreating`] branch above.
+                unreachable!("Importing timelines hold the creation guard")
+            }
             Err(TimelineExclusionError::AlreadyExists {
                 existing: TimelineOrOffloaded::Timeline(existing),
                 arg,
@@ -5271,6 +5366,7 @@ impl TenantShard {
             l0_compaction_trigger: self.l0_compaction_trigger.clone(),
             l0_flush_global_state: self.l0_flush_global_state.clone(),
             basebackup_prepare_sender: self.basebackup_prepare_sender.clone(),
+            feature_resolver: self.feature_resolver.clone(),
         }
     }
 
@@ -5736,6 +5832,7 @@ pub(crate) mod harness {
         pub conf: &'static PageServerConf,
         pub tenant_conf: pageserver_api::models::TenantConfig,
         pub tenant_shard_id: TenantShardId,
+        pub shard_identity: ShardIdentity,
         pub generation: Generation,
         pub shard: ShardIndex,
         pub remote_storage: GenericRemoteStorage,
@@ -5803,6 +5900,7 @@ pub(crate) mod harness {
                 conf,
                 tenant_conf,
                 tenant_shard_id,
+                shard_identity,
                 generation,
                 shard,
                 remote_storage,
@@ -5864,8 +5962,7 @@ pub(crate) mod harness {
                     &ShardParameters::default(),
                 ))
                 .unwrap(),
-                // This is a legacy/test code path: sharding isn't supported here.
-                ShardIdentity::unsharded(),
+                self.shard_identity,
                 Some(walredo_mgr),
                 self.tenant_shard_id,
                 self.remote_storage.clone(),
@@ -5873,6 +5970,7 @@ pub(crate) mod harness {
                 // TODO: ideally we should run all unit tests with both configs
                 L0FlushGlobalState::new(L0FlushConfig::default()),
                 basebackup_requst_sender,
+                FeatureResolver::new_disabled(),
             ));
 
             let preload = tenant
@@ -5986,6 +6084,7 @@ mod tests {
     use timeline::compaction::{KeyHistoryRetention, KeyLogAtLsn};
     use timeline::{CompactOptions, DeltaLayerTestDesc, VersionedKeySpaceQuery};
     use utils::id::TenantId;
+    use utils::shard::{ShardCount, ShardNumber};
 
     use super::*;
     use crate::DEFAULT_PG_VERSION;
@@ -8314,10 +8413,24 @@ mod tests {
             }
 
             tline.freeze_and_flush().await?;
+            // Force layers to L1
+            tline
+                .compact(
+                    &cancel,
+                    {
+                        let mut flags = EnumSet::new();
+                        flags.insert(CompactFlags::ForceL0Compaction);
+                        flags
+                    },
+                    &ctx,
+                )
+                .await?;
 
             if iter % 5 == 0 {
+                let scan_lsn = Lsn(lsn.0 + 1);
+                info!("scanning at {}", scan_lsn);
                 let (_, before_delta_file_accessed) =
-                    scan_with_statistics(&tline, &keyspace, lsn, &ctx, io_concurrency.clone())
+                    scan_with_statistics(&tline, &keyspace, scan_lsn, &ctx, io_concurrency.clone())
                         .await?;
                 tline
                     .compact(
@@ -8326,13 +8439,14 @@ mod tests {
                             let mut flags = EnumSet::new();
                             flags.insert(CompactFlags::ForceImageLayerCreation);
                             flags.insert(CompactFlags::ForceRepartition);
+                            flags.insert(CompactFlags::ForceL0Compaction);
                             flags
                         },
                         &ctx,
                     )
                     .await?;
                 let (_, after_delta_file_accessed) =
-                    scan_with_statistics(&tline, &keyspace, lsn, &ctx, io_concurrency.clone())
+                    scan_with_statistics(&tline, &keyspace, scan_lsn, &ctx, io_concurrency.clone())
                         .await?;
                 assert!(
                     after_delta_file_accessed < before_delta_file_accessed,
@@ -8773,6 +8887,8 @@ mod tests {
 
         let cancel = CancellationToken::new();
 
+        // Image layer creation happens on the disk_consistent_lsn so we need to force set it now.
+        tline.force_set_disk_consistent_lsn(Lsn(0x40));
         tline
             .compact(
                 &cancel,
@@ -8786,8 +8902,7 @@ mod tests {
             )
             .await
             .unwrap();
-
-        // Image layers are created at last_record_lsn
+        // Image layers are created at repartition LSN
         let images = tline
             .inspect_image_layers(Lsn(0x40), &ctx, io_concurrency.clone())
             .await
@@ -9305,6 +9420,77 @@ mod tests {
         Ok(())
     }
 
+    #[tokio::test]
+    async fn test_failed_flush_should_not_update_disk_consistent_lsn() -> anyhow::Result<()> {
+        //
+        // Setup
+        //
+        let harness = TenantHarness::create_custom(
+            "test_failed_flush_should_not_upload_disk_consistent_lsn",
+            pageserver_api::models::TenantConfig::default(),
+            TenantId::generate(),
+            ShardIdentity::new(ShardNumber(0), ShardCount(4), ShardStripeSize(128)).unwrap(),
+            Generation::new(1),
+        )
+        .await?;
+        let (tenant, ctx) = harness.load().await;
+
+        let timeline = tenant
+            .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
+            .await?;
+        assert_eq!(timeline.get_shard_identity().count, ShardCount(4));
+        let mut writer = timeline.writer().await;
+        writer
+            .put(
+                *TEST_KEY,
+                Lsn(0x20),
+                &Value::Image(test_img("foo at 0x20")),
+                &ctx,
+            )
+            .await?;
+        writer.finish_write(Lsn(0x20));
+        drop(writer);
+        timeline.freeze_and_flush().await.unwrap();
+
+        timeline.remote_client.wait_completion().await.unwrap();
+        let disk_consistent_lsn = timeline.get_disk_consistent_lsn();
+        let remote_consistent_lsn = timeline.get_remote_consistent_lsn_projected();
+        assert_eq!(Some(disk_consistent_lsn), remote_consistent_lsn);
+
+        //
+        // Test
+        //
+
+        let mut writer = timeline.writer().await;
+        writer
+            .put(
+                *TEST_KEY,
+                Lsn(0x30),
+                &Value::Image(test_img("foo at 0x30")),
+                &ctx,
+            )
+            .await?;
+        writer.finish_write(Lsn(0x30));
+        drop(writer);
+
+        fail::cfg(
+            "flush-layer-before-update-remote-consistent-lsn",
+            "return()",
+        )
+        .unwrap();
+
+        let flush_res = timeline.freeze_and_flush().await;
+        // if flush failed, the disk/remote consistent LSN should not be updated
+        assert!(flush_res.is_err());
+        assert_eq!(disk_consistent_lsn, timeline.get_disk_consistent_lsn());
+        assert_eq!(
+            remote_consistent_lsn,
+            timeline.get_remote_consistent_lsn_projected()
+        );
+
+        Ok(())
+    }
+
     #[cfg(feature = "testing")]
     #[tokio::test]
     async fn test_simple_bottom_most_compaction_deltas_1() -> anyhow::Result<()> {

pageserver/src/tenant/remote_timeline_client.rs

@@ -1348,6 +1348,21 @@ impl RemoteTimelineClient {
         Ok(())
     }
 
+    pub(crate) fn schedule_unlinking_of_layers_from_index_part<I>(
+        self: &Arc<Self>,
+        names: I,
+    ) -> Result<(), NotInitialized>
+    where
+        I: IntoIterator<Item = LayerName>,
+    {
+        let mut guard = self.upload_queue.lock().unwrap();
+        let upload_queue = guard.initialized_mut()?;
+
+        self.schedule_unlinking_of_layers_from_index_part0(upload_queue, names);
+
+        Ok(())
+    }
+
     /// Update the remote index file, removing the to-be-deleted files from the index,
     /// allowing scheduling of actual deletions later.
     fn schedule_unlinking_of_layers_from_index_part0<I>(

pageserver/src/tenant/timeline.rs

@@ -103,6 +103,7 @@ use crate::context::{
     DownloadBehavior, PerfInstrumentFutureExt, RequestContext, RequestContextBuilder,
 };
 use crate::disk_usage_eviction_task::{DiskUsageEvictionInfo, EvictionCandidate, finite_f32};
+use crate::feature_resolver::FeatureResolver;
 use crate::keyspace::{KeyPartitioning, KeySpace};
 use crate::l0_flush::{self, L0FlushGlobalState};
 use crate::metrics::{
@@ -198,6 +199,7 @@ pub struct TimelineResources {
     pub l0_compaction_trigger: Arc<Notify>,
     pub l0_flush_global_state: l0_flush::L0FlushGlobalState,
     pub basebackup_prepare_sender: BasebackupPrepareSender,
+    pub feature_resolver: FeatureResolver,
 }
 
 pub struct Timeline {
@@ -444,6 +446,8 @@ pub struct Timeline {
 
     /// A channel to send async requests to prepare a basebackup for the basebackup cache.
     basebackup_prepare_sender: BasebackupPrepareSender,
+
+    feature_resolver: FeatureResolver,
 }
 
 pub(crate) enum PreviousHeatmap {
@@ -3072,6 +3076,8 @@ impl Timeline {
                 wait_lsn_log_slow: tokio::sync::Semaphore::new(1),
 
                 basebackup_prepare_sender: resources.basebackup_prepare_sender,
+
+                feature_resolver: resources.feature_resolver,
             };
 
             result.repartition_threshold =
@@ -4761,7 +4767,10 @@ impl Timeline {
                     || !flushed_to_lsn.is_valid()
             );
 
-            if flushed_to_lsn < frozen_to_lsn && self.shard_identity.count.count() > 1 {
+            if flushed_to_lsn < frozen_to_lsn
+                && self.shard_identity.count.count() > 1
+                && result.is_ok()
+            {
                 // If our layer flushes didn't carry disk_consistent_lsn up to the `to_lsn` advertised
                 // to us via layer_flush_start_rx, then advance it here.
                 //
@@ -4906,6 +4915,7 @@ impl Timeline {
                     LastImageLayerCreationStatus::Initial,
                     false, // don't yield for L0, we're flushing L0
                 )
+                .instrument(info_span!("create_image_layers", mode = %ImageLayerCreationMode::Initial, partition_mode = "initial", lsn = %self.initdb_lsn))
                 .await?;
             debug_assert!(
                 matches!(is_complete, LastImageLayerCreationStatus::Complete),
@@ -4939,6 +4949,10 @@ impl Timeline {
             return Err(FlushLayerError::Cancelled);
         }
 
+        fail_point!("flush-layer-before-update-remote-consistent-lsn", |_| {
+            Err(FlushLayerError::Other(anyhow!("failpoint").into()))
+        });
+
         let disk_consistent_lsn = Lsn(lsn_range.end.0 - 1);
 
         // The new on-disk layers are now in the layer map. We can remove the
@@ -5462,7 +5476,8 @@ impl Timeline {
 
     /// Returns the image layers generated and an enum indicating whether the process is fully completed.
     /// true = we have generate all image layers, false = we preempt the process for L0 compaction.
-    #[tracing::instrument(skip_all, fields(%lsn, %mode))]
+    ///
+    /// `partition_mode` is only for logging purpose and is not used anywhere in this function.
     async fn create_image_layers(
         self: &Arc<Timeline>,
         partitioning: &KeyPartitioning,

pageserver/src/tenant/timeline/compaction.rs

@@ -206,8 +206,8 @@ pub struct GcCompactionQueue {
 }
 
 static CONCURRENT_GC_COMPACTION_TASKS: Lazy<Arc<Semaphore>> = Lazy::new(|| {
-    // Only allow two timelines on one pageserver to run gc compaction at a time.
-    Arc::new(Semaphore::new(2))
+    // Only allow one timeline on one pageserver to run gc compaction at a time.
+    Arc::new(Semaphore::new(1))
 });
 
 impl GcCompactionQueue {
@@ -1278,11 +1278,55 @@ impl Timeline {
         }
 
         let gc_cutoff = *self.applied_gc_cutoff_lsn.read();
+        let l0_l1_boundary_lsn = {
+            // We do the repartition on the L0-L1 boundary. All data below the boundary
+            // are compacted by L0 with low read amplification, thus making the `repartition`
+            // function run fast.
+            let guard = self.layers.read().await;
+            guard
+                .all_persistent_layers()
+                .iter()
+                .map(|x| {
+                    // Use the end LSN of delta layers OR the start LSN of image layers.
+                    if x.is_delta {
+                        x.lsn_range.end
+                    } else {
+                        x.lsn_range.start
+                    }
+                })
+                .max()
+        };
+
+        let (partition_mode, partition_lsn) = if cfg!(test)
+            || cfg!(feature = "testing")
+            || self
+                .feature_resolver
+                .evaluate_boolean("image-compaction-boundary", self.tenant_shard_id.tenant_id)
+                .is_ok()
+        {
+            let last_repartition_lsn = self.partitioning.read().1;
+            let lsn = match l0_l1_boundary_lsn {
+                Some(boundary) => gc_cutoff
+                    .max(boundary)
+                    .max(last_repartition_lsn)
+                    .max(self.initdb_lsn)
+                    .max(self.ancestor_lsn),
+                None => self.get_last_record_lsn(),
+            };
+            if lsn <= self.initdb_lsn || lsn <= self.ancestor_lsn {
+                // Do not attempt to create image layers below the initdb or ancestor LSN -- no data below it
+                ("l0_l1_boundary", self.get_last_record_lsn())
+            } else {
+                ("l0_l1_boundary", lsn)
+            }
+        } else {
+            ("latest_record", self.get_last_record_lsn())
+        };
 
         // 2. Repartition and create image layers if necessary
         match self
             .repartition(
-                self.get_last_record_lsn(),
+                partition_lsn,
                 self.get_compaction_target_size(),
                 options.flags,
                 ctx,
@@ -1301,18 +1345,19 @@ impl Timeline {
                     .extend(sparse_partitioning.into_dense().parts);
 
                 // 3. Create new image layers for partitions that have been modified "enough".
+                let mode = if options
+                    .flags
+                    .contains(CompactFlags::ForceImageLayerCreation)
+                {
+                    ImageLayerCreationMode::Force
+                } else {
+                    ImageLayerCreationMode::Try
+                };
                 let (image_layers, outcome) = self
                     .create_image_layers(
                         &partitioning,
                         lsn,
-                        if options
-                            .flags
-                            .contains(CompactFlags::ForceImageLayerCreation)
-                        {
-                            ImageLayerCreationMode::Force
-                        } else {
-                            ImageLayerCreationMode::Try
-                        },
+                        mode,
                         &image_ctx,
                         self.last_image_layer_creation_status
                             .load()
@@ -1320,6 +1365,7 @@ impl Timeline {
                             .clone(),
                         options.flags.contains(CompactFlags::YieldForL0),
                     )
+                    .instrument(info_span!("create_image_layers", mode = %mode, partition_mode = %partition_mode, lsn = %lsn))
                     .await
                     .inspect_err(|err| {
                         if let CreateImageLayersError::GetVectoredError(
@@ -1344,7 +1390,8 @@ impl Timeline {
             }
 
             Ok(_) => {
-                info!("skipping repartitioning due to image compaction LSN being below GC cutoff");
+                // This happens very frequently so we don't want to log it.
+                debug!("skipping repartitioning due to image compaction LSN being below GC cutoff");
             }
 
             // Suppress errors when cancelled.

pageserver/src/tenant/timeline/delete.rs

@@ -121,6 +121,7 @@ async fn remove_maybe_offloaded_timeline_from_tenant(
     // This observes the locking order between timelines and timelines_offloaded
     let mut timelines = tenant.timelines.lock().unwrap();
     let mut timelines_offloaded = tenant.timelines_offloaded.lock().unwrap();
+    let mut timelines_importing = tenant.timelines_importing.lock().unwrap();
     let offloaded_children_exist = timelines_offloaded
         .iter()
         .any(|(_, entry)| entry.ancestor_timeline_id == Some(timeline.timeline_id()));
@@ -150,8 +151,12 @@ async fn remove_maybe_offloaded_timeline_from_tenant(
                 .expect("timeline that we were deleting was concurrently removed from 'timelines_offloaded' map");
             offloaded_timeline.delete_from_ancestor_with_timelines(&timelines);
         }
+        TimelineOrOffloaded::Importing(importing) => {
+            timelines_importing.remove(&importing.timeline.timeline_id);
+        }
     }
 
+    drop(timelines_importing);
     drop(timelines_offloaded);
     drop(timelines);
 
@@ -203,8 +208,17 @@ impl DeleteTimelineFlow {
         guard.mark_in_progress()?;
 
         // Now that the Timeline is in Stopping state, request all the related tasks to shut down.
-        if let TimelineOrOffloaded::Timeline(timeline) = &timeline {
-            timeline.shutdown(super::ShutdownMode::Hard).await;
+        // TODO(vlad): shut down imported timeline here
+        match &timeline {
+            TimelineOrOffloaded::Timeline(timeline) => {
+                timeline.shutdown(super::ShutdownMode::Hard).await;
+            }
+            TimelineOrOffloaded::Importing(importing) => {
+                importing.shutdown().await;
+            }
+            TimelineOrOffloaded::Offloaded(_offloaded) => {
+                // Nothing to shut down in this case
+            }
         }
 
         tenant.gc_block.before_delete(&timeline.timeline_id());
@@ -389,10 +403,18 @@ impl DeleteTimelineFlow {
             Err(anyhow::anyhow!("failpoint: timeline-delete-before-rm"))?
         });
 
-        // Offloaded timelines have no local state
-        // TODO: once we persist offloaded information, delete the timeline from there, too
-        if let TimelineOrOffloaded::Timeline(timeline) = timeline {
-            delete_local_timeline_directory(conf, tenant.tenant_shard_id, timeline).await;
+        match timeline {
+            TimelineOrOffloaded::Timeline(timeline) => {
+                delete_local_timeline_directory(conf, tenant.tenant_shard_id, timeline).await;
+            }
+            TimelineOrOffloaded::Importing(importing) => {
+                delete_local_timeline_directory(conf, tenant.tenant_shard_id, &importing.timeline)
+                    .await;
+            }
+            TimelineOrOffloaded::Offloaded(_offloaded) => {
+                // Offloaded timelines have no local state
+                // TODO: once we persist offloaded information, delete the timeline from there, too
+            }
         }
 
         fail::fail_point!("timeline-delete-after-rm", |_| {
@@ -451,12 +473,16 @@ pub(super) fn make_timeline_delete_guard(
     // For more context see this discussion: `https://github.com/neondatabase/neon/pull/4552#discussion_r1253437346`
     let timelines = tenant.timelines.lock().unwrap();
     let timelines_offloaded = tenant.timelines_offloaded.lock().unwrap();
+    let timelines_importing = tenant.timelines_importing.lock().unwrap();
 
     let timeline = match timelines.get(&timeline_id) {
         Some(t) => TimelineOrOffloaded::Timeline(Arc::clone(t)),
         None => match timelines_offloaded.get(&timeline_id) {
             Some(t) => TimelineOrOffloaded::Offloaded(Arc::clone(t)),
-            None => return Err(DeleteTimelineError::NotFound),
+            None => match timelines_importing.get(&timeline_id) {
+                Some(t) => TimelineOrOffloaded::Importing(Arc::clone(t)),
+                None => return Err(DeleteTimelineError::NotFound),
+            },
         },
     };
 

pageserver/src/tenant/timeline/import_pgdata.rs

@@ -8,8 +8,10 @@ use tokio::task::JoinHandle;
 use tokio_util::sync::CancellationToken;
 use tracing::info;
 use utils::lsn::Lsn;
+use utils::pausable_failpoint;
+use utils::sync::gate::Gate;
 
-use super::Timeline;
+use super::{Timeline, TimelineDeleteProgress};
 use crate::context::RequestContext;
 use crate::controller_upcall_client::{StorageControllerUpcallApi, StorageControllerUpcallClient};
 use crate::tenant::metadata::TimelineMetadata;
@@ -19,14 +21,25 @@ mod importbucket_client;
 mod importbucket_format;
 pub(crate) mod index_part_format;
 
-pub(crate) struct ImportingTimeline {
+pub struct ImportingTimeline {
     pub import_task_handle: JoinHandle<()>,
+    pub import_task_gate: Gate,
     pub timeline: Arc<Timeline>,
+    pub delete_progress: TimelineDeleteProgress,
+}
+
+impl std::fmt::Debug for ImportingTimeline {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        write!(f, "ImportingTimeline<{}>", self.timeline.timeline_id)
+    }
 }
 
 impl ImportingTimeline {
-    pub(crate) fn shutdown(self) {
+    pub async fn shutdown(&self) {
         self.import_task_handle.abort();
+        self.import_task_gate.close().await;
+
+        self.timeline.remote_client.shutdown().await;
     }
 }
 
@@ -93,6 +106,13 @@ pub async fn doit(
                 );
             }
 
+            timeline
+                .remote_client
+                .schedule_index_upload_for_file_changes()?;
+            timeline.remote_client.wait_completion().await?;
+
+            pausable_failpoint!("import-timeline-pre-success-notify-pausable");
+
             // Communicate that shard is done.
             // Ensure at-least-once delivery of the upcall to storage controller
             // before we mark the task as done and never come here again.

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

@@ -11,25 +11,14 @@
 //! - => S3 as the source for the PGDATA instead of local filesystem
 //!
 //! TODOs before productionization:
-//! - ChunkProcessingJob size / ImportJob::total_size does not account for sharding.
-//!   => produced image layers likely too small.
 //! - ChunkProcessingJob should cut up an ImportJob to hit exactly target image layer size.
-//! - asserts / unwraps need to be replaced with errors
-//! - don't trust remote objects will be small (=prevent OOMs in those cases)
-//!     - limit all in-memory buffers in size, or download to disk and read from there
-//! - limit task concurrency
-//! - generally play nice with other tenants in the system
-//!   - importbucket is different bucket than main pageserver storage, so, should be fine wrt S3 rate limits
-//!   - but concerns like network bandwidth, local disk write bandwidth, local disk capacity, etc
-//! - integrate with layer eviction system
-//! - audit for Tenant::cancel nor Timeline::cancel responsivity
-//! - audit for Tenant/Timeline gate holding (we spawn tokio tasks during this flow!)
 //!
 //! An incomplete set of TODOs from the Hackathon:
 //! - version-specific CheckPointData (=> pgv abstraction, already exists for regular walingest)
 
 use std::collections::HashSet;
 use std::hash::{Hash, Hasher};
+use std::num::NonZeroUsize;
 use std::ops::Range;
 use std::sync::Arc;
 
@@ -43,7 +32,7 @@ use pageserver_api::key::{
     rel_dir_to_key, rel_size_to_key, relmap_file_key, slru_block_to_key, slru_dir_to_key,
     slru_segment_size_to_key,
 };
-use pageserver_api::keyspace::{contiguous_range_len, is_contiguous_range, singleton_range};
+use pageserver_api::keyspace::{ShardedRange, singleton_range};
 use pageserver_api::models::{ShardImportProgress, ShardImportProgressV1, ShardImportStatus};
 use pageserver_api::reltag::{RelTag, SlruKind};
 use pageserver_api::shard::ShardIdentity;
@@ -100,8 +89,24 @@ async fn run_v1(
         tasks: Vec::default(),
     };
 
-    let import_config = &timeline.conf.timeline_import_config;
-    let plan = planner.plan(import_config).await?;
+    // Use the job size limit encoded in the progress if we are resuming an import.
+    // This ensures that imports have stable plans even if the pageserver config changes.
+    let import_config = {
+        match &import_progress {
+            Some(progress) => {
+                let base = &timeline.conf.timeline_import_config;
+                TimelineImportConfig {
+                    import_job_soft_size_limit: NonZeroUsize::new(progress.job_soft_size_limit)
+                        .unwrap(),
+                    import_job_concurrency: base.import_job_concurrency,
+                    import_job_checkpoint_threshold: base.import_job_checkpoint_threshold,
+                }
+            }
+            None => timeline.conf.timeline_import_config.clone(),
+        }
+    };
+
+    let plan = planner.plan(&import_config).await?;
 
     // Hash the plan and compare with the hash of the plan we got back from the storage controller.
     // If the two match, it means that the planning stage had the same output.
@@ -113,20 +118,20 @@ async fn run_v1(
     let plan_hash = hasher.finish();
 
     if let Some(progress) = &import_progress {
-        if plan_hash != progress.import_plan_hash {
-            anyhow::bail!("Import plan does not match storcon metadata");
-        }
-
         // Handle collisions on jobs of unequal length
         if progress.jobs != plan.jobs.len() {
             anyhow::bail!("Import plan job length does not match storcon metadata")
         }
+
+        if plan_hash != progress.import_plan_hash {
+            anyhow::bail!("Import plan does not match storcon metadata");
+        }
     }
 
     pausable_failpoint!("import-timeline-pre-execute-pausable");
 
     let start_from_job_idx = import_progress.map(|progress| progress.completed);
-    plan.execute(timeline, start_from_job_idx, plan_hash, import_config, ctx)
+    plan.execute(timeline, start_from_job_idx, plan_hash, &import_config, ctx)
         .await
 }
 
@@ -150,6 +155,7 @@ impl Planner {
     /// This function is and must remain pure: given the same input, it will generate the same import plan.
     async fn plan(mut self, import_config: &TimelineImportConfig) -> anyhow::Result<Plan> {
         let pgdata_lsn = Lsn(self.control_file.control_file_data().checkPoint).align();
+        anyhow::ensure!(pgdata_lsn.is_valid());
 
         let datadir = PgDataDir::new(&self.storage).await?;
 
@@ -218,15 +224,36 @@ impl Planner {
                 checkpoint_buf,
             )));
 
+        // Sort the tasks by the key ranges they handle.
+        // The plan being generated here needs to be stable across invocations
+        // of this method.
+        self.tasks.sort_by_key(|task| match task {
+            AnyImportTask::SingleKey(key) => (key.key, key.key.next()),
+            AnyImportTask::RelBlocks(rel_blocks) => {
+                (rel_blocks.key_range.start, rel_blocks.key_range.end)
+            }
+            AnyImportTask::SlruBlocks(slru_blocks) => {
+                (slru_blocks.key_range.start, slru_blocks.key_range.end)
+            }
+        });
+
         // Assigns parts of key space to later parallel jobs
+        // Note: The image layers produced here may have gaps, meaning,
+        //       there is not an image for each key in the layer's key range.
+        //       The read path stops traversal at the first image layer, regardless
+        //       of whether a base image has been found for a key or not.
+        //       (Concept of sparse image layers doesn't exist.)
+        //       This behavior is exactly right for the base image layers we're producing here.
+        //       But, since no other place in the code currently produces image layers with gaps,
+        //       it seems noteworthy.
         let mut last_end_key = Key::MIN;
         let mut current_chunk = Vec::new();
         let mut current_chunk_size: usize = 0;
         let mut jobs = Vec::new();
         for task in std::mem::take(&mut self.tasks).into_iter() {
-            if current_chunk_size + task.total_size()
-                > import_config.import_job_soft_size_limit.into()
-            {
+            let task_size = task.total_size(&self.shard);
+            let projected_chunk_size = current_chunk_size.saturating_add(task_size);
+            if projected_chunk_size > import_config.import_job_soft_size_limit.into() {
                 let key_range = last_end_key..task.key_range().start;
                 jobs.push(ChunkProcessingJob::new(
                     key_range.clone(),
@@ -236,7 +263,7 @@ impl Planner {
                 last_end_key = key_range.end;
                 current_chunk_size = 0;
             }
-            current_chunk_size += task.total_size();
+            current_chunk_size = current_chunk_size.saturating_add(task_size);
             current_chunk.push(task);
         }
         jobs.push(ChunkProcessingJob::new(
@@ -426,6 +453,8 @@ impl Plan {
                     }));
                 },
                 maybe_complete_job_idx = work.next() => {
+                    pausable_failpoint!("import-task-complete-pausable");
+
                     match maybe_complete_job_idx {
                         Some(Ok((job_idx, res))) => {
                             assert!(last_completed_job_idx.checked_add(1).unwrap() == job_idx);
@@ -438,8 +467,12 @@ impl Plan {
                                     jobs: jobs_in_plan,
                                     completed: last_completed_job_idx,
                                     import_plan_hash,
+                                    job_soft_size_limit: import_config.import_job_soft_size_limit.into(),
                                 };
 
+                                timeline.remote_client.schedule_index_upload_for_file_changes()?;
+                                timeline.remote_client.wait_completion().await?;
+
                                 storcon_client.put_timeline_import_status(
                                     timeline.tenant_shard_id,
                                     timeline.timeline_id,
@@ -568,18 +601,18 @@ impl PgDataDirDb {
                 };
 
                 let path = datadir_path.join(rel_tag.to_segfile_name(segno));
-                assert!(filesize % BLCKSZ as usize == 0); // TODO: this should result in an error
+                anyhow::ensure!(filesize % BLCKSZ as usize == 0);
                 let nblocks = filesize / BLCKSZ as usize;
 
-                PgDataDirDbFile {
+                Ok(PgDataDirDbFile {
                     path,
                     filesize,
                     rel_tag,
                     segno,
                     nblocks: Some(nblocks), // first non-cummulative sizes
-                }
+                })
             })
-            .collect();
+            .collect::<anyhow::Result<_, _>>()?;
 
         // Set cummulative sizes. Do all of that math here, so that later we could easier
         // parallelize over segments and know with which segments we need to write relsize
@@ -614,12 +647,22 @@ impl PgDataDirDb {
 trait ImportTask {
     fn key_range(&self) -> Range<Key>;
 
-    fn total_size(&self) -> usize {
-        // TODO: revisit this
-        if is_contiguous_range(&self.key_range()) {
-            contiguous_range_len(&self.key_range()) as usize * 8192
+    fn total_size(&self, shard_identity: &ShardIdentity) -> usize {
+        let range = ShardedRange::new(self.key_range(), shard_identity);
+        let page_count = range.page_count();
+        if page_count == u32::MAX {
+            tracing::warn!(
+                "Import task has non contiguous key range: {}..{}",
+                self.key_range().start,
+                self.key_range().end
+            );
+
+            // Tasks should operate on contiguous ranges. It is unexpected for
+            // ranges to violate this assumption. Calling code handles this by mapping
+            // any task on a non contiguous range to its own image layer.
+            usize::MAX
         } else {
-            u32::MAX as usize
+            page_count as usize * 8192
         }
     }
 
@@ -640,7 +683,11 @@ impl Hash for ImportSingleKeyTask {
         let ImportSingleKeyTask { key, buf } = self;
 
         key.hash(state);
-        buf.hash(state);
+        // The key value might not have a stable binary representation.
+        // For instance, the db directory uses an unstable hash-map.
+        // To work around this we are a bit lax here and only hash the
+        // size of the buffer which must be consistent.
+        buf.len().hash(state);
     }
 }
 
@@ -713,6 +760,8 @@ impl ImportTask for ImportRelBlocksTask {
         layer_writer: &mut ImageLayerWriter,
         ctx: &RequestContext,
     ) -> anyhow::Result<usize> {
+        const MAX_BYTE_RANGE_SIZE: usize = 128 * 1024 * 1024;
+
         debug!("Importing relation file");
 
         let (rel_tag, start_blk) = self.key_range.start.to_rel_block()?;
@@ -737,7 +786,7 @@ impl ImportTask for ImportRelBlocksTask {
                 assert_eq!(key.len(), 1);
                 assert!(!acc.is_empty());
                 assert!(acc_end > acc_start);
-                if acc_end == start /* TODO additional max range check here, to limit memory consumption per task to X */ {
+                if acc_end == start && end - acc_start <= MAX_BYTE_RANGE_SIZE {
                     acc.push(key.pop().unwrap());
                     Ok((acc, acc_start, end))
                 } else {
@@ -752,8 +801,8 @@ impl ImportTask for ImportRelBlocksTask {
                 .get_range(&self.path, range_start.into_u64(), range_end.into_u64())
                 .await?;
             let mut buf = Bytes::from(range_buf);
-            // TODO: batched writes
             for key in keys {
+                // The writer buffers writes internally
                 let image = buf.split_to(8192);
                 layer_writer.put_image(key, image, ctx).await?;
                 nimages += 1;
@@ -806,6 +855,9 @@ impl ImportTask for ImportSlruBlocksTask {
         debug!("Importing SLRU segment file {}", self.path);
         let buf = self.storage.get(&self.path).await?;
 
+        // TODO(vlad): Does timestamp to LSN work for imported timelines?
+        // Probably not since we don't append the `xact_time` to it as in
+        // [`WalIngest::ingest_xact_record`].
         let (kind, segno, start_blk) = self.key_range.start.to_slru_block()?;
         let (_kind, _segno, end_blk) = self.key_range.end.to_slru_block()?;
         let mut blknum = start_blk;
@@ -915,7 +967,7 @@ impl ChunkProcessingJob {
                 let guard = timeline.layers.read().await;
                 let existing_layer = guard.try_get_from_key(&desc.key());
                 if let Some(layer) = existing_layer {
-                    if layer.metadata().generation != timeline.generation {
+                    if layer.metadata().generation == timeline.generation {
                         return Err(anyhow::anyhow!(
                             "Import attempted to rewrite layer file in the same generation: {}",
                             layer.local_path()
@@ -942,6 +994,15 @@ impl ChunkProcessingJob {
             .cloned();
         match existing_layer {
             Some(existing) => {
+                // Unlink the remote layer from the index without scheduling its deletion.
+                // When `existing_layer` drops [`LayerInner::drop`] will schedule its deletion from
+                // remote storage, but that assumes that the layer was unlinked from the index first.
+                timeline
+                    .remote_client
+                    .schedule_unlinking_of_layers_from_index_part(std::iter::once(
+                        existing.layer_desc().layer_name(),
+                    ))?;
+
                 guard.open_mut()?.rewrite_layers(
                     &[(existing.clone(), resident_layer.clone())],
                     &[],

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

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

pgxn/neon/walproposer.c

@@ -155,8 +155,9 @@ WalProposerCreate(WalProposerConfig *config, walproposer_api api)
 			int			written = 0;
 
 			written = snprintf((char *) &sk->conninfo, MAXCONNINFO,
-							   "host=%s port=%s dbname=replication options='-c timeline_id=%s tenant_id=%s'",
-							   sk->host, sk->port, wp->config->neon_timeline, wp->config->neon_tenant);
+							   "%s host=%s port=%s dbname=replication options='-c timeline_id=%s tenant_id=%s'",
+							   wp->config->safekeeper_conninfo_options, sk->host, sk->port,
+							   wp->config->neon_timeline, wp->config->neon_tenant);
 			if (written > MAXCONNINFO || written < 0)
 				wp_log(FATAL, "could not create connection string for safekeeper %s:%s", sk->host, sk->port);
 		}

pgxn/neon/walproposer.h

@@ -714,6 +714,9 @@ typedef struct WalProposerConfig
 	 */
 	char	   *safekeepers_list;
 
+	/* libpq connection info options. */
+	char	   *safekeeper_conninfo_options;
+
 	/*
 	 * WalProposer reconnects to offline safekeepers once in this interval.
 	 * Time is in milliseconds.

pgxn/neon/walproposer_pg.c

@@ -64,6 +64,7 @@ char	   *wal_acceptors_list = "";
 int			wal_acceptor_reconnect_timeout = 1000;
 int			wal_acceptor_connection_timeout = 10000;
 int			safekeeper_proto_version = 3;
+char	   *safekeeper_conninfo_options = "";
 
 /* Set to true in the walproposer bgw. */
 static bool am_walproposer;
@@ -119,6 +120,7 @@ init_walprop_config(bool syncSafekeepers)
 	walprop_config.neon_timeline = neon_timeline;
 	/* WalProposerCreate scribbles directly on it, so pstrdup */
 	walprop_config.safekeepers_list = pstrdup(wal_acceptors_list);
+	walprop_config.safekeeper_conninfo_options = pstrdup(safekeeper_conninfo_options);
 	walprop_config.safekeeper_reconnect_timeout = wal_acceptor_reconnect_timeout;
 	walprop_config.safekeeper_connection_timeout = wal_acceptor_connection_timeout;
 	walprop_config.wal_segment_size = wal_segment_size;
@@ -203,6 +205,16 @@ nwp_register_gucs(void)
 												 * GUC_LIST_QUOTE */
 							   NULL, assign_neon_safekeepers, NULL);
 
+	DefineCustomStringVariable(
+							   "neon.safekeeper_conninfo_options",
+							   "libpq keyword parameters and values to apply to safekeeper connections",
+							   NULL,
+							   &safekeeper_conninfo_options,
+							   "",
+							   PGC_POSTMASTER,
+							   0,
+							   NULL, NULL, NULL);
+
 	DefineCustomIntVariable(
 							"neon.safekeeper_reconnect_timeout",
 							"Walproposer reconnects to offline safekeepers once in this interval.",

proxy/src/auth/backend/classic.rs

@@ -17,35 +17,27 @@ pub(super) async fn authenticate(
     config: &'static AuthenticationConfig,
     secret: AuthSecret,
 ) -> auth::Result<ComputeCredentials> {
-    let flow = AuthFlow::new(client);
     let scram_keys = match secret {
         #[cfg(any(test, feature = "testing"))]
         AuthSecret::Md5(_) => {
             debug!("auth endpoint chooses MD5");
-            return Err(auth::AuthError::bad_auth_method("MD5"));
+            return Err(auth::AuthError::MalformedPassword("MD5 not supported"));
         }
         AuthSecret::Scram(secret) => {
             debug!("auth endpoint chooses SCRAM");
             let scram = auth::Scram(&secret, ctx);
 
-            let auth_outcome = tokio::time::timeout(
-                config.scram_protocol_timeout,
-                async {
-
-                    flow.begin(scram).await.map_err(|error| {
-                        warn!(?error, "error sending scram acknowledgement");
-                        error
-                    })?.authenticate().await.map_err(|error| {
+            let auth_outcome = tokio::time::timeout(config.scram_protocol_timeout, async {
+                AuthFlow::new(client, scram)
+                    .authenticate()
+                    .await
+                    .inspect_err(|error| {
                         warn!(?error, "error processing scram messages");
-                        error
                     })
-                }
-            )
+            })
             .await
-            .map_err(|e| {
-                warn!("error processing scram messages error = authentication timed out, execution time exceeded {} seconds", config.scram_protocol_timeout.as_secs());
-                auth::AuthError::user_timeout(e)
-            })??;
+            .inspect_err(|_| warn!("error processing scram messages error = authentication timed out, execution time exceeded {} seconds", config.scram_protocol_timeout.as_secs()))
+            .map_err(auth::AuthError::user_timeout)??;
 
             let client_key = match auth_outcome {
                 sasl::Outcome::Success(key) => key,

proxy/src/auth/backend/console_redirect.rs

@@ -2,7 +2,6 @@ use std::fmt;
 
 use async_trait::async_trait;
 use postgres_client::config::SslMode;
-use pq_proto::BeMessage as Be;
 use thiserror::Error;
 use tokio::io::{AsyncRead, AsyncWrite};
 use tracing::{info, info_span};
@@ -16,6 +15,7 @@ use crate::context::RequestContext;
 use crate::control_plane::client::cplane_proxy_v1;
 use crate::control_plane::{self, CachedNodeInfo, NodeInfo};
 use crate::error::{ReportableError, UserFacingError};
+use crate::pqproto::BeMessage;
 use crate::proxy::NeonOptions;
 use crate::proxy::connect_compute::ComputeConnectBackend;
 use crate::stream::PqStream;
@@ -154,11 +154,13 @@ async fn authenticate(
 
     // Give user a URL to spawn a new database.
     info!(parent: &span, "sending the auth URL to the user");
-    client
-        .write_message_noflush(&Be::AuthenticationOk)?
-        .write_message_noflush(&Be::CLIENT_ENCODING)?
-        .write_message(&Be::NoticeResponse(&greeting))
-        .await?;
+    client.write_message(BeMessage::AuthenticationOk);
+    client.write_message(BeMessage::ParameterStatus {
+        name: b"client_encoding",
+        value: b"UTF8",
+    });
+    client.write_message(BeMessage::NoticeResponse(&greeting));
+    client.flush().await?;
 
     // Wait for console response via control plane (see `mgmt`).
     info!(parent: &span, "waiting for console's reply...");
@@ -188,7 +190,7 @@ async fn authenticate(
         }
     }
 
-    client.write_message_noflush(&Be::NoticeResponse("Connecting to database."))?;
+    client.write_message(BeMessage::NoticeResponse("Connecting to database."));
 
     // This config should be self-contained, because we won't
     // take username or dbname from client's startup message.

proxy/src/auth/backend/hacks.rs

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

proxy/src/auth/backend/mod.rs

@@ -31,6 +31,7 @@ use crate::control_plane::{
 };
 use crate::intern::EndpointIdInt;
 use crate::metrics::Metrics;
+use crate::pqproto::BeMessage;
 use crate::protocol2::ConnectionInfoExtra;
 use crate::proxy::NeonOptions;
 use crate::proxy::connect_compute::ComputeConnectBackend;
@@ -402,7 +403,7 @@ async fn authenticate_with_secret(
         };
 
         // we have authenticated the password
-        client.write_message_noflush(&pq_proto::BeMessage::AuthenticationOk)?;
+        client.write_message(BeMessage::AuthenticationOk);
 
         return Ok(ComputeCredentials { info, keys });
     }
@@ -702,7 +703,7 @@ mod tests {
     #[tokio::test]
     async fn auth_quirks_scram() {
         let (mut client, server) = tokio::io::duplex(1024);
-        let mut stream = PqStream::new(Stream::from_raw(server));
+        let mut stream = PqStream::new_skip_handshake(Stream::from_raw(server));
 
         let ctx = RequestContext::test();
         let api = Auth {
@@ -784,7 +785,7 @@ mod tests {
     #[tokio::test]
     async fn auth_quirks_cleartext() {
         let (mut client, server) = tokio::io::duplex(1024);
-        let mut stream = PqStream::new(Stream::from_raw(server));
+        let mut stream = PqStream::new_skip_handshake(Stream::from_raw(server));
 
         let ctx = RequestContext::test();
         let api = Auth {
@@ -838,7 +839,7 @@ mod tests {
     #[tokio::test]
     async fn auth_quirks_password_hack() {
         let (mut client, server) = tokio::io::duplex(1024);
-        let mut stream = PqStream::new(Stream::from_raw(server));
+        let mut stream = PqStream::new_skip_handshake(Stream::from_raw(server));
 
         let ctx = RequestContext::test();
         let api = Auth {

proxy/src/auth/credentials.rs

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

proxy/src/auth/flow.rs

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

proxy/src/binary/pg_sni_router.rs

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

proxy/src/binary/proxy.rs

@@ -476,8 +476,7 @@ pub async fn run() -> anyhow::Result<()> {
         let key_path = args.tls_key.expect("already asserted it is set");
         let cert_path = args.tls_cert.expect("already asserted it is set");
 
-        let (tls_config, tls_server_end_point) =
-            super::pg_sni_router::parse_tls(&key_path, &cert_path)?;
+        let tls_config = super::pg_sni_router::parse_tls(&key_path, &cert_path)?;
 
         let dest = Arc::new(dest);
 
@@ -485,7 +484,6 @@ pub async fn run() -> anyhow::Result<()> {
             dest.clone(),
             tls_config.clone(),
             None,
-            tls_server_end_point,
             listen,
             cancellation_token.clone(),
         ));
@@ -494,7 +492,6 @@ pub async fn run() -> anyhow::Result<()> {
             dest,
             tls_config,
             Some(config.connect_to_compute.tls.clone()),
-            tls_server_end_point,
             listen_tls,
             cancellation_token.clone(),
         ));

proxy/src/cancellation.rs

@@ -5,7 +5,6 @@ use anyhow::{Context, anyhow};
 use ipnet::{IpNet, Ipv4Net, Ipv6Net};
 use postgres_client::CancelToken;
 use postgres_client::tls::MakeTlsConnect;
-use pq_proto::CancelKeyData;
 use redis::{Cmd, FromRedisValue, Value};
 use serde::{Deserialize, Serialize};
 use thiserror::Error;
@@ -21,6 +20,7 @@ use crate::control_plane::ControlPlaneApi;
 use crate::error::ReportableError;
 use crate::ext::LockExt;
 use crate::metrics::{CancelChannelSizeGuard, CancellationRequest, Metrics, RedisMsgKind};
+use crate::pqproto::CancelKeyData;
 use crate::protocol2::ConnectionInfoExtra;
 use crate::rate_limiter::LeakyBucketRateLimiter;
 use crate::redis::keys::KeyPrefix;

proxy/src/compute.rs

@@ -8,7 +8,6 @@ use itertools::Itertools;
 use postgres_client::tls::MakeTlsConnect;
 use postgres_client::{CancelToken, RawConnection};
 use postgres_protocol::message::backend::NoticeResponseBody;
-use pq_proto::StartupMessageParams;
 use rustls::pki_types::InvalidDnsNameError;
 use thiserror::Error;
 use tokio::net::{TcpStream, lookup_host};
@@ -24,6 +23,7 @@ use crate::control_plane::errors::WakeComputeError;
 use crate::control_plane::messages::MetricsAuxInfo;
 use crate::error::{ReportableError, UserFacingError};
 use crate::metrics::{Metrics, NumDbConnectionsGuard};
+use crate::pqproto::StartupMessageParams;
 use crate::proxy::neon_option;
 use crate::tls::postgres_rustls::MakeRustlsConnect;
 use crate::types::Host;

proxy/src/console_redirect_proxy.rs

@@ -1,7 +1,7 @@
 use std::sync::Arc;
 
 use futures::{FutureExt, TryFutureExt};
-use tokio::io::{AsyncRead, AsyncWrite, AsyncWriteExt};
+use tokio::io::{AsyncRead, AsyncWrite};
 use tokio_util::sync::CancellationToken;
 use tracing::{Instrument, debug, error, info};
 
@@ -221,12 +221,10 @@ pub(crate) async fn handle_client<S: AsyncRead + AsyncWrite + Unpin>(
         .await
     {
         Ok(auth_result) => auth_result,
-        Err(e) => {
-            return stream.throw_error(e, Some(ctx)).await?;
-        }
+        Err(e) => Err(stream.throw_error(e, Some(ctx)).await)?,
     };
 
-    let mut node = connect_to_compute(
+    let node = connect_to_compute(
         ctx,
         &TcpMechanism {
             user_info,
@@ -238,7 +236,7 @@ pub(crate) async fn handle_client<S: AsyncRead + AsyncWrite + Unpin>(
         config.wake_compute_retry_config,
         &config.connect_to_compute,
     )
-    .or_else(|e| stream.throw_error(e, Some(ctx)))
+    .or_else(|e| async { Err(stream.throw_error(e, Some(ctx)).await) })
     .await?;
 
     let cancellation_handler_clone = Arc::clone(&cancellation_handler);
@@ -246,14 +244,8 @@ pub(crate) async fn handle_client<S: AsyncRead + AsyncWrite + Unpin>(
 
     session.write_cancel_key(node.cancel_closure.clone())?;
 
-    prepare_client_connection(&node, *session.key(), &mut stream).await?;
-
-    // Before proxy passing, forward to compute whatever data is left in the
-    // PqStream input buffer. Normally there is none, but our serverless npm
-    // driver in pipeline mode sends startup, password and first query
-    // immediately after opening the connection.
-    let (stream, read_buf) = stream.into_inner();
-    node.stream.write_all(&read_buf).await?;
+    prepare_client_connection(&node, *session.key(), &mut stream);
+    let stream = stream.flush_and_into_inner().await?;
 
     Ok(Some(ProxyPassthrough {
         client: stream,

proxy/src/context/mod.rs

@@ -4,7 +4,6 @@ use std::net::IpAddr;
 
 use chrono::Utc;
 use once_cell::sync::OnceCell;
-use pq_proto::StartupMessageParams;
 use smol_str::SmolStr;
 use tokio::sync::mpsc;
 use tracing::field::display;
@@ -20,6 +19,7 @@ use crate::metrics::{
     ConnectOutcome, InvalidEndpointsGroup, LatencyAccumulated, LatencyTimer, Metrics, Protocol,
     Waiting,
 };
+use crate::pqproto::StartupMessageParams;
 use crate::protocol2::{ConnectionInfo, ConnectionInfoExtra};
 use crate::types::{DbName, EndpointId, RoleName};
 

proxy/src/context/parquet.rs

@@ -11,7 +11,6 @@ use parquet::file::metadata::RowGroupMetaDataPtr;
 use parquet::file::properties::{DEFAULT_PAGE_SIZE, WriterProperties, WriterPropertiesPtr};
 use parquet::file::writer::SerializedFileWriter;
 use parquet::record::RecordWriter;
-use pq_proto::StartupMessageParams;
 use remote_storage::{GenericRemoteStorage, RemotePath, RemoteStorageConfig, TimeoutOrCancel};
 use serde::ser::SerializeMap;
 use tokio::sync::mpsc;
@@ -24,6 +23,7 @@ use super::{LOG_CHAN, RequestContextInner};
 use crate::config::remote_storage_from_toml;
 use crate::context::LOG_CHAN_DISCONNECT;
 use crate::ext::TaskExt;
+use crate::pqproto::StartupMessageParams;
 
 #[derive(clap::Args, Clone, Debug)]
 pub struct ParquetUploadArgs {

proxy/src/lib.rs

@@ -92,6 +92,7 @@ mod logging;
 mod metrics;
 mod parse;
 mod pglb;
+mod pqproto;
 mod protocol2;
 mod proxy;
 mod rate_limiter;

proxy/src/pqproto.rs

@@ -0,0 +1,693 @@
+//! Postgres protocol codec
+//!
+//! <https://www.postgresql.org/docs/current/protocol-message-formats.html>
+
+use std::fmt;
+use std::io::{self, Cursor};
+
+use bytes::{Buf, BufMut};
+use itertools::Itertools;
+use rand::distributions::{Distribution, Standard};
+use tokio::io::{AsyncRead, AsyncReadExt};
+use zerocopy::{FromBytes, Immutable, IntoBytes, big_endian};
+
+pub type ErrorCode = [u8; 5];
+
+pub const FE_PASSWORD_MESSAGE: u8 = b'p';
+
+pub const SQLSTATE_INTERNAL_ERROR: [u8; 5] = *b"XX000";
+
+/// The protocol version number.
+///
+/// The most significant 16 bits are the major version number (3 for the protocol described here).
+/// The least significant 16 bits are the minor version number (0 for the protocol described here).
+/// <https://www.postgresql.org/docs/current/protocol-message-formats.html#PROTOCOL-MESSAGE-FORMATS-STARTUPMESSAGE>
+#[derive(Clone, Copy, PartialEq, PartialOrd, FromBytes, IntoBytes, Immutable)]
+#[repr(C)]
+pub struct ProtocolVersion {
+    major: big_endian::U16,
+    minor: big_endian::U16,
+}
+
+impl ProtocolVersion {
+    pub const fn new(major: u16, minor: u16) -> Self {
+        Self {
+            major: big_endian::U16::new(major),
+            minor: big_endian::U16::new(minor),
+        }
+    }
+    pub const fn minor(self) -> u16 {
+        self.minor.get()
+    }
+    pub const fn major(self) -> u16 {
+        self.major.get()
+    }
+}
+
+impl fmt::Debug for ProtocolVersion {
+    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
+        f.debug_list()
+            .entry(&self.major())
+            .entry(&self.minor())
+            .finish()
+    }
+}
+
+/// read the type from the stream using zerocopy.
+///
+/// not cancel safe.
+macro_rules! read {
+    ($s:expr => $t:ty) => {{
+        // cannot be implemented as a function due to lack of const-generic-expr
+        let mut buf = [0; size_of::<$t>()];
+        $s.read_exact(&mut buf).await?;
+        let res: $t = zerocopy::transmute!(buf);
+        res
+    }};
+}
+
+pub async fn read_startup<S>(stream: &mut S) -> io::Result<FeStartupPacket>
+where
+    S: AsyncRead + Unpin,
+{
+    /// <https://github.com/postgres/postgres/blob/ca481d3c9ab7bf69ff0c8d71ad3951d407f6a33c/src/include/libpq/pqcomm.h#L118>
+    const MAX_STARTUP_PACKET_LENGTH: usize = 10000;
+    const RESERVED_INVALID_MAJOR_VERSION: u16 = 1234;
+    /// <https://github.com/postgres/postgres/blob/ca481d3c9ab7bf69ff0c8d71ad3951d407f6a33c/src/include/libpq/pqcomm.h#L132>
+    const CANCEL_REQUEST_CODE: ProtocolVersion = ProtocolVersion::new(1234, 5678);
+    /// <https://github.com/postgres/postgres/blob/ca481d3c9ab7bf69ff0c8d71ad3951d407f6a33c/src/include/libpq/pqcomm.h#L166>
+    const NEGOTIATE_SSL_CODE: ProtocolVersion = ProtocolVersion::new(1234, 5679);
+    /// <https://github.com/postgres/postgres/blob/ca481d3c9ab7bf69ff0c8d71ad3951d407f6a33c/src/include/libpq/pqcomm.h#L167>
+    const NEGOTIATE_GSS_CODE: ProtocolVersion = ProtocolVersion::new(1234, 5680);
+
+    /// This first reads the startup message header, is 8 bytes.
+    /// The first 4 bytes is a big-endian message length, and the next 4 bytes is a version number.
+    ///
+    /// The length value is inclusive of the header. For example,
+    /// an empty message will always have length 8.
+    #[derive(Clone, Copy, FromBytes, IntoBytes, Immutable)]
+    #[repr(C)]
+    struct StartupHeader {
+        len: big_endian::U32,
+        version: ProtocolVersion,
+    }
+
+    let header = read!(stream => StartupHeader);
+
+    // <https://github.com/postgres/postgres/blob/04bcf9e19a4261fe9c7df37c777592c2e10c32a7/src/backend/tcop/backend_startup.c#L378-L382>
+    // First byte indicates standard SSL handshake message
+    // (It can't be a Postgres startup length because in network byte order
+    // that would be a startup packet hundreds of megabytes long)
+    if header.as_bytes()[0] == 0x16 {
+        return Ok(FeStartupPacket::SslRequest {
+            // The bytes we read for the header are actually part of a TLS ClientHello.
+            // In theory, if the ClientHello was < 8 bytes we would fail with EOF before we get here.
+            // In practice though, I see no world where a ClientHello is less than 8 bytes
+            // since it includes ephemeral keys etc.
+            direct: Some(zerocopy::transmute!(header)),
+        });
+    }
+
+    let Some(len) = (header.len.get() as usize).checked_sub(8) else {
+        return Err(io::Error::other(format!(
+            "invalid startup message length {}, must be at least 8.",
+            header.len,
+        )));
+    };
+
+    // TODO: add a histogram for startup packet lengths
+    if len > MAX_STARTUP_PACKET_LENGTH {
+        tracing::warn!("large startup message detected: {len} bytes");
+        return Err(io::Error::other(format!(
+            "invalid startup message length {len}"
+        )));
+    }
+
+    match header.version {
+        // <https://www.postgresql.org/docs/current/protocol-message-formats.html#PROTOCOL-MESSAGE-FORMATS-CANCELREQUEST>
+        CANCEL_REQUEST_CODE => {
+            if len != 8 {
+                return Err(io::Error::other(
+                    "CancelRequest message is malformed, backend PID / secret key missing",
+                ));
+            }
+
+            Ok(FeStartupPacket::CancelRequest(
+                read!(stream => CancelKeyData),
+            ))
+        }
+        // <https://www.postgresql.org/docs/current/protocol-message-formats.html#PROTOCOL-MESSAGE-FORMATS-SSLREQUEST>
+        NEGOTIATE_SSL_CODE => {
+            // Requested upgrade to SSL (aka TLS)
+            Ok(FeStartupPacket::SslRequest { direct: None })
+        }
+        NEGOTIATE_GSS_CODE => {
+            // Requested upgrade to GSSAPI
+            Ok(FeStartupPacket::GssEncRequest)
+        }
+        version if version.major() == RESERVED_INVALID_MAJOR_VERSION => Err(io::Error::other(
+            format!("Unrecognized request code {version:?}"),
+        )),
+        // StartupMessage
+        version => {
+            // The protocol version number is followed by one or more pairs of parameter name and value strings.
+            // A zero byte is required as a terminator after the last name/value pair.
+            // Parameters can appear in any order. user is required, others are optional.
+
+            let mut buf = vec![0; len];
+            stream.read_exact(&mut buf).await?;
+
+            if buf.pop() != Some(b'\0') {
+                return Err(io::Error::other(
+                    "StartupMessage params: missing null terminator",
+                ));
+            }
+
+            // TODO: Don't do this.
+            // There's no guarantee that these messages are utf8,
+            // but they usually happen to be simple ascii.
+            let params = String::from_utf8(buf)
+                .map_err(|_| io::Error::other("StartupMessage params: invalid utf-8"))?;
+
+            Ok(FeStartupPacket::StartupMessage {
+                version,
+                params: StartupMessageParams { params },
+            })
+        }
+    }
+}
+
+/// Read a raw postgres packet, which will respect the max length requested.
+///
+/// This returns the message tag, as well as the message body. The message
+/// body is written into `buf`, and it is otherwise completely overwritten.
+///
+/// This is not cancel safe.
+pub async fn read_message<'a, S>(
+    stream: &mut S,
+    buf: &'a mut Vec<u8>,
+    max: usize,
+) -> io::Result<(u8, &'a mut [u8])>
+where
+    S: AsyncRead + Unpin,
+{
+    /// This first reads the header, which for regular messages in the 3.0 protocol is 5 bytes.
+    /// The first byte is a message tag, and the next 4 bytes is a big-endian length.
+    ///
+    /// Awkwardly, the length value is inclusive of itself, but not of the tag. For example,
+    /// an empty message will always have length 4.
+    #[derive(Clone, Copy, FromBytes)]
+    #[repr(C)]
+    struct Header {
+        tag: u8,
+        len: big_endian::U32,
+    }
+
+    let header = read!(stream => Header);
+
+    // as described above, the length must be at least 4.
+    let Some(len) = (header.len.get() as usize).checked_sub(4) else {
+        return Err(io::Error::other(format!(
+            "invalid startup message length {}, must be at least 4.",
+            header.len,
+        )));
+    };
+
+    // TODO: add a histogram for message lengths
+
+    // check if the message exceeds our desired max.
+    if len > max {
+        tracing::warn!("large postgres message detected: {len} bytes");
+        return Err(io::Error::other(format!("invalid message length {len}")));
+    }
+
+    // read in our entire message.
+    buf.resize(len, 0);
+    stream.read_exact(buf).await?;
+
+    Ok((header.tag, buf))
+}
+
+pub struct WriteBuf(Cursor<Vec<u8>>);
+
+impl Buf for WriteBuf {
+    #[inline]
+    fn remaining(&self) -> usize {
+        self.0.remaining()
+    }
+
+    #[inline]
+    fn chunk(&self) -> &[u8] {
+        self.0.chunk()
+    }
+
+    #[inline]
+    fn advance(&mut self, cnt: usize) {
+        self.0.advance(cnt);
+    }
+}
+
+impl WriteBuf {
+    pub const fn new() -> Self {
+        Self(Cursor::new(Vec::new()))
+    }
+
+    /// Use a heuristic to determine if we should shrink the write buffer.
+    #[inline]
+    fn should_shrink(&self) -> bool {
+        let n = self.0.position() as usize;
+        let len = self.0.get_ref().len();
+
+        // the unused space at the front of our buffer is 2x the size of our filled portion.
+        n + n > len
+    }
+
+    /// Shrink the write buffer so that subsequent writes have more spare capacity.
+    #[cold]
+    fn shrink(&mut self) {
+        let n = self.0.position() as usize;
+        let buf = self.0.get_mut();
+
+        // buf repr:
+        // [----unused------|-----filled-----|-----uninit-----]
+        //                  ^ n              ^ buf.len()      ^ buf.capacity()
+        let filled = n..buf.len();
+        let filled_len = filled.len();
+        buf.copy_within(filled, 0);
+        buf.truncate(filled_len);
+        self.0.set_position(0);
+    }
+
+    /// clear the write buffer.
+    pub fn reset(&mut self) {
+        let buf = self.0.get_mut();
+        buf.clear();
+        self.0.set_position(0);
+    }
+
+    /// Write a raw message to the internal buffer.
+    ///
+    /// The size_hint value is only a hint for reserving space. It's ok if it's incorrect, since
+    /// we calculate the length after the fact.
+    pub fn write_raw(&mut self, size_hint: usize, tag: u8, f: impl FnOnce(&mut Vec<u8>)) {
+        if self.should_shrink() {
+            self.shrink();
+        }
+
+        let buf = self.0.get_mut();
+        buf.reserve(5 + size_hint);
+
+        buf.push(tag);
+        let start = buf.len();
+        buf.extend_from_slice(&[0, 0, 0, 0]);
+
+        f(buf);
+
+        let end = buf.len();
+        let len = (end - start) as u32;
+        buf[start..start + 4].copy_from_slice(&len.to_be_bytes());
+    }
+
+    /// Write an encryption response message.
+    pub fn encryption(&mut self, m: u8) {
+        self.0.get_mut().push(m);
+    }
+
+    pub fn write_error(&mut self, msg: &str, error_code: ErrorCode) {
+        self.shrink();
+
+        // <https://www.postgresql.org/docs/current/protocol-message-formats.html#PROTOCOL-MESSAGE-FORMATS-ERRORRESPONSE>
+        // <https://www.postgresql.org/docs/current/protocol-error-fields.html>
+        // "SERROR\0CXXXXX\0M\0\0".len() == 17
+        self.write_raw(17 + msg.len(), b'E', |buf| {
+            // Severity: ERROR
+            buf.put_slice(b"SERROR\0");
+
+            // Code: error_code
+            buf.put_u8(b'C');
+            buf.put_slice(&error_code);
+            buf.put_u8(0);
+
+            // Message: msg
+            buf.put_u8(b'M');
+            buf.put_slice(msg.as_bytes());
+            buf.put_u8(0);
+
+            // End.
+            buf.put_u8(0);
+        });
+    }
+}
+
+#[derive(Debug)]
+pub enum FeStartupPacket {
+    CancelRequest(CancelKeyData),
+    SslRequest {
+        direct: Option<[u8; 8]>,
+    },
+    GssEncRequest,
+    StartupMessage {
+        version: ProtocolVersion,
+        params: StartupMessageParams,
+    },
+}
+
+#[derive(Debug, Clone, Default)]
+pub struct StartupMessageParams {
+    pub params: String,
+}
+
+impl StartupMessageParams {
+    /// Get parameter's value by its name.
+    pub fn get(&self, name: &str) -> Option<&str> {
+        self.iter().find_map(|(k, v)| (k == name).then_some(v))
+    }
+
+    /// Split command-line options according to PostgreSQL's logic,
+    /// taking into account all escape sequences but leaving them as-is.
+    /// [`None`] means that there's no `options` in [`Self`].
+    pub fn options_raw(&self) -> Option<impl Iterator<Item = &str>> {
+        self.get("options").map(Self::parse_options_raw)
+    }
+
+    /// Split command-line options according to PostgreSQL's logic,
+    /// taking into account all escape sequences but leaving them as-is.
+    pub fn parse_options_raw(input: &str) -> impl Iterator<Item = &str> {
+        // See `postgres: pg_split_opts`.
+        let mut last_was_escape = false;
+        input
+            .split(move |c: char| {
+                // We split by non-escaped whitespace symbols.
+                let should_split = c.is_ascii_whitespace() && !last_was_escape;
+                last_was_escape = c == '\\' && !last_was_escape;
+                should_split
+            })
+            .filter(|s| !s.is_empty())
+    }
+
+    /// Iterate through key-value pairs in an arbitrary order.
+    pub fn iter(&self) -> impl Iterator<Item = (&str, &str)> {
+        self.params.split_terminator('\0').tuples()
+    }
+
+    // This function is mostly useful in tests.
+    #[cfg(test)]
+    pub fn new<'a, const N: usize>(pairs: [(&'a str, &'a str); N]) -> Self {
+        let mut b = Self {
+            params: String::new(),
+        };
+        for (k, v) in pairs {
+            b.insert(k, v);
+        }
+        b
+    }
+
+    /// Set parameter's value by its name.
+    /// name and value must not contain a \0 byte
+    pub fn insert(&mut self, name: &str, value: &str) {
+        self.params.reserve(name.len() + value.len() + 2);
+        self.params.push_str(name);
+        self.params.push('\0');
+        self.params.push_str(value);
+        self.params.push('\0');
+    }
+}
+
+/// Cancel keys usually are represented as PID+SecretKey, but to proxy they're just
+/// opaque bytes.
+#[derive(Debug, Hash, PartialEq, Eq, Clone, Copy, FromBytes, IntoBytes, Immutable)]
+pub struct CancelKeyData(pub big_endian::U64);
+
+pub fn id_to_cancel_key(id: u64) -> CancelKeyData {
+    CancelKeyData(big_endian::U64::new(id))
+}
+
+impl fmt::Display for CancelKeyData {
+    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
+        let id = self.0;
+        f.debug_tuple("CancelKeyData")
+            .field(&format_args!("{id:x}"))
+            .finish()
+    }
+}
+impl Distribution<CancelKeyData> for Standard {
+    fn sample<R: rand::Rng + ?Sized>(&self, rng: &mut R) -> CancelKeyData {
+        id_to_cancel_key(rng.r#gen())
+    }
+}
+
+pub enum BeMessage<'a> {
+    AuthenticationOk,
+    AuthenticationSasl(BeAuthenticationSaslMessage<'a>),
+    AuthenticationCleartextPassword,
+    BackendKeyData(CancelKeyData),
+    ParameterStatus {
+        name: &'a [u8],
+        value: &'a [u8],
+    },
+    ReadyForQuery,
+    NoticeResponse(&'a str),
+    NegotiateProtocolVersion {
+        version: ProtocolVersion,
+        options: &'a [&'a str],
+    },
+}
+
+#[derive(Debug)]
+pub enum BeAuthenticationSaslMessage<'a> {
+    Methods(&'a [&'a str]),
+    Continue(&'a [u8]),
+    Final(&'a [u8]),
+}
+
+impl BeMessage<'_> {
+    /// Write the message into an internal buffer
+    pub fn write_message(self, buf: &mut WriteBuf) {
+        match self {
+            // <https://www.postgresql.org/docs/current/protocol-message-formats.html#PROTOCOL-MESSAGE-FORMATS-AUTHENTICATIONCLEARTEXTPASSWORD>
+            BeMessage::AuthenticationOk => {
+                buf.write_raw(1, b'R', |buf| buf.put_i32(0));
+            }
+            // <https://www.postgresql.org/docs/current/protocol-message-formats.html#PROTOCOL-MESSAGE-FORMATS-AUTHENTICATIONCLEARTEXTPASSWORD>
+            BeMessage::AuthenticationCleartextPassword => {
+                buf.write_raw(1, b'R', |buf| buf.put_i32(3));
+            }
+
+            // <https://www.postgresql.org/docs/current/protocol-message-formats.html#PROTOCOL-MESSAGE-FORMATS-AUTHENTICATIONSASL>
+            BeMessage::AuthenticationSasl(BeAuthenticationSaslMessage::Methods(methods)) => {
+                let len: usize = methods.iter().map(|m| m.len() + 1).sum();
+                buf.write_raw(len + 2, b'R', |buf| {
+                    buf.put_i32(10); // Specifies that SASL auth method is used.
+                    for method in methods {
+                        buf.put_slice(method.as_bytes());
+                        buf.put_u8(0);
+                    }
+                    buf.put_u8(0); // zero terminator for the list
+                });
+            }
+            // <https://www.postgresql.org/docs/current/protocol-message-formats.html#PROTOCOL-MESSAGE-FORMATS-AUTHENTICATIONSASL>
+            BeMessage::AuthenticationSasl(BeAuthenticationSaslMessage::Continue(extra)) => {
+                buf.write_raw(extra.len() + 1, b'R', |buf| {
+                    buf.put_i32(11); // Continue SASL auth.
+                    buf.put_slice(extra);
+                });
+            }
+            // <https://www.postgresql.org/docs/current/protocol-message-formats.html#PROTOCOL-MESSAGE-FORMATS-AUTHENTICATIONSASL>
+            BeMessage::AuthenticationSasl(BeAuthenticationSaslMessage::Final(extra)) => {
+                buf.write_raw(extra.len() + 1, b'R', |buf| {
+                    buf.put_i32(12); // Send final SASL message.
+                    buf.put_slice(extra);
+                });
+            }
+
+            // <https://www.postgresql.org/docs/current/protocol-message-formats.html#PROTOCOL-MESSAGE-FORMATS-BACKENDKEYDATA>
+            BeMessage::BackendKeyData(key_data) => {
+                buf.write_raw(8, b'K', |buf| buf.put_slice(key_data.as_bytes()));
+            }
+
+            // <https://www.postgresql.org/docs/current/protocol-message-formats.html#PROTOCOL-MESSAGE-FORMATS-NOTICERESPONSE>
+            // <https://www.postgresql.org/docs/current/protocol-error-fields.html>
+            BeMessage::NoticeResponse(msg) => {
+                // 'N' signalizes NoticeResponse messages
+                buf.write_raw(18 + msg.len(), b'N', |buf| {
+                    // Severity: NOTICE
+                    buf.put_slice(b"SNOTICE\0");
+
+                    // Code: XX000 (ignored for notice, but still required)
+                    buf.put_slice(b"CXX000\0");
+
+                    // Message: msg
+                    buf.put_u8(b'M');
+                    buf.put_slice(msg.as_bytes());
+                    buf.put_u8(0);
+
+                    // End notice.
+                    buf.put_u8(0);
+                });
+            }
+
+            // <https://www.postgresql.org/docs/current/protocol-message-formats.html#PROTOCOL-MESSAGE-FORMATS-PARAMETERSTATUS>
+            BeMessage::ParameterStatus { name, value } => {
+                buf.write_raw(name.len() + value.len() + 2, b'S', |buf| {
+                    buf.put_slice(name.as_bytes());
+                    buf.put_u8(0);
+                    buf.put_slice(value.as_bytes());
+                    buf.put_u8(0);
+                });
+            }
+
+            // <https://www.postgresql.org/docs/current/protocol-message-formats.html#PROTOCOL-MESSAGE-FORMATS-NEGOTIATEPROTOCOLVERSION>
+            BeMessage::ReadyForQuery => {
+                buf.write_raw(1, b'Z', |buf| buf.put_u8(b'I'));
+            }
+
+            // <https://www.postgresql.org/docs/current/protocol-message-formats.html#PROTOCOL-MESSAGE-FORMATS-NEGOTIATEPROTOCOLVERSION>
+            BeMessage::NegotiateProtocolVersion { version, options } => {
+                let len: usize = options.iter().map(|o| o.len() + 1).sum();
+                buf.write_raw(8 + len, b'v', |buf| {
+                    buf.put_slice(version.as_bytes());
+                    buf.put_u32(options.len() as u32);
+                    for option in options {
+                        buf.put_slice(option.as_bytes());
+                        buf.put_u8(0);
+                    }
+                });
+            }
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::io::Cursor;
+
+    use tokio::io::{AsyncWriteExt, duplex};
+    use zerocopy::IntoBytes;
+
+    use crate::pqproto::{FeStartupPacket, read_message, read_startup};
+
+    use super::ProtocolVersion;
+
+    #[tokio::test]
+    async fn reject_large_startup() {
+        // we're going to define a v3.0 startup message with far too many parameters.
+        let mut payload = vec![];
+        // 10001 + 8 bytes.
+        payload.extend_from_slice(&10009_u32.to_be_bytes());
+        payload.extend_from_slice(ProtocolVersion::new(3, 0).as_bytes());
+        payload.resize(10009, b'a');
+
+        let (mut server, mut client) = duplex(128);
+        #[rustfmt::skip]
+        let (server, client) = tokio::join!(
+            async move { read_startup(&mut server).await.unwrap_err() },
+            async move { client.write_all(&payload).await.unwrap_err() },
+        );
+
+        assert_eq!(server.to_string(), "invalid startup message length 10001");
+        assert_eq!(client.to_string(), "broken pipe");
+    }
+
+    #[tokio::test]
+    async fn reject_large_password() {
+        // we're going to define a password message that is far too long.
+        let mut payload = vec![];
+        payload.push(b'p');
+        payload.extend_from_slice(&517_u32.to_be_bytes());
+        payload.resize(518, b'a');
+
+        let (mut server, mut client) = duplex(128);
+        #[rustfmt::skip]
+        let (server, client) = tokio::join!(
+            async move { read_message(&mut server, &mut vec![], 512).await.unwrap_err() },
+            async move { client.write_all(&payload).await.unwrap_err() },
+        );
+
+        assert_eq!(server.to_string(), "invalid message length 513");
+        assert_eq!(client.to_string(), "broken pipe");
+    }
+
+    #[tokio::test]
+    async fn read_startup_message() {
+        let mut payload = vec![];
+        payload.extend_from_slice(&17_u32.to_be_bytes());
+        payload.extend_from_slice(ProtocolVersion::new(3, 0).as_bytes());
+        payload.extend_from_slice(b"abc\0def\0\0");
+
+        let startup = read_startup(&mut Cursor::new(&payload)).await.unwrap();
+        let FeStartupPacket::StartupMessage { version, params } = startup else {
+            panic!("unexpected startup message: {startup:?}");
+        };
+
+        assert_eq!(version.major(), 3);
+        assert_eq!(version.minor(), 0);
+        assert_eq!(params.params, "abc\0def\0");
+    }
+
+    #[tokio::test]
+    async fn read_ssl_message() {
+        let mut payload = vec![];
+        payload.extend_from_slice(&8_u32.to_be_bytes());
+        payload.extend_from_slice(ProtocolVersion::new(1234, 5679).as_bytes());
+
+        let startup = read_startup(&mut Cursor::new(&payload)).await.unwrap();
+        let FeStartupPacket::SslRequest { direct: None } = startup else {
+            panic!("unexpected startup message: {startup:?}");
+        };
+    }
+
+    #[tokio::test]
+    async fn read_tls_message() {
+        // sample client hello taken from <https://tls13.xargs.org/#client-hello>
+        let client_hello = [
+            0x16, 0x03, 0x01, 0x00, 0xf8, 0x01, 0x00, 0x00, 0xf4, 0x03, 0x03, 0x00, 0x01, 0x02,
+            0x03, 0x04, 0x05, 0x06, 0x07, 0x08, 0x09, 0x0a, 0x0b, 0x0c, 0x0d, 0x0e, 0x0f, 0x10,
+            0x11, 0x12, 0x13, 0x14, 0x15, 0x16, 0x17, 0x18, 0x19, 0x1a, 0x1b, 0x1c, 0x1d, 0x1e,
+            0x1f, 0x20, 0xe0, 0xe1, 0xe2, 0xe3, 0xe4, 0xe5, 0xe6, 0xe7, 0xe8, 0xe9, 0xea, 0xeb,
+            0xec, 0xed, 0xee, 0xef, 0xf0, 0xf1, 0xf2, 0xf3, 0xf4, 0xf5, 0xf6, 0xf7, 0xf8, 0xf9,
+            0xfa, 0xfb, 0xfc, 0xfd, 0xfe, 0xff, 0x00, 0x08, 0x13, 0x02, 0x13, 0x03, 0x13, 0x01,
+            0x00, 0xff, 0x01, 0x00, 0x00, 0xa3, 0x00, 0x00, 0x00, 0x18, 0x00, 0x16, 0x00, 0x00,
+            0x13, 0x65, 0x78, 0x61, 0x6d, 0x70, 0x6c, 0x65, 0x2e, 0x75, 0x6c, 0x66, 0x68, 0x65,
+            0x69, 0x6d, 0x2e, 0x6e, 0x65, 0x74, 0x00, 0x0b, 0x00, 0x04, 0x03, 0x00, 0x01, 0x02,
+            0x00, 0x0a, 0x00, 0x16, 0x00, 0x14, 0x00, 0x1d, 0x00, 0x17, 0x00, 0x1e, 0x00, 0x19,
+            0x00, 0x18, 0x01, 0x00, 0x01, 0x01, 0x01, 0x02, 0x01, 0x03, 0x01, 0x04, 0x00, 0x23,
+            0x00, 0x00, 0x00, 0x16, 0x00, 0x00, 0x00, 0x17, 0x00, 0x00, 0x00, 0x0d, 0x00, 0x1e,
+            0x00, 0x1c, 0x04, 0x03, 0x05, 0x03, 0x06, 0x03, 0x08, 0x07, 0x08, 0x08, 0x08, 0x09,
+            0x08, 0x0a, 0x08, 0x0b, 0x08, 0x04, 0x08, 0x05, 0x08, 0x06, 0x04, 0x01, 0x05, 0x01,
+            0x06, 0x01, 0x00, 0x2b, 0x00, 0x03, 0x02, 0x03, 0x04, 0x00, 0x2d, 0x00, 0x02, 0x01,
+            0x01, 0x00, 0x33, 0x00, 0x26, 0x00, 0x24, 0x00, 0x1d, 0x00, 0x20, 0x35, 0x80, 0x72,
+            0xd6, 0x36, 0x58, 0x80, 0xd1, 0xae, 0xea, 0x32, 0x9a, 0xdf, 0x91, 0x21, 0x38, 0x38,
+            0x51, 0xed, 0x21, 0xa2, 0x8e, 0x3b, 0x75, 0xe9, 0x65, 0xd0, 0xd2, 0xcd, 0x16, 0x62,
+            0x54,
+        ];
+
+        let mut cursor = Cursor::new(&client_hello);
+
+        let startup = read_startup(&mut cursor).await.unwrap();
+        let FeStartupPacket::SslRequest {
+            direct: Some(prefix),
+        } = startup
+        else {
+            panic!("unexpected startup message: {startup:?}");
+        };
+
+        // check that no data is lost.
+        assert_eq!(prefix, [0x16, 0x03, 0x01, 0x00, 0xf8, 0x01, 0x00, 0x00]);
+        assert_eq!(cursor.position(), 8);
+    }
+
+    #[tokio::test]
+    async fn read_message_success() {
+        let query = b"Q\0\0\0\x0cSELECT 1Q\0\0\0\x0cSELECT 2";
+        let mut cursor = Cursor::new(&query);
+
+        let mut buf = vec![];
+        let (tag, message) = read_message(&mut cursor, &mut buf, 100).await.unwrap();
+        assert_eq!(tag, b'Q');
+        assert_eq!(message, b"SELECT 1");
+
+        let (tag, message) = read_message(&mut cursor, &mut buf, 100).await.unwrap();
+        assert_eq!(tag, b'Q');
+        assert_eq!(message, b"SELECT 2");
+    }
+}

proxy/src/proxy/connect_compute.rs

@@ -1,5 +1,4 @@
 use async_trait::async_trait;
-use pq_proto::StartupMessageParams;
 use tokio::time;
 use tracing::{debug, info, warn};
 
@@ -15,6 +14,7 @@ use crate::error::ReportableError;
 use crate::metrics::{
     ConnectOutcome, ConnectionFailureKind, Metrics, RetriesMetricGroup, RetryType,
 };
+use crate::pqproto::StartupMessageParams;
 use crate::proxy::retry::{CouldRetry, retry_after, should_retry};
 use crate::proxy::wake_compute::wake_compute;
 use crate::types::Host;

proxy/src/proxy/handshake.rs

@@ -1,8 +1,3 @@
-use bytes::Buf;
-use pq_proto::framed::Framed;
-use pq_proto::{
-    BeMessage as Be, CancelKeyData, FeStartupPacket, ProtocolVersion, StartupMessageParams,
-};
 use thiserror::Error;
 use tokio::io::{AsyncRead, AsyncWrite};
 use tracing::{debug, info, warn};
@@ -12,7 +7,10 @@ use crate::config::TlsConfig;
 use crate::context::RequestContext;
 use crate::error::ReportableError;
 use crate::metrics::Metrics;
-use crate::proxy::ERR_INSECURE_CONNECTION;
+use crate::pqproto::{
+    BeMessage, CancelKeyData, FeStartupPacket, ProtocolVersion, StartupMessageParams,
+};
+use crate::proxy::TlsRequired;
 use crate::stream::{PqStream, Stream, StreamUpgradeError};
 use crate::tls::PG_ALPN_PROTOCOL;
 
@@ -71,33 +69,25 @@ pub(crate) async fn handshake<S: AsyncRead + AsyncWrite + Unpin>(
     const PG_PROTOCOL_EARLIEST: ProtocolVersion = ProtocolVersion::new(3, 0);
     const PG_PROTOCOL_LATEST: ProtocolVersion = ProtocolVersion::new(3, 0);
 
-    let mut stream = PqStream::new(Stream::from_raw(stream));
+    let (mut stream, mut msg) = PqStream::parse_startup(Stream::from_raw(stream)).await?;
     loop {
-        let msg = stream.read_startup_packet().await?;
         match msg {
             FeStartupPacket::SslRequest { direct } => match stream.get_ref() {
                 Stream::Raw { .. } if !tried_ssl => {
                     tried_ssl = true;
 
-                    // We can't perform TLS handshake without a config
-                    let have_tls = tls.is_some();
-                    if !direct {
-                        stream
-                            .write_message(&Be::EncryptionResponse(have_tls))
-                            .await?;
-                    } else if !have_tls {
-                        return Err(HandshakeError::ProtocolViolation);
-                    }
-
                     if let Some(tls) = tls.take() {
                         // Upgrade raw stream into a secure TLS-backed stream.
                         // NOTE: We've consumed `tls`; this fact will be used later.
 
-                        let Framed {
-                            stream: raw,
-                            read_buf,
-                            write_buf,
-                        } = stream.framed;
+                        let mut read_buf;
+                        let raw = if let Some(direct) = &direct {
+                            read_buf = &direct[..];
+                            stream.accept_direct_tls()
+                        } else {
+                            read_buf = &[];
+                            stream.accept_tls().await?
+                        };
 
                         let Stream::Raw { raw } = raw else {
                             return Err(HandshakeError::StreamUpgradeError(
@@ -105,12 +95,11 @@ pub(crate) async fn handshake<S: AsyncRead + AsyncWrite + Unpin>(
                             ));
                         };
 
-                        let mut read_buf = read_buf.reader();
                         let mut res = Ok(());
                         let accept = tokio_rustls::TlsAcceptor::from(tls.pg_config.clone())
                             .accept_with(raw, |session| {
                                 // push the early data to the tls session
-                                while !read_buf.get_ref().is_empty() {
+                                while !read_buf.is_empty() {
                                     match session.read_tls(&mut read_buf) {
                                         Ok(_) => {}
                                         Err(e) => {
@@ -123,7 +112,6 @@ pub(crate) async fn handshake<S: AsyncRead + AsyncWrite + Unpin>(
 
                         res?;
 
-                        let read_buf = read_buf.into_inner();
                         if !read_buf.is_empty() {
                             return Err(HandshakeError::EarlyData);
                         }
@@ -157,16 +145,17 @@ pub(crate) async fn handshake<S: AsyncRead + AsyncWrite + Unpin>(
                         let (_, tls_server_end_point) =
                             tls.cert_resolver.resolve(conn_info.server_name());
 
-                        stream = PqStream {
-                            framed: Framed {
-                                stream: Stream::Tls {
-                                    tls: Box::new(tls_stream),
-                                    tls_server_end_point,
-                                },
-                                read_buf,
-                                write_buf,
-                            },
+                        let tls = Stream::Tls {
+                            tls: Box::new(tls_stream),
+                            tls_server_end_point,
                         };
+                        (stream, msg) = PqStream::parse_startup(tls).await?;
+                    } else {
+                        if direct.is_some() {
+                            // client sent us a ClientHello already, we can't do anything with it.
+                            return Err(HandshakeError::ProtocolViolation);
+                        }
+                        msg = stream.reject_encryption().await?;
                     }
                 }
                 _ => return Err(HandshakeError::ProtocolViolation),
@@ -176,7 +165,7 @@ pub(crate) async fn handshake<S: AsyncRead + AsyncWrite + Unpin>(
                     tried_gss = true;
 
                     // Currently, we don't support GSSAPI
-                    stream.write_message(&Be::EncryptionResponse(false)).await?;
+                    msg = stream.reject_encryption().await?;
                 }
                 _ => return Err(HandshakeError::ProtocolViolation),
             },
@@ -186,13 +175,7 @@ pub(crate) async fn handshake<S: AsyncRead + AsyncWrite + Unpin>(
                 // Check that the config has been consumed during upgrade
                 // OR we didn't provide it at all (for dev purposes).
                 if tls.is_some() {
-                    return stream
-                        .throw_error_str(
-                            ERR_INSECURE_CONNECTION,
-                            crate::error::ErrorKind::User,
-                            None,
-                        )
-                        .await?;
+                    Err(stream.throw_error(TlsRequired, None).await)?;
                 }
 
                 // This log highlights the start of the connection.
@@ -214,20 +197,21 @@ pub(crate) async fn handshake<S: AsyncRead + AsyncWrite + Unpin>(
                 // no protocol extensions are supported.
                 // <https://github.com/postgres/postgres/blob/ca481d3c9ab7bf69ff0c8d71ad3951d407f6a33c/src/backend/tcop/backend_startup.c#L744-L753>
                 let mut unsupported = vec![];
-                for (k, _) in params.iter() {
+                let mut supported = StartupMessageParams::default();
+
+                for (k, v) in params.iter() {
                     if k.starts_with("_pq_.") {
                         unsupported.push(k);
+                    } else {
+                        supported.insert(k, v);
                     }
                 }
 
-                // TODO: remove unsupported options so we don't send them to compute.
-
-                stream
-                    .write_message(&Be::NegotiateProtocolVersion {
-                        version: PG_PROTOCOL_LATEST,
-                        options: &unsupported,
-                    })
-                    .await?;
+                stream.write_message(BeMessage::NegotiateProtocolVersion {
+                    version: PG_PROTOCOL_LATEST,
+                    options: &unsupported,
+                });
+                stream.flush().await?;
 
                 info!(
                     ?version,
@@ -235,7 +219,7 @@ pub(crate) async fn handshake<S: AsyncRead + AsyncWrite + Unpin>(
                     session_type = "normal",
                     "successful handshake; unsupported minor version requested"
                 );
-                break Ok(HandshakeData::Startup(stream, params));
+                break Ok(HandshakeData::Startup(stream, supported));
             }
             FeStartupPacket::StartupMessage { version, params } => {
                 warn!(

proxy/src/proxy/mod.rs

@@ -10,15 +10,14 @@ pub(crate) mod wake_compute;
 use std::sync::Arc;
 
 pub use copy_bidirectional::{ErrorSource, copy_bidirectional_client_compute};
-use futures::{FutureExt, TryFutureExt};
+use futures::FutureExt;
 use itertools::Itertools;
 use once_cell::sync::OnceCell;
-use pq_proto::{BeMessage as Be, CancelKeyData, StartupMessageParams};
 use regex::Regex;
 use serde::{Deserialize, Serialize};
 use smol_str::{SmolStr, ToSmolStr, format_smolstr};
 use thiserror::Error;
-use tokio::io::{AsyncRead, AsyncWrite, AsyncWriteExt};
+use tokio::io::{AsyncRead, AsyncWrite};
 use tokio_util::sync::CancellationToken;
 use tracing::{Instrument, debug, error, info, warn};
 
@@ -27,8 +26,9 @@ use self::passthrough::ProxyPassthrough;
 use crate::cancellation::{self, CancellationHandler};
 use crate::config::{ProxyConfig, ProxyProtocolV2, TlsConfig};
 use crate::context::RequestContext;
-use crate::error::ReportableError;
+use crate::error::{ReportableError, UserFacingError};
 use crate::metrics::{Metrics, NumClientConnectionsGuard};
+use crate::pqproto::{BeMessage, CancelKeyData, StartupMessageParams};
 use crate::protocol2::{ConnectHeader, ConnectionInfo, ConnectionInfoExtra, read_proxy_protocol};
 use crate::proxy::handshake::{HandshakeData, handshake};
 use crate::rate_limiter::EndpointRateLimiter;
@@ -38,6 +38,18 @@ use crate::{auth, compute};
 
 const ERR_INSECURE_CONNECTION: &str = "connection is insecure (try using `sslmode=require`)";
 
+#[derive(Error, Debug)]
+#[error("{ERR_INSECURE_CONNECTION}")]
+pub struct TlsRequired;
+
+impl ReportableError for TlsRequired {
+    fn get_error_kind(&self) -> crate::error::ErrorKind {
+        crate::error::ErrorKind::User
+    }
+}
+
+impl UserFacingError for TlsRequired {}
+
 pub async fn run_until_cancelled<F: std::future::Future>(
     f: F,
     cancellation_token: &CancellationToken,
@@ -329,7 +341,7 @@ pub(crate) async fn handle_client<S: AsyncRead + AsyncWrite + Unpin>(
 
     let user_info = match result {
         Ok(user_info) => user_info,
-        Err(e) => stream.throw_error(e, Some(ctx)).await?,
+        Err(e) => Err(stream.throw_error(e, Some(ctx)).await)?,
     };
 
     let user = user_info.get_user().to_owned();
@@ -349,10 +361,10 @@ pub(crate) async fn handle_client<S: AsyncRead + AsyncWrite + Unpin>(
             let app = params.get("application_name");
             let params_span = tracing::info_span!("", ?user, ?db, ?app);
 
-            return stream
+            return Err(stream
                 .throw_error(e, Some(ctx))
                 .instrument(params_span)
-                .await?;
+                .await)?;
         }
     };
 
@@ -365,7 +377,7 @@ pub(crate) async fn handle_client<S: AsyncRead + AsyncWrite + Unpin>(
         .get(NeonOptions::PARAMS_COMPAT)
         .is_some();
 
-    let mut node = connect_to_compute(
+    let res = connect_to_compute(
         ctx,
         &TcpMechanism {
             user_info: compute_user_info.clone(),
@@ -377,22 +389,19 @@ pub(crate) async fn handle_client<S: AsyncRead + AsyncWrite + Unpin>(
         config.wake_compute_retry_config,
         &config.connect_to_compute,
     )
-    .or_else(|e| stream.throw_error(e, Some(ctx)))
-    .await?;
+    .await;
+
+    let node = match res {
+        Ok(node) => node,
+        Err(e) => Err(stream.throw_error(e, Some(ctx)).await)?,
+    };
 
     let cancellation_handler_clone = Arc::clone(&cancellation_handler);
     let session = cancellation_handler_clone.get_key();
 
     session.write_cancel_key(node.cancel_closure.clone())?;
-
-    prepare_client_connection(&node, *session.key(), &mut stream).await?;
-
-    // Before proxy passing, forward to compute whatever data is left in the
-    // PqStream input buffer. Normally there is none, but our serverless npm
-    // driver in pipeline mode sends startup, password and first query
-    // immediately after opening the connection.
-    let (stream, read_buf) = stream.into_inner();
-    node.stream.write_all(&read_buf).await?;
+    prepare_client_connection(&node, *session.key(), &mut stream);
+    let stream = stream.flush_and_into_inner().await?;
 
     let private_link_id = match ctx.extra() {
         Some(ConnectionInfoExtra::Aws { vpce_id }) => Some(vpce_id.clone()),
@@ -413,31 +422,28 @@ pub(crate) async fn handle_client<S: AsyncRead + AsyncWrite + Unpin>(
 }
 
 /// Finish client connection initialization: confirm auth success, send params, etc.
-#[tracing::instrument(skip_all)]
-pub(crate) async fn prepare_client_connection(
+pub(crate) fn prepare_client_connection(
     node: &compute::PostgresConnection,
     cancel_key_data: CancelKeyData,
     stream: &mut PqStream<impl AsyncRead + AsyncWrite + Unpin>,
-) -> Result<(), std::io::Error> {
+) {
     // Forward all deferred notices to the client.
     for notice in &node.delayed_notice {
-        stream.write_message_noflush(&Be::Raw(b'N', notice.as_bytes()))?;
+        stream.write_raw(notice.as_bytes().len(), b'N', |buf| {
+            buf.extend_from_slice(notice.as_bytes());
+        });
     }
 
     // Forward all postgres connection params to the client.
     for (name, value) in &node.params {
-        stream.write_message_noflush(&Be::ParameterStatus {
+        stream.write_message(BeMessage::ParameterStatus {
             name: name.as_bytes(),
             value: value.as_bytes(),
-        })?;
+        });
     }
 
-    stream
-        .write_message_noflush(&Be::BackendKeyData(cancel_key_data))?
-        .write_message(&Be::ReadyForQuery)
-        .await?;
-
-    Ok(())
+    stream.write_message(BeMessage::BackendKeyData(cancel_key_data));
+    stream.write_message(BeMessage::ReadyForQuery);
 }
 
 #[derive(Debug, Clone, PartialEq, Eq, Default, Serialize, Deserialize)]

proxy/src/proxy/retry.rs

@@ -125,9 +125,10 @@ pub(crate) fn retry_after(num_retries: u32, config: RetryConfig) -> time::Durati
 
 #[cfg(test)]
 mod tests {
-    use super::ShouldRetryWakeCompute;
     use postgres_client::error::{DbError, SqlState};
 
+    use super::ShouldRetryWakeCompute;
+
     #[test]
     fn should_retry_wake_compute_for_db_error() {
         // These SQLStates should NOT trigger a wake_compute retry.

proxy/src/proxy/tests/mitm.rs

@@ -10,7 +10,7 @@ use bytes::{Bytes, BytesMut};
 use futures::{SinkExt, StreamExt};
 use postgres_client::tls::TlsConnect;
 use postgres_protocol::message::frontend;
-use tokio::io::{AsyncReadExt, DuplexStream};
+use tokio::io::{AsyncReadExt, AsyncWriteExt, DuplexStream};
 use tokio_util::codec::{Decoder, Encoder};
 
 use super::*;
@@ -49,15 +49,14 @@ async fn proxy_mitm(
         };
 
         let mut end_server = tokio_util::codec::Framed::new(end_server, PgFrame);
-        let (end_client, buf) = end_client.framed.into_inner();
-        assert!(buf.is_empty());
+        let end_client = end_client.flush_and_into_inner().await.unwrap();
         let mut end_client = tokio_util::codec::Framed::new(end_client, PgFrame);
 
         // give the end_server the startup parameters
         let mut buf = BytesMut::new();
         frontend::startup_message(
             &postgres_protocol::message::frontend::StartupMessageParams {
-                params: startup.params.into(),
+                params: startup.params.as_bytes().into(),
             },
             &mut buf,
         )

proxy/src/proxy/tests/mod.rs

@@ -128,7 +128,7 @@ trait TestAuth: Sized {
         self,
         stream: &mut PqStream<Stream<S>>,
     ) -> anyhow::Result<()> {
-        stream.write_message_noflush(&Be::AuthenticationOk)?;
+        stream.write_message(BeMessage::AuthenticationOk);
         Ok(())
     }
 }
@@ -157,9 +157,7 @@ impl TestAuth for Scram {
         self,
         stream: &mut PqStream<Stream<S>>,
     ) -> anyhow::Result<()> {
-        let outcome = auth::AuthFlow::new(stream)
-            .begin(auth::Scram(&self.0, &RequestContext::test()))
-            .await?
+        let outcome = auth::AuthFlow::new(stream, auth::Scram(&self.0, &RequestContext::test()))
             .authenticate()
             .await?;
 
@@ -185,10 +183,12 @@ async fn dummy_proxy(
 
     auth.authenticate(&mut stream).await?;
 
-    stream
-        .write_message_noflush(&Be::CLIENT_ENCODING)?
-        .write_message(&Be::ReadyForQuery)
-        .await?;
+    stream.write_message(BeMessage::ParameterStatus {
+        name: b"client_encoding",
+        value: b"UTF8",
+    });
+    stream.write_message(BeMessage::ReadyForQuery);
+    stream.flush().await?;
 
     Ok(())
 }

proxy/src/redis/cancellation_publisher.rs

@@ -1,10 +1,11 @@
 use core::net::IpAddr;
 use std::sync::Arc;
 
-use pq_proto::CancelKeyData;
 use tokio::sync::Mutex;
 use uuid::Uuid;
 
+use crate::pqproto::CancelKeyData;
+
 pub trait CancellationPublisherMut: Send + Sync + 'static {
     #[allow(async_fn_in_trait)]
     async fn try_publish(

proxy/src/redis/keys.rs

@@ -1,16 +1,15 @@
 use std::io::ErrorKind;
 
 use anyhow::Ok;
-use pq_proto::{CancelKeyData, id_to_cancel_key};
-use serde::{Deserialize, Serialize};
+
+use crate::pqproto::{CancelKeyData, id_to_cancel_key};
 
 pub mod keyspace {
     pub const CANCEL_PREFIX: &str = "cancel";
 }
 
-#[derive(Clone, Debug, Serialize, Deserialize, Eq, PartialEq)]
+#[derive(Clone, Debug, Eq, PartialEq)]
 pub(crate) enum KeyPrefix {
-    #[serde(untagged)]
     Cancel(CancelKeyData),
 }
 
@@ -18,9 +17,7 @@ impl KeyPrefix {
     pub(crate) fn build_redis_key(&self) -> String {
         match self {
             KeyPrefix::Cancel(key) => {
-                let hi = (key.backend_pid as u64) << 32;
-                let lo = (key.cancel_key as u64) & 0xffff_ffff;
-                let id = hi | lo;
+                let id = key.0.get();
                 let keyspace = keyspace::CANCEL_PREFIX;
                 format!("{keyspace}:{id:x}")
             }
@@ -63,10 +60,7 @@ mod tests {
 
     #[test]
     fn test_build_redis_key() {
-        let cancel_key: KeyPrefix = KeyPrefix::Cancel(CancelKeyData {
-            backend_pid: 12345,
-            cancel_key: 54321,
-        });
+        let cancel_key: KeyPrefix = KeyPrefix::Cancel(id_to_cancel_key(12345 << 32 | 54321));
 
         let redis_key = cancel_key.build_redis_key();
         assert_eq!(redis_key, "cancel:30390000d431");
@@ -77,10 +71,7 @@ mod tests {
         let redis_key = "cancel:30390000d431";
         let key: KeyPrefix = parse_redis_key(redis_key).expect("Failed to parse key");
 
-        let ref_key = CancelKeyData {
-            backend_pid: 12345,
-            cancel_key: 54321,
-        };
+        let ref_key = id_to_cancel_key(12345 << 32 | 54321);
 
         assert_eq!(key.as_str(), KeyPrefix::Cancel(ref_key).as_str());
         let KeyPrefix::Cancel(cancel_key) = key;

proxy/src/redis/notifications.rs

@@ -2,11 +2,9 @@ use std::convert::Infallible;
 use std::sync::Arc;
 
 use futures::StreamExt;
-use pq_proto::CancelKeyData;
 use redis::aio::PubSub;
 use serde::{Deserialize, Serialize};
 use tokio_util::sync::CancellationToken;
-use uuid::Uuid;
 
 use super::connection_with_credentials_provider::ConnectionWithCredentialsProvider;
 use crate::cache::project_info::ProjectInfoCache;
@@ -100,14 +98,6 @@ pub(crate) struct PasswordUpdate {
     role_name: RoleNameInt,
 }
 
-#[derive(Clone, Debug, Serialize, Deserialize, Eq, PartialEq)]
-pub(crate) struct CancelSession {
-    pub(crate) region_id: Option<String>,
-    pub(crate) cancel_key_data: CancelKeyData,
-    pub(crate) session_id: Uuid,
-    pub(crate) peer_addr: Option<std::net::IpAddr>,
-}
-
 fn deserialize_json_string<'de, D, T>(deserializer: D) -> Result<T, D::Error>
 where
     T: for<'de2> serde::Deserialize<'de2>,

proxy/src/sasl/messages.rs

@@ -1,7 +1,5 @@
 //! Definitions for SASL messages.
 
-use pq_proto::{BeAuthenticationSaslMessage, BeMessage};
-
 use crate::parse::split_cstr;
 
 /// SASL-specific payload of [`PasswordMessage`](pq_proto::FeMessage::PasswordMessage).
@@ -30,26 +28,6 @@ impl<'a> FirstMessage<'a> {
     }
 }
 
-/// A single SASL message.
-/// This struct is deliberately decoupled from lower-level
-/// [`BeAuthenticationSaslMessage`].
-#[derive(Debug)]
-pub(super) enum ServerMessage<T> {
-    /// We expect to see more steps.
-    Continue(T),
-    /// This is the final step.
-    Final(T),
-}
-
-impl<'a> ServerMessage<&'a str> {
-    pub(super) fn to_reply(&self) -> BeMessage<'a> {
-        BeMessage::AuthenticationSasl(match self {
-            ServerMessage::Continue(s) => BeAuthenticationSaslMessage::Continue(s.as_bytes()),
-            ServerMessage::Final(s) => BeAuthenticationSaslMessage::Final(s.as_bytes()),
-        })
-    }
-}
-
 #[cfg(test)]
 mod tests {
     use super::*;

proxy/src/sasl/mod.rs

@@ -14,7 +14,7 @@ use std::io;
 
 pub(crate) use channel_binding::ChannelBinding;
 pub(crate) use messages::FirstMessage;
-pub(crate) use stream::{Outcome, SaslStream};
+pub(crate) use stream::{Outcome, authenticate};
 use thiserror::Error;
 
 use crate::error::{ReportableError, UserFacingError};
@@ -22,6 +22,9 @@ use crate::error::{ReportableError, UserFacingError};
 /// Fine-grained auth errors help in writing tests.
 #[derive(Error, Debug)]
 pub(crate) enum Error {
+    #[error("Unsupported authentication method: {0}")]
+    BadAuthMethod(Box<str>),
+
     #[error("Channel binding failed: {0}")]
     ChannelBindingFailed(&'static str),
 
@@ -54,6 +57,7 @@ impl UserFacingError for Error {
 impl ReportableError for Error {
     fn get_error_kind(&self) -> crate::error::ErrorKind {
         match self {
+            Error::BadAuthMethod(_) => crate::error::ErrorKind::User,
             Error::ChannelBindingFailed(_) => crate::error::ErrorKind::User,
             Error::ChannelBindingBadMethod(_) => crate::error::ErrorKind::User,
             Error::BadClientMessage(_) => crate::error::ErrorKind::User,

proxy/src/sasl/stream.rs

@@ -3,61 +3,12 @@
 use std::io;
 
 use tokio::io::{AsyncRead, AsyncWrite};
-use tracing::info;
 
-use super::Mechanism;
-use super::messages::ServerMessage;
+use super::{Mechanism, Step};
+use crate::context::RequestContext;
+use crate::pqproto::{BeAuthenticationSaslMessage, BeMessage};
 use crate::stream::PqStream;
 
-/// Abstracts away all peculiarities of the libpq's protocol.
-pub(crate) struct SaslStream<'a, S> {
-    /// The underlying stream.
-    stream: &'a mut PqStream<S>,
-    /// Current password message we received from client.
-    current: bytes::Bytes,
-    /// First SASL message produced by client.
-    first: Option<&'a str>,
-}
-
-impl<'a, S> SaslStream<'a, S> {
-    pub(crate) fn new(stream: &'a mut PqStream<S>, first: &'a str) -> Self {
-        Self {
-            stream,
-            current: bytes::Bytes::new(),
-            first: Some(first),
-        }
-    }
-}
-
-impl<S: AsyncRead + Unpin> SaslStream<'_, S> {
-    // Receive a new SASL message from the client.
-    async fn recv(&mut self) -> io::Result<&str> {
-        if let Some(first) = self.first.take() {
-            return Ok(first);
-        }
-
-        self.current = self.stream.read_password_message().await?;
-        let s = std::str::from_utf8(&self.current)
-            .map_err(|_| io::Error::new(io::ErrorKind::InvalidData, "bad encoding"))?;
-
-        Ok(s)
-    }
-}
-
-impl<S: AsyncWrite + Unpin> SaslStream<'_, S> {
-    // Send a SASL message to the client.
-    async fn send(&mut self, msg: &ServerMessage<&str>) -> io::Result<()> {
-        self.stream.write_message(&msg.to_reply()).await?;
-        Ok(())
-    }
-
-    // Queue a SASL message for the client.
-    fn send_noflush(&mut self, msg: &ServerMessage<&str>) -> io::Result<()> {
-        self.stream.write_message_noflush(&msg.to_reply())?;
-        Ok(())
-    }
-}
-
 /// SASL authentication outcome.
 /// It's much easier to match on those two variants
 /// than to peek into a noisy protocol error type.
@@ -69,33 +20,62 @@ pub(crate) enum Outcome<R> {
     Failure(&'static str),
 }
 
-impl<S: AsyncRead + AsyncWrite + Unpin> SaslStream<'_, S> {
-    /// Perform SASL message exchange according to the underlying algorithm
-    /// until user is either authenticated or denied access.
-    pub(crate) async fn authenticate<M: Mechanism>(
-        mut self,
-        mut mechanism: M,
-    ) -> super::Result<Outcome<M::Output>> {
-        loop {
-            let input = self.recv().await?;
-            let step = mechanism.exchange(input).map_err(|error| {
-                info!(?error, "error during SASL exchange");
-                error
-            })?;
+pub async fn authenticate<S, F, M>(
+    ctx: &RequestContext,
+    stream: &mut PqStream<S>,
+    mechanism: F,
+) -> super::Result<Outcome<M::Output>>
+where
+    S: AsyncRead + AsyncWrite + Unpin,
+    F: FnOnce(&str) -> super::Result<M>,
+    M: Mechanism,
+{
+    let sasl = {
+        // pause the timer while we communicate with the client
+        let _paused = ctx.latency_timer_pause(crate::metrics::Waiting::Client);
 
-            use super::Step;
-            return Ok(match step {
-                Step::Continue(moved_mechanism, reply) => {
-                    self.send(&ServerMessage::Continue(&reply)).await?;
-                    mechanism = moved_mechanism;
-                    continue;
-                }
-                Step::Success(result, reply) => {
-                    self.send_noflush(&ServerMessage::Final(&reply))?;
-                    Outcome::Success(result)
-                }
-                Step::Failure(reason) => Outcome::Failure(reason),
-            });
+        // Initial client message contains the chosen auth method's name.
+        let msg = stream.read_password_message().await?;
+        super::FirstMessage::parse(msg).ok_or(super::Error::BadClientMessage("bad sasl message"))?
+    };
+
+    let mut mechanism = mechanism(sasl.method)?;
+    let mut input = sasl.message;
+    loop {
+        let step = mechanism
+            .exchange(input)
+            .inspect_err(|error| tracing::info!(?error, "error during SASL exchange"))?;
+
+        match step {
+            Step::Continue(moved_mechanism, reply) => {
+                mechanism = moved_mechanism;
+
+                // pause the timer while we communicate with the client
+                let _paused = ctx.latency_timer_pause(crate::metrics::Waiting::Client);
+
+                // write reply
+                let sasl_msg = BeAuthenticationSaslMessage::Continue(reply.as_bytes());
+                stream.write_message(BeMessage::AuthenticationSasl(sasl_msg));
+
+                // get next input
+                stream.flush().await?;
+                let msg = stream.read_password_message().await?;
+                input = std::str::from_utf8(msg)
+                    .map_err(|_| io::Error::new(io::ErrorKind::InvalidData, "bad encoding"))?;
+            }
+            Step::Success(result, reply) => {
+                // pause the timer while we communicate with the client
+                let _paused = ctx.latency_timer_pause(crate::metrics::Waiting::Client);
+
+                // write reply
+                let sasl_msg = BeAuthenticationSaslMessage::Final(reply.as_bytes());
+                stream.write_message(BeMessage::AuthenticationSasl(sasl_msg));
+                stream.write_message(BeMessage::AuthenticationOk);
+                // exit with success
+                break Ok(Outcome::Success(result));
+            }
+            // exit with failure
+            Step::Failure(reason) => break Ok(Outcome::Failure(reason)),
         }
     }
 }

proxy/src/serverless/json.rs

@@ -70,34 +70,6 @@ pub(crate) enum JsonConversionError {
     ParseJsonError(#[from] serde_json::Error),
     #[error("unbalanced array")]
     UnbalancedArray,
-    #[error("unbalanced quoted string")]
-    UnbalancedString,
-}
-
-enum OutputMode {
-    Array(Vec<Value>),
-    Object(Map<String, Value>),
-}
-
-impl OutputMode {
-    fn key(&mut self, key: &str) -> &mut Value {
-        match self {
-            OutputMode::Array(values) => push_entry(values, Value::Null),
-            OutputMode::Object(map) => map.entry(key.to_string()).or_insert(Value::Null),
-        }
-    }
-
-    fn finish(self) -> Value {
-        match self {
-            OutputMode::Array(values) => Value::Array(values),
-            OutputMode::Object(map) => Value::Object(map),
-        }
-    }
-}
-
-fn push_entry<T>(arr: &mut Vec<T>, t: T) -> &mut T {
-    arr.push(t);
-    arr.last_mut().expect("a value was just inserted")
 }
 
 //
@@ -105,276 +77,182 @@ fn push_entry<T>(arr: &mut Vec<T>, t: T) -> &mut T {
 //
 pub(crate) fn pg_text_row_to_json(
     row: &Row,
+    columns: &[Type],
     raw_output: bool,
     array_mode: bool,
 ) -> Result<Value, JsonConversionError> {
-    let mut entries = if array_mode {
-        OutputMode::Array(Vec::with_capacity(row.columns().len()))
+    let iter = row
+        .columns()
+        .iter()
+        .zip(columns)
+        .enumerate()
+        .map(|(i, (column, typ))| {
+            let name = column.name();
+            let pg_value = row.as_text(i).map_err(JsonConversionError::AsTextError)?;
+            let json_value = if raw_output {
+                match pg_value {
+                    Some(v) => Value::String(v.to_string()),
+                    None => Value::Null,
+                }
+            } else {
+                pg_text_to_json(pg_value, typ)?
+            };
+            Ok((name.to_string(), json_value))
+        });
+
+    if array_mode {
+        // drop keys and aggregate into array
+        let arr = iter
+            .map(|r| r.map(|(_key, val)| val))
+            .collect::<Result<Vec<Value>, JsonConversionError>>()?;
+        Ok(Value::Array(arr))
     } else {
-        OutputMode::Object(Map::with_capacity(row.columns().len()))
-    };
-
-    for (i, column) in row.columns().iter().enumerate() {
-        let pg_value = row.as_text(i).map_err(JsonConversionError::AsTextError)?;
-
-        let value = entries.key(column.name());
-
-        match pg_value {
-            Some(v) if raw_output => *value = Value::String(v.to_string()),
-            Some(v) => pg_text_to_json(value, v, column.type_())?,
-            None => *value = Value::Null,
-        }
+        let obj = iter.collect::<Result<Map<String, Value>, JsonConversionError>>()?;
+        Ok(Value::Object(obj))
     }
-
-    Ok(entries.finish())
 }
 
 //
 // Convert postgres text-encoded value to JSON value
 //
-fn pg_text_to_json(
-    output: &mut Value,
-    val: &str,
-    pg_type: &Type,
-) -> Result<(), JsonConversionError> {
-    if let Kind::Array(elem_type) = pg_type.kind() {
-        // todo: we should fetch this from postgres.
-        let delimiter = ',';
-
-        let mut array = vec![];
-        pg_array_parse(&mut array, val, elem_type, delimiter)?;
-        *output = Value::Array(array);
-        return Ok(());
-    }
-
-    match *pg_type {
-        Type::BOOL => *output = Value::Bool(val == "t"),
-        Type::INT2 | Type::INT4 => {
-            let val = val.parse::<i32>()?;
-            *output = Value::Number(serde_json::Number::from(val));
+fn pg_text_to_json(pg_value: Option<&str>, pg_type: &Type) -> Result<Value, JsonConversionError> {
+    if let Some(val) = pg_value {
+        if let Kind::Array(elem_type) = pg_type.kind() {
+            return pg_array_parse(val, elem_type);
         }
-        Type::FLOAT4 | Type::FLOAT8 => {
-            let fval = val.parse::<f64>()?;
-            let num = serde_json::Number::from_f64(fval);
-            if let Some(num) = num {
-                *output = Value::Number(num);
-            } else {
-                // Pass Nan, Inf, -Inf as strings
-                // JS JSON.stringify() does converts them to null, but we
-                // want to preserve them, so we pass them as strings
-                *output = Value::String(val.to_string());
+
+        match *pg_type {
+            Type::BOOL => Ok(Value::Bool(val == "t")),
+            Type::INT2 | Type::INT4 => {
+                let val = val.parse::<i32>()?;
+                Ok(Value::Number(serde_json::Number::from(val)))
+            }
+            Type::FLOAT4 | Type::FLOAT8 => {
+                let fval = val.parse::<f64>()?;
+                let num = serde_json::Number::from_f64(fval);
+                if let Some(num) = num {
+                    Ok(Value::Number(num))
+                } else {
+                    // Pass Nan, Inf, -Inf as strings
+                    // JS JSON.stringify() does converts them to null, but we
+                    // want to preserve them, so we pass them as strings
+                    Ok(Value::String(val.to_string()))
+                }
+            }
+            Type::JSON | Type::JSONB => Ok(serde_json::from_str(val)?),
+            _ => Ok(Value::String(val.to_string())),
+        }
+    } else {
+        Ok(Value::Null)
+    }
+}
+
+//
+// Parse postgres array into JSON array.
+//
+// This is a bit involved because we need to handle nested arrays and quoted
+// values. Unlike postgres we don't check that all nested arrays have the same
+// dimensions, we just return them as is.
+//
+fn pg_array_parse(pg_array: &str, elem_type: &Type) -> Result<Value, JsonConversionError> {
+    pg_array_parse_inner(pg_array, elem_type, false).map(|(v, _)| v)
+}
+
+fn pg_array_parse_inner(
+    pg_array: &str,
+    elem_type: &Type,
+    nested: bool,
+) -> Result<(Value, usize), JsonConversionError> {
+    let mut pg_array_chr = pg_array.char_indices();
+    let mut level = 0;
+    let mut quote = false;
+    let mut entries: Vec<Value> = Vec::new();
+    let mut entry = String::new();
+
+    // skip bounds decoration
+    if let Some('[') = pg_array.chars().next() {
+        for (_, c) in pg_array_chr.by_ref() {
+            if c == '=' {
+                break;
             }
         }
-        Type::JSON | Type::JSONB => *output = serde_json::from_str(val)?,
-        _ => *output = Value::String(val.to_string()),
     }
 
-    Ok(())
-}
+    fn push_checked(
+        entry: &mut String,
+        entries: &mut Vec<Value>,
+        elem_type: &Type,
+    ) -> Result<(), JsonConversionError> {
+        if !entry.is_empty() {
+            // While in usual postgres response we get nulls as None and everything else
+            // as Some(&str), in arrays we get NULL as unquoted 'NULL' string (while
+            // string with value 'NULL' will be represented by '"NULL"'). So catch NULLs
+            // here while we have quotation info and convert them to None.
+            if entry == "NULL" {
+                entries.push(pg_text_to_json(None, elem_type)?);
+            } else {
+                entries.push(pg_text_to_json(Some(entry), elem_type)?);
+            }
+            entry.clear();
+        }
 
-/// Parse postgres array into JSON array.
-///
-/// This is a bit involved because we need to handle nested arrays and quoted
-/// values. Unlike postgres we don't check that all nested arrays have the same
-/// dimensions, we just return them as is.
-///
-/// <https://www.postgresql.org/docs/current/arrays.html#ARRAYS-IO>
-///
-/// The external text representation of an array value consists of items that are interpreted
-/// according to the I/O conversion rules for the array's element type, plus decoration that
-/// indicates the array structure. The decoration consists of curly braces (`{` and `}`) around
-/// the array value plus delimiter characters between adjacent items. The delimiter character
-/// is usually a comma (,) but can be something else: it is determined by the typdelim setting
-/// for the array's element type. Among the standard data types provided in the PostgreSQL
-/// distribution, all use a comma, except for type box, which uses a semicolon (;).
-///
-/// In a multidimensional array, each dimension (row, plane, cube, etc.)
-/// gets its own level of curly braces, and delimiters must be written between adjacent
-/// curly-braced entities of the same level.
-fn pg_array_parse(
-    elements: &mut Vec<Value>,
-    mut pg_array: &str,
-    elem: &Type,
-    delim: char,
-) -> Result<(), JsonConversionError> {
-    // skip bounds decoration, eg:
-    // `[1:1][-2:-1][3:5]={{{1,2,3},{4,5,6}}}`
-    // technically these are significant, but we have no way to represent them in json.
-    if let Some('[') = pg_array.chars().next() {
-        let Some((_bounds, array)) = pg_array.split_once('=') else {
-            return Err(JsonConversionError::UnbalancedArray);
-        };
-        pg_array = array;
+        Ok(())
     }
 
-    // whitespace might preceed a `{`.
-    let pg_array = pg_array.trim_start();
+    while let Some((mut i, mut c)) = pg_array_chr.next() {
+        let mut escaped = false;
 
-    let rest = pg_array_parse_inner(elements, pg_array, elem, delim)?;
-    if !rest.is_empty() {
-        return Err(JsonConversionError::UnbalancedArray);
-    }
+        if c == '\\' {
+            escaped = true;
+            let Some(x) = pg_array_chr.next() else {
+                return Err(JsonConversionError::UnbalancedArray);
+            };
+            (i, c) = x;
+        }
 
-    Ok(())
-}
-
-/// reads a single array from the `pg_array` string and pushes each values to `elements`.
-/// returns the rest of the `pg_array` string that was not read.
-fn pg_array_parse_inner<'a>(
-    elements: &mut Vec<Value>,
-    mut pg_array: &'a str,
-    elem: &Type,
-    delim: char,
-) -> Result<&'a str, JsonConversionError> {
-    // array should have a `{` prefix.
-    pg_array = pg_array
-        .strip_prefix('{')
-        .ok_or(JsonConversionError::UnbalancedArray)?;
-
-    let mut q = String::new();
-
-    loop {
-        let value = push_entry(elements, Value::Null);
-        pg_array = pg_array_parse_item(value, &mut q, pg_array, elem, delim)?;
-
-        // check for separator.
-        if let Some(next) = pg_array.strip_prefix(delim) {
-            // next item.
-            pg_array = next;
-        } else {
-            break;
+        match c {
+            '{' if !quote => {
+                level += 1;
+                if level > 1 {
+                    let (res, off) = pg_array_parse_inner(&pg_array[i..], elem_type, true)?;
+                    entries.push(res);
+                    for _ in 0..off - 1 {
+                        pg_array_chr.next();
+                    }
+                }
+            }
+            '}' if !quote => {
+                level -= 1;
+                if level == 0 {
+                    push_checked(&mut entry, &mut entries, elem_type)?;
+                    if nested {
+                        return Ok((Value::Array(entries), i));
+                    }
+                }
+            }
+            '"' if !escaped => {
+                if quote {
+                    // end of quoted string, so push it manually without any checks
+                    // for emptiness or nulls
+                    entries.push(pg_text_to_json(Some(&entry), elem_type)?);
+                    entry.clear();
+                }
+                quote = !quote;
+            }
+            ',' if !quote => {
+                push_checked(&mut entry, &mut entries, elem_type)?;
+            }
+            _ => {
+                entry.push(c);
+            }
         }
     }
 
-    let Some(next) = pg_array.strip_prefix('}') else {
-        // missing `}` terminator.
+    if level != 0 {
         return Err(JsonConversionError::UnbalancedArray);
-    };
-
-    // whitespace might follow a `}`.
-    Ok(next.trim_start())
-}
-
-/// reads a single item from the `pg_array` string.
-/// returns the rest of the `pg_array` string that was not read.
-///
-/// `quoted` is a scratch allocation that has no defined output.
-fn pg_array_parse_item<'a>(
-    output: &mut Value,
-    quoted: &mut String,
-    mut pg_array: &'a str,
-    elem: &Type,
-    delim: char,
-) -> Result<&'a str, JsonConversionError> {
-    // We are trying to parse an array item.
-    // This could be a new array, if this is a multi-dimentional array.
-    // This could be a quoted string representing `elem`.
-    // This could be an unquoted string representing `elem`.
-
-    // whitespace might preceed an item.
-    pg_array = pg_array.trim_start();
-
-    if pg_array.strip_prefix('{').is_some() {
-        // nested array.
-        let mut nested = vec![];
-        pg_array = pg_array_parse_inner(&mut nested, pg_array, elem, delim)?;
-        *output = Value::Array(nested);
-        return Ok(pg_array);
     }
 
-    if let Some(mut pg_array) = pg_array.strip_prefix('"') {
-        pg_array = pg_array_parse_quoted(quoted, pg_array)?;
-
-        // we have unquoted an item string:
-        pg_text_to_json(output, quoted, elem)?;
-
-        quoted.clear();
-
-        return Ok(pg_array);
-    }
-
-    // we need to parse an item. read until we find a delimiter or `}`.
-    let index = pg_array
-        .find([delim, '}'])
-        .ok_or(JsonConversionError::UnbalancedArray)?;
-
-    let item;
-    (item, pg_array) = pg_array.split_at(index);
-
-    // item might have trailing whitespace that we need to ignore.
-    let item = item.trim_end();
-
-    // we might have an item string:
-    // check for null
-    if item == "NULL" {
-        *output = Value::Null;
-    } else {
-        pg_text_to_json(output, item, elem)?;
-    }
-
-    Ok(pg_array)
-}
-
-/// reads a single quoted item from the `pg_array` string.
-///
-/// Returns the rest of the `pg_array` string that was not read.
-/// The output is written into `quoted`.
-///
-/// The pg_array string must have a `"` terminator, but the `"` initial value
-/// must have already been removed from the input. The terminator is removed.
-fn pg_array_parse_quoted<'a>(
-    quoted: &mut String,
-    mut pg_array: &'a str,
-) -> Result<&'a str, JsonConversionError> {
-    // The array output routine will put double quotes around element values if they are empty strings,
-    // contain curly braces, delimiter characters, double quotes, backslashes, or white space,
-    // or match the word `NULL`. Double quotes and backslashes embedded in element values will be backslash-escaped.
-    // For numeric data types it is safe to assume that double quotes will never appear,
-    // but for textual data types one should be prepared to cope with either the presence or absence of quotes.
-
-    // We write to quoted in chunks terminated by an escape character.
-    // Eg if we have the input `foo\"bar"`, then we write `foo`, then `"`, then finally `bar`.
-
-    loop {
-        // we need to parse an chunk. read until we find a '\\' or `"`.
-        let i = pg_array
-            .find(['\\', '"'])
-            .ok_or(JsonConversionError::UnbalancedString)?;
-
-        let chunk: &str;
-        (chunk, pg_array) = pg_array
-            .split_at_checked(i)
-            .expect("i is guaranteed to be in-bounds of pg_array");
-
-        // push the chunk.
-        quoted.push_str(chunk);
-
-        // consume the chunk_end character.
-        let chunk_end: char;
-        (chunk_end, pg_array) =
-            split_first_char(pg_array).expect("pg_array should start with either '\\\\' or '\"'");
-
-        // finished.
-        if chunk_end == '"' {
-            // whitespace might follow the '"'.
-            pg_array = pg_array.trim_start();
-
-            break Ok(pg_array);
-        }
-
-        // consume the escaped character.
-        let escaped: char;
-        (escaped, pg_array) =
-            split_first_char(pg_array).ok_or(JsonConversionError::UnbalancedString)?;
-
-        quoted.push(escaped);
-    }
-}
-
-fn split_first_char(s: &str) -> Option<(char, &str)> {
-    let mut chars = s.chars();
-    let c = chars.next()?;
-    Some((c, chars.as_str()))
+    Ok((Value::Array(entries), 0))
 }
 
 #[cfg(test)]
@@ -438,33 +316,37 @@ mod tests {
         );
     }
 
-    fn pg_text_to_json(val: &str, pg_type: &Type) -> Value {
-        let mut v = Value::Null;
-        super::pg_text_to_json(&mut v, val, pg_type).unwrap();
-        v
-    }
-
-    fn pg_array_parse(pg_array: &str, pg_type: &Type) -> Value {
-        let mut array = vec![];
-        super::pg_array_parse(&mut array, pg_array, pg_type, ',').unwrap();
-        Value::Array(array)
-    }
-
     #[test]
     fn test_atomic_types_parse() {
-        assert_eq!(pg_text_to_json("foo", &Type::TEXT), json!("foo"));
-        assert_eq!(pg_text_to_json("42", &Type::INT4), json!(42));
-        assert_eq!(pg_text_to_json("42", &Type::INT2), json!(42));
-        assert_eq!(pg_text_to_json("42", &Type::INT8), json!("42"));
-        assert_eq!(pg_text_to_json("42.42", &Type::FLOAT8), json!(42.42));
-        assert_eq!(pg_text_to_json("42.42", &Type::FLOAT4), json!(42.42));
-        assert_eq!(pg_text_to_json("NaN", &Type::FLOAT4), json!("NaN"));
         assert_eq!(
-            pg_text_to_json("Infinity", &Type::FLOAT4),
+            pg_text_to_json(Some("foo"), &Type::TEXT).unwrap(),
+            json!("foo")
+        );
+        assert_eq!(pg_text_to_json(None, &Type::TEXT).unwrap(), json!(null));
+        assert_eq!(pg_text_to_json(Some("42"), &Type::INT4).unwrap(), json!(42));
+        assert_eq!(pg_text_to_json(Some("42"), &Type::INT2).unwrap(), json!(42));
+        assert_eq!(
+            pg_text_to_json(Some("42"), &Type::INT8).unwrap(),
+            json!("42")
+        );
+        assert_eq!(
+            pg_text_to_json(Some("42.42"), &Type::FLOAT8).unwrap(),
+            json!(42.42)
+        );
+        assert_eq!(
+            pg_text_to_json(Some("42.42"), &Type::FLOAT4).unwrap(),
+            json!(42.42)
+        );
+        assert_eq!(
+            pg_text_to_json(Some("NaN"), &Type::FLOAT4).unwrap(),
+            json!("NaN")
+        );
+        assert_eq!(
+            pg_text_to_json(Some("Infinity"), &Type::FLOAT4).unwrap(),
             json!("Infinity")
         );
         assert_eq!(
-            pg_text_to_json("-Infinity", &Type::FLOAT4),
+            pg_text_to_json(Some("-Infinity"), &Type::FLOAT4).unwrap(),
             json!("-Infinity")
         );
 
@@ -473,9 +355,10 @@ mod tests {
                 .unwrap();
         assert_eq!(
             pg_text_to_json(
-                r#"{"s":"str","n":42,"f":4.2,"a":[null,3,"a"]}"#,
+                Some(r#"{"s":"str","n":42,"f":4.2,"a":[null,3,"a"]}"#),
                 &Type::JSONB
-            ),
+            )
+            .unwrap(),
             json
         );
     }
@@ -483,7 +366,7 @@ mod tests {
     #[test]
     fn test_pg_array_parse_text() {
         fn pt(pg_arr: &str) -> Value {
-            pg_array_parse(pg_arr, &Type::TEXT)
+            pg_array_parse(pg_arr, &Type::TEXT).unwrap()
         }
         assert_eq!(
             pt(r#"{"aa\"\\\,a",cha,"bbbb"}"#),
@@ -506,7 +389,7 @@ mod tests {
     #[test]
     fn test_pg_array_parse_bool() {
         fn pb(pg_arr: &str) -> Value {
-            pg_array_parse(pg_arr, &Type::BOOL)
+            pg_array_parse(pg_arr, &Type::BOOL).unwrap()
         }
         assert_eq!(pb(r#"{t,f,t}"#), json!([true, false, true]));
         assert_eq!(pb(r#"{{t,f,t}}"#), json!([[true, false, true]]));
@@ -523,7 +406,7 @@ mod tests {
     #[test]
     fn test_pg_array_parse_numbers() {
         fn pn(pg_arr: &str, ty: &Type) -> Value {
-            pg_array_parse(pg_arr, ty)
+            pg_array_parse(pg_arr, ty).unwrap()
         }
         assert_eq!(pn(r#"{1,2,3}"#, &Type::INT4), json!([1, 2, 3]));
         assert_eq!(pn(r#"{1,2,3}"#, &Type::INT2), json!([1, 2, 3]));
@@ -551,7 +434,7 @@ mod tests {
     #[test]
     fn test_pg_array_with_decoration() {
         fn p(pg_arr: &str) -> Value {
-            pg_array_parse(pg_arr, &Type::INT2)
+            pg_array_parse(pg_arr, &Type::INT2).unwrap()
         }
         assert_eq!(
             p(r#"[1:1][-2:-1][3:5]={{{1,2,3},{4,5,6}}}"#),
@@ -562,7 +445,7 @@ mod tests {
     #[test]
     fn test_pg_array_parse_json() {
         fn pt(pg_arr: &str) -> Value {
-            pg_array_parse(pg_arr, &Type::JSONB)
+            pg_array_parse(pg_arr, &Type::JSONB).unwrap()
         }
         assert_eq!(pt(r#"{"{}"}"#), json!([{}]));
         assert_eq!(

proxy/src/serverless/sql_over_http.rs

@@ -17,7 +17,6 @@ use postgres_client::error::{DbError, ErrorPosition, SqlState};
 use postgres_client::{
     GenericClient, IsolationLevel, NoTls, ReadyForQueryStatus, RowStream, Transaction,
 };
-use pq_proto::StartupMessageParamsBuilder;
 use serde::Serialize;
 use serde_json::Value;
 use serde_json::value::RawValue;
@@ -41,6 +40,7 @@ use crate::context::RequestContext;
 use crate::error::{ErrorKind, ReportableError, UserFacingError};
 use crate::http::{ReadBodyError, read_body_with_limit};
 use crate::metrics::{HttpDirection, Metrics, SniGroup, SniKind};
+use crate::pqproto::StartupMessageParams;
 use crate::proxy::{NeonOptions, run_until_cancelled};
 use crate::serverless::backend::HttpConnError;
 use crate::types::{DbName, RoleName};
@@ -219,7 +219,7 @@ fn get_conn_info(
 
     let mut options = Option::None;
 
-    let mut params = StartupMessageParamsBuilder::default();
+    let mut params = StartupMessageParams::default();
     params.insert("user", &username);
     params.insert("database", &dbname);
     for (key, value) in pairs {
@@ -1102,6 +1102,7 @@ async fn query_to_json<T: GenericClient>(
 
     let columns_len = row_stream.statement.columns().len();
     let mut fields = Vec::with_capacity(columns_len);
+    let mut types = Vec::with_capacity(columns_len);
 
     for c in row_stream.statement.columns() {
         fields.push(json!({
@@ -1113,6 +1114,8 @@ async fn query_to_json<T: GenericClient>(
             "dataTypeModifier": c.type_modifier(),
             "format": "text",
         }));
+
+        types.push(c.type_().clone());
     }
 
     let raw_output = parsed_headers.raw_output;
@@ -1134,7 +1137,7 @@ async fn query_to_json<T: GenericClient>(
             ));
         }
 
-        let row = pg_text_row_to_json(&row, raw_output, array_mode)?;
+        let row = pg_text_row_to_json(&row, &types, raw_output, array_mode)?;
         rows.push(row);
 
         // assumption: parsing pg text and converting to json takes CPU time.

proxy/src/stream.rs

@@ -2,19 +2,17 @@ use std::pin::Pin;
 use std::sync::Arc;
 use std::{io, task};
 
-use bytes::BytesMut;
-use pq_proto::framed::{ConnectionError, Framed};
-use pq_proto::{BeMessage, FeMessage, FeStartupPacket, ProtocolError};
 use rustls::ServerConfig;
-use serde::{Deserialize, Serialize};
 use thiserror::Error;
-use tokio::io::{AsyncRead, AsyncWrite, ReadBuf};
+use tokio::io::{AsyncRead, AsyncWrite, AsyncWriteExt, ReadBuf};
 use tokio_rustls::server::TlsStream;
-use tracing::debug;
 
-use crate::control_plane::messages::ColdStartInfo;
 use crate::error::{ErrorKind, ReportableError, UserFacingError};
 use crate::metrics::Metrics;
+use crate::pqproto::{
+    BeMessage, FE_PASSWORD_MESSAGE, FeStartupPacket, SQLSTATE_INTERNAL_ERROR, WriteBuf,
+    read_message, read_startup,
+};
 use crate::tls::TlsServerEndPoint;
 
 /// Stream wrapper which implements libpq's protocol.
@@ -23,58 +21,77 @@ use crate::tls::TlsServerEndPoint;
 /// or [`AsyncWrite`] to prevent subtle errors (e.g. trying
 /// to pass random malformed bytes through the connection).
 pub struct PqStream<S> {
-    pub(crate) framed: Framed<S>,
+    stream: S,
+    read: Vec<u8>,
+    write: WriteBuf,
 }
 
 impl<S> PqStream<S> {
-    /// Construct a new libpq protocol wrapper.
-    pub fn new(stream: S) -> Self {
+    pub fn get_ref(&self) -> &S {
+        &self.stream
+    }
+
+    /// Construct a new libpq protocol wrapper over a stream without the first startup message.
+    #[cfg(test)]
+    pub fn new_skip_handshake(stream: S) -> Self {
         Self {
-            framed: Framed::new(stream),
+            stream,
+            read: Vec::new(),
+            write: WriteBuf::new(),
         }
     }
-
-    /// Extract the underlying stream and read buffer.
-    pub fn into_inner(self) -> (S, BytesMut) {
-        self.framed.into_inner()
-    }
-
-    /// Get a shared reference to the underlying stream.
-    pub(crate) fn get_ref(&self) -> &S {
-        self.framed.get_ref()
-    }
 }
 
-fn err_connection() -> io::Error {
-    io::Error::new(io::ErrorKind::ConnectionAborted, "connection is lost")
+impl<S: AsyncRead + AsyncWrite + Unpin> PqStream<S> {
+    /// Construct a new libpq protocol wrapper and read the first startup message.
+    ///
+    /// This is not cancel safe.
+    pub async fn parse_startup(mut stream: S) -> io::Result<(Self, FeStartupPacket)> {
+        let startup = read_startup(&mut stream).await?;
+        Ok((
+            Self {
+                stream,
+                read: Vec::new(),
+                write: WriteBuf::new(),
+            },
+            startup,
+        ))
+    }
+
+    /// Tell the client that encryption is not supported.
+    ///
+    /// This is not cancel safe
+    pub async fn reject_encryption(&mut self) -> io::Result<FeStartupPacket> {
+        // N for No.
+        self.write.encryption(b'N');
+        self.flush().await?;
+        read_startup(&mut self.stream).await
+    }
 }
 
 impl<S: AsyncRead + Unpin> PqStream<S> {
-    /// Receive [`FeStartupPacket`], which is a first packet sent by a client.
-    pub async fn read_startup_packet(&mut self) -> io::Result<FeStartupPacket> {
-        self.framed
-            .read_startup_message()
-            .await
-            .map_err(ConnectionError::into_io_error)?
-            .ok_or_else(err_connection)
-    }
-
-    async fn read_message(&mut self) -> io::Result<FeMessage> {
-        self.framed
-            .read_message()
-            .await
-            .map_err(ConnectionError::into_io_error)?
-            .ok_or_else(err_connection)
-    }
-
-    pub(crate) async fn read_password_message(&mut self) -> io::Result<bytes::Bytes> {
-        match self.read_message().await? {
-            FeMessage::PasswordMessage(msg) => Ok(msg),
-            bad => Err(io::Error::new(
-                io::ErrorKind::InvalidData,
-                format!("unexpected message type: {bad:?}"),
-            )),
+    /// Read a raw postgres packet, which will respect the max length requested.
+    /// This is not cancel safe.
+    async fn read_raw_expect(&mut self, tag: u8, max: usize) -> io::Result<&mut [u8]> {
+        let (actual_tag, msg) = read_message(&mut self.stream, &mut self.read, max).await?;
+        if actual_tag != tag {
+            return Err(io::Error::other(format!(
+                "incorrect message tag, expected {:?}, got {:?}",
+                tag as char, actual_tag as char,
+            )));
         }
+        Ok(msg)
+    }
+
+    /// Read a postgres password message, which will respect the max length requested.
+    /// This is not cancel safe.
+    pub async fn read_password_message(&mut self) -> io::Result<&mut [u8]> {
+        // passwords are usually pretty short
+        // and SASL SCRAM messages are no longer than 256 bytes in my testing
+        // (a few hashes and random bytes, encoded into base64).
+        const MAX_PASSWORD_LENGTH: usize = 512;
+        self.read_raw_expect(FE_PASSWORD_MESSAGE, MAX_PASSWORD_LENGTH)
+            .await
     }
 }
 
@@ -84,6 +101,16 @@ pub struct ReportedError {
     error_kind: ErrorKind,
 }
 
+impl ReportedError {
+    pub fn new(e: (impl UserFacingError + Into<anyhow::Error>)) -> Self {
+        let error_kind = e.get_error_kind();
+        Self {
+            source: e.into(),
+            error_kind,
+        }
+    }
+}
+
 impl std::fmt::Display for ReportedError {
     fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
         self.source.fmt(f)
@@ -102,109 +129,65 @@ impl ReportableError for ReportedError {
     }
 }
 
-#[derive(Serialize, Deserialize, Debug)]
-enum ErrorTag {
-    #[serde(rename = "proxy")]
-    Proxy,
-    #[serde(rename = "compute")]
-    Compute,
-    #[serde(rename = "client")]
-    Client,
-    #[serde(rename = "controlplane")]
-    ControlPlane,
-    #[serde(rename = "other")]
-    Other,
-}
-
-impl From<ErrorKind> for ErrorTag {
-    fn from(error_kind: ErrorKind) -> Self {
-        match error_kind {
-            ErrorKind::User => Self::Client,
-            ErrorKind::ClientDisconnect => Self::Client,
-            ErrorKind::RateLimit => Self::Proxy,
-            ErrorKind::ServiceRateLimit => Self::Proxy, // considering rate limit as proxy error for SLI
-            ErrorKind::Quota => Self::Proxy,
-            ErrorKind::Service => Self::Proxy,
-            ErrorKind::ControlPlane => Self::ControlPlane,
-            ErrorKind::Postgres => Self::Other,
-            ErrorKind::Compute => Self::Compute,
-        }
-    }
-}
-
-#[derive(Serialize, Deserialize, Debug)]
-#[serde(rename_all = "snake_case")]
-struct ProbeErrorData {
-    tag: ErrorTag,
-    msg: String,
-    cold_start_info: Option<ColdStartInfo>,
-}
-
 impl<S: AsyncWrite + Unpin> PqStream<S> {
-    /// Write the message into an internal buffer, but don't flush the underlying stream.
-    pub(crate) fn write_message_noflush(
-        &mut self,
-        message: &BeMessage<'_>,
-    ) -> io::Result<&mut Self> {
-        self.framed
-            .write_message(message)
-            .map_err(ProtocolError::into_io_error)?;
-        Ok(self)
+    /// Tell the client that we are willing to accept SSL.
+    /// This is not cancel safe
+    pub async fn accept_tls(mut self) -> io::Result<S> {
+        // S for SSL.
+        self.write.encryption(b'S');
+        self.flush().await?;
+        Ok(self.stream)
     }
 
-    /// Write the message into an internal buffer and flush it.
-    pub async fn write_message(&mut self, message: &BeMessage<'_>) -> io::Result<&mut Self> {
-        self.write_message_noflush(message)?;
-        self.flush().await?;
-        Ok(self)
+    /// Assert that we are using direct TLS.
+    pub fn accept_direct_tls(self) -> S {
+        self.stream
+    }
+
+    /// Write a raw message to the internal buffer.
+    pub fn write_raw(&mut self, size_hint: usize, tag: u8, f: impl FnOnce(&mut Vec<u8>)) {
+        self.write.write_raw(size_hint, tag, f);
+    }
+
+    /// Write the message into an internal buffer
+    pub fn write_message(&mut self, message: BeMessage<'_>) {
+        message.write_message(&mut self.write);
     }
 
     /// Flush the output buffer into the underlying stream.
-    pub(crate) async fn flush(&mut self) -> io::Result<&mut Self> {
-        self.framed.flush().await?;
-        Ok(self)
+    ///
+    /// This is cancel safe.
+    pub async fn flush(&mut self) -> io::Result<()> {
+        self.stream.write_all_buf(&mut self.write).await?;
+        self.write.reset();
+
+        self.stream.flush().await?;
+
+        Ok(())
     }
 
-    /// Writes message with the given error kind to the stream.
-    /// Used only for probe queries
-    async fn write_format_message(
-        &mut self,
-        msg: &str,
-        error_kind: ErrorKind,
-        ctx: Option<&crate::context::RequestContext>,
-    ) -> String {
-        let formatted_msg = match ctx {
-            Some(ctx) if ctx.get_testodrome_id().is_some() => {
-                serde_json::to_string(&ProbeErrorData {
-                    tag: ErrorTag::from(error_kind),
-                    msg: msg.to_string(),
-                    cold_start_info: Some(ctx.cold_start_info()),
-                })
-                .unwrap_or_default()
-            }
-            _ => msg.to_string(),
-        };
-
-        // already error case, ignore client IO error
-        self.write_message(&BeMessage::ErrorResponse(&formatted_msg, None))
-            .await
-            .inspect_err(|e| debug!("write_message failed: {e}"))
-            .ok();
-
-        formatted_msg
+    /// Flush the output buffer into the underlying stream.
+    ///
+    /// This is cancel safe.
+    pub async fn flush_and_into_inner(mut self) -> io::Result<S> {
+        self.flush().await?;
+        Ok(self.stream)
     }
 
-    /// Write the error message using [`Self::write_format_message`], then re-throw it.
-    /// Allowing string literals is safe under the assumption they might not contain any runtime info.
-    /// This method exists due to `&str` not implementing `Into<anyhow::Error>`.
+    /// Write the error message to the client, then re-throw it.
+    ///
+    /// Trait [`UserFacingError`] acts as an allowlist for error types.
     /// If `ctx` is provided and has testodrome_id set, error messages will be prefixed according to error kind.
-    pub async fn throw_error_str<T>(
+    pub(crate) async fn throw_error<E>(
         &mut self,
-        msg: &'static str,
-        error_kind: ErrorKind,
+        error: E,
         ctx: Option<&crate::context::RequestContext>,
-    ) -> Result<T, ReportedError> {
-        self.write_format_message(msg, error_kind, ctx).await;
+    ) -> ReportedError
+    where
+        E: UserFacingError + Into<anyhow::Error>,
+    {
+        let error_kind = error.get_error_kind();
+        let msg = error.to_string_client();
 
         if error_kind != ErrorKind::RateLimit && error_kind != ErrorKind::User {
             tracing::info!(
@@ -214,39 +197,39 @@ impl<S: AsyncWrite + Unpin> PqStream<S> {
             );
         }
 
-        Err(ReportedError {
-            source: anyhow::anyhow!(msg),
-            error_kind,
-        })
-    }
-
-    /// Write the error message using [`Self::write_format_message`], then re-throw it.
-    /// Trait [`UserFacingError`] acts as an allowlist for error types.
-    /// If `ctx` is provided and has testodrome_id set, error messages will be prefixed according to error kind.
-    pub(crate) async fn throw_error<T, E>(
-        &mut self,
-        error: E,
-        ctx: Option<&crate::context::RequestContext>,
-    ) -> Result<T, ReportedError>
-    where
-        E: UserFacingError + Into<anyhow::Error>,
-    {
-        let error_kind = error.get_error_kind();
-        let msg = error.to_string_client();
-        self.write_format_message(&msg, error_kind, ctx).await;
-        if error_kind != ErrorKind::RateLimit && error_kind != ErrorKind::User {
-            tracing::info!(
-                kind=error_kind.to_metric_label(),
-                error=%error,
-                msg,
-                "forwarding error to user",
-            );
+        let probe_msg;
+        let mut msg = &*msg;
+        if let Some(ctx) = ctx {
+            if ctx.get_testodrome_id().is_some() {
+                let tag = match error_kind {
+                    ErrorKind::User => "client",
+                    ErrorKind::ClientDisconnect => "client",
+                    ErrorKind::RateLimit => "proxy",
+                    ErrorKind::ServiceRateLimit => "proxy",
+                    ErrorKind::Quota => "proxy",
+                    ErrorKind::Service => "proxy",
+                    ErrorKind::ControlPlane => "controlplane",
+                    ErrorKind::Postgres => "other",
+                    ErrorKind::Compute => "compute",
+                };
+                probe_msg = typed_json::json!({
+                    "tag": tag,
+                    "msg": msg,
+                    "cold_start_info": ctx.cold_start_info(),
+                })
+                .to_string();
+                msg = &probe_msg;
+            }
         }
 
-        Err(ReportedError {
-            source: anyhow::anyhow!(error),
-            error_kind,
-        })
+        // TODO: either preserve the error code from postgres, or assign error codes to proxy errors.
+        self.write.write_error(msg, SQLSTATE_INTERNAL_ERROR);
+
+        self.flush()
+            .await
+            .unwrap_or_else(|e| tracing::debug!("write_message failed: {e}"));
+
+        ReportedError::new(error)
     }
 }
 

safekeeper/src/timelines_global_map.rs

@@ -44,6 +44,7 @@ struct GlobalTimelinesState {
     // on-demand timeline creation from recreating deleted timelines.  This is only soft-enforced, as
     // this map is dropped on restart.
     tombstones: HashMap<TenantTimelineId, Instant>,
+    tenant_tombstones: HashMap<TenantId, Instant>,
 
     conf: Arc<SafeKeeperConf>,
     broker_active_set: Arc<TimelinesSet>,
@@ -81,10 +82,25 @@ impl GlobalTimelinesState {
         }
     }
 
+    fn has_tombstone(&self, ttid: &TenantTimelineId) -> bool {
+        self.tombstones.contains_key(ttid) || self.tenant_tombstones.contains_key(&ttid.tenant_id)
+    }
+
+    /// Removes all blocking tombstones for the given timeline ID.
+    /// Returns `true` if there have been actual changes.
+    fn remove_tombstone(&mut self, ttid: &TenantTimelineId) -> bool {
+        self.tombstones.remove(ttid).is_some()
+            || self.tenant_tombstones.remove(&ttid.tenant_id).is_some()
+    }
+
     fn delete(&mut self, ttid: TenantTimelineId) {
         self.timelines.remove(&ttid);
         self.tombstones.insert(ttid, Instant::now());
     }
+
+    fn add_tenant_tombstone(&mut self, tenant_id: TenantId) {
+        self.tenant_tombstones.insert(tenant_id, Instant::now());
+    }
 }
 
 /// A struct used to manage access to the global timelines map.
@@ -99,6 +115,7 @@ impl GlobalTimelines {
             state: Mutex::new(GlobalTimelinesState {
                 timelines: HashMap::new(),
                 tombstones: HashMap::new(),
+                tenant_tombstones: HashMap::new(),
                 conf,
                 broker_active_set: Arc::new(TimelinesSet::default()),
                 global_rate_limiter: RateLimiter::new(1, 1),
@@ -245,7 +262,7 @@ impl GlobalTimelines {
                 return Ok(timeline);
             }
 
-            if state.tombstones.contains_key(&ttid) {
+            if state.has_tombstone(&ttid) {
                 anyhow::bail!("Timeline {ttid} is deleted, refusing to recreate");
             }
 
@@ -295,13 +312,14 @@ impl GlobalTimelines {
                 _ => {}
             }
             if check_tombstone {
-                if state.tombstones.contains_key(&ttid) {
+                if state.has_tombstone(&ttid) {
                     anyhow::bail!("timeline {ttid} is deleted, refusing to recreate");
                 }
             } else {
                 // We may be have been asked to load a timeline that was previously deleted (e.g. from `pull_timeline.rs`).  We trust
                 // that the human doing this manual intervention knows what they are doing, and remove its tombstone.
-                if state.tombstones.remove(&ttid).is_some() {
+                // It's also possible that we enter this when the tenant has been deleted, even if the timeline itself has never existed.
+                if state.remove_tombstone(&ttid) {
                     warn!("un-deleted timeline {ttid}");
                 }
             }
@@ -482,6 +500,7 @@ impl GlobalTimelines {
         let tli_res = {
             let state = self.state.lock().unwrap();
 
+            // Do NOT check tenant tombstones here: those were set earlier
             if state.tombstones.contains_key(ttid) {
                 // Presence of a tombstone guarantees that a previous deletion has completed and there is no work to do.
                 info!("Timeline {ttid} was already deleted");
@@ -557,6 +576,10 @@ impl GlobalTimelines {
         action: DeleteOrExclude,
     ) -> Result<HashMap<TenantTimelineId, TimelineDeleteResult>> {
         info!("deleting all timelines for tenant {}", tenant_id);
+
+        // Adding a tombstone before getting the timelines to prevent new timeline additions
+        self.state.lock().unwrap().add_tenant_tombstone(*tenant_id);
+
         let to_delete = self.get_all_for_tenant(*tenant_id);
 
         let mut err = None;
@@ -600,6 +623,9 @@ impl GlobalTimelines {
         state
             .tombstones
             .retain(|_, v| now.duration_since(*v) < *tombstone_ttl);
+        state
+            .tenant_tombstones
+            .retain(|_, v| now.duration_since(*v) < *tombstone_ttl);
     }
 }
 

safekeeper/tests/walproposer_sim/simulation.rs

@@ -87,6 +87,7 @@ impl WalProposer {
         let config = Config {
             ttid,
             safekeepers_list: addrs,
+            safekeeper_conninfo_options: String::new(),
             safekeeper_reconnect_timeout: 1000,
             safekeeper_connection_timeout: 5000,
             sync_safekeepers,

storage_controller/src/http.rs

@@ -482,6 +482,10 @@ async fn handle_tenant_timeline_delete(
         ForwardOutcome::NotForwarded(_req) => {}
     };
 
+    service
+        .maybe_delete_timeline_import(tenant_id, timeline_id)
+        .await?;
+
     // For timeline deletions, which both implement an "initially return 202, then 404 once
     // we're done" semantic, we wrap with a retry loop to expose a simpler API upstream.
     async fn deletion_wrapper<R, F>(service: Arc<Service>, f: F) -> Result<Response<Body>, ApiError>

storage_controller/src/metrics.rs

@@ -139,6 +139,14 @@ pub(crate) struct StorageControllerMetricGroup {
     /// HTTP request status counters for handled requests
     pub(crate) storage_controller_reconcile_long_running:
         measured::CounterVec<ReconcileLongRunningLabelGroupSet>,
+
+    /// Indicator of safekeeper reconciler queue depth, broken down by safekeeper, excluding ongoing reconciles.
+    pub(crate) storage_controller_safkeeper_reconciles_queued:
+        measured::GaugeVec<SafekeeperReconcilerLabelGroupSet>,
+
+    /// Indicator of completed safekeeper reconciles, broken down by safekeeper.
+    pub(crate) storage_controller_safkeeper_reconciles_complete:
+        measured::CounterVec<SafekeeperReconcilerLabelGroupSet>,
 }
 
 impl StorageControllerMetrics {
@@ -257,6 +265,17 @@ pub(crate) enum Method {
     Other,
 }
 
+#[derive(measured::LabelGroup, Clone)]
+#[label(set = SafekeeperReconcilerLabelGroupSet)]
+pub(crate) struct SafekeeperReconcilerLabelGroup<'a> {
+    #[label(dynamic_with = lasso::ThreadedRodeo, default)]
+    pub(crate) sk_az: &'a str,
+    #[label(dynamic_with = lasso::ThreadedRodeo, default)]
+    pub(crate) sk_node_id: &'a str,
+    #[label(dynamic_with = lasso::ThreadedRodeo, default)]
+    pub(crate) sk_hostname: &'a str,
+}
+
 impl From<hyper::Method> for Method {
     fn from(value: hyper::Method) -> Self {
         if value == hyper::Method::GET {

storage_controller/src/service.rs

@@ -99,8 +99,8 @@ use crate::tenant_shard::{
     ScheduleOptimization, ScheduleOptimizationAction, TenantShard,
 };
 use crate::timeline_import::{
-    ImportResult, ShardImportStatuses, TimelineImport, TimelineImportFinalizeError,
-    TimelineImportState, UpcallClient,
+    FinalizingImport, ImportResult, ShardImportStatuses, TimelineImport,
+    TimelineImportFinalizeError, TimelineImportState, UpcallClient,
 };
 
 const WAITER_FILL_DRAIN_POLL_TIMEOUT: Duration = Duration::from_millis(500);
@@ -232,6 +232,9 @@ struct ServiceState {
 
     /// Queue of tenants who are waiting for concurrency limits to permit them to reconcile
     delayed_reconcile_rx: tokio::sync::mpsc::Receiver<TenantShardId>,
+
+    /// Tracks ongoing timeline import finalization tasks
+    imports_finalizing: BTreeMap<(TenantId, TimelineId), FinalizingImport>,
 }
 
 /// Transform an error from a pageserver into an error to return to callers of a storage
@@ -308,6 +311,7 @@ impl ServiceState {
             scheduler,
             ongoing_operation: None,
             delayed_reconcile_rx,
+            imports_finalizing: Default::default(),
         }
     }
 
@@ -3823,6 +3827,13 @@ impl Service {
         .await;
         failpoint_support::sleep_millis_async!("tenant-create-timeline-shared-lock");
         let is_import = create_req.is_import();
+        let read_only = matches!(
+            create_req.mode,
+            models::TimelineCreateRequestMode::Branch {
+                read_only: true,
+                ..
+            }
+        );
 
         if is_import {
             // Ensure that there is no split on-going.
@@ -3895,13 +3906,13 @@ impl Service {
             }
 
             None
-        } else if safekeepers {
+        } else if safekeepers || read_only {
             // Note that for imported timelines, we do not create the timeline on the safekeepers
             // straight away. Instead, we do it once the import finalized such that we know what
             // start LSN to provide for the safekeepers. This is done in
             // [`Self::finalize_timeline_import`].
             let res = self
-                .tenant_timeline_create_safekeepers(tenant_id, &timeline_info)
+                .tenant_timeline_create_safekeepers(tenant_id, &timeline_info, read_only)
                 .instrument(tracing::info_span!("timeline_create_safekeepers", %tenant_id, timeline_id=%timeline_info.timeline_id))
                 .await?;
             Some(res)
@@ -3915,6 +3926,11 @@ impl Service {
         })
     }
 
+    #[instrument(skip_all, fields(
+        tenant_id=%req.tenant_shard_id.tenant_id,
+        shard_id=%req.tenant_shard_id.shard_slug(),
+        timeline_id=%req.timeline_id,
+    ))]
     pub(crate) async fn handle_timeline_shard_import_progress(
         self: &Arc<Self>,
         req: TimelineImportStatusRequest,
@@ -3964,6 +3980,11 @@ impl Service {
             })
     }
 
+    #[instrument(skip_all, fields(
+        tenant_id=%req.tenant_shard_id.tenant_id,
+        shard_id=%req.tenant_shard_id.shard_slug(),
+        timeline_id=%req.timeline_id,
+    ))]
     pub(crate) async fn handle_timeline_shard_import_progress_upcall(
         self: &Arc<Self>,
         req: PutTimelineImportStatusRequest,
@@ -4080,13 +4101,58 @@ impl Service {
     ///
     /// If this method gets pre-empted by shut down, it will be called again at start-up (on-going
     /// imports are stored in the database).
+    ///
+    /// # Cancel-Safety
+    /// Not cancel safe.
+    /// If the caller stops polling, the import will not be removed from
+    /// [`ServiceState::imports_finalizing`].
     #[instrument(skip_all, fields(
         tenant_id=%import.tenant_id,
         timeline_id=%import.timeline_id,
     ))]
+
     async fn finalize_timeline_import(
         self: &Arc<Self>,
         import: TimelineImport,
+    ) -> Result<(), TimelineImportFinalizeError> {
+        let tenant_timeline = (import.tenant_id, import.timeline_id);
+
+        let (_finalize_import_guard, cancel) = {
+            let mut locked = self.inner.write().unwrap();
+            let gate = Gate::default();
+            let cancel = CancellationToken::default();
+
+            let guard = gate.enter().unwrap();
+
+            locked.imports_finalizing.insert(
+                tenant_timeline,
+                FinalizingImport {
+                    gate,
+                    cancel: cancel.clone(),
+                },
+            );
+
+            (guard, cancel)
+        };
+
+        let res = tokio::select! {
+            res = self.finalize_timeline_import_impl(import) => {
+                res
+            },
+            _ = cancel.cancelled() => {
+                Err(TimelineImportFinalizeError::Cancelled)
+            }
+        };
+
+        let mut locked = self.inner.write().unwrap();
+        locked.imports_finalizing.remove(&tenant_timeline);
+
+        res
+    }
+
+    async fn finalize_timeline_import_impl(
+        self: &Arc<Self>,
+        import: TimelineImport,
     ) -> Result<(), TimelineImportFinalizeError> {
         tracing::info!("Finalizing timeline import");
 
@@ -4286,6 +4352,46 @@ impl Service {
         .await;
     }
 
+    /// Delete a timeline import if it exists
+    ///
+    /// Firstly, delete the entry from the database. Any updates
+    /// from pageservers after the update will fail with a 404, so the
+    /// import cannot progress into finalizing state if it's not there already.
+    /// Secondly, cancel the finalization if one is in progress.
+    pub(crate) async fn maybe_delete_timeline_import(
+        self: &Arc<Self>,
+        tenant_id: TenantId,
+        timeline_id: TimelineId,
+    ) -> Result<(), DatabaseError> {
+        let tenant_has_ongoing_import = {
+            let locked = self.inner.read().unwrap();
+            locked
+                .tenants
+                .range(TenantShardId::tenant_range(tenant_id))
+                .any(|(_tid, shard)| shard.importing == TimelineImportState::Importing)
+        };
+
+        if !tenant_has_ongoing_import {
+            return Ok(());
+        }
+
+        self.persistence
+            .delete_timeline_import(tenant_id, timeline_id)
+            .await?;
+
+        let maybe_finalizing = {
+            let mut locked = self.inner.write().unwrap();
+            locked.imports_finalizing.remove(&(tenant_id, timeline_id))
+        };
+
+        if let Some(finalizing) = maybe_finalizing {
+            finalizing.cancel.cancel();
+            finalizing.gate.close().await;
+        }
+
+        Ok(())
+    }
+
     pub(crate) async fn tenant_timeline_archival_config(
         &self,
         tenant_id: TenantId,
@@ -8521,8 +8627,9 @@ impl Service {
         Some(ShardCount(new_shard_count))
     }
 
-    /// Fetches the top tenant shards from every node, in descending order of
-    /// max logical size. Any node errors will be logged and ignored.
+    /// Fetches the top tenant shards from every available node, in descending order of
+    /// max logical size. Offline nodes are skipped, and any errors from available nodes
+    /// will be logged and ignored.
     async fn get_top_tenant_shards(
         &self,
         request: &TopTenantShardsRequest,
@@ -8533,6 +8640,7 @@ impl Service {
             .unwrap()
             .nodes
             .values()
+            .filter(|node| node.is_available())
             .cloned()
             .collect_vec();
 

storage_controller/src/service/safekeeper_reconciler.rs

@@ -20,7 +20,9 @@ use utils::{
 };
 
 use crate::{
-    persistence::SafekeeperTimelineOpKind, safekeeper::Safekeeper,
+    metrics::{METRICS_REGISTRY, SafekeeperReconcilerLabelGroup},
+    persistence::SafekeeperTimelineOpKind,
+    safekeeper::Safekeeper,
     safekeeper_client::SafekeeperClient,
 };
 
@@ -218,7 +220,26 @@ impl ReconcilerHandle {
     fn schedule_reconcile(&self, req: ScheduleRequest) {
         let (cancel, token_id) = self.new_token_slot(req.tenant_id, req.timeline_id);
         let hostname = req.safekeeper.skp.host.clone();
+        let sk_az = req.safekeeper.skp.availability_zone_id.clone();
+        let sk_node_id = req.safekeeper.get_id().to_string();
+
+        // We don't have direct access to the queue depth here, so increase it blindly by 1.
+        // We know that putting into the queue increases the queue depth. The receiver will
+        // update with the correct value once it processes the next item. To avoid races where we
+        // reduce before we increase, leaving the gauge with a 1 value for a long time, we
+        // increase it before putting into the queue.
+        let queued_gauge = &METRICS_REGISTRY
+            .metrics_group
+            .storage_controller_safkeeper_reconciles_queued;
+        let label_group = SafekeeperReconcilerLabelGroup {
+            sk_az: &sk_az,
+            sk_node_id: &sk_node_id,
+            sk_hostname: &hostname,
+        };
+        queued_gauge.inc(label_group.clone());
+
         if let Err(err) = self.tx.send((req, cancel, token_id)) {
+            queued_gauge.set(label_group, 0);
             tracing::info!("scheduling request onto {hostname} returned error: {err}");
         }
     }
@@ -283,6 +304,18 @@ impl SafekeeperReconciler {
                 continue;
             }
 
+            let queued_gauge = &METRICS_REGISTRY
+                .metrics_group
+                .storage_controller_safkeeper_reconciles_queued;
+            queued_gauge.set(
+                SafekeeperReconcilerLabelGroup {
+                    sk_az: &req.safekeeper.skp.availability_zone_id,
+                    sk_node_id: &req.safekeeper.get_id().to_string(),
+                    sk_hostname: &req.safekeeper.skp.host,
+                },
+                self.rx.len() as i64,
+            );
+
             tokio::task::spawn(async move {
                 let kind = req.kind;
                 let tenant_id = req.tenant_id;
@@ -511,6 +544,16 @@ impl SafekeeperReconcilerInner {
                             req.generation,
                         )
                         .await;
+
+                    let complete_counter = &METRICS_REGISTRY
+                        .metrics_group
+                        .storage_controller_safkeeper_reconciles_complete;
+                    complete_counter.inc(SafekeeperReconcilerLabelGroup {
+                        sk_az: &req.safekeeper.skp.availability_zone_id,
+                        sk_node_id: &req.safekeeper.get_id().to_string(),
+                        sk_hostname: &req.safekeeper.skp.host,
+                    });
+
                     if let Err(err) = res {
                         tracing::info!(
                             "couldn't remove reconciliation request onto {} from persistence: {err:?}",

storage_controller/src/service/safekeeper_service.rs

@@ -208,6 +208,7 @@ impl Service {
         self: &Arc<Self>,
         tenant_id: TenantId,
         timeline_info: &TimelineInfo,
+        read_only: bool,
     ) -> Result<SafekeepersInfo, ApiError> {
         let timeline_id = timeline_info.timeline_id;
         let pg_version = timeline_info.pg_version * 10000;
@@ -220,7 +221,11 @@ impl Service {
         let start_lsn = timeline_info.last_record_lsn;
 
         // Choose initial set of safekeepers respecting affinity
-        let sks = self.safekeepers_for_new_timeline().await?;
+        let sks = if !read_only {
+            self.safekeepers_for_new_timeline().await?
+        } else {
+            Vec::new()
+        };
         let sks_persistence = sks.iter().map(|sk| sk.id.0 as i64).collect::<Vec<_>>();
         // Add timeline to db
         let mut timeline_persist = TimelinePersistence {
@@ -253,6 +258,16 @@ impl Service {
                 )));
             }
         }
+        let ret = SafekeepersInfo {
+            generation: timeline_persist.generation as u32,
+            safekeepers: sks.clone(),
+            tenant_id,
+            timeline_id,
+        };
+        if read_only {
+            return Ok(ret);
+        }
+
         // Create the timeline on a quorum of safekeepers
         let remaining = self
             .tenant_timeline_create_safekeepers_quorum(
@@ -316,12 +331,7 @@ impl Service {
             }
         }
 
-        Ok(SafekeepersInfo {
-            generation: timeline_persist.generation as u32,
-            safekeepers: sks,
-            tenant_id,
-            timeline_id,
-        })
+        Ok(ret)
     }
 
     pub(crate) async fn tenant_timeline_create_safekeepers_until_success(
@@ -336,8 +346,10 @@ impl Service {
                 return Err(TimelineImportFinalizeError::ShuttingDown);
             }
 
+            // This function is only used in non-read-only scenarios
+            let read_only = false;
             let res = self
-                .tenant_timeline_create_safekeepers(tenant_id, &timeline_info)
+                .tenant_timeline_create_safekeepers(tenant_id, &timeline_info, read_only)
                 .await;
 
             match res {
@@ -410,6 +422,18 @@ impl Service {
             .chain(tl.sk_set.iter())
             .collect::<HashSet<_>>();
 
+        // The timeline has no safekeepers: we need to delete it from the db manually,
+        // as no safekeeper reconciler will get to it
+        if all_sks.is_empty() {
+            if let Err(err) = self
+                .persistence
+                .delete_timeline(tenant_id, timeline_id)
+                .await
+            {
+                tracing::warn!(%tenant_id, %timeline_id, "couldn't delete timeline from db: {err}");
+            }
+        }
+
         // Schedule reconciliations
         for &sk_id in all_sks.iter() {
             let pending_op = TimelinePendingOpPersistence {

storage_controller/src/timeline_import.rs

@@ -7,6 +7,7 @@ use serde::{Deserialize, Serialize};
 
 use pageserver_api::models::{ShardImportProgress, ShardImportStatus};
 use tokio_util::sync::CancellationToken;
+use utils::sync::gate::Gate;
 use utils::{
     id::{TenantId, TimelineId},
     shard::ShardIndex,
@@ -55,6 +56,8 @@ pub(crate) enum TimelineImportUpdateFollowUp {
 pub(crate) enum TimelineImportFinalizeError {
     #[error("Shut down interrupted import finalize")]
     ShuttingDown,
+    #[error("Import finalization was cancelled")]
+    Cancelled,
     #[error("Mismatched shard detected during import finalize: {0}")]
     MismatchedShards(ShardIndex),
 }
@@ -164,6 +167,11 @@ impl TimelineImport {
     }
 }
 
+pub(crate) struct FinalizingImport {
+    pub(crate) gate: Gate,
+    pub(crate) cancel: CancellationToken,
+}
+
 pub(crate) type ImportResult = Result<(), String>;
 
 pub(crate) struct UpcallClient {

test_runner/fixtures/fast_import.py

@@ -1,3 +1,4 @@
+import json
 import os
 import shutil
 import subprocess
@@ -11,6 +12,7 @@ from _pytest.config import Config
 
 from fixtures.log_helper import log
 from fixtures.neon_cli import AbstractNeonCli
+from fixtures.neon_fixtures import Endpoint, VanillaPostgres
 from fixtures.pg_version import PgVersion
 from fixtures.remote_storage import MockS3Server
 
@@ -161,3 +163,57 @@ def fast_import(
             f.write(fi.cmd.stderr)
 
         log.info("Written logs to %s", test_output_dir)
+
+
+def mock_import_bucket(vanilla_pg: VanillaPostgres, path: Path):
+    """
+    Mock the import S3 bucket into a local directory for a provided vanilla PG instance.
+    """
+    assert not vanilla_pg.is_running()
+
+    path.mkdir()
+    # what cplane writes before scheduling fast_import
+    specpath = path / "spec.json"
+    specpath.write_text(json.dumps({"branch_id": "somebranch", "project_id": "someproject"}))
+    # what fast_import writes
+    vanilla_pg.pgdatadir.rename(path / "pgdata")
+    statusdir = path / "status"
+    statusdir.mkdir()
+    (statusdir / "pgdata").write_text(json.dumps({"done": True}))
+    (statusdir / "fast_import").write_text(json.dumps({"command": "pgdata", "done": True}))
+
+
+def populate_vanilla_pg(vanilla_pg: VanillaPostgres, target_relblock_size: int) -> int:
+    assert vanilla_pg.is_running()
+
+    vanilla_pg.safe_psql("create user cloud_admin with password 'postgres' superuser")
+    # fillfactor so we don't need to produce that much data
+    # 900 byte per row is > 10% => 1 row per page
+    vanilla_pg.safe_psql("""create table t (data char(900)) with (fillfactor = 10)""")
+
+    nrows = 0
+    while True:
+        relblock_size = vanilla_pg.safe_psql_scalar("select pg_relation_size('t')")
+        log.info(
+            f"relblock size: {relblock_size / 8192} pages (target: {target_relblock_size // 8192}) pages"
+        )
+        if relblock_size >= target_relblock_size:
+            break
+        addrows = int((target_relblock_size - relblock_size) // 8192)
+        assert addrows >= 1, "forward progress"
+        vanilla_pg.safe_psql(
+            f"insert into t select generate_series({nrows + 1}, {nrows + addrows})"
+        )
+        nrows += addrows
+
+    return nrows
+
+
+def validate_import_from_vanilla_pg(endpoint: Endpoint, nrows: int):
+    assert endpoint.safe_psql_many(
+        [
+            "set effective_io_concurrency=32;",
+            "SET statement_timeout='300s';",
+            "select count(*), sum(data::bigint)::bigint from t",
+        ]
+    ) == [[], [], [(nrows, nrows * (nrows + 1) // 2)]]

test_runner/fixtures/neon_fixtures.py

@@ -404,6 +404,29 @@ class PageserverTracingConfig:
         return ("tracing", value)
 
 
+@dataclass
+class PageserverImportConfig:
+    import_job_concurrency: int
+    import_job_soft_size_limit: int
+    import_job_checkpoint_threshold: int
+
+    @staticmethod
+    def default() -> PageserverImportConfig:
+        return PageserverImportConfig(
+            import_job_concurrency=4,
+            import_job_soft_size_limit=512 * 1024,
+            import_job_checkpoint_threshold=4,
+        )
+
+    def to_config_key_value(self) -> tuple[str, dict[str, Any]]:
+        value = {
+            "import_job_concurrency": self.import_job_concurrency,
+            "import_job_soft_size_limit": self.import_job_soft_size_limit,
+            "import_job_checkpoint_threshold": self.import_job_checkpoint_threshold,
+        }
+        return ("timeline_import_config", value)
+
+
 class NeonEnvBuilder:
     """
     Builder object to create a Neon runtime environment
@@ -454,6 +477,7 @@ class NeonEnvBuilder:
         pageserver_wal_receiver_protocol: PageserverWalReceiverProtocol | None = None,
         pageserver_get_vectored_concurrent_io: str | None = None,
         pageserver_tracing_config: PageserverTracingConfig | None = None,
+        pageserver_import_config: PageserverImportConfig | None = None,
     ):
         self.repo_dir = repo_dir
         self.rust_log_override = rust_log_override
@@ -511,6 +535,7 @@ class NeonEnvBuilder:
         )
 
         self.pageserver_tracing_config = pageserver_tracing_config
+        self.pageserver_import_config = pageserver_import_config
 
         self.pageserver_default_tenant_config_compaction_algorithm: dict[str, Any] | None = (
             pageserver_default_tenant_config_compaction_algorithm
@@ -1179,6 +1204,10 @@ class NeonEnv:
         self.pageserver_wal_receiver_protocol = config.pageserver_wal_receiver_protocol
         self.pageserver_get_vectored_concurrent_io = config.pageserver_get_vectored_concurrent_io
         self.pageserver_tracing_config = config.pageserver_tracing_config
+        if config.pageserver_import_config is None:
+            self.pageserver_import_config = PageserverImportConfig.default()
+        else:
+            self.pageserver_import_config = config.pageserver_import_config
 
         # Create the neon_local's `NeonLocalInitConf`
         cfg: dict[str, Any] = {
@@ -1258,12 +1287,6 @@ class NeonEnv:
                 "no_sync": True,
                 # Look for gaps in WAL received from safekeepeers
                 "validate_wal_contiguity": True,
-                # TODO(vlad): make these configurable through the builder
-                "timeline_import_config": {
-                    "import_job_concurrency": 4,
-                    "import_job_soft_size_limit": 512 * 1024,
-                    "import_job_checkpoint_threshold": 4,
-                },
             }
 
             # Batching (https://github.com/neondatabase/neon/issues/9377):
@@ -1325,6 +1348,12 @@ class NeonEnv:
 
                 ps_cfg[key] = value
 
+            if self.pageserver_import_config is not None:
+                key, value = self.pageserver_import_config.to_config_key_value()
+
+                if key not in ps_cfg:
+                    ps_cfg[key] = value
+
             # Create a corresponding NeonPageserver object
             ps = NeonPageserver(
                 self, ps_id, port=pageserver_port, az_id=ps_cfg["availability_zone"]
@@ -2308,6 +2337,22 @@ class NeonStorageController(MetricsGetter, LogUtils):
             headers=self.headers(TokenScope.ADMIN),
         )
 
+    def import_status(
+        self, tenant_shard_id: TenantShardId, timeline_id: TimelineId, generation: int
+    ):
+        payload = {
+            "tenant_shard_id": str(tenant_shard_id),
+            "timeline_id": str(timeline_id),
+            "generation": generation,
+        }
+
+        self.request(
+            "GET",
+            f"{self.api}/upcall/v1/timeline_import_status",
+            headers=self.headers(TokenScope.GENERATIONS_API),
+            json=payload,
+        )
+
     def reconcile_all(self):
         r = self.request(
             "POST",
@@ -2784,6 +2829,11 @@ class NeonPageserver(PgProtocol, LogUtils):
         if self.running:
             self.http_client().configure_failpoints([(name, action)])
 
+    def clear_persistent_failpoint(self, name: str):
+        del self._persistent_failpoints[name]
+        if self.running:
+            self.http_client().configure_failpoints([(name, "off")])
+
     def timeline_dir(
         self,
         tenant_shard_id: TenantId | TenantShardId,

test_runner/fixtures/pageserver/http.py

@@ -675,7 +675,7 @@ class PageserverHttpClient(requests.Session, MetricsGetter):
 
     def timeline_delete(
         self, tenant_id: TenantId | TenantShardId, timeline_id: TimelineId, **kwargs
-    ):
+    ) -> int:
         """
         Note that deletion is not instant, it is scheduled and performed mostly in the background.
         So if you need to wait for it to complete use `timeline_delete_wait_completed`.
@@ -688,6 +688,8 @@ class PageserverHttpClient(requests.Session, MetricsGetter):
         res_json = res.json()
         assert res_json is None
 
+        return res.status_code
+
     def timeline_gc(
         self,
         tenant_id: TenantId | TenantShardId,

test_runner/regress/test_disk_usage_eviction.py

@@ -1,31 +1,41 @@
 from __future__ import annotations
 
 import enum
+import json
 import time
 from collections import Counter
 from dataclasses import dataclass
 from enum import StrEnum
+from threading import Event
 from typing import TYPE_CHECKING
 
 import pytest
 from fixtures.common_types import Lsn, TenantId, TimelineId
+from fixtures.fast_import import mock_import_bucket, populate_vanilla_pg
 from fixtures.log_helper import log
 from fixtures.neon_fixtures import (
     NeonEnv,
     NeonEnvBuilder,
     NeonPageserver,
     PgBin,
+    VanillaPostgres,
     wait_for_last_flush_lsn,
 )
+from fixtures.pageserver.http import (
+    ImportPgdataIdemptencyKey,
+)
 from fixtures.pageserver.utils import wait_for_upload_queue_empty
 from fixtures.remote_storage import RemoteStorageKind
-from fixtures.utils import human_bytes, wait_until
+from fixtures.utils import human_bytes, run_only_on_default_postgres, wait_until
+from werkzeug.wrappers.response import Response
 
 if TYPE_CHECKING:
     from collections.abc import Iterable
     from typing import Any
 
     from fixtures.pageserver.http import PageserverHttpClient
+    from pytest_httpserver import HTTPServer
+    from werkzeug.wrappers.request import Request
 
 
 GLOBAL_LRU_LOG_LINE = "tenant_min_resident_size-respecting LRU would not relieve pressure, evicting more following global LRU policy"
@@ -164,6 +174,7 @@ class EvictionEnv:
         min_avail_bytes,
         mock_behavior,
         eviction_order: EvictionOrder,
+        wait_logical_size: bool = True,
     ):
         """
         Starts pageserver up with mocked statvfs setup. The startup is
@@ -201,11 +212,12 @@ class EvictionEnv:
         pageserver.start()
 
         # we now do initial logical size calculation on startup, which on debug builds can fight with disk usage based eviction
-        for tenant_id, timeline_id in self.timelines:
-            tenant_ps = self.neon_env.get_tenant_pageserver(tenant_id)
-            # Pageserver may be none if we are currently not attached anywhere, e.g. during secondary eviction test
-            if tenant_ps is not None:
-                tenant_ps.http_client().timeline_wait_logical_size(tenant_id, timeline_id)
+        if wait_logical_size:
+            for tenant_id, timeline_id in self.timelines:
+                tenant_ps = self.neon_env.get_tenant_pageserver(tenant_id)
+                # Pageserver may be none if we are currently not attached anywhere, e.g. during secondary eviction test
+                if tenant_ps is not None:
+                    tenant_ps.http_client().timeline_wait_logical_size(tenant_id, timeline_id)
 
         def statvfs_called():
             pageserver.assert_log_contains(".*running mocked statvfs.*")
@@ -882,3 +894,121 @@ def test_secondary_mode_eviction(eviction_env_ha: EvictionEnv):
     assert total_size - post_eviction_total_size >= evict_bytes, (
         "we requested at least evict_bytes worth of free space"
     )
+
+
+@run_only_on_default_postgres(reason="PG version is irrelevant here")
+def test_import_timeline_disk_pressure_eviction(
+    neon_env_builder: NeonEnvBuilder,
+    vanilla_pg: VanillaPostgres,
+    make_httpserver: HTTPServer,
+    pg_bin: PgBin,
+):
+    """
+    TODO
+    """
+    # Set up mock control plane HTTP server to listen for import completions
+    import_completion_signaled = Event()
+
+    def handler(request: Request) -> Response:
+        log.info(f"control plane /import_complete request: {request.json}")
+        import_completion_signaled.set()
+        return Response(json.dumps({}), status=200)
+
+    cplane_mgmt_api_server = make_httpserver
+    cplane_mgmt_api_server.expect_request(
+        "/storage/api/v1/import_complete", method="PUT"
+    ).respond_with_handler(handler)
+
+    # Plug the cplane mock in
+    neon_env_builder.control_plane_hooks_api = (
+        f"http://{cplane_mgmt_api_server.host}:{cplane_mgmt_api_server.port}/storage/api/v1/"
+    )
+
+    # The import will specifiy a local filesystem path mocking remote storage
+    neon_env_builder.enable_pageserver_remote_storage(RemoteStorageKind.LOCAL_FS)
+
+    vanilla_pg.start()
+    target_relblock_size = 1024 * 1024 * 128
+    populate_vanilla_pg(vanilla_pg, target_relblock_size)
+    vanilla_pg.stop()
+
+    env = neon_env_builder.init_configs()
+    env.start()
+
+    importbucket_path = neon_env_builder.repo_dir / "test_import_completion_bucket"
+    mock_import_bucket(vanilla_pg, importbucket_path)
+
+    tenant_id = TenantId.generate()
+    timeline_id = TimelineId.generate()
+    idempotency = ImportPgdataIdemptencyKey.random()
+
+    eviction_env = EvictionEnv(
+        timelines=[(tenant_id, timeline_id)],
+        neon_env=env,
+        pageserver_http=env.pageserver.http_client(),
+        layer_size=5 * 1024 * 1024,  # Doesn't apply here
+        pg_bin=pg_bin,  # Not used here
+        pgbench_init_lsns={},  # Not used here
+    )
+
+    # Pause before delivering the final notification to storcon.
+    # This keeps the import in progress.
+    failpoint_name = "import-timeline-pre-success-notify-pausable"
+    env.pageserver.add_persistent_failpoint(failpoint_name, "pause")
+
+    env.storage_controller.tenant_create(tenant_id)
+    env.storage_controller.timeline_create(
+        tenant_id,
+        {
+            "new_timeline_id": str(timeline_id),
+            "import_pgdata": {
+                "idempotency_key": str(idempotency),
+                "location": {"LocalFs": {"path": str(importbucket_path.absolute())}},
+            },
+        },
+    )
+
+    def hit_failpoint():
+        log.info("Checking log for pattern...")
+        try:
+            assert env.pageserver.log_contains(f".*at failpoint {failpoint_name}.*")
+        except Exception:
+            log.exception("Failed to find pattern in log")
+            raise
+
+    wait_until(hit_failpoint)
+    assert not import_completion_signaled.is_set()
+
+    env.pageserver.stop()
+
+    total_size, _, _ = eviction_env.timelines_du(env.pageserver)
+    blocksize = 512
+    total_blocks = (total_size + (blocksize - 1)) // blocksize
+
+    eviction_env.pageserver_start_with_disk_usage_eviction(
+        env.pageserver,
+        period="1s",
+        max_usage_pct=33,
+        min_avail_bytes=0,
+        mock_behavior={
+            "type": "Success",
+            "blocksize": blocksize,
+            "total_blocks": total_blocks,
+            # Only count layer files towards used bytes in the mock_statvfs.
+            # This avoids accounting for metadata files & tenant conf in the tests.
+            "name_filter": ".*__.*",
+        },
+        eviction_order=EvictionOrder.RELATIVE_ORDER_SPARE,
+        wait_logical_size=False,
+    )
+
+    wait_until(lambda: env.pageserver.assert_log_contains(".*disk usage pressure relieved"))
+
+    env.pageserver.clear_persistent_failpoint(failpoint_name)
+
+    def cplane_notified():
+        assert import_completion_signaled.is_set()
+
+    wait_until(cplane_notified)
+
+    env.pageserver.allowed_errors.append(r".* running disk usage based eviction due to pressure.*")

test_runner/regress/test_import_pgdata.py

@@ -1,7 +1,10 @@
 import base64
+import concurrent.futures
 import json
+import random
+import threading
 import time
-from enum import Enum
+from enum import Enum, StrEnum
 from pathlib import Path
 from threading import Event
 
@@ -9,9 +12,22 @@ import psycopg2
 import psycopg2.errors
 import pytest
 from fixtures.common_types import Lsn, TenantId, TenantShardId, TimelineId
-from fixtures.fast_import import FastImport
+from fixtures.fast_import import (
+    FastImport,
+    mock_import_bucket,
+    populate_vanilla_pg,
+    validate_import_from_vanilla_pg,
+)
 from fixtures.log_helper import log
-from fixtures.neon_fixtures import NeonEnvBuilder, PgBin, PgProtocol, VanillaPostgres
+from fixtures.neon_fixtures import (
+    NeonEnvBuilder,
+    PageserverImportConfig,
+    PgBin,
+    PgProtocol,
+    StorageControllerApiException,
+    StorageControllerMigrationConfig,
+    VanillaPostgres,
+)
 from fixtures.pageserver.http import (
     ImportPgdataIdemptencyKey,
 )
@@ -49,24 +65,6 @@ smoke_params = [
 ]
 
 
-def mock_import_bucket(vanilla_pg: VanillaPostgres, path: Path):
-    """
-    Mock the import S3 bucket into a local directory for a provided vanilla PG instance.
-    """
-    assert not vanilla_pg.is_running()
-
-    path.mkdir()
-    # what cplane writes before scheduling fast_import
-    specpath = path / "spec.json"
-    specpath.write_text(json.dumps({"branch_id": "somebranch", "project_id": "someproject"}))
-    # what fast_import writes
-    vanilla_pg.pgdatadir.rename(path / "pgdata")
-    statusdir = path / "status"
-    statusdir.mkdir()
-    (statusdir / "pgdata").write_text(json.dumps({"done": True}))
-    (statusdir / "fast_import").write_text(json.dumps({"command": "pgdata", "done": True}))
-
-
 @skip_in_debug_build("MULTIPLE_RELATION_SEGMENTS has non trivial amount of data")
 @pytest.mark.parametrize("shard_count,stripe_size,rel_block_size", smoke_params)
 def test_pgdata_import_smoke(
@@ -121,10 +119,6 @@ def test_pgdata_import_smoke(
     # Put data in vanilla pg
     #
 
-    vanilla_pg.start()
-    vanilla_pg.safe_psql("create user cloud_admin with password 'postgres' superuser")
-
-    log.info("create relblock data")
     if rel_block_size == RelBlockSize.ONE_STRIPE_SIZE:
         target_relblock_size = stripe_size * 8192
     elif rel_block_size == RelBlockSize.TWO_STRPES_PER_SHARD:
@@ -135,45 +129,8 @@ def test_pgdata_import_smoke(
     else:
         raise ValueError
 
-    # fillfactor so we don't need to produce that much data
-    # 900 byte per row is > 10% => 1 row per page
-    vanilla_pg.safe_psql("""create table t (data char(900)) with (fillfactor = 10)""")
-
-    nrows = 0
-    while True:
-        relblock_size = vanilla_pg.safe_psql_scalar("select pg_relation_size('t')")
-        log.info(
-            f"relblock size: {relblock_size / 8192} pages (target: {target_relblock_size // 8192}) pages"
-        )
-        if relblock_size >= target_relblock_size:
-            break
-        addrows = int((target_relblock_size - relblock_size) // 8192)
-        assert addrows >= 1, "forward progress"
-        vanilla_pg.safe_psql(
-            f"insert into t select generate_series({nrows + 1}, {nrows + addrows})"
-        )
-        nrows += addrows
-    expect_nrows = nrows
-    expect_sum = (
-        (nrows) * (nrows + 1) // 2
-    )  # https://stackoverflow.com/questions/43901484/sum-of-the-integers-from-1-to-n
-
-    def validate_vanilla_equivalence(ep):
-        # TODO: would be nicer to just compare pgdump
-
-        # Enable IO concurrency for batching on large sequential scan, to avoid making
-        # this test unnecessarily onerous on CPU. Especially on debug mode, it's still
-        # pretty onerous though, so increase statement_timeout to avoid timeouts.
-        assert ep.safe_psql_many(
-            [
-                "set effective_io_concurrency=32;",
-                "SET statement_timeout='300s';",
-                "select count(*), sum(data::bigint)::bigint from t",
-            ]
-        ) == [[], [], [(expect_nrows, expect_sum)]]
-
-    validate_vanilla_equivalence(vanilla_pg)
-
+    vanilla_pg.start()
+    rows_inserted = populate_vanilla_pg(vanilla_pg, target_relblock_size)
     vanilla_pg.stop()
 
     #
@@ -264,14 +221,14 @@ def test_pgdata_import_smoke(
         config_lines=ep_config,
     )
 
-    validate_vanilla_equivalence(ro_endpoint)
+    validate_import_from_vanilla_pg(ro_endpoint, rows_inserted)
 
     # ensure the import survives restarts
     ro_endpoint.stop()
     env.pageserver.stop(immediate=True)
     env.pageserver.start()
     ro_endpoint.start()
-    validate_vanilla_equivalence(ro_endpoint)
+    validate_import_from_vanilla_pg(ro_endpoint, rows_inserted)
 
     #
     # validate the layer files in each shard only have the shard-specific data
@@ -311,7 +268,7 @@ def test_pgdata_import_smoke(
     child_workload = workload.branch(timeline_id=child_timeline_id, branch_name="br-tip")
     child_workload.validate()
 
-    validate_vanilla_equivalence(child_workload.endpoint())
+    validate_import_from_vanilla_pg(child_workload.endpoint(), rows_inserted)
 
     # ... at the initdb lsn
     _ = env.create_branch(
@@ -326,10 +283,21 @@ def test_pgdata_import_smoke(
         tenant_id=tenant_id,
         config_lines=ep_config,
     )
-    validate_vanilla_equivalence(br_initdb_endpoint)
+    validate_import_from_vanilla_pg(br_initdb_endpoint, rows_inserted)
     with pytest.raises(psycopg2.errors.UndefinedTable):
         br_initdb_endpoint.safe_psql(f"select * from {workload.table}")
 
+    # The storage controller might be overly eager and attempt to finalize
+    # the import before the task got a chance to exit.
+    env.storage_controller.allowed_errors.extend(
+        [
+            ".*Call to node.*management API.*failed.*Import task still running.*",
+        ]
+    )
+
+    for ps in env.pageservers:
+        ps.allowed_errors.extend([".*Error processing HTTP request.*Import task not done yet.*"])
+
 
 @run_only_on_default_postgres(reason="PG version is irrelevant here")
 def test_import_completion_on_restart(
@@ -413,8 +381,12 @@ def test_import_completion_on_restart(
 
 
 @run_only_on_default_postgres(reason="PG version is irrelevant here")
-def test_import_respects_tenant_shutdown(
-    neon_env_builder: NeonEnvBuilder, vanilla_pg: VanillaPostgres, make_httpserver: HTTPServer
+@pytest.mark.parametrize("action", ["restart", "delete"])
+def test_import_respects_timeline_lifecycle(
+    neon_env_builder: NeonEnvBuilder,
+    vanilla_pg: VanillaPostgres,
+    make_httpserver: HTTPServer,
+    action: str,
 ):
     """
     Validate that importing timelines respect the usual timeline life cycle:
@@ -482,16 +454,276 @@ def test_import_respects_tenant_shutdown(
     wait_until(hit_failpoint)
     assert not import_completion_signaled.is_set()
 
-    # Restart the pageserver while an import job is in progress.
-    # This clears the failpoint and we expect that the import starts up afresh
-    # after the restart and eventually completes.
-    env.pageserver.stop()
-    env.pageserver.start()
+    if action == "restart":
+        # Restart the pageserver while an import job is in progress.
+        # This clears the failpoint and we expect that the import starts up afresh
+        # after the restart and eventually completes.
+        env.pageserver.stop()
+        env.pageserver.start()
 
-    def cplane_notified():
-        assert import_completion_signaled.is_set()
+        def cplane_notified():
+            assert import_completion_signaled.is_set()
 
-    wait_until(cplane_notified)
+        wait_until(cplane_notified)
+    elif action == "delete":
+        status = env.storage_controller.pageserver_api().timeline_delete(tenant_id, timeline_id)
+        assert status == 200
+
+        timeline_path = env.pageserver.timeline_dir(tenant_id, timeline_id)
+        assert not timeline_path.exists(), "Timeline dir exists after deletion"
+
+        shard_zero = TenantShardId(tenant_id, 0, 0)
+        location = env.storage_controller.inspect(shard_zero)
+        assert location is not None
+        generation = location[0]
+
+        with pytest.raises(StorageControllerApiException, match="not found"):
+            env.storage_controller.import_status(shard_zero, timeline_id, generation)
+    else:
+        raise RuntimeError(f"{action} param not recognized")
+
+    # The storage controller might be overly eager and attempt to finalize
+    # the import before the task got a chance to exit.
+    env.storage_controller.allowed_errors.extend(
+        [
+            ".*Call to node.*management API.*failed.*Import task still running.*",
+        ]
+    )
+
+    for ps in env.pageservers:
+        ps.allowed_errors.extend([".*Error processing HTTP request.*Import task not done yet.*"])
+
+
+@skip_in_debug_build("Validation query takes too long in debug builds")
+def test_import_chaos(
+    neon_env_builder: NeonEnvBuilder, vanilla_pg: VanillaPostgres, make_httpserver: HTTPServer
+):
+    """
+    Perform a timeline import while injecting chaos in the environment.
+    We expect that the import completes eventually, that it passes validation and
+    the resulting timeline can be written to.
+    """
+    TARGET_RELBOCK_SIZE = 512 * 1024 * 1024  # 512 MiB
+    ALLOWED_IMPORT_RUNTIME = 90  # seconds
+    SHARD_COUNT = 4
+
+    neon_env_builder.num_pageservers = SHARD_COUNT
+    neon_env_builder.pageserver_import_config = PageserverImportConfig(
+        import_job_concurrency=1,
+        import_job_soft_size_limit=64 * 1024,
+        import_job_checkpoint_threshold=4,
+    )
+
+    # Set up mock control plane HTTP server to listen for import completions
+    import_completion_signaled = Event()
+    # There's some Python magic at play here. A list can be updated from the
+    # handler thread, but an optional cannot. Hence, use a list with one element.
+    import_error = []
+
+    def handler(request: Request) -> Response:
+        assert request.json is not None
+
+        body = request.json
+        if "error" in body:
+            if body["error"]:
+                import_error.append(body["error"])
+
+        log.info(f"control plane /import_complete request: {request.json}")
+        import_completion_signaled.set()
+        return Response(json.dumps({}), status=200)
+
+    cplane_mgmt_api_server = make_httpserver
+    cplane_mgmt_api_server.expect_request(
+        "/storage/api/v1/import_complete", method="PUT"
+    ).respond_with_handler(handler)
+
+    # Plug the cplane mock in
+    neon_env_builder.control_plane_hooks_api = (
+        f"http://{cplane_mgmt_api_server.host}:{cplane_mgmt_api_server.port}/storage/api/v1/"
+    )
+
+    # The import will specifiy a local filesystem path mocking remote storage
+    neon_env_builder.enable_pageserver_remote_storage(RemoteStorageKind.LOCAL_FS)
+
+    vanilla_pg.start()
+
+    inserted_rows = populate_vanilla_pg(vanilla_pg, TARGET_RELBOCK_SIZE)
+
+    vanilla_pg.stop()
+
+    env = neon_env_builder.init_configs()
+    env.start()
+
+    # Pause after every import task to extend the test runtime and allow
+    # for more chaos injection.
+    for ps in env.pageservers:
+        ps.add_persistent_failpoint("import-task-complete-pausable", "sleep(5)")
+
+    env.storage_controller.allowed_errors.extend(
+        [
+            # The shard might have moved or the pageserver hosting the shard restarted
+            ".*Call to node.*management API.*failed.*",
+            # Migrations have their time outs set to 0
+            ".*Timed out after.*downloading layers.*",
+            ".*Failed to prepare by downloading layers.*",
+            # The test may kill the storage controller or pageservers
+            ".*request was dropped before completing.*",
+        ]
+    )
+    for ps in env.pageservers:
+        ps.allowed_errors.extend(
+            [
+                # We might re-write a layer in a different generation if the import
+                # needs to redo some of the progress since not each job is checkpointed.
+                ".*was unlinked but was not dangling.*",
+                # The test may kill the storage controller or pageservers
+                ".*request was dropped before completing.*",
+                # Test can SIGTERM pageserver while it is downloading
+                ".*removing local file.*temp_download.*",
+                ".*Failed to flush heatmap.*",
+                # Test can SIGTERM the storage controller while pageserver
+                # is attempting to upcall.
+                ".*storage controller upcall failed.*timeline_import_status.*",
+                # TODO(vlad): TenantManager::reset_tenant returns a blanked anyhow error.
+                # It should return ResourceUnavailable or something that doesn't error log.
+                ".*activate_post_import.*InternalServerError.*tenant map is shutting down.*",
+                # TODO(vlad): How can this happen?
+                ".*Failed to download a remote file: deserialize index part file.*",
+                ".*Cancelled request finished with an error.*",
+            ]
+        )
+
+    importbucket_path = neon_env_builder.repo_dir / "test_import_chaos_bucket"
+    mock_import_bucket(vanilla_pg, importbucket_path)
+
+    tenant_id = TenantId.generate()
+    timeline_id = TimelineId.generate()
+    idempotency = ImportPgdataIdemptencyKey.random()
+
+    env.storage_controller.tenant_create(
+        tenant_id, shard_count=SHARD_COUNT, placement_policy={"Attached": 1}
+    )
+    env.storage_controller.reconcile_until_idle()
+
+    env.storage_controller.timeline_create(
+        tenant_id,
+        {
+            "new_timeline_id": str(timeline_id),
+            "import_pgdata": {
+                "idempotency_key": str(idempotency),
+                "location": {"LocalFs": {"path": str(importbucket_path.absolute())}},
+            },
+        },
+    )
+
+    def chaos(stop_chaos: threading.Event):
+        class ChaosType(StrEnum):
+            MIGRATE_SHARD = "migrate_shard"
+            RESTART_IMMEDIATE = "restart_immediate"
+            RESTART = "restart"
+            STORCON_RESTART_IMMEDIATE = "storcon_restart_immediate"
+
+        while not stop_chaos.is_set():
+            chaos_type = random.choices(
+                population=[
+                    ChaosType.MIGRATE_SHARD,
+                    ChaosType.RESTART,
+                    ChaosType.RESTART_IMMEDIATE,
+                    ChaosType.STORCON_RESTART_IMMEDIATE,
+                ],
+                weights=[0.25, 0.25, 0.25, 0.25],
+                k=1,
+            )[0]
+
+            try:
+                if chaos_type == ChaosType.MIGRATE_SHARD:
+                    target_shard_number = random.randint(0, SHARD_COUNT - 1)
+                    target_shard = TenantShardId(tenant_id, target_shard_number, SHARD_COUNT)
+
+                    placements = env.storage_controller.get_tenants_placement()
+                    log.info(f"{placements=}")
+                    target_ps = placements[str(target_shard)]["intent"]["attached"]
+                    if len(placements[str(target_shard)]["intent"]["secondary"]) == 0:
+                        dest_ps = None
+                    else:
+                        dest_ps = placements[str(target_shard)]["intent"]["secondary"][0]
+
+                    if target_ps is None or dest_ps is None:
+                        continue
+
+                    config = StorageControllerMigrationConfig(
+                        secondary_warmup_timeout="0s",
+                        secondary_download_request_timeout="0s",
+                        prewarm=False,
+                    )
+                    env.storage_controller.tenant_shard_migrate(target_shard, dest_ps, config)
+
+                    log.info(
+                        f"CHAOS: Migrating shard {target_shard} from pageserver {target_ps} to {dest_ps}"
+                    )
+                elif chaos_type == ChaosType.RESTART_IMMEDIATE:
+                    target_ps = random.choice(env.pageservers)
+                    log.info(f"CHAOS: Immediate restart of pageserver {target_ps.id}")
+                    target_ps.stop(immediate=True)
+                    target_ps.start()
+                elif chaos_type == ChaosType.RESTART:
+                    target_ps = random.choice(env.pageservers)
+                    log.info(f"CHAOS: Normal restart of pageserver {target_ps.id}")
+                    target_ps.stop(immediate=False)
+                    target_ps.start()
+                elif chaos_type == ChaosType.STORCON_RESTART_IMMEDIATE:
+                    log.info("CHAOS: Immediate restart of storage controller")
+                    env.storage_controller.stop(immediate=True)
+                    env.storage_controller.start()
+            except Exception as e:
+                log.warning(f"CHAOS: Error during chaos operation {chaos_type}: {e}")
+
+            # Sleep before next chaos event
+            time.sleep(1)
+
+        log.info("Chaos injector stopped")
+
+    def wait_for_import_completion():
+        start = time.time()
+        done = import_completion_signaled.wait(ALLOWED_IMPORT_RUNTIME)
+        if not done:
+            raise TimeoutError(f"Import did not signal completion within {ALLOWED_IMPORT_RUNTIME}")
+
+        end = time.time()
+
+        log.info(f"Import completion signalled after {end - start}s {import_error=}")
+
+        if import_error:
+            raise RuntimeError(f"Import error: {import_error}")
+
+    with concurrent.futures.ThreadPoolExecutor() as executor:
+        stop_chaos = threading.Event()
+
+        wait_for_import_completion_fut = executor.submit(wait_for_import_completion)
+        chaos_fut = executor.submit(chaos, stop_chaos)
+
+        try:
+            wait_for_import_completion_fut.result()
+        except Exception as e:
+            raise e
+        finally:
+            stop_chaos.set()
+            chaos_fut.result()
+
+    import_branch_name = "imported"
+    env.neon_cli.mappings_map_branch(import_branch_name, tenant_id, timeline_id)
+    endpoint = env.endpoints.create_start(branch_name=import_branch_name, tenant_id=tenant_id)
+
+    # Validate the imported data is legit
+    validate_import_from_vanilla_pg(endpoint, inserted_rows)
+
+    endpoint.stop()
+
+    # Validate writes
+    workload = Workload(env, tenant_id, timeline_id, branch_name=import_branch_name)
+    workload.init()
+    workload.write_rows(64)
+    workload.validate()
 
 
 def test_fast_import_with_pageserver_ingest(

test_runner/regress/test_layers_from_future.py

@@ -20,6 +20,9 @@ from fixtures.remote_storage import LocalFsStorage, RemoteStorageKind
 from fixtures.utils import query_scalar, wait_until
 
 
+@pytest.mark.skip(
+    reason="We won't create future layers any more after https://github.com/neondatabase/neon/pull/10548"
+)
 @pytest.mark.parametrize(
     "attach_mode",
     ["default_generation", "same_generation"],

test_runner/regress/test_pageserver_secondary.py

@@ -124,6 +124,9 @@ def test_location_conf_churn(neon_env_builder: NeonEnvBuilder, make_httpserver,
                 ".*downloading failed, possibly for shutdown",
                 # {tenant_id=... timeline_id=...}:handle_pagerequests:handle_get_page_at_lsn_request{rel=1664/0/1260 blkno=0 req_lsn=0/149F0D8}: error reading relation or page version: Not found: will not become active.  Current state: Stopping\n'
                 ".*page_service.*will not become active.*",
+                # the following errors are possible when pageserver tries to ingest wal records despite being in unreadable state
+                ".*wal_connection_manager.*layer file download failed: No file found.*",
+                ".*wal_connection_manager.*could not ingest record.*",
             ]
         )
 
@@ -156,6 +159,45 @@ def test_location_conf_churn(neon_env_builder: NeonEnvBuilder, make_httpserver,
         env.pageservers[2].id: ("Detached", None),
     }
 
+    # Track all the attached locations with mode and generation
+    history: list[tuple[int, str, int | None]] = []
+
+    def may_read(pageserver: NeonPageserver, mode: str, generation: int | None) -> bool:
+        # Rules for when a pageserver may read:
+        # - our generation is higher than any previous
+        # - our generation is equal to previous, but no other pageserver
+        #   in that generation has been AttachedSingle (i.e. allowed to compact/GC)
+        # - our generation is equal to previous, and the previous holder of this
+        #   generation was the same node as we're attaching now.
+        #
+        # If these conditions are not met, then a read _might_ work, but the pageserver might
+        # also hit errors trying to download layers.
+        highest_historic_generation = max([i[2] for i in history if i[2] is not None], default=None)
+
+        if generation is None:
+            # We're not in an attached state, we may not read
+            return False
+        elif highest_historic_generation is not None and generation < highest_historic_generation:
+            # We are in an outdated generation, we may not read
+            return False
+        elif highest_historic_generation is not None and generation == highest_historic_generation:
+            # We are re-using a generation: if any pageserver other than this one
+            # has held AttachedSingle mode, this node may not read (because some other
+            # node may be doing GC/compaction).
+            if any(
+                i[1] == "AttachedSingle"
+                and i[2] == highest_historic_generation
+                and i[0] != pageserver.id
+                for i in history
+            ):
+                log.info(
+                    f"Skipping read on {pageserver.id} because other pageserver has been in AttachedSingle mode in generation {highest_historic_generation}"
+                )
+                return False
+
+        # Fall through: we have passed conditions for readability
+        return True
+
     latest_attached = env.pageservers[0].id
 
     for _i in range(0, 64):
@@ -199,9 +241,10 @@ def test_location_conf_churn(neon_env_builder: NeonEnvBuilder, make_httpserver,
                 assert len(tenants) == 1
                 assert tenants[0]["generation"] == new_generation
 
-                log.info("Entering postgres...")
-                workload.churn_rows(rng.randint(128, 256), pageserver.id)
-                workload.validate(pageserver.id)
+                if may_read(pageserver, last_state_ps[0], last_state_ps[1]):
+                    log.info("Entering postgres...")
+                    workload.churn_rows(rng.randint(128, 256), pageserver.id)
+                    workload.validate(pageserver.id)
             elif last_state_ps[0].startswith("Attached"):
                 # The `storage_controller` will only re-attach on startup when a pageserver was the
                 # holder of the latest generation: otherwise the pageserver will revert to detached
@@ -241,18 +284,16 @@ def test_location_conf_churn(neon_env_builder: NeonEnvBuilder, make_httpserver,
             location_conf["generation"] = generation
 
             pageserver.tenant_location_configure(tenant_id, location_conf)
+
             last_state[pageserver.id] = (mode, generation)
 
-            # It's only valid to connect to the last generation. Newer generations may yank layer
-            # files used in older generations.
-            last_generation = max(
-                [s[1] for s in last_state.values() if s[1] is not None], default=None
-            )
+            may_read_this_generation = may_read(pageserver, mode, generation)
+            history.append((pageserver.id, mode, generation))
 
-            if mode.startswith("Attached") and generation == last_generation:
-                # This is a basic test: we are validating that he endpoint works properly _between_
-                # configuration changes.  A stronger test would be to validate that clients see
-                # no errors while we are making the changes.
+            # This is a basic test: we are validating that he endpoint works properly _between_
+            # configuration changes.  A stronger test would be to validate that clients see
+            # no errors while we are making the changes.
+            if may_read_this_generation:
                 workload.churn_rows(
                     rng.randint(128, 256), pageserver.id, upload=mode != "AttachedStale"
                 )
@@ -265,9 +306,16 @@ def test_location_conf_churn(neon_env_builder: NeonEnvBuilder, make_httpserver,
     assert gc_summary["remote_storage_errors"] == 0
     assert gc_summary["indices_deleted"] > 0
 
-    # Attach all pageservers
+    # Attach all pageservers, in a higher generation than any previous.  We will use the same
+    # gen for all, and AttachedMulti mode so that they do not interfere with one another.
+    generation = env.storage_controller.attach_hook_issue(tenant_id, env.pageservers[0].id)
     for ps in env.pageservers:
-        location_conf = {"mode": "AttachedMulti", "secondary_conf": None, "tenant_conf": {}}
+        location_conf = {
+            "mode": "AttachedMulti",
+            "secondary_conf": None,
+            "tenant_conf": {},
+            "generation": generation,
+        }
         ps.tenant_location_configure(tenant_id, location_conf)
 
     # Confirm that all are readable

test_runner/regress/test_storage_controller.py

@@ -4158,17 +4158,12 @@ def test_storcon_create_delete_sk_down(
         env.storage_controller.stop()
         env.storage_controller.start()
 
-    config_lines = [
-        "neon.safekeeper_proto_version = 3",
-    ]
-    with env.endpoints.create("main", tenant_id=tenant_id, config_lines=config_lines) as ep:
+    with env.endpoints.create("main", tenant_id=tenant_id) as ep:
         # endpoint should start.
         ep.start(safekeeper_generation=1, safekeepers=[1, 2, 3])
         ep.safe_psql("CREATE TABLE IF NOT EXISTS t(key int, value text)")
 
-    with env.endpoints.create(
-        "child_of_main", tenant_id=tenant_id, config_lines=config_lines
-    ) as ep:
+    with env.endpoints.create("child_of_main", tenant_id=tenant_id) as ep:
         # endpoint should start.
         ep.start(safekeeper_generation=1, safekeepers=[1, 2, 3])
         ep.safe_psql("CREATE TABLE IF NOT EXISTS t(key int, value text)")
@@ -4197,10 +4192,10 @@ def test_storcon_create_delete_sk_down(
     # ensure the safekeeper deleted the timeline
     def timeline_deleted_on_active_sks():
         env.safekeepers[0].assert_log_contains(
-            f"deleting timeline {tenant_id}/{child_timeline_id} from disk"
+            f"((deleting timeline|Timeline) {tenant_id}/{child_timeline_id} (from disk|was already deleted)|DELETE.*tenant/{tenant_id} .*status: 200 OK)"
         )
         env.safekeepers[2].assert_log_contains(
-            f"deleting timeline {tenant_id}/{child_timeline_id} from disk"
+            f"((deleting timeline|Timeline) {tenant_id}/{child_timeline_id} (from disk|was already deleted)|DELETE.*tenant/{tenant_id} .*status: 200 OK)"
         )
 
     wait_until(timeline_deleted_on_active_sks)
@@ -4215,7 +4210,7 @@ def test_storcon_create_delete_sk_down(
     # ensure that there is log msgs for the third safekeeper too
     def timeline_deleted_on_sk():
         env.safekeepers[1].assert_log_contains(
-            f"deleting timeline {tenant_id}/{child_timeline_id} from disk"
+            f"((deleting timeline|Timeline) {tenant_id}/{child_timeline_id} (from disk|was already deleted)|DELETE.*tenant/{tenant_id} .*status: 200 OK)"
         )
 
     wait_until(timeline_deleted_on_sk)
@@ -4249,17 +4244,12 @@ def test_storcon_few_sk(
 
     env.safekeepers[0].assert_log_contains(f"creating new timeline {tenant_id}/{timeline_id}")
 
-    config_lines = [
-        "neon.safekeeper_proto_version = 3",
-    ]
-    with env.endpoints.create("main", tenant_id=tenant_id, config_lines=config_lines) as ep:
+    with env.endpoints.create("main", tenant_id=tenant_id) as ep:
         # endpoint should start.
         ep.start(safekeeper_generation=1, safekeepers=safekeeper_list)
         ep.safe_psql("CREATE TABLE IF NOT EXISTS t(key int, value text)")
 
-    with env.endpoints.create(
-        "child_of_main", tenant_id=tenant_id, config_lines=config_lines
-    ) as ep:
+    with env.endpoints.create("child_of_main", tenant_id=tenant_id) as ep:
         # endpoint should start.
         ep.start(safekeeper_generation=1, safekeepers=safekeeper_list)
         ep.safe_psql("CREATE TABLE IF NOT EXISTS t(key int, value text)")

test_runner/regress/test_timeline_detach_ancestor.py

@@ -10,6 +10,7 @@ from queue import Empty, Queue
 from threading import Barrier
 
 import pytest
+import requests
 from fixtures.common_types import Lsn, TimelineArchivalState, TimelineId
 from fixtures.log_helper import log
 from fixtures.neon_fixtures import (
@@ -401,8 +402,25 @@ def test_ancestor_detach_behavior_v2(neon_env_builder: NeonEnvBuilder, snapshots
         "earlier", ancestor_branch_name="main", ancestor_start_lsn=branchpoint_pipe
     )
 
-    snapshot_branchpoint_old = env.create_branch(
-        "snapshot_branchpoint_old", ancestor_branch_name="main", ancestor_start_lsn=branchpoint_y
+    snapshot_branchpoint_old = TimelineId.generate()
+
+    env.storage_controller.timeline_create(
+        env.initial_tenant,
+        {
+            "new_timeline_id": str(snapshot_branchpoint_old),
+            "ancestor_start_lsn": str(branchpoint_y),
+            "ancestor_timeline_id": str(env.initial_timeline),
+            "read_only": True,
+        },
+    )
+    sk = env.safekeepers[0]
+    assert sk
+    with pytest.raises(requests.exceptions.HTTPError, match="Not Found"):
+        sk.http_client().timeline_status(
+            tenant_id=env.initial_tenant, timeline_id=snapshot_branchpoint_old
+        )
+    env.neon_cli.mappings_map_branch(
+        "snapshot_branchpoint_old", env.initial_tenant, snapshot_branchpoint_old
     )
 
     snapshot_branchpoint = env.create_branch(

test_runner/regress/test_wal_acceptor.py

@@ -2012,10 +2012,7 @@ def test_explicit_timeline_creation(neon_env_builder: NeonEnvBuilder):
     tenant_id = env.initial_tenant
     timeline_id = env.initial_timeline
 
-    config_lines = [
-        "neon.safekeeper_proto_version = 3",
-    ]
-    ep = env.endpoints.create("main", config_lines=config_lines)
+    ep = env.endpoints.create("main")
 
     # expected to fail because timeline is not created on safekeepers
     with pytest.raises(Exception, match=r".*timed out.*"):
@@ -2043,10 +2040,7 @@ def test_explicit_timeline_creation_storcon(neon_env_builder: NeonEnvBuilder):
     }
     env = neon_env_builder.init_start()
 
-    config_lines = [
-        "neon.safekeeper_proto_version = 3",
-    ]
-    ep = env.endpoints.create("main", config_lines=config_lines)
+    ep = env.endpoints.create("main")
 
     # endpoint should start.
     ep.start(safekeeper_generation=1, safekeepers=[1, 2, 3])

test_runner/regress/test_wal_acceptor_async.py

@@ -637,10 +637,7 @@ async def quorum_sanity_single(
     # create timeline on `members_sks`
     Safekeeper.create_timeline(tenant_id, timeline_id, env.pageservers[0], mconf, members_sks)
 
-    config_lines = [
-        "neon.safekeeper_proto_version = 3",
-    ]
-    ep = env.endpoints.create(branch_name, config_lines=config_lines)
+    ep = env.endpoints.create(branch_name)
     ep.start(safekeeper_generation=1, safekeepers=compute_sks_ids)
     ep.safe_psql("create table t(key int, value text)")
 

workspace_hack/Cargo.toml

@@ -41,10 +41,8 @@ env_logger = { version = "0.11" }
 fail = { version = "0.5", default-features = false, features = ["failpoints"] }
 form_urlencoded = { version = "1" }
 futures-channel = { version = "0.3", features = ["sink"] }
-futures-core = { version = "0.3" }
 futures-executor = { version = "0.3" }
 futures-io = { version = "0.3" }
-futures-task = { version = "0.3", default-features = false, features = ["std"] }
 futures-util = { version = "0.3", features = ["channel", "io", "sink"] }
 generic-array = { version = "0.14", default-features = false, features = ["more_lengths", "zeroize"] }
 getrandom = { version = "0.2", default-features = false, features = ["std"] }
@@ -74,7 +72,6 @@ num-traits = { version = "0.2", features = ["i128", "libm"] }
 once_cell = { version = "1" }
 p256 = { version = "0.13", features = ["jwk"] }
 parquet = { version = "53", default-features = false, features = ["zstd"] }
-percent-encoding = { version = "2" }
 prost = { version = "0.13", features = ["no-recursion-limit", "prost-derive"] }
 rand = { version = "0.8", features = ["small_rng"] }
 regex = { version = "1" }