Merge branch 'problame/async-cleanup-on-drop-for-writers' into yuchen/direct-io-delta-image-layer-write

In #10063 we will switch BlobWriter, which underlies delta and image
layer writers, to use the owned buffers IO buffered writer.

That buffered writer implements double-buffering by virtue of a background task
that performs the flushing -- it owns the VirtualFile and both
DeltaLayerWriter and ImageLayerWriter are mere clients to it.

The implication is that it's no longer true that dropping these client
objects guarantees that all IO activity is complete. We must wait for the
flush task to exit.

In preparation for that new world, this PR moves the cleanup to a short-lived
task that is spawned from the Drop impl, and adds appropriate gate guard
holdings to hook it into the Timeline lifecycle.

We must (theoretically) worry that there will be a retry inbetween Drop
completing and the spawned task completing. It could collide on the
randomly generated temporary file name. We avoid this by switching to a
global monotonic counter.

Refs
- extracted from https://github.com/neondatabase/neon/pull/10063
- epic https://github.com/neondatabase/neon/issues/9868

.dockerignore

@@ -19,6 +19,7 @@
 !pageserver/
 !pgxn/
 !proxy/
+!object_storage/
 !storage_scrubber/
 !safekeeper/
 !storage_broker/

.gitignore

@@ -1,3 +1,4 @@
+/artifact_cache
 /pg_install
 /target
 /tmp_check

Cargo.lock

@@ -2837,6 +2837,7 @@ dependencies = [
  "utils",
  "uuid",
  "workspace_hack",
+ "x509-cert",
 ]
 
 [[package]]
@@ -3991,6 +3992,33 @@ dependencies = [
  "memchr",
 ]
 
+[[package]]
+name = "object_storage"
+version = "0.0.1"
+dependencies = [
+ "anyhow",
+ "axum",
+ "axum-extra",
+ "camino",
+ "camino-tempfile",
+ "futures",
+ "http-body-util",
+ "itertools 0.10.5",
+ "jsonwebtoken",
+ "prometheus",
+ "rand 0.8.5",
+ "remote_storage",
+ "serde",
+ "serde_json",
+ "test-log",
+ "tokio",
+ "tokio-util",
+ "tower 0.5.2",
+ "tracing",
+ "utils",
+ "workspace_hack",
+]
+
 [[package]]
 name = "once_cell"
 version = "1.20.2"
@@ -4693,7 +4721,7 @@ dependencies = [
 [[package]]
 name = "postgres-protocol"
 version = "0.6.6"
-source = "git+https://github.com/neondatabase/rust-postgres.git?branch=neon#1f21e7959a96a34dcfbfce1b14b73286cdadffe9"
+source = "git+https://github.com/neondatabase/rust-postgres.git?branch=neon#f3cf448febde5fd298071d54d568a9c875a7a62b"
 dependencies = [
  "base64 0.22.1",
  "byteorder",
@@ -4727,7 +4755,7 @@ dependencies = [
 [[package]]
 name = "postgres-types"
 version = "0.2.6"
-source = "git+https://github.com/neondatabase/rust-postgres.git?branch=neon#1f21e7959a96a34dcfbfce1b14b73286cdadffe9"
+source = "git+https://github.com/neondatabase/rust-postgres.git?branch=neon#f3cf448febde5fd298071d54d568a9c875a7a62b"
 dependencies = [
  "bytes",
  "chrono",
@@ -6925,6 +6953,28 @@ dependencies = [
  "syn 2.0.100",
 ]
 
+[[package]]
+name = "test-log"
+version = "0.2.17"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "e7f46083d221181166e5b6f6b1e5f1d499f3a76888826e6cb1d057554157cd0f"
+dependencies = [
+ "env_logger",
+ "test-log-macros",
+ "tracing-subscriber",
+]
+
+[[package]]
+name = "test-log-macros"
+version = "0.2.17"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "888d0c3c6db53c0fdab160d2ed5e12ba745383d3e85813f2ea0f2b1475ab553f"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn 2.0.100",
+]
+
 [[package]]
 name = "thiserror"
 version = "1.0.69"
@@ -7116,9 +7166,9 @@ dependencies = [
 
 [[package]]
 name = "tokio"
-version = "1.43.0"
+version = "1.43.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "3d61fa4ffa3de412bfea335c6ecff681de2b609ba3c77ef3e00e521813a9ed9e"
+checksum = "492a604e2fd7f814268a378409e6c92b5525d747d10db9a229723f55a417958c"
 dependencies = [
  "backtrace",
  "bytes",
@@ -7172,7 +7222,7 @@ dependencies = [
 [[package]]
 name = "tokio-postgres"
 version = "0.7.10"
-source = "git+https://github.com/neondatabase/rust-postgres.git?branch=neon#1f21e7959a96a34dcfbfce1b14b73286cdadffe9"
+source = "git+https://github.com/neondatabase/rust-postgres.git?branch=neon#f3cf448febde5fd298071d54d568a9c875a7a62b"
 dependencies = [
  "async-trait",
  "byteorder",

Cargo.toml

@@ -40,6 +40,7 @@ members = [
     "libs/proxy/postgres-protocol2",
     "libs/proxy/postgres-types2",
     "libs/proxy/tokio-postgres2",
+    "object_storage",
 ]
 
 [workspace.package]
@@ -183,7 +184,7 @@ test-context = "0.3"
 thiserror = "1.0"
 tikv-jemallocator = { version = "0.6", features = ["profiling", "stats", "unprefixed_malloc_on_supported_platforms"] }
 tikv-jemalloc-ctl = { version = "0.6", features = ["stats"] }
-tokio = { version = "1.41", features = ["macros"] }
+tokio = { version = "1.43.1", features = ["macros"] }
 tokio-epoll-uring = { git = "https://github.com/neondatabase/tokio-epoll-uring.git" , branch = "main" }
 tokio-io-timeout = "1.2.0"
 tokio-postgres-rustls = "0.12.0"
@@ -208,6 +209,7 @@ tracing-opentelemetry = "0.28"
 tracing-serde = "0.2.0"
 tracing-subscriber = { version = "0.3", default-features = false, features = ["smallvec", "fmt", "tracing-log", "std", "env-filter", "json"] }
 try-lock = "0.2.5"
+test-log = { version = "0.2.17", default-features = false, features = ["log"] }
 twox-hash = { version = "1.6.3", default-features = false }
 typed-json = "0.1"
 url = "2.2"

Dockerfile

@@ -89,6 +89,7 @@ RUN set -e \
       --bin storage_broker  \
       --bin storage_controller  \
       --bin proxy  \
+      --bin object_storage \
       --bin neon_local \
       --bin storage_scrubber \
       --locked --release
@@ -121,6 +122,7 @@ COPY --from=build --chown=neon:neon /home/nonroot/target/release/safekeeper
 COPY --from=build --chown=neon:neon /home/nonroot/target/release/storage_broker      /usr/local/bin
 COPY --from=build --chown=neon:neon /home/nonroot/target/release/storage_controller  /usr/local/bin
 COPY --from=build --chown=neon:neon /home/nonroot/target/release/proxy               /usr/local/bin
+COPY --from=build --chown=neon:neon /home/nonroot/target/release/object_storage      /usr/local/bin
 COPY --from=build --chown=neon:neon /home/nonroot/target/release/neon_local          /usr/local/bin
 COPY --from=build --chown=neon:neon /home/nonroot/target/release/storage_scrubber    /usr/local/bin
 

compute_tools/src/bin/compute_ctl.rs

@@ -118,16 +118,18 @@ struct Cli {
     #[arg(long)]
     pub set_disk_quota_for_fs: Option<String>,
 
-    #[arg(short = 's', long = "spec", group = "spec")]
-    pub spec_json: Option<String>,
-
     #[arg(short = 'S', long, group = "spec-path")]
     pub spec_path: Option<OsString>,
 
     #[arg(short = 'i', long, group = "compute-id")]
     pub compute_id: String,
 
-    #[arg(short = 'p', long, conflicts_with_all = ["spec", "spec-path"], value_name = "CONTROL_PLANE_API_BASE_URL")]
+    #[arg(
+        short = 'p',
+        long,
+        conflicts_with = "spec-path",
+        value_name = "CONTROL_PLANE_API_BASE_URL"
+    )]
     pub control_plane_uri: Option<String>,
 }
 
@@ -172,7 +174,6 @@ fn main() -> Result<()> {
             cgroup: cli.cgroup,
             #[cfg(target_os = "linux")]
             vm_monitor_addr: cli.vm_monitor_addr,
-            live_config_allowed: cli_spec.live_config_allowed,
         },
         cli_spec.spec,
         cli_spec.compute_ctl_config,
@@ -201,23 +202,12 @@ async fn init() -> Result<()> {
 }
 
 fn try_spec_from_cli(cli: &Cli) -> Result<CliSpecParams> {
-    // First, try to get cluster spec from the cli argument
-    if let Some(ref spec_json) = cli.spec_json {
-        info!("got spec from cli argument {}", spec_json);
-        return Ok(CliSpecParams {
-            spec: Some(serde_json::from_str(spec_json)?),
-            compute_ctl_config: ComputeCtlConfig::default(),
-            live_config_allowed: false,
-        });
-    }
-
-    // Second, try to read it from the file if path is provided
+    // First, read spec from the path if provided
     if let Some(ref spec_path) = cli.spec_path {
         let file = File::open(Path::new(spec_path))?;
         return Ok(CliSpecParams {
             spec: Some(serde_json::from_reader(file)?),
             compute_ctl_config: ComputeCtlConfig::default(),
-            live_config_allowed: true,
         });
     }
 
@@ -225,11 +215,12 @@ fn try_spec_from_cli(cli: &Cli) -> Result<CliSpecParams> {
         panic!("must specify --control-plane-uri");
     };
 
+    // If the spec wasn't provided in the CLI arguments, then retrieve it from
+    // the control plane
     match get_spec_from_control_plane(cli.control_plane_uri.as_ref().unwrap(), &cli.compute_id) {
         Ok(resp) => Ok(CliSpecParams {
             spec: resp.0,
             compute_ctl_config: resp.1,
-            live_config_allowed: true,
         }),
         Err(e) => {
             error!(
@@ -247,7 +238,6 @@ struct CliSpecParams {
     spec: Option<ComputeSpec>,
     #[allow(dead_code)]
     compute_ctl_config: ComputeCtlConfig,
-    live_config_allowed: bool,
 }
 
 fn deinit_and_exit(exit_code: Option<i32>) -> ! {

compute_tools/src/catalog.rs

@@ -98,13 +98,15 @@ pub async fn get_database_schema(
         .kill_on_drop(true)
         .spawn()?;
 
-    let stdout = cmd.stdout.take().ok_or_else(|| {
-        std::io::Error::new(std::io::ErrorKind::Other, "Failed to capture stdout.")
-    })?;
+    let stdout = cmd
+        .stdout
+        .take()
+        .ok_or_else(|| std::io::Error::other("Failed to capture stdout."))?;
 
-    let stderr = cmd.stderr.take().ok_or_else(|| {
-        std::io::Error::new(std::io::ErrorKind::Other, "Failed to capture stderr.")
-    })?;
+    let stderr = cmd
+        .stderr
+        .take()
+        .ok_or_else(|| std::io::Error::other("Failed to capture stderr."))?;
 
     let mut stdout_reader = FramedRead::new(stdout, BytesCodec::new());
     let stderr_reader = BufReader::new(stderr);
@@ -128,8 +130,7 @@ pub async fn get_database_schema(
                 }
             });
 
-            return Err(SchemaDumpError::IO(std::io::Error::new(
-                std::io::ErrorKind::Other,
+            return Err(SchemaDumpError::IO(std::io::Error::other(
                 "failed to start pg_dump",
             )));
         }

compute_tools/src/compute.rs

@@ -93,20 +93,6 @@ pub struct ComputeNodeParams {
 
     /// the address of extension storage proxy gateway
     pub ext_remote_storage: Option<String>,
-
-    /// We should only allow live re- / configuration of the compute node if
-    /// it uses 'pull model', i.e. it can go to control-plane and fetch
-    /// the latest configuration. Otherwise, there could be a case:
-    /// - we start compute with some spec provided as argument
-    /// - we push new spec and it does reconfiguration
-    /// - but then something happens and compute pod / VM is destroyed,
-    ///   so k8s controller starts it again with the **old** spec
-    ///
-    /// and the same for empty computes:
-    /// - we started compute without any spec
-    /// - we push spec and it does configuration
-    /// - but then it is restarted without any spec again
-    pub live_config_allowed: bool,
 }
 
 /// Compute node info shared across several `compute_ctl` threads.
@@ -537,11 +523,14 @@ impl ComputeNode {
 
         let pspec = compute_state.pspec.as_ref().expect("spec must be set");
         info!(
-            "starting compute for project {}, operation {}, tenant {}, timeline {}, features {:?}, spec.remote_extensions {:?}",
+            "starting compute for project {}, operation {}, tenant {}, timeline {}, project {}, branch {}, endpoint {}, features {:?}, spec.remote_extensions {:?}",
             pspec.spec.cluster.cluster_id.as_deref().unwrap_or("None"),
             pspec.spec.operation_uuid.as_deref().unwrap_or("None"),
             pspec.tenant_id,
             pspec.timeline_id,
+            pspec.spec.project_id.as_deref().unwrap_or("None"),
+            pspec.spec.branch_id.as_deref().unwrap_or("None"),
+            pspec.spec.endpoint_id.as_deref().unwrap_or("None"),
             pspec.spec.features,
             pspec.spec.remote_extensions,
         );
@@ -645,31 +634,28 @@ impl ComputeNode {
             });
         }
 
-        // Configure and start rsyslog for HIPAA if necessary
-        if let ComputeAudit::Hipaa = pspec.spec.audit_log_level {
-            let remote_endpoint = std::env::var("AUDIT_LOGGING_ENDPOINT").unwrap_or("".to_string());
-            if remote_endpoint.is_empty() {
-                anyhow::bail!("AUDIT_LOGGING_ENDPOINT is empty");
+        // Configure and start rsyslog for compliance audit logging
+        match pspec.spec.audit_log_level {
+            ComputeAudit::Hipaa | ComputeAudit::Extended | ComputeAudit::Full => {
+                let remote_endpoint =
+                    std::env::var("AUDIT_LOGGING_ENDPOINT").unwrap_or("".to_string());
+                if remote_endpoint.is_empty() {
+                    anyhow::bail!("AUDIT_LOGGING_ENDPOINT is empty");
+                }
+
+                let log_directory_path = Path::new(&self.params.pgdata).join("log");
+                let log_directory_path = log_directory_path.to_string_lossy().to_string();
+                configure_audit_rsyslog(log_directory_path.clone(), "hipaa", &remote_endpoint)?;
+
+                // Launch a background task to clean up the audit logs
+                launch_pgaudit_gc(log_directory_path);
             }
-
-            let log_directory_path = Path::new(&self.params.pgdata).join("log");
-            let log_directory_path = log_directory_path.to_string_lossy().to_string();
-            configure_audit_rsyslog(log_directory_path.clone(), "hipaa", &remote_endpoint)?;
-
-            // Launch a background task to clean up the audit logs
-            launch_pgaudit_gc(log_directory_path);
+            _ => {}
         }
 
         // Configure and start rsyslog for Postgres logs export
-        if self.has_feature(ComputeFeature::PostgresLogsExport) {
-            if let Some(ref project_id) = pspec.spec.cluster.cluster_id {
-                let host = PostgresLogsRsyslogConfig::default_host(project_id);
-                let conf = PostgresLogsRsyslogConfig::new(Some(&host));
-                configure_postgres_logs_export(conf)?;
-            } else {
-                warn!("not configuring rsyslog for Postgres logs export: project ID is missing")
-            }
-        }
+        let conf = PostgresLogsRsyslogConfig::new(pspec.spec.logs_export_host.as_deref());
+        configure_postgres_logs_export(conf)?;
 
         // Launch remaining service threads
         let _monitor_handle = launch_monitor(self);
@@ -1573,6 +1559,10 @@ impl ComputeNode {
             });
         }
 
+        // Reconfigure rsyslog for Postgres logs export
+        let conf = PostgresLogsRsyslogConfig::new(spec.logs_export_host.as_deref());
+        configure_postgres_logs_export(conf)?;
+
         // Write new config
         let pgdata_path = Path::new(&self.params.pgdata);
         config::write_postgres_conf(

compute_tools/src/config.rs

@@ -7,7 +7,7 @@ use std::io::prelude::*;
 use std::path::Path;
 
 use compute_api::responses::TlsConfig;
-use compute_api::spec::{ComputeAudit, ComputeFeature, ComputeMode, ComputeSpec, GenericOption};
+use compute_api::spec::{ComputeAudit, ComputeMode, ComputeSpec, GenericOption};
 
 use crate::pg_helpers::{
     GenericOptionExt, GenericOptionsSearch, PgOptionsSerialize, escape_conf_value,
@@ -89,6 +89,15 @@ pub fn write_postgres_conf(
             escape_conf_value(&s.to_string())
         )?;
     }
+    if let Some(s) = &spec.project_id {
+        writeln!(file, "neon.project_id={}", escape_conf_value(s))?;
+    }
+    if let Some(s) = &spec.branch_id {
+        writeln!(file, "neon.branch_id={}", escape_conf_value(s))?;
+    }
+    if let Some(s) = &spec.endpoint_id {
+        writeln!(file, "neon.endpoint_id={}", escape_conf_value(s))?;
+    }
 
     // tls
     if let Some(tls_config) = tls_config {
@@ -169,7 +178,7 @@ pub fn write_postgres_conf(
     // and don't allow the user or the control plane admin to change them.
     match spec.audit_log_level {
         ComputeAudit::Disabled => {}
-        ComputeAudit::Log => {
+        ComputeAudit::Log | ComputeAudit::Base => {
             writeln!(file, "# Managed by compute_ctl base audit settings: start")?;
             writeln!(file, "pgaudit.log='ddl,role'")?;
             // Disable logging of catalog queries to reduce the noise
@@ -193,16 +202,20 @@ pub fn write_postgres_conf(
             }
             writeln!(file, "# Managed by compute_ctl base audit settings: end")?;
         }
-        ComputeAudit::Hipaa => {
+        ComputeAudit::Hipaa | ComputeAudit::Extended | ComputeAudit::Full => {
             writeln!(
                 file,
                 "# Managed by compute_ctl compliance audit settings: begin"
             )?;
-            // This log level is very verbose
-            // but this is necessary for HIPAA compliance.
-            // Exclude 'misc' category, because it doesn't contain anythig relevant.
-            writeln!(file, "pgaudit.log='all, -misc'")?;
-            writeln!(file, "pgaudit.log_parameter=on")?;
+            // Enable logging of parameters.
+            // This is very verbose and may contain sensitive data.
+            if spec.audit_log_level == ComputeAudit::Full {
+                writeln!(file, "pgaudit.log_parameter=on")?;
+                writeln!(file, "pgaudit.log='all'")?;
+            } else {
+                writeln!(file, "pgaudit.log_parameter=off")?;
+                writeln!(file, "pgaudit.log='all, -misc'")?;
+            }
             // Disable logging of catalog queries
             // The catalog doesn't contain sensitive data, so we don't need to audit it.
             writeln!(file, "pgaudit.log_catalog=off")?;
@@ -255,7 +268,7 @@ pub fn write_postgres_conf(
 
     // We need Postgres to send logs to rsyslog so that we can forward them
     // further to customers' log aggregation systems.
-    if spec.features.contains(&ComputeFeature::PostgresLogsExport) {
+    if spec.logs_export_host.is_some() {
         writeln!(file, "log_destination='stderr,syslog'")?;
     }
 

compute_tools/src/http/middleware/authorize.rs

@@ -6,20 +6,15 @@ use axum_extra::{
     TypedHeader,
     headers::{Authorization, authorization::Bearer},
 };
+use compute_api::requests::ComputeClaims;
 use futures::future::BoxFuture;
 use http::{Request, Response, StatusCode};
 use jsonwebtoken::{Algorithm, DecodingKey, TokenData, Validation, jwk::JwkSet};
-use serde::Deserialize;
 use tower_http::auth::AsyncAuthorizeRequest;
 use tracing::warn;
 
 use crate::http::{JsonResponse, extract::RequestId};
 
-#[derive(Clone, Debug, Deserialize)]
-pub(in crate::http) struct Claims {
-    compute_id: String,
-}
-
 #[derive(Clone, Debug)]
 pub(in crate::http) struct Authorize {
     compute_id: String,
@@ -112,7 +107,11 @@ impl AsyncAuthorizeRequest<Body> for Authorize {
 
 impl Authorize {
     /// Verify the token using the JSON Web Key set and return the token data.
-    fn verify(jwks: &JwkSet, token: &str, validation: &Validation) -> Result<TokenData<Claims>> {
+    fn verify(
+        jwks: &JwkSet,
+        token: &str,
+        validation: &Validation,
+    ) -> Result<TokenData<ComputeClaims>> {
         for jwk in jwks.keys.iter() {
             let decoding_key = match DecodingKey::from_jwk(jwk) {
                 Ok(key) => key,
@@ -127,7 +126,7 @@ impl Authorize {
                 }
             };
 
-            match jsonwebtoken::decode::<Claims>(token, &decoding_key, validation) {
+            match jsonwebtoken::decode::<ComputeClaims>(token, &decoding_key, validation) {
                 Ok(data) => return Ok(data),
                 Err(e) => {
                     warn!(

compute_tools/src/http/openapi_spec.yaml

@@ -306,36 +306,6 @@ paths:
               schema:
                 $ref: "#/components/schemas/GenericError"
 
-  /configure_telemetry:
-    post:
-      tags:
-        - Configure
-      summary: Configure rsyslog
-      description: |
-        This API endpoint configures rsyslog to forward Postgres logs
-        to a specified otel collector.
-      operationId: configureTelemetry
-      requestBody:
-        required: true
-        content:
-          application/json:
-            schema:
-              type: object
-              properties:
-                logs_export_host:
-                  type: string
-                  description: |
-                    Hostname and the port of the otel collector. Leave empty to disable logs forwarding.
-                    Example: config-shy-breeze-123-collector-monitoring.neon-telemetry.svc.cluster.local:54526
-      responses:
-        204:
-          description: "Telemetry configured successfully"
-        500:
-          content:
-            application/json:
-              schema:
-                $ref: "#/components/schemas/GenericError"
-
 components:
   securitySchemes:
     JWT:

compute_tools/src/http/routes/configure.rs

@@ -1,11 +1,9 @@
 use std::sync::Arc;
 
-use axum::body::Body;
 use axum::extract::State;
 use axum::response::Response;
-use compute_api::requests::{ConfigurationRequest, ConfigureTelemetryRequest};
+use compute_api::requests::ConfigurationRequest;
 use compute_api::responses::{ComputeStatus, ComputeStatusResponse};
-use compute_api::spec::ComputeFeature;
 use http::StatusCode;
 use tokio::task;
 use tracing::info;
@@ -13,7 +11,6 @@ use tracing::info;
 use crate::compute::{ComputeNode, ParsedSpec};
 use crate::http::JsonResponse;
 use crate::http::extract::Json;
-use crate::rsyslog::{PostgresLogsRsyslogConfig, configure_postgres_logs_export};
 
 // Accept spec in JSON format and request compute configuration. If anything
 // goes wrong after we set the compute status to `ConfigurationPending` and
@@ -25,13 +22,6 @@ pub(in crate::http) async fn configure(
     State(compute): State<Arc<ComputeNode>>,
     request: Json<ConfigurationRequest>,
 ) -> Response {
-    if !compute.params.live_config_allowed {
-        return JsonResponse::error(
-            StatusCode::PRECONDITION_FAILED,
-            "live configuration is not allowed for this compute node".to_string(),
-        );
-    }
-
     let pspec = match ParsedSpec::try_from(request.spec.clone()) {
         Ok(p) => p,
         Err(e) => return JsonResponse::error(StatusCode::BAD_REQUEST, e),
@@ -95,25 +85,3 @@ pub(in crate::http) async fn configure(
 
     JsonResponse::success(StatusCode::OK, body)
 }
-
-pub(in crate::http) async fn configure_telemetry(
-    State(compute): State<Arc<ComputeNode>>,
-    request: Json<ConfigureTelemetryRequest>,
-) -> Response {
-    if !compute.has_feature(ComputeFeature::PostgresLogsExport) {
-        return JsonResponse::error(
-            StatusCode::PRECONDITION_FAILED,
-            "Postgres logs export feature is not enabled".to_string(),
-        );
-    }
-
-    let conf = PostgresLogsRsyslogConfig::new(request.logs_export_host.as_deref());
-    if let Err(err) = configure_postgres_logs_export(conf) {
-        return JsonResponse::error(StatusCode::INTERNAL_SERVER_ERROR, err.to_string());
-    }
-
-    Response::builder()
-        .status(StatusCode::NO_CONTENT)
-        .body(Body::from(""))
-        .unwrap()
-}

compute_tools/src/http/server.rs

@@ -87,7 +87,6 @@ impl From<&Server> for Router<Arc<ComputeNode>> {
                 let authenticated_router = Router::<Arc<ComputeNode>>::new()
                     .route("/check_writability", post(check_writability::is_writable))
                     .route("/configure", post(configure::configure))
-                    .route("/configure_telemetry", post(configure::configure_telemetry))
                     .route("/database_schema", get(database_schema::get_schema_dump))
                     .route("/dbs_and_roles", get(dbs_and_roles::get_catalog_objects))
                     .route("/insights", get(insights::get_insights))

compute_tools/src/rsyslog.rs

@@ -119,16 +119,9 @@ impl<'a> PostgresLogsRsyslogConfig<'a> {
         };
         Ok(config_content)
     }
-
-    /// Returns the default host for otel collector that receives Postgres logs
-    pub fn default_host(project_id: &str) -> String {
-        format!(
-            "config-{}-collector.neon-telemetry.svc.cluster.local:10514",
-            project_id
-        )
-    }
 }
 
+/// Writes rsyslogd configuration for Postgres logs export and restarts rsyslog.
 pub fn configure_postgres_logs_export(conf: PostgresLogsRsyslogConfig) -> Result<()> {
     let new_config = conf.build()?;
     let current_config = PostgresLogsRsyslogConfig::current_config()?;
@@ -261,16 +254,5 @@ mod tests {
             let res = conf.build();
             assert!(res.is_err());
         }
-
-        {
-            // Verify config with default host
-            let host = PostgresLogsRsyslogConfig::default_host("shy-breeze-123");
-            let conf = PostgresLogsRsyslogConfig::new(Some(&host));
-            let res = conf.build();
-            assert!(res.is_ok());
-            let conf_str = res.unwrap();
-            assert!(conf_str.contains(r#"shy-breeze-123"#));
-            assert!(conf_str.contains(r#"port="10514""#));
-        }
     }
 }

compute_tools/src/spec_apply.rs

@@ -278,12 +278,12 @@ impl ComputeNode {
             // so that all config operations are audit logged.
             match spec.audit_log_level
             {
-                ComputeAudit::Hipaa => {
+                ComputeAudit::Hipaa | ComputeAudit::Extended | ComputeAudit::Full => {
                     phases.push(CreatePgauditExtension);
                     phases.push(CreatePgauditlogtofileExtension);
                     phases.push(DisablePostgresDBPgAudit);
                 }
-                ComputeAudit::Log => {
+                ComputeAudit::Log | ComputeAudit::Base => {
                     phases.push(CreatePgauditExtension);
                     phases.push(DisablePostgresDBPgAudit);
                 }

control_plane/src/bin/neon_local.rs

@@ -20,8 +20,10 @@ use compute_api::spec::ComputeMode;
 use control_plane::endpoint::ComputeControlPlane;
 use control_plane::local_env::{
     InitForceMode, LocalEnv, NeonBroker, NeonLocalInitConf, NeonLocalInitPageserverConf,
-    SafekeeperConf,
+    ObjectStorageConf, SafekeeperConf,
 };
+use control_plane::object_storage::OBJECT_STORAGE_DEFAULT_PORT;
+use control_plane::object_storage::ObjectStorage;
 use control_plane::pageserver::PageServerNode;
 use control_plane::safekeeper::SafekeeperNode;
 use control_plane::storage_controller::{
@@ -39,7 +41,7 @@ use pageserver_api::controller_api::{
 use pageserver_api::models::{
     ShardParameters, TenantConfigRequest, TimelineCreateRequest, TimelineInfo,
 };
-use pageserver_api::shard::{ShardCount, ShardStripeSize, TenantShardId};
+use pageserver_api::shard::{DEFAULT_STRIPE_SIZE, ShardCount, ShardStripeSize, TenantShardId};
 use postgres_backend::AuthType;
 use postgres_connection::parse_host_port;
 use safekeeper_api::membership::SafekeeperGeneration;
@@ -91,6 +93,8 @@ enum NeonLocalCmd {
     #[command(subcommand)]
     Safekeeper(SafekeeperCmd),
     #[command(subcommand)]
+    ObjectStorage(ObjectStorageCmd),
+    #[command(subcommand)]
     Endpoint(EndpointCmd),
     #[command(subcommand)]
     Mappings(MappingsCmd),
@@ -454,6 +458,32 @@ enum SafekeeperCmd {
     Restart(SafekeeperRestartCmdArgs),
 }
 
+#[derive(clap::Subcommand)]
+#[clap(about = "Manage object storage")]
+enum ObjectStorageCmd {
+    Start(ObjectStorageStartCmd),
+    Stop(ObjectStorageStopCmd),
+}
+
+#[derive(clap::Args)]
+#[clap(about = "Start object storage")]
+struct ObjectStorageStartCmd {
+    #[clap(short = 't', long, help = "timeout until we fail the command")]
+    #[arg(default_value = "10s")]
+    start_timeout: humantime::Duration,
+}
+
+#[derive(clap::Args)]
+#[clap(about = "Stop object storage")]
+struct ObjectStorageStopCmd {
+    #[arg(value_enum, default_value = "fast")]
+    #[clap(
+        short = 'm',
+        help = "If 'immediate', don't flush repository data at shutdown"
+    )]
+    stop_mode: StopMode,
+}
+
 #[derive(clap::Args)]
 #[clap(about = "Start local safekeeper")]
 struct SafekeeperStartCmdArgs {
@@ -759,6 +789,7 @@ fn main() -> Result<()> {
             }
             NeonLocalCmd::StorageBroker(subcmd) => rt.block_on(handle_storage_broker(&subcmd, env)),
             NeonLocalCmd::Safekeeper(subcmd) => rt.block_on(handle_safekeeper(&subcmd, env)),
+            NeonLocalCmd::ObjectStorage(subcmd) => rt.block_on(handle_object_storage(&subcmd, env)),
             NeonLocalCmd::Endpoint(subcmd) => rt.block_on(handle_endpoint(&subcmd, env)),
             NeonLocalCmd::Mappings(subcmd) => handle_mappings(&subcmd, env),
         };
@@ -975,6 +1006,9 @@ fn handle_init(args: &InitCmdArgs) -> anyhow::Result<LocalEnv> {
                     }
                 })
                 .collect(),
+            object_storage: ObjectStorageConf {
+                port: OBJECT_STORAGE_DEFAULT_PORT,
+            },
             pg_distrib_dir: None,
             neon_distrib_dir: None,
             default_tenant_id: TenantId::from_array(std::array::from_fn(|_| 0)),
@@ -1083,7 +1117,7 @@ async fn handle_tenant(subcmd: &TenantCmd, env: &mut local_env::LocalEnv) -> any
                         stripe_size: args
                             .shard_stripe_size
                             .map(ShardStripeSize)
-                            .unwrap_or(ShardParameters::DEFAULT_STRIPE_SIZE),
+                            .unwrap_or(DEFAULT_STRIPE_SIZE),
                     },
                     placement_policy: args.placement_policy.clone(),
                     config: tenant_conf,
@@ -1396,7 +1430,7 @@ async fn handle_endpoint(subcmd: &EndpointCmd, env: &local_env::LocalEnv) -> Res
                     vec![(parsed.0, parsed.1.unwrap_or(5432))],
                     // If caller is telling us what pageserver to use, this is not a tenant which is
                     // full managed by storage controller, therefore not sharded.
-                    ShardParameters::DEFAULT_STRIPE_SIZE,
+                    DEFAULT_STRIPE_SIZE,
                 )
             } else {
                 // Look up the currently attached location of the tenant, and its striping metadata,
@@ -1683,6 +1717,41 @@ async fn handle_safekeeper(subcmd: &SafekeeperCmd, env: &local_env::LocalEnv) ->
     Ok(())
 }
 
+async fn handle_object_storage(subcmd: &ObjectStorageCmd, env: &local_env::LocalEnv) -> Result<()> {
+    use ObjectStorageCmd::*;
+    let storage = ObjectStorage::from_env(env);
+
+    // In tests like test_forward_compatibility or test_graceful_cluster_restart
+    // old neon binaries (without object_storage) are present
+    if !storage.bin.exists() {
+        eprintln!(
+            "{} binary not found. Ignore if this is a compatibility test",
+            storage.bin
+        );
+        return Ok(());
+    }
+
+    match subcmd {
+        Start(ObjectStorageStartCmd { start_timeout }) => {
+            if let Err(e) = storage.start(start_timeout).await {
+                eprintln!("object_storage start failed: {e}");
+                exit(1);
+            }
+        }
+        Stop(ObjectStorageStopCmd { stop_mode }) => {
+            let immediate = match stop_mode {
+                StopMode::Fast => false,
+                StopMode::Immediate => true,
+            };
+            if let Err(e) = storage.stop(immediate) {
+                eprintln!("proxy stop failed: {e}");
+                exit(1);
+            }
+        }
+    };
+    Ok(())
+}
+
 async fn handle_storage_broker(subcmd: &StorageBrokerCmd, env: &local_env::LocalEnv) -> Result<()> {
     match subcmd {
         StorageBrokerCmd::Start(args) => {
@@ -1777,6 +1846,13 @@ async fn handle_start_all_impl(
                     .map_err(|e| e.context(format!("start safekeeper {}", safekeeper.id)))
             });
         }
+
+        js.spawn(async move {
+            ObjectStorage::from_env(env)
+                .start(&retry_timeout)
+                .await
+                .map_err(|e| e.context("start object_storage"))
+        });
     })();
 
     let mut errors = Vec::new();
@@ -1874,6 +1950,11 @@ async fn try_stop_all(env: &local_env::LocalEnv, immediate: bool) {
         }
     }
 
+    let storage = ObjectStorage::from_env(env);
+    if let Err(e) = storage.stop(immediate) {
+        eprintln!("object_storage stop failed: {:#}", e);
+    }
+
     for ps_conf in &env.pageservers {
         let pageserver = PageServerNode::from_env(env, ps_conf);
         if let Err(e) = pageserver.stop(immediate) {

control_plane/src/endpoint.rs

@@ -658,6 +658,9 @@ impl Endpoint {
             delta_operations: None,
             tenant_id: Some(self.tenant_id),
             timeline_id: Some(self.timeline_id),
+            project_id: None,
+            branch_id: None,
+            endpoint_id: Some(self.endpoint_id.clone()),
             mode: self.mode,
             pageserver_connstring: Some(pageserver_connstring),
             safekeepers_generation: safekeepers_generation.map(|g| g.into_inner()),
@@ -670,6 +673,7 @@ impl Endpoint {
             reconfigure_concurrency: self.reconfigure_concurrency,
             drop_subscriptions_before_start: self.drop_subscriptions_before_start,
             audit_log_level: ComputeAudit::Disabled,
+            logs_export_host: None::<String>,
         };
 
         // this strange code is needed to support respec() in tests

control_plane/src/lib.rs

@@ -10,6 +10,7 @@ mod background_process;
 pub mod broker;
 pub mod endpoint;
 pub mod local_env;
+pub mod object_storage;
 pub mod pageserver;
 pub mod postgresql_conf;
 pub mod safekeeper;

control_plane/src/local_env.rs

@@ -15,9 +15,10 @@ use clap::ValueEnum;
 use postgres_backend::AuthType;
 use reqwest::Url;
 use serde::{Deserialize, Serialize};
-use utils::auth::{Claims, encode_from_key_file};
+use utils::auth::encode_from_key_file;
 use utils::id::{NodeId, TenantId, TenantTimelineId, TimelineId};
 
+use crate::object_storage::{OBJECT_STORAGE_REMOTE_STORAGE_DIR, ObjectStorage};
 use crate::pageserver::{PAGESERVER_REMOTE_STORAGE_DIR, PageServerNode};
 use crate::safekeeper::SafekeeperNode;
 
@@ -55,6 +56,7 @@ pub struct LocalEnv {
 
     // used to issue tokens during e.g pg start
     pub private_key_path: PathBuf,
+    pub public_key_path: PathBuf,
 
     pub broker: NeonBroker,
 
@@ -68,6 +70,8 @@ pub struct LocalEnv {
 
     pub safekeepers: Vec<SafekeeperConf>,
 
+    pub object_storage: ObjectStorageConf,
+
     // Control plane upcall API for pageserver: if None, we will not run storage_controller  If set, this will
     // be propagated into each pageserver's configuration.
     pub control_plane_api: Url,
@@ -95,6 +99,7 @@ pub struct OnDiskConfig {
     pub neon_distrib_dir: PathBuf,
     pub default_tenant_id: Option<TenantId>,
     pub private_key_path: PathBuf,
+    pub public_key_path: PathBuf,
     pub broker: NeonBroker,
     pub storage_controller: NeonStorageControllerConf,
     #[serde(
@@ -103,6 +108,7 @@ pub struct OnDiskConfig {
     )]
     pub pageservers: Vec<PageServerConf>,
     pub safekeepers: Vec<SafekeeperConf>,
+    pub object_storage: ObjectStorageConf,
     pub control_plane_api: Option<Url>,
     pub control_plane_hooks_api: Option<Url>,
     pub control_plane_compute_hook_api: Option<Url>,
@@ -136,11 +142,18 @@ pub struct NeonLocalInitConf {
     pub storage_controller: Option<NeonStorageControllerConf>,
     pub pageservers: Vec<NeonLocalInitPageserverConf>,
     pub safekeepers: Vec<SafekeeperConf>,
+    pub object_storage: ObjectStorageConf,
     pub control_plane_api: Option<Url>,
     pub control_plane_hooks_api: Option<Url>,
     pub generate_local_ssl_certs: bool,
 }
 
+#[derive(Serialize, Default, Deserialize, PartialEq, Eq, Clone, Debug)]
+#[serde(default)]
+pub struct ObjectStorageConf {
+    pub port: u16,
+}
+
 /// Broker config for cluster internal communication.
 #[derive(Serialize, Deserialize, PartialEq, Eq, Clone, Debug)]
 #[serde(default)]
@@ -398,6 +411,10 @@ impl LocalEnv {
         self.pg_dir(pg_version, "lib")
     }
 
+    pub fn object_storage_bin(&self) -> PathBuf {
+        self.neon_distrib_dir.join("object_storage")
+    }
+
     pub fn pageserver_bin(&self) -> PathBuf {
         self.neon_distrib_dir.join("pageserver")
     }
@@ -431,6 +448,10 @@ impl LocalEnv {
         self.base_data_dir.join("safekeepers").join(data_dir_name)
     }
 
+    pub fn object_storage_data_dir(&self) -> PathBuf {
+        self.base_data_dir.join("object_storage")
+    }
+
     pub fn get_pageserver_conf(&self, id: NodeId) -> anyhow::Result<&PageServerConf> {
         if let Some(conf) = self.pageservers.iter().find(|node| node.id == id) {
             Ok(conf)
@@ -582,6 +603,7 @@ impl LocalEnv {
                 neon_distrib_dir,
                 default_tenant_id,
                 private_key_path,
+                public_key_path,
                 broker,
                 storage_controller,
                 pageservers,
@@ -591,6 +613,7 @@ impl LocalEnv {
                 control_plane_compute_hook_api: _,
                 branch_name_mappings,
                 generate_local_ssl_certs,
+                object_storage,
             } = on_disk_config;
             LocalEnv {
                 base_data_dir: repopath.to_owned(),
@@ -598,6 +621,7 @@ impl LocalEnv {
                 neon_distrib_dir,
                 default_tenant_id,
                 private_key_path,
+                public_key_path,
                 broker,
                 storage_controller,
                 pageservers,
@@ -606,6 +630,7 @@ impl LocalEnv {
                 control_plane_hooks_api,
                 branch_name_mappings,
                 generate_local_ssl_certs,
+                object_storage,
             }
         };
 
@@ -705,6 +730,7 @@ impl LocalEnv {
                 neon_distrib_dir: self.neon_distrib_dir.clone(),
                 default_tenant_id: self.default_tenant_id,
                 private_key_path: self.private_key_path.clone(),
+                public_key_path: self.public_key_path.clone(),
                 broker: self.broker.clone(),
                 storage_controller: self.storage_controller.clone(),
                 pageservers: vec![], // it's skip_serializing anyway
@@ -714,6 +740,7 @@ impl LocalEnv {
                 control_plane_compute_hook_api: None,
                 branch_name_mappings: self.branch_name_mappings.clone(),
                 generate_local_ssl_certs: self.generate_local_ssl_certs,
+                object_storage: self.object_storage.clone(),
             },
         )
     }
@@ -730,7 +757,7 @@ impl LocalEnv {
     }
 
     // this function is used only for testing purposes in CLI e g generate tokens during init
-    pub fn generate_auth_token(&self, claims: &Claims) -> anyhow::Result<String> {
+    pub fn generate_auth_token<S: Serialize>(&self, claims: &S) -> anyhow::Result<String> {
         let private_key_path = self.get_private_key_path();
         let key_data = fs::read(private_key_path)?;
         encode_from_key_file(claims, &key_data)
@@ -797,6 +824,7 @@ impl LocalEnv {
             control_plane_api,
             generate_local_ssl_certs,
             control_plane_hooks_api,
+            object_storage,
         } = conf;
 
         // Find postgres binaries.
@@ -828,6 +856,7 @@ impl LocalEnv {
         )
         .context("generate auth keys")?;
         let private_key_path = PathBuf::from("auth_private_key.pem");
+        let public_key_path = PathBuf::from("auth_public_key.pem");
 
         // create the runtime type because the remaining initialization code below needs
         // a LocalEnv instance op operation
@@ -838,6 +867,7 @@ impl LocalEnv {
             neon_distrib_dir,
             default_tenant_id: Some(default_tenant_id),
             private_key_path,
+            public_key_path,
             broker,
             storage_controller: storage_controller.unwrap_or_default(),
             pageservers: pageservers.iter().map(Into::into).collect(),
@@ -846,6 +876,7 @@ impl LocalEnv {
             control_plane_hooks_api,
             branch_name_mappings: Default::default(),
             generate_local_ssl_certs,
+            object_storage,
         };
 
         if generate_local_ssl_certs {
@@ -873,8 +904,13 @@ impl LocalEnv {
                 .context("pageserver init failed")?;
         }
 
+        ObjectStorage::from_env(&env)
+            .init()
+            .context("object storage init failed")?;
+
         // setup remote remote location for default LocalFs remote storage
         std::fs::create_dir_all(env.base_data_dir.join(PAGESERVER_REMOTE_STORAGE_DIR))?;
+        std::fs::create_dir_all(env.base_data_dir.join(OBJECT_STORAGE_REMOTE_STORAGE_DIR))?;
 
         env.persist_config()
     }

control_plane/src/object_storage.rs

@@ -0,0 +1,107 @@
+use crate::background_process::{self, start_process, stop_process};
+use crate::local_env::LocalEnv;
+use anyhow::anyhow;
+use anyhow::{Context, Result};
+use camino::Utf8PathBuf;
+use std::io::Write;
+use std::time::Duration;
+
+/// Directory within .neon which will be used by default for LocalFs remote storage.
+pub const OBJECT_STORAGE_REMOTE_STORAGE_DIR: &str = "local_fs_remote_storage/object_storage";
+pub const OBJECT_STORAGE_DEFAULT_PORT: u16 = 9993;
+
+pub struct ObjectStorage {
+    pub bin: Utf8PathBuf,
+    pub data_dir: Utf8PathBuf,
+    pub pemfile: Utf8PathBuf,
+    pub port: u16,
+}
+
+impl ObjectStorage {
+    pub fn from_env(env: &LocalEnv) -> ObjectStorage {
+        ObjectStorage {
+            bin: Utf8PathBuf::from_path_buf(env.object_storage_bin()).unwrap(),
+            data_dir: Utf8PathBuf::from_path_buf(env.object_storage_data_dir()).unwrap(),
+            pemfile: Utf8PathBuf::from_path_buf(env.public_key_path.clone()).unwrap(),
+            port: env.object_storage.port,
+        }
+    }
+
+    fn config_path(&self) -> Utf8PathBuf {
+        self.data_dir.join("object_storage.json")
+    }
+
+    fn listen_addr(&self) -> Utf8PathBuf {
+        format!("127.0.0.1:{}", self.port).into()
+    }
+
+    pub fn init(&self) -> Result<()> {
+        println!("Initializing object storage in {:?}", self.data_dir);
+        let parent = self.data_dir.parent().unwrap();
+
+        #[derive(serde::Serialize)]
+        struct Cfg {
+            listen: Utf8PathBuf,
+            pemfile: Utf8PathBuf,
+            local_path: Utf8PathBuf,
+            r#type: String,
+        }
+        let cfg = Cfg {
+            listen: self.listen_addr(),
+            pemfile: parent.join(self.pemfile.clone()),
+            local_path: parent.join(OBJECT_STORAGE_REMOTE_STORAGE_DIR),
+            r#type: "LocalFs".to_string(),
+        };
+        std::fs::create_dir_all(self.config_path().parent().unwrap())?;
+        std::fs::write(self.config_path(), serde_json::to_string(&cfg)?)
+            .context("write object storage config")?;
+        Ok(())
+    }
+
+    pub async fn start(&self, retry_timeout: &Duration) -> Result<()> {
+        println!("Starting s3 proxy at {}", self.listen_addr());
+        std::io::stdout().flush().context("flush stdout")?;
+
+        let process_status_check = || async {
+            tokio::time::sleep(Duration::from_millis(500)).await;
+            let res = reqwest::Client::new()
+                .get(format!("http://{}/metrics", self.listen_addr()))
+                .send()
+                .await;
+            match res {
+                Ok(response) if response.status().is_success() => Ok(true),
+                Ok(_) => Err(anyhow!("Failed to query /metrics")),
+                Err(e) => Err(anyhow!("Failed to check node status: {e}")),
+            }
+        };
+
+        let res = start_process(
+            "object_storage",
+            &self.data_dir.clone().into_std_path_buf(),
+            &self.bin.clone().into_std_path_buf(),
+            vec![self.config_path().to_string()],
+            vec![("RUST_LOG".into(), "debug".into())],
+            background_process::InitialPidFile::Create(self.pid_file()),
+            retry_timeout,
+            process_status_check,
+        )
+        .await;
+        if res.is_err() {
+            eprintln!("Logs:\n{}", std::fs::read_to_string(self.log_file())?);
+        }
+
+        res
+    }
+
+    pub fn stop(&self, immediate: bool) -> anyhow::Result<()> {
+        stop_process(immediate, "object_storage", &self.pid_file())
+    }
+
+    fn log_file(&self) -> Utf8PathBuf {
+        self.data_dir.join("object_storage.log")
+    }
+
+    fn pid_file(&self) -> Utf8PathBuf {
+        self.data_dir.join("object_storage.pid")
+    }
+}

control_plane/src/storage_controller.rs

@@ -13,7 +13,9 @@ use pageserver_api::controller_api::{
     NodeConfigureRequest, NodeDescribeResponse, NodeRegisterRequest, TenantCreateRequest,
     TenantCreateResponse, TenantLocateResponse,
 };
-use pageserver_api::models::{TenantConfigRequest, TimelineCreateRequest, TimelineInfo};
+use pageserver_api::models::{
+    TenantConfig, TenantConfigRequest, TimelineCreateRequest, TimelineInfo,
+};
 use pageserver_api::shard::TenantShardId;
 use pageserver_client::mgmt_api::ResponseErrorMessageExt;
 use postgres_backend::AuthType;
@@ -82,7 +84,8 @@ impl NeonStorageControllerStopArgs {
 pub struct AttachHookRequest {
     pub tenant_shard_id: TenantShardId,
     pub node_id: Option<NodeId>,
-    pub generation_override: Option<i32>,
+    pub generation_override: Option<i32>, // only new tenants
+    pub config: Option<TenantConfig>,     // only new tenants
 }
 
 #[derive(Serialize, Deserialize)]
@@ -805,6 +808,7 @@ impl StorageController {
             tenant_shard_id,
             node_id: Some(pageserver_id),
             generation_override: None,
+            config: None,
         };
 
         let response = self

control_plane/storcon_cli/src/main.rs

@@ -941,7 +941,7 @@ async fn main() -> anyhow::Result<()> {
             let mut node_to_fill_descs = Vec::new();
 
             for desc in node_descs {
-                let to_drain = nodes.iter().any(|id| *id == desc.id);
+                let to_drain = nodes.contains(&desc.id);
                 if to_drain {
                     node_to_drain_descs.push(desc);
                 } else {

docs/storage_controller.md

@@ -151,7 +151,7 @@ Example body:
 ```
 {
   "tenant_id": "1f359dd625e519a1a4e8d7509690f6fc",
-  "stripe_size": 32768,
+  "stripe_size": 2048,
   "shards": [
       {"node_id": 344, "shard_number": 0},
       {"node_id": 722, "shard_number": 1},

libs/compute_api/src/requests.rs

@@ -5,6 +5,14 @@ use crate::privilege::Privilege;
 use crate::responses::ComputeCtlConfig;
 use crate::spec::{ComputeSpec, ExtVersion, PgIdent};
 
+/// When making requests to the `compute_ctl` external HTTP server, the client
+/// must specify a set of claims in `Authorization` header JWTs such that
+/// `compute_ctl` can authorize the request.
+#[derive(Clone, Debug, Deserialize, Serialize)]
+pub struct ComputeClaims {
+    pub compute_id: String,
+}
+
 /// Request of the /configure API
 ///
 /// We now pass only `spec` in the configuration request, but later we can
@@ -30,9 +38,3 @@ pub struct SetRoleGrantsRequest {
     pub privileges: Vec<Privilege>,
     pub role: PgIdent,
 }
-
-/// Request of the /configure_telemetry API
-#[derive(Debug, Deserialize, Serialize)]
-pub struct ConfigureTelemetryRequest {
-    pub logs_export_host: Option<String>,
-}

libs/compute_api/src/spec.rs

@@ -104,6 +104,12 @@ pub struct ComputeSpec {
     pub timeline_id: Option<TimelineId>,
     pub pageserver_connstring: Option<String>,
 
+    // More neon ids that we expose to the compute_ctl
+    // and to postgres as neon extension GUCs.
+    pub project_id: Option<String>,
+    pub branch_id: Option<String>,
+    pub endpoint_id: Option<String>,
+
     /// Safekeeper membership config generation. It is put in
     /// neon.safekeepers GUC and serves two purposes:
     /// 1) Non zero value forces walproposer to use membership configurations.
@@ -159,15 +165,13 @@ pub struct ComputeSpec {
     #[serde(default)] // Default false
     pub drop_subscriptions_before_start: bool,
 
-    /// Log level for audit logging:
-    ///
-    /// Disabled - no audit logging. This is the default.
-    /// log - log masked statements to the postgres log using pgaudit extension
-    /// hipaa - log unmasked statements to the file using pgaudit and pgauditlogtofile extension
-    ///
-    /// Extensions should be present in shared_preload_libraries
+    /// Log level for compute audit logging
     #[serde(default)]
     pub audit_log_level: ComputeAudit,
+
+    /// Hostname and the port of the otel collector. Leave empty to disable Postgres logs forwarding.
+    /// Example: config-shy-breeze-123-collector-monitoring.neon-telemetry.svc.cluster.local:10514
+    pub logs_export_host: Option<String>,
 }
 
 /// Feature flag to signal `compute_ctl` to enable certain experimental functionality.
@@ -179,9 +183,6 @@ pub enum ComputeFeature {
     /// track short-lived connections as user activity.
     ActivityMonitorExperimental,
 
-    /// Allow to configure rsyslog for Postgres logs export
-    PostgresLogsExport,
-
     /// This is a special feature flag that is used to represent unknown feature flags.
     /// Basically all unknown to enum flags are represented as this one. See unit test
     /// `parse_unknown_features()` for more details.
@@ -288,14 +289,25 @@ impl ComputeMode {
 }
 
 /// Log level for audit logging
-/// Disabled, log, hipaa
-/// Default is Disabled
 #[derive(Clone, Debug, Default, Eq, PartialEq, Deserialize, Serialize)]
 pub enum ComputeAudit {
     #[default]
     Disabled,
+    // Deprecated, use Base instead
     Log,
+    // (pgaudit.log = 'ddl', pgaudit.log_parameter='off')
+    // logged to the standard postgresql log stream
+    Base,
+    // Deprecated, use Full or Extended instead
     Hipaa,
+    // (pgaudit.log = 'all, -misc', pgaudit.log_parameter='off')
+    // logged to separate files collected by rsyslog
+    // into dedicated log storage with strict access
+    Extended,
+    // (pgaudit.log='all', pgaudit.log_parameter='on'),
+    // logged to separate files collected by rsyslog
+    // into dedicated log storage with strict access.
+    Full,
 }
 
 #[derive(Clone, Debug, Default, Deserialize, Serialize, PartialEq, Eq)]

libs/http-utils/Cargo.toml

@@ -30,6 +30,7 @@ tokio.workspace = true
 tracing.workspace = true
 url.workspace = true
 uuid.workspace = true
+x509-cert.workspace = true
 
 # to use tokio channels as streams, this is faster to compile than async_stream
 # why is it only here? no other crate should use it, streams are rarely needed.

libs/http-utils/src/server.rs

@@ -4,6 +4,8 @@ use futures::StreamExt;
 use futures::stream::FuturesUnordered;
 use hyper0::Body;
 use hyper0::server::conn::Http;
+use metrics::{IntCounterVec, register_int_counter_vec};
+use once_cell::sync::Lazy;
 use routerify::{RequestService, RequestServiceBuilder};
 use tokio::io::{AsyncRead, AsyncWrite};
 use tokio_rustls::TlsAcceptor;
@@ -26,6 +28,24 @@ pub struct Server {
     tls_acceptor: Option<TlsAcceptor>,
 }
 
+static CONNECTION_STARTED_COUNT: Lazy<IntCounterVec> = Lazy::new(|| {
+    register_int_counter_vec!(
+        "http_server_connection_started_total",
+        "Number of established http/https connections",
+        &["scheme"]
+    )
+    .expect("failed to define a metric")
+});
+
+static CONNECTION_ERROR_COUNT: Lazy<IntCounterVec> = Lazy::new(|| {
+    register_int_counter_vec!(
+        "http_server_connection_errors_total",
+        "Number of occured connection errors by type",
+        &["type"]
+    )
+    .expect("failed to define a metric")
+});
+
 impl Server {
     pub fn new(
         request_service: Arc<RequestServiceBuilder<Body, ApiError>>,
@@ -60,6 +80,15 @@ impl Server {
             false
         }
 
+        let tcp_error_cnt = CONNECTION_ERROR_COUNT.with_label_values(&["tcp"]);
+        let tls_error_cnt = CONNECTION_ERROR_COUNT.with_label_values(&["tls"]);
+        let http_error_cnt = CONNECTION_ERROR_COUNT.with_label_values(&["http"]);
+        let https_error_cnt = CONNECTION_ERROR_COUNT.with_label_values(&["https"]);
+        let panic_error_cnt = CONNECTION_ERROR_COUNT.with_label_values(&["panic"]);
+
+        let http_connection_cnt = CONNECTION_STARTED_COUNT.with_label_values(&["http"]);
+        let https_connection_cnt = CONNECTION_STARTED_COUNT.with_label_values(&["https"]);
+
         let mut connections = FuturesUnordered::new();
         loop {
             tokio::select! {
@@ -67,6 +96,7 @@ impl Server {
                     let (tcp_stream, remote_addr) = match stream {
                         Ok(stream) => stream,
                         Err(err) => {
+                            tcp_error_cnt.inc();
                             if !suppress_io_error(&err) {
                                 info!("Failed to accept TCP connection: {err:#}");
                             }
@@ -78,11 +108,18 @@ impl Server {
                     let tls_acceptor = self.tls_acceptor.clone();
                     let cancel = cancel.clone();
 
+                    let tls_error_cnt = tls_error_cnt.clone();
+                    let http_error_cnt = http_error_cnt.clone();
+                    let https_error_cnt = https_error_cnt.clone();
+                    let http_connection_cnt = http_connection_cnt.clone();
+                    let https_connection_cnt = https_connection_cnt.clone();
+
                     connections.push(tokio::spawn(
                         async move {
                             match tls_acceptor {
                                 Some(tls_acceptor) => {
                                     // Handle HTTPS connection.
+                                    https_connection_cnt.inc();
                                     let tls_stream = tokio::select! {
                                         tls_stream = tls_acceptor.accept(tcp_stream) => tls_stream,
                                         _ = cancel.cancelled() => return,
@@ -90,6 +127,7 @@ impl Server {
                                     let tls_stream = match tls_stream {
                                         Ok(tls_stream) => tls_stream,
                                         Err(err) => {
+                                            tls_error_cnt.inc();
                                             if !suppress_io_error(&err) {
                                                 info!(%remote_addr, "Failed to accept TLS connection: {err:#}");
                                             }
@@ -97,6 +135,7 @@ impl Server {
                                         }
                                     };
                                     if let Err(err) = Self::serve_connection(tls_stream, service, cancel).await {
+                                        https_error_cnt.inc();
                                         if !suppress_hyper_error(&err) {
                                             info!(%remote_addr, "Failed to serve HTTPS connection: {err:#}");
                                         }
@@ -104,7 +143,9 @@ impl Server {
                                 }
                                 None => {
                                     // Handle HTTP connection.
+                                    http_connection_cnt.inc();
                                     if let Err(err) = Self::serve_connection(tcp_stream, service, cancel).await {
+                                        http_error_cnt.inc();
                                         if !suppress_hyper_error(&err) {
                                             info!(%remote_addr, "Failed to serve HTTP connection: {err:#}");
                                         }
@@ -115,6 +156,7 @@ impl Server {
                  }
                 Some(conn) = connections.next() => {
                     if let Err(err) = conn {
+                        panic_error_cnt.inc();
                         error!("Connection panicked: {err:#}");
                     }
                 }
@@ -122,6 +164,7 @@ impl Server {
                     // Wait for graceful shutdown of all connections.
                     while let Some(conn) = connections.next().await {
                         if let Err(err) = conn {
+                            panic_error_cnt.inc();
                             error!("Connection panicked: {err:#}");
                         }
                     }

libs/http-utils/src/tls_certs.rs

@@ -3,11 +3,14 @@ use std::{sync::Arc, time::Duration};
 use anyhow::Context;
 use arc_swap::ArcSwap;
 use camino::Utf8Path;
+use metrics::{IntCounterVec, UIntGaugeVec, register_int_counter_vec, register_uint_gauge_vec};
+use once_cell::sync::Lazy;
 use rustls::{
-    pki_types::{CertificateDer, PrivateKeyDer},
+    pki_types::{CertificateDer, PrivateKeyDer, UnixTime},
     server::{ClientHello, ResolvesServerCert},
     sign::CertifiedKey,
 };
+use x509_cert::der::Reader;
 
 pub async fn load_cert_chain(filename: &Utf8Path) -> anyhow::Result<Vec<CertificateDer<'static>>> {
     let cert_data = tokio::fs::read(filename)
@@ -53,6 +56,76 @@ pub async fn load_certified_key(
     Ok(certified_key)
 }
 
+/// rustls's CertifiedKey with extra parsed fields used for metrics.
+struct ParsedCertifiedKey {
+    certified_key: CertifiedKey,
+    expiration_time: UnixTime,
+}
+
+/// Parse expiration time from an X509 certificate.
+fn parse_expiration_time(cert: &CertificateDer<'_>) -> anyhow::Result<UnixTime> {
+    let parsed_cert = x509_cert::der::SliceReader::new(cert)
+        .context("Failed to parse cerficiate")?
+        .decode::<x509_cert::Certificate>()
+        .context("Failed to parse cerficiate")?;
+
+    Ok(UnixTime::since_unix_epoch(
+        parsed_cert
+            .tbs_certificate
+            .validity
+            .not_after
+            .to_unix_duration(),
+    ))
+}
+
+async fn load_and_parse_certified_key(
+    key_filename: &Utf8Path,
+    cert_filename: &Utf8Path,
+) -> anyhow::Result<ParsedCertifiedKey> {
+    let certified_key = load_certified_key(key_filename, cert_filename).await?;
+    let expiration_time = parse_expiration_time(certified_key.end_entity_cert()?)?;
+    Ok(ParsedCertifiedKey {
+        certified_key,
+        expiration_time,
+    })
+}
+
+static CERT_EXPIRATION_TIME: Lazy<UIntGaugeVec> = Lazy::new(|| {
+    register_uint_gauge_vec!(
+        "tls_certs_expiration_time_seconds",
+        "Expiration time of the loaded certificate since unix epoch in seconds",
+        &["resolver_name"]
+    )
+    .expect("failed to define a metric")
+});
+
+static CERT_RELOAD_STARTED_COUNTER: Lazy<IntCounterVec> = Lazy::new(|| {
+    register_int_counter_vec!(
+        "tls_certs_reload_started_total",
+        "Number of certificate reload loop iterations started",
+        &["resolver_name"]
+    )
+    .expect("failed to define a metric")
+});
+
+static CERT_RELOAD_UPDATED_COUNTER: Lazy<IntCounterVec> = Lazy::new(|| {
+    register_int_counter_vec!(
+        "tls_certs_reload_updated_total",
+        "Number of times the certificate was updated to the new one",
+        &["resolver_name"]
+    )
+    .expect("failed to define a metric")
+});
+
+static CERT_RELOAD_FAILED_COUNTER: Lazy<IntCounterVec> = Lazy::new(|| {
+    register_int_counter_vec!(
+        "tls_certs_reload_failed_total",
+        "Number of times the certificate reload failed",
+        &["resolver_name"]
+    )
+    .expect("failed to define a metric")
+});
+
 /// Implementation of [`rustls::server::ResolvesServerCert`] which reloads certificates from
 /// the disk periodically.
 #[derive(Debug)]
@@ -63,16 +136,28 @@ pub struct ReloadingCertificateResolver {
 impl ReloadingCertificateResolver {
     /// Creates a new Resolver by loading certificate and private key from FS and
     /// creating tokio::task to reload them with provided reload_period.
+    /// resolver_name is used as metric's label.
     pub async fn new(
+        resolver_name: &str,
         key_filename: &Utf8Path,
         cert_filename: &Utf8Path,
         reload_period: Duration,
     ) -> anyhow::Result<Arc<Self>> {
+        // Create metrics for current resolver.
+        let cert_expiration_time = CERT_EXPIRATION_TIME.with_label_values(&[resolver_name]);
+        let cert_reload_started_counter =
+            CERT_RELOAD_STARTED_COUNTER.with_label_values(&[resolver_name]);
+        let cert_reload_updated_counter =
+            CERT_RELOAD_UPDATED_COUNTER.with_label_values(&[resolver_name]);
+        let cert_reload_failed_counter =
+            CERT_RELOAD_FAILED_COUNTER.with_label_values(&[resolver_name]);
+
+        let parsed_key = load_and_parse_certified_key(key_filename, cert_filename).await?;
+
         let this = Arc::new(Self {
-            certified_key: ArcSwap::from_pointee(
-                load_certified_key(key_filename, cert_filename).await?,
-            ),
+            certified_key: ArcSwap::from_pointee(parsed_key.certified_key),
         });
+        cert_expiration_time.set(parsed_key.expiration_time.as_secs());
 
         tokio::spawn({
             let weak_this = Arc::downgrade(&this);
@@ -88,17 +173,22 @@ impl ReloadingCertificateResolver {
                         Some(this) => this,
                         None => break, // Resolver has been destroyed, exit.
                     };
-                    match load_certified_key(&key_filename, &cert_filename).await {
-                        Ok(new_certified_key) => {
-                            if new_certified_key.cert == this.certified_key.load().cert {
+                    cert_reload_started_counter.inc();
+
+                    match load_and_parse_certified_key(&key_filename, &cert_filename).await {
+                        Ok(parsed_key) => {
+                            if parsed_key.certified_key.cert == this.certified_key.load().cert {
                                 tracing::debug!("Certificate has not changed since last reloading");
                             } else {
                                 tracing::info!("Certificate has been reloaded");
-                                this.certified_key.store(Arc::new(new_certified_key));
+                                this.certified_key.store(Arc::new(parsed_key.certified_key));
+                                cert_expiration_time.set(parsed_key.expiration_time.as_secs());
+                                cert_reload_updated_counter.inc();
                             }
                             last_reload_failed = false;
                         }
                         Err(err) => {
+                            cert_reload_failed_counter.inc();
                             // Note: Reloading certs may fail if it conflicts with the script updating
                             // the files at the same time. Warn only if the error is persistent.
                             if last_reload_failed {

libs/pageserver_api/src/config.rs

@@ -180,6 +180,7 @@ pub struct ConfigToml {
     #[serde(skip_serializing_if = "Option::is_none")]
     pub generate_unarchival_heatmap: Option<bool>,
     pub tracing: Option<Tracing>,
+    pub enable_tls_page_service_api: bool,
 }
 
 #[derive(Debug, Clone, PartialEq, Eq, serde::Serialize, serde::Deserialize)]
@@ -631,6 +632,7 @@ impl Default for ConfigToml {
             load_previous_heatmap: None,
             generate_unarchival_heatmap: None,
             tracing: None,
+            enable_tls_page_service_api: false,
         }
     }
 }

libs/pageserver_api/src/controller_api.rs

@@ -7,7 +7,8 @@ use std::time::{Duration, Instant};
 /// API (`/control/v1` prefix).  Implemented by the server
 /// in [`storage_controller::http`]
 use serde::{Deserialize, Serialize};
-use utils::id::{NodeId, TenantId};
+use utils::id::{NodeId, TenantId, TimelineId};
+use utils::lsn::Lsn;
 
 use crate::models::{PageserverUtilization, ShardParameters, TenantConfig};
 use crate::shard::{ShardStripeSize, TenantShardId};
@@ -499,6 +500,15 @@ pub struct SafekeeperSchedulingPolicyRequest {
     pub scheduling_policy: SkSchedulingPolicy,
 }
 
+/// Import request for safekeeper timelines.
+#[derive(Serialize, Deserialize, Clone)]
+pub struct TimelineImportRequest {
+    pub tenant_id: TenantId,
+    pub timeline_id: TimelineId,
+    pub start_lsn: Lsn,
+    pub sk_set: Vec<NodeId>,
+}
+
 #[cfg(test)]
 mod test {
     use serde_json;

libs/pageserver_api/src/key.rs

@@ -927,7 +927,7 @@ impl Key {
 
     /// Guaranteed to return `Ok()` if [`Self::is_rel_block_key`] returns `true` for `key`.
     #[inline(always)]
-    pub fn to_rel_block(self) -> anyhow::Result<(RelTag, BlockNumber)> {
+    pub fn to_rel_block(self) -> Result<(RelTag, BlockNumber), ToRelBlockError> {
         Ok(match self.field1 {
             0x00 => (
                 RelTag {
@@ -938,7 +938,7 @@ impl Key {
                 },
                 self.field6,
             ),
-            _ => anyhow::bail!("unexpected value kind 0x{:02x}", self.field1),
+            _ => return Err(ToRelBlockError(self.field1)),
         })
     }
 }
@@ -951,6 +951,17 @@ impl std::str::FromStr for Key {
     }
 }
 
+#[derive(Debug)]
+pub struct ToRelBlockError(u8);
+
+impl fmt::Display for ToRelBlockError {
+    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
+        write!(f, "unexpected value kind 0x{:02x}", self.0)
+    }
+}
+
+impl std::error::Error for ToRelBlockError {}
+
 #[cfg(test)]
 mod tests {
     use std::str::FromStr;

libs/pageserver_api/src/keyspace.rs

@@ -613,8 +613,7 @@ mod tests {
     use rand::{RngCore, SeedableRng};
 
     use super::*;
-    use crate::models::ShardParameters;
-    use crate::shard::{ShardCount, ShardNumber};
+    use crate::shard::{DEFAULT_STRIPE_SIZE, ShardCount, ShardNumber, ShardStripeSize};
 
     // Helper function to create a key range.
     //
@@ -964,12 +963,8 @@ mod tests {
     }
     #[test]
     fn sharded_range_relation_gap() {
-        let shard_identity = ShardIdentity::new(
-            ShardNumber(0),
-            ShardCount::new(4),
-            ShardParameters::DEFAULT_STRIPE_SIZE,
-        )
-        .unwrap();
+        let shard_identity =
+            ShardIdentity::new(ShardNumber(0), ShardCount::new(4), DEFAULT_STRIPE_SIZE).unwrap();
 
         let range = ShardedRange::new(
             Range {
@@ -985,12 +980,8 @@ mod tests {
 
     #[test]
     fn shard_identity_keyspaces_single_key() {
-        let shard_identity = ShardIdentity::new(
-            ShardNumber(1),
-            ShardCount::new(4),
-            ShardParameters::DEFAULT_STRIPE_SIZE,
-        )
-        .unwrap();
+        let shard_identity =
+            ShardIdentity::new(ShardNumber(1), ShardCount::new(4), DEFAULT_STRIPE_SIZE).unwrap();
 
         let range = ShardedRange::new(
             Range {
@@ -1034,12 +1025,8 @@ mod tests {
 
     #[test]
     fn shard_identity_keyspaces_forkno_gap() {
-        let shard_identity = ShardIdentity::new(
-            ShardNumber(1),
-            ShardCount::new(4),
-            ShardParameters::DEFAULT_STRIPE_SIZE,
-        )
-        .unwrap();
+        let shard_identity =
+            ShardIdentity::new(ShardNumber(1), ShardCount::new(4), DEFAULT_STRIPE_SIZE).unwrap();
 
         let range = ShardedRange::new(
             Range {
@@ -1061,7 +1048,7 @@ mod tests {
             let shard_identity = ShardIdentity::new(
                 ShardNumber(shard_number),
                 ShardCount::new(4),
-                ShardParameters::DEFAULT_STRIPE_SIZE,
+                DEFAULT_STRIPE_SIZE,
             )
             .unwrap();
 
@@ -1144,37 +1131,44 @@ mod tests {
     /// for a single tenant.
     #[test]
     fn sharded_range_fragment_simple() {
+        const SHARD_COUNT: u8 = 4;
+        const STRIPE_SIZE: u32 = DEFAULT_STRIPE_SIZE.0;
+
         let shard_identity = ShardIdentity::new(
             ShardNumber(0),
-            ShardCount::new(4),
-            ShardParameters::DEFAULT_STRIPE_SIZE,
+            ShardCount::new(SHARD_COUNT),
+            ShardStripeSize(STRIPE_SIZE),
         )
         .unwrap();
 
         // A range which we happen to know covers exactly one stripe which belongs to this shard
         let input_start = Key::from_hex("000000067f00000001000000ae0000000000").unwrap();
-        let input_end = Key::from_hex("000000067f00000001000000ae0000008000").unwrap();
+        let mut input_end = input_start;
+        input_end.field6 += STRIPE_SIZE; // field6 is block number
 
         // Ask for stripe_size blocks, we get the whole stripe
         assert_eq!(
-            do_fragment(input_start, input_end, &shard_identity, 32768),
-            (32768, vec![(32768, input_start..input_end)])
+            do_fragment(input_start, input_end, &shard_identity, STRIPE_SIZE),
+            (STRIPE_SIZE, vec![(STRIPE_SIZE, input_start..input_end)])
         );
 
         // Ask for more, we still get the whole stripe
         assert_eq!(
-            do_fragment(input_start, input_end, &shard_identity, 10000000),
-            (32768, vec![(32768, input_start..input_end)])
+            do_fragment(input_start, input_end, &shard_identity, 10 * STRIPE_SIZE),
+            (STRIPE_SIZE, vec![(STRIPE_SIZE, input_start..input_end)])
         );
 
         // Ask for target_nblocks of half the stripe size, we get two halves
         assert_eq!(
-            do_fragment(input_start, input_end, &shard_identity, 16384),
+            do_fragment(input_start, input_end, &shard_identity, STRIPE_SIZE / 2),
             (
-                32768,
+                STRIPE_SIZE,
                 vec![
-                    (16384, input_start..input_start.add(16384)),
-                    (16384, input_start.add(16384)..input_end)
+                    (
+                        STRIPE_SIZE / 2,
+                        input_start..input_start.add(STRIPE_SIZE / 2)
+                    ),
+                    (STRIPE_SIZE / 2, input_start.add(STRIPE_SIZE / 2)..input_end)
                 ]
             )
         );
@@ -1182,40 +1176,53 @@ mod tests {
 
     #[test]
     fn sharded_range_fragment_multi_stripe() {
+        const SHARD_COUNT: u8 = 4;
+        const STRIPE_SIZE: u32 = DEFAULT_STRIPE_SIZE.0;
+        const RANGE_SIZE: u32 = SHARD_COUNT as u32 * STRIPE_SIZE;
+
         let shard_identity = ShardIdentity::new(
             ShardNumber(0),
-            ShardCount::new(4),
-            ShardParameters::DEFAULT_STRIPE_SIZE,
+            ShardCount::new(SHARD_COUNT),
+            ShardStripeSize(STRIPE_SIZE),
         )
         .unwrap();
 
         // A range which covers multiple stripes, exactly one of which belongs to the current shard.
         let input_start = Key::from_hex("000000067f00000001000000ae0000000000").unwrap();
-        let input_end = Key::from_hex("000000067f00000001000000ae0000020000").unwrap();
+        let mut input_end = input_start;
+        input_end.field6 += RANGE_SIZE; // field6 is block number
+
         // Ask for all the blocks, get a fragment that covers the whole range but reports
         // its size to be just the blocks belonging to our shard.
         assert_eq!(
-            do_fragment(input_start, input_end, &shard_identity, 131072),
-            (32768, vec![(32768, input_start..input_end)])
+            do_fragment(input_start, input_end, &shard_identity, RANGE_SIZE),
+            (STRIPE_SIZE, vec![(STRIPE_SIZE, input_start..input_end)])
         );
 
-        // Ask for a sub-stripe quantity
+        // Ask for a sub-stripe quantity that results in 3 fragments.
+        let limit = STRIPE_SIZE / 3 + 1;
         assert_eq!(
-            do_fragment(input_start, input_end, &shard_identity, 16000),
+            do_fragment(input_start, input_end, &shard_identity, limit),
             (
-                32768,
+                STRIPE_SIZE,
                 vec![
-                    (16000, input_start..input_start.add(16000)),
-                    (16000, input_start.add(16000)..input_start.add(32000)),
-                    (768, input_start.add(32000)..input_end),
+                    (limit, input_start..input_start.add(limit)),
+                    (limit, input_start.add(limit)..input_start.add(2 * limit)),
+                    (
+                        STRIPE_SIZE - 2 * limit,
+                        input_start.add(2 * limit)..input_end
+                    ),
                 ]
             )
         );
 
         // Try on a range that starts slightly after our owned stripe
         assert_eq!(
-            do_fragment(input_start.add(1), input_end, &shard_identity, 131072),
-            (32767, vec![(32767, input_start.add(1)..input_end)])
+            do_fragment(input_start.add(1), input_end, &shard_identity, RANGE_SIZE),
+            (
+                STRIPE_SIZE - 1,
+                vec![(STRIPE_SIZE - 1, input_start.add(1)..input_end)]
+            )
         );
     }
 
@@ -1223,32 +1230,40 @@ mod tests {
     /// a previous relation.
     #[test]
     fn sharded_range_fragment_starting_from_logical_size() {
+        const SHARD_COUNT: u8 = 4;
+        const STRIPE_SIZE: u32 = DEFAULT_STRIPE_SIZE.0;
+        const RANGE_SIZE: u32 = SHARD_COUNT as u32 * STRIPE_SIZE;
+
         let input_start = Key::from_hex("000000067f00000001000000ae00ffffffff").unwrap();
-        let input_end = Key::from_hex("000000067f00000001000000ae0100008000").unwrap();
+        let mut input_end = Key::from_hex("000000067f00000001000000ae0100000000").unwrap();
+        input_end.field6 += RANGE_SIZE; // field6 is block number
 
         // Shard 0 owns the first stripe in the relation, and the preceding logical size is shard local too
         let shard_identity = ShardIdentity::new(
             ShardNumber(0),
-            ShardCount::new(4),
-            ShardParameters::DEFAULT_STRIPE_SIZE,
+            ShardCount::new(SHARD_COUNT),
+            ShardStripeSize(STRIPE_SIZE),
         )
         .unwrap();
         assert_eq!(
-            do_fragment(input_start, input_end, &shard_identity, 0x10000),
-            (0x8001, vec![(0x8001, input_start..input_end)])
+            do_fragment(input_start, input_end, &shard_identity, 2 * STRIPE_SIZE),
+            (
+                STRIPE_SIZE + 1,
+                vec![(STRIPE_SIZE + 1, input_start..input_end)]
+            )
         );
 
         // Shard 1 does not own the first stripe in the relation, but it does own the logical size (all shards
         // store all logical sizes)
         let shard_identity = ShardIdentity::new(
             ShardNumber(1),
-            ShardCount::new(4),
-            ShardParameters::DEFAULT_STRIPE_SIZE,
+            ShardCount::new(SHARD_COUNT),
+            ShardStripeSize(STRIPE_SIZE),
         )
         .unwrap();
         assert_eq!(
-            do_fragment(input_start, input_end, &shard_identity, 0x10000),
-            (0x1, vec![(0x1, input_start..input_end)])
+            do_fragment(input_start, input_end, &shard_identity, 2 * STRIPE_SIZE),
+            (1, vec![(1, input_start..input_end)])
         );
     }
 
@@ -1284,12 +1299,8 @@ mod tests {
         );
 
         // Same, but using a sharded identity
-        let shard_identity = ShardIdentity::new(
-            ShardNumber(0),
-            ShardCount::new(4),
-            ShardParameters::DEFAULT_STRIPE_SIZE,
-        )
-        .unwrap();
+        let shard_identity =
+            ShardIdentity::new(ShardNumber(0), ShardCount::new(4), DEFAULT_STRIPE_SIZE).unwrap();
         assert_eq!(
             do_fragment(input_start, input_end, &shard_identity, 0x8000),
             (u32::MAX, vec![(u32::MAX, input_start..input_end),])
@@ -1331,7 +1342,7 @@ mod tests {
                 ShardIdentity::new(
                     ShardNumber((prng.next_u32() % shard_count) as u8),
                     ShardCount::new(shard_count as u8),
-                    ShardParameters::DEFAULT_STRIPE_SIZE,
+                    DEFAULT_STRIPE_SIZE,
                 )
                 .unwrap()
             };

libs/pageserver_api/src/models.rs

@@ -26,7 +26,7 @@ use utils::{completion, serde_system_time};
 use crate::config::Ratio;
 use crate::key::{CompactKey, Key};
 use crate::reltag::RelTag;
-use crate::shard::{ShardCount, ShardStripeSize, TenantShardId};
+use crate::shard::{DEFAULT_STRIPE_SIZE, ShardCount, ShardStripeSize, TenantShardId};
 
 /// The state of a tenant in this pageserver.
 ///
@@ -80,10 +80,22 @@ pub enum TenantState {
     ///
     /// Transitions out of this state are possible through `set_broken()`.
     Stopping {
+        /// The barrier can be used to wait for shutdown to complete. The first caller to set
+        /// Some(Barrier) is responsible for driving shutdown to completion. Subsequent callers
+        /// will wait for the first caller's existing barrier.
+        ///
+        /// None is set when an attach is cancelled, to signal to shutdown that the attach has in
+        /// fact cancelled:
+        ///
+        /// 1. `shutdown` sees `TenantState::Attaching`, and cancels the tenant.
+        /// 2. `attach` sets `TenantState::Stopping(None)` and exits.
+        /// 3. `set_stopping` waits for `TenantState::Stopping(None)` and sets
+        ///    `TenantState::Stopping(Some)` to claim the barrier as the shutdown owner.
+        //
         // Because of https://github.com/serde-rs/serde/issues/2105 this has to be a named field,
         // otherwise it will not be skipped during deserialization
         #[serde(skip)]
-        progress: completion::Barrier,
+        progress: Option<completion::Barrier>,
     },
     /// The tenant is recognized by the pageserver, but can no longer be used for
     /// any operations.
@@ -426,8 +438,6 @@ pub struct ShardParameters {
 }
 
 impl ShardParameters {
-    pub const DEFAULT_STRIPE_SIZE: ShardStripeSize = ShardStripeSize(256 * 1024 / 8);
-
     pub fn is_unsharded(&self) -> bool {
         self.count.is_unsharded()
     }
@@ -437,7 +447,7 @@ impl Default for ShardParameters {
     fn default() -> Self {
         Self {
             count: ShardCount::new(0),
-            stripe_size: Self::DEFAULT_STRIPE_SIZE,
+            stripe_size: DEFAULT_STRIPE_SIZE,
         }
     }
 }
@@ -1668,6 +1678,7 @@ pub struct SecondaryProgress {
 pub struct TenantScanRemoteStorageShard {
     pub tenant_shard_id: TenantShardId,
     pub generation: Option<u32>,
+    pub stripe_size: Option<ShardStripeSize>,
 }
 
 #[derive(Serialize, Deserialize, Debug, Default)]
@@ -2719,10 +2730,15 @@ mod tests {
                 "Activating",
             ),
             (line!(), TenantState::Active, "Active"),
+            (
+                line!(),
+                TenantState::Stopping { progress: None },
+                "Stopping",
+            ),
             (
                 line!(),
                 TenantState::Stopping {
-                    progress: utils::completion::Barrier::default(),
+                    progress: Some(completion::Barrier::default()),
                 },
                 "Stopping",
             ),

libs/pageserver_api/src/record.rs

@@ -58,6 +58,8 @@ pub enum NeonWalRecord {
         /// to true. This record does not need the history WALs to reconstruct. See [`NeonWalRecord::will_init`] and
         /// its references in `timeline.rs`.
         will_init: bool,
+        /// Only append the record if the current image is the same as the one specified in this field.
+        only_if: Option<String>,
     },
 }
 
@@ -81,6 +83,17 @@ impl NeonWalRecord {
             append: s.as_ref().to_string(),
             clear: false,
             will_init: false,
+            only_if: None,
+        }
+    }
+
+    #[cfg(feature = "testing")]
+    pub fn wal_append_conditional(s: impl AsRef<str>, only_if: impl AsRef<str>) -> Self {
+        Self::Test {
+            append: s.as_ref().to_string(),
+            clear: false,
+            will_init: false,
+            only_if: Some(only_if.as_ref().to_string()),
         }
     }
 
@@ -90,6 +103,7 @@ impl NeonWalRecord {
             append: s.as_ref().to_string(),
             clear: true,
             will_init: false,
+            only_if: None,
         }
     }
 
@@ -99,6 +113,7 @@ impl NeonWalRecord {
             append: s.as_ref().to_string(),
             clear: true,
             will_init: true,
+            only_if: None,
         }
     }
 }

libs/pageserver_api/src/shard.rs

@@ -78,6 +78,12 @@ impl Default for ShardStripeSize {
     }
 }
 
+impl std::fmt::Display for ShardStripeSize {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        self.0.fmt(f)
+    }
+}
+
 /// Layout version: for future upgrades where we might change how the key->shard mapping works
 #[derive(Clone, Copy, Serialize, Deserialize, Eq, PartialEq, Hash, Debug)]
 pub struct ShardLayout(u8);
@@ -86,8 +92,11 @@ const LAYOUT_V1: ShardLayout = ShardLayout(1);
 /// ShardIdentity uses a magic layout value to indicate if it is unusable
 const LAYOUT_BROKEN: ShardLayout = ShardLayout(255);
 
-/// Default stripe size in pages: 256MiB divided by 8kiB page size.
-const DEFAULT_STRIPE_SIZE: ShardStripeSize = ShardStripeSize(256 * 1024 / 8);
+/// The default stripe size in pages. 16 MiB divided by 8 kiB page size.
+///
+/// A lower stripe size distributes ingest load better across shards, but reduces IO amortization.
+/// 16 MiB appears to be a reasonable balance: <https://github.com/neondatabase/neon/pull/10510>.
+pub const DEFAULT_STRIPE_SIZE: ShardStripeSize = ShardStripeSize(16 * 1024 / 8);
 
 #[derive(thiserror::Error, Debug, PartialEq, Eq)]
 pub enum ShardConfigError {
@@ -537,7 +546,7 @@ mod tests {
             field6: 0x7d06,
         };
 
-        let shard = key_to_shard_number(ShardCount(10), DEFAULT_STRIPE_SIZE, &key);
+        let shard = key_to_shard_number(ShardCount(10), ShardStripeSize(32768), &key);
         assert_eq!(shard, ShardNumber(8));
     }
 

libs/postgres_backend/src/lib.rs

@@ -5,7 +5,6 @@
 #![deny(unsafe_code)]
 #![deny(clippy::undocumented_unsafe_blocks)]
 use std::future::Future;
-use std::io::ErrorKind;
 use std::net::SocketAddr;
 use std::os::fd::{AsRawFd, RawFd};
 use std::pin::Pin;
@@ -227,7 +226,7 @@ impl<IO: AsyncRead + AsyncWrite + Unpin> MaybeWriteOnly<IO> {
         match self {
             MaybeWriteOnly::Full(framed) => framed.read_startup_message().await,
             MaybeWriteOnly::WriteOnly(_) => {
-                Err(io::Error::new(ErrorKind::Other, "reading from write only half").into())
+                Err(io::Error::other("reading from write only half").into())
             }
             MaybeWriteOnly::Broken => panic!("IO on invalid MaybeWriteOnly"),
         }
@@ -237,7 +236,7 @@ impl<IO: AsyncRead + AsyncWrite + Unpin> MaybeWriteOnly<IO> {
         match self {
             MaybeWriteOnly::Full(framed) => framed.read_message().await,
             MaybeWriteOnly::WriteOnly(_) => {
-                Err(io::Error::new(ErrorKind::Other, "reading from write only half").into())
+                Err(io::Error::other("reading from write only half").into())
             }
             MaybeWriteOnly::Broken => panic!("IO on invalid MaybeWriteOnly"),
         }
@@ -975,7 +974,7 @@ impl<IO: AsyncRead + AsyncWrite + Unpin> AsyncWrite for CopyDataWriter<'_, IO> {
             .write_message_noflush(&BeMessage::CopyData(buf))
             // write_message only writes to the buffer, so it can fail iff the
             // message is invaid, but CopyData can't be invalid.
-            .map_err(|_| io::Error::new(ErrorKind::Other, "failed to serialize CopyData"))?;
+            .map_err(|_| io::Error::other("failed to serialize CopyData"))?;
 
         Poll::Ready(Ok(buf.len()))
     }

libs/postgres_backend/tests/simple_select.rs

@@ -85,8 +85,8 @@ static KEY: Lazy<rustls::pki_types::PrivateKeyDer<'static>> = Lazy::new(|| {
 
 static CERT: Lazy<rustls::pki_types::CertificateDer<'static>> = Lazy::new(|| {
     let mut cursor = Cursor::new(include_bytes!("cert.pem"));
-    let cert = rustls_pemfile::certs(&mut cursor).next().unwrap().unwrap();
-    cert
+
+    rustls_pemfile::certs(&mut cursor).next().unwrap().unwrap()
 });
 
 // test that basic select with ssl works

libs/pq_proto/src/framed.rs

@@ -35,7 +35,7 @@ impl ConnectionError {
     pub fn into_io_error(self) -> io::Error {
         match self {
             ConnectionError::Io(io) => io,
-            ConnectionError::Protocol(pe) => io::Error::new(io::ErrorKind::Other, pe.to_string()),
+            ConnectionError::Protocol(pe) => io::Error::other(pe.to_string()),
         }
     }
 }

libs/pq_proto/src/lib.rs

@@ -257,7 +257,7 @@ pub enum ProtocolError {
 impl ProtocolError {
     /// Proxy stream.rs uses only io::Error; provide it.
     pub fn into_io_error(self) -> io::Error {
-        io::Error::new(io::ErrorKind::Other, self.to_string())
+        io::Error::other(self.to_string())
     }
 }
 

libs/proxy/postgres-protocol2/src/authentication/sasl.rs

@@ -212,7 +212,7 @@ impl ScramSha256 {
                     password,
                     channel_binding,
                 } => (nonce, password, channel_binding),
-                _ => return Err(io::Error::new(io::ErrorKind::Other, "invalid SCRAM state")),
+                _ => return Err(io::Error::other("invalid SCRAM state")),
             };
 
         let message =
@@ -291,7 +291,7 @@ impl ScramSha256 {
                 server_key,
                 auth_message,
             } => (server_key, auth_message),
-            _ => return Err(io::Error::new(io::ErrorKind::Other, "invalid SCRAM state")),
+            _ => return Err(io::Error::other("invalid SCRAM state")),
         };
 
         let message =
@@ -301,10 +301,7 @@ impl ScramSha256 {
 
         let verifier = match parsed {
             ServerFinalMessage::Error(e) => {
-                return Err(io::Error::new(
-                    io::ErrorKind::Other,
-                    format!("SCRAM error: {}", e),
-                ));
+                return Err(io::Error::other(format!("SCRAM error: {}", e)));
             }
             ServerFinalMessage::Verifier(verifier) => verifier,
         };

libs/remote_storage/Cargo.toml

@@ -28,7 +28,7 @@ toml_edit.workspace = true
 tracing.workspace = true
 scopeguard.workspace = true
 metrics.workspace = true
-utils.workspace = true
+utils = { path = "../utils", default-features = false }
 pin-project-lite.workspace = true
 
 azure_core.workspace = true

libs/remote_storage/src/azure_blob.rs

@@ -801,8 +801,7 @@ where
             // that support needs to be hacked in.
             //
             // including {self:?} into the message would be useful, but unsure how to unproject.
-            _ => std::task::Poll::Ready(Err(std::io::Error::new(
-                std::io::ErrorKind::Other,
+            _ => std::task::Poll::Ready(Err(std::io::Error::other(
                 "cloned or initial values cannot be read",
             ))),
         }
@@ -855,7 +854,7 @@ where
         };
         Err(azure_core::error::Error::new(
             azure_core::error::ErrorKind::Io,
-            std::io::Error::new(std::io::ErrorKind::Other, msg),
+            std::io::Error::other(msg),
         ))
     }
 

libs/utils/Cargo.toml

@@ -5,7 +5,8 @@ edition.workspace = true
 license.workspace = true
 
 [features]
-default = []
+default = ["rename_noreplace"]
+rename_noreplace = []
 # Enables test-only APIs, incuding failpoints. In particular, enables the `fail_point!` macro,
 # which adds some runtime cost to run tests on outage conditions
 testing = ["fail/failpoints"]
@@ -35,7 +36,7 @@ serde_with.workspace = true
 serde_json.workspace = true
 signal-hook.workspace = true
 thiserror.workspace = true
-tokio.workspace = true
+tokio = { workspace = true, features = ["signal"] }
 tokio-tar.workspace = true
 tokio-util.workspace = true
 toml_edit = { workspace = true, features = ["serde"] }

libs/utils/src/auth.rs

@@ -173,7 +173,7 @@ impl std::fmt::Debug for JwtAuth {
 }
 
 // this function is used only for testing purposes in CLI e g generate tokens during init
-pub fn encode_from_key_file(claims: &Claims, key_data: &[u8]) -> Result<String> {
+pub fn encode_from_key_file<S: Serialize>(claims: &S, key_data: &[u8]) -> Result<String> {
     let key = EncodingKey::from_ed_pem(key_data)?;
     Ok(encode(&Header::new(STORAGE_TOKEN_ALGORITHM), claims, &key)?)
 }

libs/utils/src/crashsafe.rs

@@ -81,12 +81,9 @@ pub fn path_with_suffix_extension(
 }
 
 pub fn fsync_file_and_parent(file_path: &Utf8Path) -> io::Result<()> {
-    let parent = file_path.parent().ok_or_else(|| {
-        io::Error::new(
-            io::ErrorKind::Other,
-            format!("File {file_path:?} has no parent"),
-        )
-    })?;
+    let parent = file_path
+        .parent()
+        .ok_or_else(|| io::Error::other(format!("File {file_path:?} has no parent")))?;
 
     fsync(file_path)?;
     fsync(parent)?;

libs/utils/src/fs_ext.rs

@@ -3,7 +3,9 @@ use std::{fs, io, path::Path};
 
 use anyhow::Context;
 
+#[cfg(feature = "rename_noreplace")]
 mod rename_noreplace;
+#[cfg(feature = "rename_noreplace")]
 pub use rename_noreplace::rename_noreplace;
 
 pub trait PathExt {

libs/utils/src/fs_ext/rename_noreplace.rs

@@ -8,7 +8,7 @@ pub fn rename_noreplace<P1: ?Sized + NixPath, P2: ?Sized + NixPath>(
     dst: &P2,
 ) -> nix::Result<()> {
     {
-        #[cfg(target_os = "linux")]
+        #[cfg(all(target_os = "linux", target_env = "gnu"))]
         {
             nix::fcntl::renameat2(
                 None,
@@ -29,7 +29,7 @@ pub fn rename_noreplace<P1: ?Sized + NixPath, P2: ?Sized + NixPath>(
             })??;
             nix::errno::Errno::result(res).map(drop)
         }
-        #[cfg(not(any(target_os = "linux", target_os = "macos")))]
+        #[cfg(not(any(all(target_os = "linux", target_env = "gnu"), target_os = "macos")))]
         {
             std::compile_error!("OS does not support no-replace renames");
         }

libs/utils/src/signals.rs

@@ -1,6 +1,8 @@
 pub use signal_hook::consts::TERM_SIGNALS;
 pub use signal_hook::consts::signal::*;
 use signal_hook::iterator::Signals;
+use tokio::signal::unix::{SignalKind, signal};
+use tracing::info;
 
 pub enum Signal {
     Quit,
@@ -36,3 +38,30 @@ impl ShutdownSignals {
         Ok(())
     }
 }
+
+/// Runs in a loop since we want to be responsive to multiple signals
+/// even after triggering shutdown (e.g. a SIGQUIT after a slow SIGTERM shutdown)
+/// <https://github.com/neondatabase/neon/issues/9740>
+pub async fn signal_handler(token: tokio_util::sync::CancellationToken) {
+    let mut sigint = signal(SignalKind::interrupt()).unwrap();
+    let mut sigterm = signal(SignalKind::terminate()).unwrap();
+    let mut sigquit = signal(SignalKind::quit()).unwrap();
+
+    loop {
+        let signal = tokio::select! {
+            _ = sigquit.recv() => {
+                info!("Got signal SIGQUIT. Terminating in immediate shutdown mode.");
+                std::process::exit(111);
+            }
+            _ = sigint.recv() => "SIGINT",
+            _ = sigterm.recv() => "SIGTERM",
+        };
+
+        if !token.is_cancelled() {
+            info!("Got signal {signal}. Terminating gracefully in fast shutdown mode.");
+            token.cancel();
+        } else {
+            info!("Got signal {signal}. Already shutting down.");
+        }
+    }
+}

object_storage/Cargo.toml

@@ -0,0 +1,28 @@
+[package]
+name = "object_storage"
+version = "0.0.1"
+edition.workspace = true
+license.workspace = true
+[dependencies]
+anyhow.workspace = true
+axum-extra.workspace = true
+axum.workspace = true
+camino.workspace = true
+futures.workspace = true
+jsonwebtoken.workspace = true
+prometheus.workspace = true
+remote_storage.workspace = true
+serde.workspace = true
+serde_json.workspace = true
+tokio-util.workspace = true
+tokio.workspace = true
+tracing.workspace = true
+utils = { path = "../libs/utils", default-features = false }
+workspace_hack.workspace = true
+[dev-dependencies]
+camino-tempfile.workspace = true
+http-body-util.workspace = true
+itertools.workspace = true
+rand.workspace = true
+test-log.workspace = true
+tower.workspace = true

object_storage/src/app.rs

@@ -0,0 +1,561 @@
+use anyhow::anyhow;
+use axum::body::{Body, Bytes};
+use axum::response::{IntoResponse, Response};
+use axum::{Router, http::StatusCode};
+use object_storage::{PrefixS3Path, S3Path, Storage, bad_request, internal_error, not_found, ok};
+use remote_storage::TimeoutOrCancel;
+use remote_storage::{DownloadError, DownloadOpts, GenericRemoteStorage, RemotePath};
+use std::{sync::Arc, time::SystemTime, time::UNIX_EPOCH};
+use tokio_util::sync::CancellationToken;
+use tracing::{error, info};
+use utils::backoff::retry;
+
+pub fn app(state: Arc<Storage>) -> Router<()> {
+    use axum::routing::{delete as _delete, get as _get};
+    let delete_prefix = _delete(delete_prefix);
+    Router::new()
+        .route(
+            "/{tenant_id}/{timeline_id}/{endpoint_id}/{*path}",
+            _get(get).put(set).delete(delete),
+        )
+        .route(
+            "/{tenant_id}/{timeline_id}/{endpoint_id}",
+            delete_prefix.clone(),
+        )
+        .route("/{tenant_id}/{timeline_id}", delete_prefix.clone())
+        .route("/{tenant_id}", delete_prefix)
+        .route("/metrics", _get(metrics))
+        .route("/status", _get(async || StatusCode::OK.into_response()))
+        .with_state(state)
+}
+
+type Result = anyhow::Result<Response, Response>;
+type State = axum::extract::State<Arc<Storage>>;
+
+const CONTENT_TYPE: &str = "content-type";
+const APPLICATION_OCTET_STREAM: &str = "application/octet-stream";
+const WARN_THRESHOLD: u32 = 3;
+const MAX_RETRIES: u32 = 10;
+
+async fn metrics() -> Result {
+    prometheus::TextEncoder::new()
+        .encode_to_string(&prometheus::gather())
+        .map(|s| s.into_response())
+        .map_err(|e| internal_error(e, "/metrics", "collecting metrics"))
+}
+
+async fn get(S3Path { path }: S3Path, state: State) -> Result {
+    info!(%path, "downloading");
+    let download_err = |e| {
+        if let DownloadError::NotFound = e {
+            info!(%path, %e, "downloading"); // 404 is not an issue of _this_ service
+            return not_found(&path);
+        }
+        internal_error(e, &path, "downloading")
+    };
+    let cancel = state.cancel.clone();
+    let opts = &DownloadOpts::default();
+
+    let stream = retry(
+        async || state.storage.download(&path, opts, &cancel).await,
+        DownloadError::is_permanent,
+        WARN_THRESHOLD,
+        MAX_RETRIES,
+        "downloading",
+        &cancel,
+    )
+    .await
+    .unwrap_or(Err(DownloadError::Cancelled))
+    .map_err(download_err)?
+    .download_stream;
+
+    Response::builder()
+        .status(StatusCode::OK)
+        .header(CONTENT_TYPE, APPLICATION_OCTET_STREAM)
+        .body(Body::from_stream(stream))
+        .map_err(|e| internal_error(e, path, "reading response"))
+}
+
+// Best solution for files is multipart upload, but remote_storage doesn't support it,
+// so we can either read Bytes in memory and push at once or forward BodyDataStream to
+// remote_storage. The latter may seem more peformant, but BodyDataStream doesn't have a
+// guaranteed size() which may produce issues while uploading to s3.
+// So, currently we're going with an in-memory copy plus a boundary to prevent uploading
+// very large files.
+async fn set(S3Path { path }: S3Path, state: State, bytes: Bytes) -> Result {
+    info!(%path, "uploading");
+    let request_len = bytes.len();
+    let max_len = state.max_upload_file_limit;
+    if request_len > max_len {
+        return Err(bad_request(
+            anyhow!("File size {request_len} exceeds max {max_len}"),
+            "uploading",
+        ));
+    }
+
+    let cancel = state.cancel.clone();
+    let fun = async || {
+        let stream = bytes_to_stream(bytes.clone());
+        state
+            .storage
+            .upload(stream, request_len, &path, None, &cancel)
+            .await
+    };
+    retry(
+        fun,
+        TimeoutOrCancel::caused_by_cancel,
+        WARN_THRESHOLD,
+        MAX_RETRIES,
+        "uploading",
+        &cancel,
+    )
+    .await
+    .unwrap_or(Err(anyhow!("uploading cancelled")))
+    .map_err(|e| internal_error(e, path, "reading response"))?;
+    Ok(ok())
+}
+
+async fn delete(S3Path { path }: S3Path, state: State) -> Result {
+    info!(%path, "deleting");
+    let cancel = state.cancel.clone();
+    retry(
+        async || state.storage.delete(&path, &cancel).await,
+        TimeoutOrCancel::caused_by_cancel,
+        WARN_THRESHOLD,
+        MAX_RETRIES,
+        "deleting",
+        &cancel,
+    )
+    .await
+    .unwrap_or(Err(anyhow!("deleting cancelled")))
+    .map_err(|e| internal_error(e, path, "deleting"))?;
+    Ok(ok())
+}
+
+async fn delete_prefix(PrefixS3Path { path }: PrefixS3Path, state: State) -> Result {
+    info!(%path, "deleting prefix");
+    let cancel = state.cancel.clone();
+    retry(
+        async || state.storage.delete_prefix(&path, &cancel).await,
+        TimeoutOrCancel::caused_by_cancel,
+        WARN_THRESHOLD,
+        MAX_RETRIES,
+        "deleting prefix",
+        &cancel,
+    )
+    .await
+    .unwrap_or(Err(anyhow!("deleting prefix cancelled")))
+    .map_err(|e| internal_error(e, path, "deleting prefix"))?;
+    Ok(ok())
+}
+
+pub async fn check_storage_permissions(
+    client: &GenericRemoteStorage,
+    cancel: CancellationToken,
+) -> anyhow::Result<()> {
+    info!("storage permissions check");
+
+    // as_nanos() as multiple instances proxying same bucket may be started at once
+    let now = SystemTime::now()
+        .duration_since(UNIX_EPOCH)?
+        .as_nanos()
+        .to_string();
+
+    let path = RemotePath::from_string(&format!("write_access_{now}"))?;
+    info!(%path, "uploading");
+
+    let body = now.to_string();
+    let stream = bytes_to_stream(Bytes::from(body.clone()));
+    client
+        .upload(stream, body.len(), &path, None, &cancel)
+        .await?;
+
+    use tokio::io::AsyncReadExt;
+    info!(%path, "downloading");
+    let download_opts = DownloadOpts {
+        kind: remote_storage::DownloadKind::Small,
+        ..Default::default()
+    };
+    let mut body_read_buf = Vec::new();
+    let stream = client
+        .download(&path, &download_opts, &cancel)
+        .await?
+        .download_stream;
+    tokio_util::io::StreamReader::new(stream)
+        .read_to_end(&mut body_read_buf)
+        .await?;
+    let body_read = String::from_utf8(body_read_buf)?;
+    if body != body_read {
+        error!(%body, %body_read, "File contents do not match");
+        anyhow::bail!("Read back file doesn't match original")
+    }
+
+    info!(%path, "removing");
+    client.delete(&path, &cancel).await
+}
+
+fn bytes_to_stream(bytes: Bytes) -> impl futures::Stream<Item = std::io::Result<Bytes>> {
+    futures::stream::once(futures::future::ready(Ok(bytes)))
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use axum::{body::Body, extract::Request, response::Response};
+    use http_body_util::BodyExt;
+    use itertools::iproduct;
+    use std::env::var;
+    use std::sync::Arc;
+    use std::time::Duration;
+    use test_log::test as testlog;
+    use tower::{Service, util::ServiceExt};
+    use utils::id::{TenantId, TimelineId};
+
+    // see libs/remote_storage/tests/test_real_s3.rs
+    const REAL_S3_ENV: &str = "ENABLE_REAL_S3_REMOTE_STORAGE";
+    const REAL_S3_BUCKET: &str = "REMOTE_STORAGE_S3_BUCKET";
+    const REAL_S3_REGION: &str = "REMOTE_STORAGE_S3_REGION";
+
+    async fn proxy() -> (Storage, Option<camino_tempfile::Utf8TempDir>) {
+        let cancel = CancellationToken::new();
+        let (dir, storage) = if var(REAL_S3_ENV).is_err() {
+            // tests execute in parallel and we need a new directory for each of them
+            let dir = camino_tempfile::tempdir().unwrap();
+            let fs =
+                remote_storage::LocalFs::new(dir.path().into(), Duration::from_secs(5)).unwrap();
+            (Some(dir), GenericRemoteStorage::LocalFs(fs))
+        } else {
+            // test_real_s3::create_s3_client is hard to reference, reimplementing here
+            let millis = SystemTime::now()
+                .duration_since(UNIX_EPOCH)
+                .unwrap()
+                .as_millis();
+            use rand::Rng;
+            let random = rand::thread_rng().r#gen::<u32>();
+
+            let s3_config = remote_storage::S3Config {
+                bucket_name: var(REAL_S3_BUCKET).unwrap(),
+                bucket_region: var(REAL_S3_REGION).unwrap(),
+                prefix_in_bucket: Some(format!("test_{millis}_{random:08x}/")),
+                endpoint: None,
+                concurrency_limit: std::num::NonZeroUsize::new(100).unwrap(),
+                max_keys_per_list_response: None,
+                upload_storage_class: None,
+            };
+            let bucket = remote_storage::S3Bucket::new(&s3_config, Duration::from_secs(1))
+                .await
+                .unwrap();
+            (None, GenericRemoteStorage::AwsS3(Arc::new(bucket)))
+        };
+
+        let proxy = Storage {
+            auth: object_storage::JwtAuth::new(TEST_PUB_KEY_ED25519).unwrap(),
+            storage,
+            cancel: cancel.clone(),
+            max_upload_file_limit: usize::MAX,
+        };
+        check_storage_permissions(&proxy.storage, cancel)
+            .await
+            .unwrap();
+        (proxy, dir)
+    }
+
+    // see libs/utils/src/auth.rs
+    const TEST_PUB_KEY_ED25519: &[u8] = b"
+-----BEGIN PUBLIC KEY-----
+MCowBQYDK2VwAyEARYwaNBayR+eGI0iXB4s3QxE3Nl2g1iWbr6KtLWeVD/w=
+-----END PUBLIC KEY-----
+";
+
+    const TEST_PRIV_KEY_ED25519: &[u8] = br#"
+-----BEGIN PRIVATE KEY-----
+MC4CAQAwBQYDK2VwBCIEID/Drmc1AA6U/znNRWpF3zEGegOATQxfkdWxitcOMsIH
+-----END PRIVATE KEY-----
+"#;
+
+    async fn request(req: Request<Body>) -> Response<Body> {
+        let (proxy, _) = proxy().await;
+        app(Arc::new(proxy))
+            .into_service()
+            .oneshot(req)
+            .await
+            .unwrap()
+    }
+
+    #[testlog(tokio::test)]
+    async fn status() {
+        let res = Request::builder()
+            .uri("/status")
+            .body(Body::empty())
+            .map(request)
+            .unwrap()
+            .await;
+        assert_eq!(res.status(), StatusCode::OK);
+    }
+
+    fn routes() -> impl Iterator<Item = (&'static str, &'static str)> {
+        iproduct!(
+            vec!["/1", "/1/2", "/1/2/3", "/1/2/3/4"],
+            vec!["GET", "PUT", "DELETE"]
+        )
+    }
+
+    #[testlog(tokio::test)]
+    async fn no_token() {
+        for (uri, method) in routes() {
+            info!(%uri, %method);
+            let res = Request::builder()
+                .uri(uri)
+                .method(method)
+                .body(Body::empty())
+                .map(request)
+                .unwrap()
+                .await;
+            assert!(matches!(
+                res.status(),
+                StatusCode::METHOD_NOT_ALLOWED | StatusCode::BAD_REQUEST
+            ));
+        }
+    }
+
+    #[testlog(tokio::test)]
+    async fn invalid_token() {
+        for (uri, method) in routes() {
+            info!(%uri, %method);
+            let status = Request::builder()
+                .uri(uri)
+                .header("Authorization", "Bearer 123")
+                .method(method)
+                .body(Body::empty())
+                .map(request)
+                .unwrap()
+                .await;
+            assert!(matches!(
+                status.status(),
+                StatusCode::METHOD_NOT_ALLOWED | StatusCode::BAD_REQUEST
+            ));
+        }
+    }
+
+    const TENANT_ID: TenantId =
+        TenantId::from_array([0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 1, 2, 3, 4, 5, 6]);
+    const TIMELINE_ID: TimelineId =
+        TimelineId::from_array([0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 1, 2, 3, 4, 5, 7]);
+    const ENDPOINT_ID: &str = "ep-winter-frost-a662z3vg";
+    fn token() -> String {
+        let claims = object_storage::Claims {
+            tenant_id: TENANT_ID,
+            timeline_id: TIMELINE_ID,
+            endpoint_id: ENDPOINT_ID.into(),
+            exp: u64::MAX,
+        };
+        let key = jsonwebtoken::EncodingKey::from_ed_pem(TEST_PRIV_KEY_ED25519).unwrap();
+        let header = jsonwebtoken::Header::new(object_storage::VALIDATION_ALGO);
+        jsonwebtoken::encode(&header, &claims, &key).unwrap()
+    }
+
+    #[testlog(tokio::test)]
+    async fn unauthorized() {
+        let (proxy, _) = proxy().await;
+        let mut app = app(Arc::new(proxy)).into_service();
+        let token = token();
+        let args = itertools::iproduct!(
+            vec![TENANT_ID.to_string(), TenantId::generate().to_string()],
+            vec![TIMELINE_ID.to_string(), TimelineId::generate().to_string()],
+            vec![ENDPOINT_ID, "ep-ololo"]
+        )
+        .skip(1);
+
+        for ((uri, method), (tenant, timeline, endpoint)) in iproduct!(routes(), args) {
+            info!(%uri, %method, %tenant, %timeline, %endpoint);
+            let request = Request::builder()
+                .uri(format!("/{tenant}/{timeline}/{endpoint}/sub/path/key"))
+                .method(method)
+                .header("Authorization", format!("Bearer {}", token))
+                .body(Body::empty())
+                .unwrap();
+            let status = ServiceExt::ready(&mut app)
+                .await
+                .unwrap()
+                .call(request)
+                .await
+                .unwrap()
+                .status();
+            assert_eq!(status, StatusCode::UNAUTHORIZED);
+        }
+    }
+
+    #[testlog(tokio::test)]
+    async fn method_not_allowed() {
+        let token = token();
+        let iter = iproduct!(vec!["", "/.."], vec!["GET", "PUT"]);
+        for (key, method) in iter {
+            let status = Request::builder()
+                .uri(format!("/{TENANT_ID}/{TIMELINE_ID}/{ENDPOINT_ID}{key}"))
+                .method(method)
+                .header("Authorization", format!("Bearer {token}"))
+                .body(Body::empty())
+                .map(request)
+                .unwrap()
+                .await
+                .status();
+            assert!(matches!(
+                status,
+                StatusCode::BAD_REQUEST | StatusCode::METHOD_NOT_ALLOWED
+            ));
+        }
+    }
+
+    async fn requests_chain(
+        chain: impl Iterator<Item = (String, &str, &'static str, StatusCode, bool)>,
+        token: impl Fn(&str) -> String,
+    ) {
+        let (proxy, _) = proxy().await;
+        let mut app = app(Arc::new(proxy)).into_service();
+        for (uri, method, body, expected_status, compare_body) in chain {
+            info!(%uri, %method, %body, %expected_status);
+            let bearer = format!("Bearer {}", token(&uri));
+            let request = Request::builder()
+                .uri(uri)
+                .method(method)
+                .header("Authorization", &bearer)
+                .body(Body::from(body))
+                .unwrap();
+            let response = ServiceExt::ready(&mut app)
+                .await
+                .unwrap()
+                .call(request)
+                .await
+                .unwrap();
+            assert_eq!(response.status(), expected_status);
+            if !compare_body {
+                continue;
+            }
+            let read_body = response.into_body().collect().await.unwrap().to_bytes();
+            assert_eq!(body, read_body);
+        }
+    }
+
+    #[testlog(tokio::test)]
+    async fn metrics() {
+        let uri = format!("/{TENANT_ID}/{TIMELINE_ID}/{ENDPOINT_ID}/key");
+        let req = vec![
+            (uri.clone(), "PUT", "body", StatusCode::OK, false),
+            (uri.clone(), "DELETE", "", StatusCode::OK, false),
+        ];
+        requests_chain(req.into_iter(), |_| token()).await;
+
+        let res = Request::builder()
+            .uri("/metrics")
+            .body(Body::empty())
+            .map(request)
+            .unwrap()
+            .await;
+        assert_eq!(res.status(), StatusCode::OK);
+        let body = res.into_body().collect().await.unwrap().to_bytes();
+        let body = String::from_utf8_lossy(&body);
+        tracing::debug!(%body);
+        // Storage metrics are not gathered for LocalFs
+        if var(REAL_S3_ENV).is_ok() {
+            assert!(body.contains("remote_storage_s3_deleted_objects_total"));
+        }
+        assert!(body.contains("process_threads"));
+    }
+
+    #[testlog(tokio::test)]
+    async fn insert_retrieve_remove() {
+        let uri = format!("/{TENANT_ID}/{TIMELINE_ID}/{ENDPOINT_ID}/key");
+        let chain = vec![
+            (uri.clone(), "GET", "", StatusCode::NOT_FOUND, false),
+            (uri.clone(), "PUT", "пыщьпыщь", StatusCode::OK, false),
+            (uri.clone(), "GET", "пыщьпыщь", StatusCode::OK, true),
+            (uri.clone(), "DELETE", "", StatusCode::OK, false),
+            (uri, "GET", "", StatusCode::NOT_FOUND, false),
+        ];
+        requests_chain(chain.into_iter(), |_| token()).await;
+    }
+
+    fn delete_prefix_token(uri: &str) -> String {
+        use serde::Serialize;
+        let parts = uri.split("/").collect::<Vec<&str>>();
+        #[derive(Serialize)]
+        struct PrefixClaims {
+            tenant_id: TenantId,
+            timeline_id: Option<TimelineId>,
+            endpoint_id: Option<object_storage::EndpointId>,
+            exp: u64,
+        }
+        let claims = PrefixClaims {
+            tenant_id: parts.get(1).map(|c| c.parse().unwrap()).unwrap(),
+            timeline_id: parts.get(2).map(|c| c.parse().unwrap()),
+            endpoint_id: parts.get(3).map(ToString::to_string),
+            exp: u64::MAX,
+        };
+        let key = jsonwebtoken::EncodingKey::from_ed_pem(TEST_PRIV_KEY_ED25519).unwrap();
+        let header = jsonwebtoken::Header::new(object_storage::VALIDATION_ALGO);
+        jsonwebtoken::encode(&header, &claims, &key).unwrap()
+    }
+
+    // Can't use single digit numbers as they won't be validated as TimelineId and EndpointId
+    #[testlog(tokio::test)]
+    async fn delete_prefix() {
+        let tenant_id =
+            TenantId::from_array([1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]).to_string();
+        let t2 = TimelineId::from_array([2, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]);
+        let t3 = TimelineId::from_array([3, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]);
+        let t4 = TimelineId::from_array([4, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]);
+        let f = |timeline, path| format!("/{tenant_id}/{timeline}{path}");
+        // Why extra slash in string literals? Axum is weird with URIs:
+        // /1/2 and 1/2/ match different routes, thus first yields OK and second NOT_FOUND
+        //  as it matches /tenant/timeline/endpoint, see https://stackoverflow.com/a/75355932
+        // The cost of removing trailing slash is suprisingly hard:
+        // * Add tower dependency with NormalizePath layer
+        // * wrap Router<()> in this layer https://github.com/tokio-rs/axum/discussions/2377
+        // * Rewrite make_service() -> into_make_service()
+        // * Rewrite oneshot() (not available for NormalizePath)
+        // I didn't manage to get it working correctly
+        let chain = vec![
+            // create 1/2/3/4, 1/2/3/5, delete prefix 1/2/3 -> empty
+            (f(t2, "/3/4"), "PUT", "", StatusCode::OK, false),
+            (f(t2, "/3/4"), "PUT", "", StatusCode::OK, false), // we can override file contents
+            (f(t2, "/3/5"), "PUT", "", StatusCode::OK, false),
+            (f(t2, "/3"), "DELETE", "", StatusCode::OK, false),
+            (f(t2, "/3/4"), "GET", "", StatusCode::NOT_FOUND, false),
+            (f(t2, "/3/5"), "GET", "", StatusCode::NOT_FOUND, false),
+            // create 1/2/3/4, 1/2/5/6, delete prefix 1/2/3 -> 1/2/5/6
+            (f(t2, "/3/4"), "PUT", "", StatusCode::OK, false),
+            (f(t2, "/5/6"), "PUT", "", StatusCode::OK, false),
+            (f(t2, "/3"), "DELETE", "", StatusCode::OK, false),
+            (f(t2, "/3/4"), "GET", "", StatusCode::NOT_FOUND, false),
+            (f(t2, "/5/6"), "GET", "", StatusCode::OK, false),
+            // create 1/2/3/4, 1/2/7/8, delete prefix 1/2 -> empty
+            (f(t2, "/3/4"), "PUT", "", StatusCode::OK, false),
+            (f(t2, "/7/8"), "PUT", "", StatusCode::OK, false),
+            (f(t2, ""), "DELETE", "", StatusCode::OK, false),
+            (f(t2, "/3/4"), "GET", "", StatusCode::NOT_FOUND, false),
+            (f(t2, "/7/8"), "GET", "", StatusCode::NOT_FOUND, false),
+            // create 1/2/3/4, 1/2/5/6, 1/3/8/9, delete prefix 1/2/3 -> 1/2/5/6, 1/3/8/9
+            (f(t2, "/3/4"), "PUT", "", StatusCode::OK, false),
+            (f(t2, "/5/6"), "PUT", "", StatusCode::OK, false),
+            (f(t3, "/8/9"), "PUT", "", StatusCode::OK, false),
+            (f(t2, "/3"), "DELETE", "", StatusCode::OK, false),
+            (f(t2, "/3/4"), "GET", "", StatusCode::NOT_FOUND, false),
+            (f(t2, "/5/6"), "GET", "", StatusCode::OK, false),
+            (f(t3, "/8/9"), "GET", "", StatusCode::OK, false),
+            // create 1/4/5/6, delete prefix 1/2 -> 1/3/8/9, 1/4/5/6
+            (f(t4, "/5/6"), "PUT", "", StatusCode::OK, false),
+            (f(t2, ""), "DELETE", "", StatusCode::OK, false),
+            (f(t2, "/3/4"), "GET", "", StatusCode::NOT_FOUND, false),
+            (f(t2, "/5/6"), "GET", "", StatusCode::NOT_FOUND, false),
+            (f(t3, "/8/9"), "GET", "", StatusCode::OK, false),
+            (f(t4, "/5/6"), "GET", "", StatusCode::OK, false),
+            // delete prefix 1 -> empty
+            (format!("/{tenant_id}"), "DELETE", "", StatusCode::OK, false),
+            (f(t2, "/3/4"), "GET", "", StatusCode::NOT_FOUND, false),
+            (f(t2, "/5/6"), "GET", "", StatusCode::NOT_FOUND, false),
+            (f(t3, "/8/9"), "GET", "", StatusCode::NOT_FOUND, false),
+            (f(t4, "/5/6"), "GET", "", StatusCode::NOT_FOUND, false),
+        ];
+        requests_chain(chain.into_iter(), delete_prefix_token).await;
+    }
+}

object_storage/src/lib.rs

@@ -0,0 +1,344 @@
+use anyhow::Result;
+use axum::extract::{FromRequestParts, Path};
+use axum::response::{IntoResponse, Response};
+use axum::{RequestPartsExt, http::StatusCode, http::request::Parts};
+use axum_extra::TypedHeader;
+use axum_extra::headers::{Authorization, authorization::Bearer};
+use camino::Utf8PathBuf;
+use jsonwebtoken::{DecodingKey, Validation};
+use remote_storage::{GenericRemoteStorage, RemotePath};
+use serde::{Deserialize, Serialize};
+use std::fmt::Display;
+use std::result::Result as StdResult;
+use std::sync::Arc;
+use tokio_util::sync::CancellationToken;
+use tracing::{debug, error};
+use utils::id::{TenantId, TimelineId};
+
+// simplified version of utils::auth::JwtAuth
+pub struct JwtAuth {
+    decoding_key: DecodingKey,
+    validation: Validation,
+}
+
+pub const VALIDATION_ALGO: jsonwebtoken::Algorithm = jsonwebtoken::Algorithm::EdDSA;
+impl JwtAuth {
+    pub fn new(key: &[u8]) -> Result<Self> {
+        Ok(Self {
+            decoding_key: DecodingKey::from_ed_pem(key)?,
+            validation: Validation::new(VALIDATION_ALGO),
+        })
+    }
+
+    pub fn decode<T: serde::de::DeserializeOwned>(&self, token: &str) -> Result<T> {
+        Ok(jsonwebtoken::decode(token, &self.decoding_key, &self.validation).map(|t| t.claims)?)
+    }
+}
+
+fn normalize_key(key: &str) -> StdResult<Utf8PathBuf, String> {
+    let key = clean_utf8(&Utf8PathBuf::from(key));
+    if key.starts_with("..") || key == "." || key == "/" {
+        return Err(format!("invalid key {key}"));
+    }
+    match key.strip_prefix("/").map(Utf8PathBuf::from) {
+        Ok(p) => Ok(p),
+        _ => Ok(key),
+    }
+}
+
+// Copied from path_clean crate with PathBuf->Utf8PathBuf
+fn clean_utf8(path: &camino::Utf8Path) -> Utf8PathBuf {
+    use camino::Utf8Component as Comp;
+    let mut out = Vec::new();
+    for comp in path.components() {
+        match comp {
+            Comp::CurDir => (),
+            Comp::ParentDir => match out.last() {
+                Some(Comp::RootDir) => (),
+                Some(Comp::Normal(_)) => {
+                    out.pop();
+                }
+                None | Some(Comp::CurDir) | Some(Comp::ParentDir) | Some(Comp::Prefix(_)) => {
+                    out.push(comp)
+                }
+            },
+            comp => out.push(comp),
+        }
+    }
+    if !out.is_empty() {
+        out.iter().collect()
+    } else {
+        Utf8PathBuf::from(".")
+    }
+}
+
+pub struct Storage {
+    pub auth: JwtAuth,
+    pub storage: GenericRemoteStorage,
+    pub cancel: CancellationToken,
+    pub max_upload_file_limit: usize,
+}
+
+pub type EndpointId = String; // If needed, reuse small string from proxy/src/types.rc
+
+#[derive(Deserialize, Serialize, PartialEq)]
+pub struct Claims {
+    pub tenant_id: TenantId,
+    pub timeline_id: TimelineId,
+    pub endpoint_id: EndpointId,
+    pub exp: u64,
+}
+
+impl Display for Claims {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        write!(
+            f,
+            "Claims(tenant_id {} timeline_id {} endpoint_id {} exp {})",
+            self.tenant_id, self.timeline_id, self.endpoint_id, self.exp
+        )
+    }
+}
+
+#[derive(Deserialize, Serialize)]
+struct KeyRequest {
+    tenant_id: TenantId,
+    timeline_id: TimelineId,
+    endpoint_id: EndpointId,
+    path: String,
+}
+
+#[derive(Debug, PartialEq)]
+pub struct S3Path {
+    pub path: RemotePath,
+}
+
+impl TryFrom<&KeyRequest> for S3Path {
+    type Error = String;
+    fn try_from(req: &KeyRequest) -> StdResult<Self, Self::Error> {
+        let KeyRequest {
+            tenant_id,
+            timeline_id,
+            endpoint_id,
+            path,
+        } = &req;
+        let prefix = format!("{tenant_id}/{timeline_id}/{endpoint_id}",);
+        let path = Utf8PathBuf::from(prefix).join(normalize_key(path)?);
+        let path = RemotePath::new(&path).unwrap(); // unwrap() because the path is already relative
+        Ok(S3Path { path })
+    }
+}
+
+fn unauthorized(route: impl Display, claims: impl Display) -> Response {
+    debug!(%route, %claims, "route doesn't match claims");
+    StatusCode::UNAUTHORIZED.into_response()
+}
+
+pub fn bad_request(err: impl Display, desc: &'static str) -> Response {
+    debug!(%err, desc);
+    (StatusCode::BAD_REQUEST, err.to_string()).into_response()
+}
+
+pub fn ok() -> Response {
+    StatusCode::OK.into_response()
+}
+
+pub fn internal_error(err: impl Display, path: impl Display, desc: &'static str) -> Response {
+    error!(%err, %path, desc);
+    StatusCode::INTERNAL_SERVER_ERROR.into_response()
+}
+
+pub fn not_found(key: impl ToString) -> Response {
+    (StatusCode::NOT_FOUND, key.to_string()).into_response()
+}
+
+impl FromRequestParts<Arc<Storage>> for S3Path {
+    type Rejection = Response;
+    async fn from_request_parts(
+        parts: &mut Parts,
+        state: &Arc<Storage>,
+    ) -> Result<Self, Self::Rejection> {
+        let Path(path): Path<KeyRequest> = parts
+            .extract()
+            .await
+            .map_err(|e| bad_request(e, "invalid route"))?;
+        let TypedHeader(Authorization(bearer)) = parts
+            .extract::<TypedHeader<Authorization<Bearer>>>()
+            .await
+            .map_err(|e| bad_request(e, "invalid token"))?;
+        let claims: Claims = state
+            .auth
+            .decode(bearer.token())
+            .map_err(|e| bad_request(e, "decoding token"))?;
+        let route = Claims {
+            tenant_id: path.tenant_id,
+            timeline_id: path.timeline_id,
+            endpoint_id: path.endpoint_id.clone(),
+            exp: claims.exp,
+        };
+        if route != claims {
+            return Err(unauthorized(route, claims));
+        }
+        (&path)
+            .try_into()
+            .map_err(|e| bad_request(e, "invalid route"))
+    }
+}
+
+#[derive(Deserialize, Serialize, PartialEq)]
+pub struct PrefixKeyPath {
+    pub tenant_id: TenantId,
+    pub timeline_id: Option<TimelineId>,
+    pub endpoint_id: Option<EndpointId>,
+}
+
+impl Display for PrefixKeyPath {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        write!(
+            f,
+            "PrefixKeyPath(tenant_id {} timeline_id {} endpoint_id {})",
+            self.tenant_id,
+            self.timeline_id
+                .as_ref()
+                .map(ToString::to_string)
+                .unwrap_or("".to_string()),
+            self.endpoint_id
+                .as_ref()
+                .map(ToString::to_string)
+                .unwrap_or("".to_string())
+        )
+    }
+}
+
+#[derive(Debug, PartialEq)]
+pub struct PrefixS3Path {
+    pub path: RemotePath,
+}
+
+impl From<&PrefixKeyPath> for PrefixS3Path {
+    fn from(path: &PrefixKeyPath) -> Self {
+        let timeline_id = path
+            .timeline_id
+            .as_ref()
+            .map(ToString::to_string)
+            .unwrap_or("".to_string());
+        let endpoint_id = path
+            .endpoint_id
+            .as_ref()
+            .map(ToString::to_string)
+            .unwrap_or("".to_string());
+        let path = Utf8PathBuf::from(path.tenant_id.to_string())
+            .join(timeline_id)
+            .join(endpoint_id);
+        let path = RemotePath::new(&path).unwrap(); // unwrap() because the path is already relative
+        PrefixS3Path { path }
+    }
+}
+
+impl FromRequestParts<Arc<Storage>> for PrefixS3Path {
+    type Rejection = Response;
+    async fn from_request_parts(
+        parts: &mut Parts,
+        state: &Arc<Storage>,
+    ) -> Result<Self, Self::Rejection> {
+        let Path(path) = parts
+            .extract::<Path<PrefixKeyPath>>()
+            .await
+            .map_err(|e| bad_request(e, "invalid route"))?;
+        let TypedHeader(Authorization(bearer)) = parts
+            .extract::<TypedHeader<Authorization<Bearer>>>()
+            .await
+            .map_err(|e| bad_request(e, "invalid token"))?;
+        let claims: PrefixKeyPath = state
+            .auth
+            .decode(bearer.token())
+            .map_err(|e| bad_request(e, "invalid token"))?;
+        if path != claims {
+            return Err(unauthorized(path, claims));
+        }
+        Ok((&path).into())
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn normalize_key() {
+        let f = super::normalize_key;
+        assert_eq!(f("hello/world/..").unwrap(), Utf8PathBuf::from("hello"));
+        assert_eq!(
+            f("ololo/1/../../not_ololo").unwrap(),
+            Utf8PathBuf::from("not_ololo")
+        );
+        assert!(f("ololo/1/../../../").is_err());
+        assert!(f(".").is_err());
+        assert!(f("../").is_err());
+        assert!(f("").is_err());
+        assert_eq!(f("/1/2/3").unwrap(), Utf8PathBuf::from("1/2/3"));
+        assert!(f("/1/2/3/../../../").is_err());
+        assert!(f("/1/2/3/../../../../").is_err());
+    }
+
+    const TENANT_ID: TenantId =
+        TenantId::from_array([1, 1, 2, 3, 4, 5, 6, 7, 8, 9, 1, 2, 3, 4, 5, 6]);
+    const TIMELINE_ID: TimelineId =
+        TimelineId::from_array([1, 1, 2, 3, 4, 5, 6, 7, 8, 9, 1, 2, 3, 4, 5, 7]);
+    const ENDPOINT_ID: &str = "ep-winter-frost-a662z3vg";
+
+    #[test]
+    fn s3_path() {
+        let auth = Claims {
+            tenant_id: TENANT_ID,
+            timeline_id: TIMELINE_ID,
+            endpoint_id: ENDPOINT_ID.into(),
+            exp: u64::MAX,
+        };
+        let s3_path = |key| {
+            let path = &format!("{TENANT_ID}/{TIMELINE_ID}/{ENDPOINT_ID}/{key}");
+            let path = RemotePath::from_string(path).unwrap();
+            S3Path { path }
+        };
+
+        let path = "cache_key".to_string();
+        let mut key_path = KeyRequest {
+            path,
+            tenant_id: auth.tenant_id,
+            timeline_id: auth.timeline_id,
+            endpoint_id: auth.endpoint_id,
+        };
+        assert_eq!(S3Path::try_from(&key_path).unwrap(), s3_path(key_path.path));
+
+        key_path.path = "we/can/have/nested/paths".to_string();
+        assert_eq!(S3Path::try_from(&key_path).unwrap(), s3_path(key_path.path));
+
+        key_path.path = "../error/hello/../".to_string();
+        assert!(S3Path::try_from(&key_path).is_err());
+    }
+
+    #[test]
+    fn prefix_s3_path() {
+        let mut path = PrefixKeyPath {
+            tenant_id: TENANT_ID,
+            timeline_id: None,
+            endpoint_id: None,
+        };
+        let prefix_path = |s: String| RemotePath::from_string(&s).unwrap();
+        assert_eq!(
+            PrefixS3Path::from(&path).path,
+            prefix_path(format!("{TENANT_ID}"))
+        );
+
+        path.timeline_id = Some(TIMELINE_ID);
+        assert_eq!(
+            PrefixS3Path::from(&path).path,
+            prefix_path(format!("{TENANT_ID}/{TIMELINE_ID}"))
+        );
+
+        path.endpoint_id = Some(ENDPOINT_ID.into());
+        assert_eq!(
+            PrefixS3Path::from(&path).path,
+            prefix_path(format!("{TENANT_ID}/{TIMELINE_ID}/{ENDPOINT_ID}"))
+        );
+    }
+}

object_storage/src/main.rs

@@ -0,0 +1,65 @@
+//! `object_storage` is a service which provides API for uploading and downloading
+//! files. It is used by compute and control plane for accessing LFC prewarm data.
+//! This service is deployed either as a separate component or as part of compute image
+//! for large computes.
+mod app;
+use anyhow::Context;
+use tracing::info;
+use utils::logging;
+
+//see set()
+const fn max_upload_file_limit() -> usize {
+    100 * 1024 * 1024
+}
+
+#[derive(serde::Deserialize)]
+#[serde(tag = "type")]
+struct Config {
+    listen: std::net::SocketAddr,
+    pemfile: camino::Utf8PathBuf,
+    #[serde(flatten)]
+    storage_config: remote_storage::RemoteStorageConfig,
+    #[serde(default = "max_upload_file_limit")]
+    max_upload_file_limit: usize,
+}
+
+#[tokio::main]
+async fn main() -> anyhow::Result<()> {
+    logging::init(
+        logging::LogFormat::Plain,
+        logging::TracingErrorLayerEnablement::EnableWithRustLogFilter,
+        logging::Output::Stdout,
+    )?;
+
+    let config: String = std::env::args().skip(1).take(1).collect();
+    if config.is_empty() {
+        anyhow::bail!("Usage: object_storage config.json")
+    }
+    info!("Reading config from {config}");
+    let config = std::fs::read_to_string(config.clone())?;
+    let config: Config = serde_json::from_str(&config).context("parsing config")?;
+    info!("Reading pemfile from {}", config.pemfile.clone());
+    let pemfile = std::fs::read(config.pemfile.clone())?;
+    info!("Loading public key from {}", config.pemfile.clone());
+    let auth = object_storage::JwtAuth::new(&pemfile)?;
+
+    let listener = tokio::net::TcpListener::bind(config.listen).await.unwrap();
+    info!("listening on {}", listener.local_addr().unwrap());
+
+    let storage = remote_storage::GenericRemoteStorage::from_config(&config.storage_config).await?;
+    let cancel = tokio_util::sync::CancellationToken::new();
+    app::check_storage_permissions(&storage, cancel.clone()).await?;
+
+    let proxy = std::sync::Arc::new(object_storage::Storage {
+        auth,
+        storage,
+        cancel: cancel.clone(),
+        max_upload_file_limit: config.max_upload_file_limit,
+    });
+
+    tokio::spawn(utils::signals::signal_handler(cancel.clone()));
+    axum::serve(listener, app::app(proxy))
+        .with_graceful_shutdown(async move { cancel.cancelled().await })
+        .await?;
+    Ok(())
+}

pageserver/benches/bench_ingest.rs

@@ -126,7 +126,7 @@ async fn ingest(
             max_concurrency: NonZeroUsize::new(1).unwrap(),
         });
         let (_desc, path) = layer
-            .write_to_disk(&ctx, None, l0_flush_state.inner())
+            .write_to_disk(&ctx, None, l0_flush_state.inner(), &gate, cancel.clone())
             .await?
             .unwrap();
         tokio::fs::remove_file(path).await?;

pageserver/benches/bench_walredo.rs

@@ -65,7 +65,7 @@ use bytes::{Buf, Bytes};
 use criterion::{BenchmarkId, Criterion};
 use once_cell::sync::Lazy;
 use pageserver::config::PageServerConf;
-use pageserver::walredo::PostgresRedoManager;
+use pageserver::walredo::{PostgresRedoManager, RedoAttemptType};
 use pageserver_api::key::Key;
 use pageserver_api::record::NeonWalRecord;
 use pageserver_api::shard::TenantShardId;
@@ -223,7 +223,14 @@ impl Request {
 
         // TODO: avoid these clones
         manager
-            .request_redo(*key, *lsn, base_img.clone(), records.clone(), *pg_version)
+            .request_redo(
+                *key,
+                *lsn,
+                base_img.clone(),
+                records.clone(),
+                *pg_version,
+                RedoAttemptType::ReadPage,
+            )
             .await
             .context("request_redo")
     }

pageserver/src/bin/pageserver.rs

@@ -31,7 +31,6 @@ use pageserver::{
 };
 use postgres_backend::AuthType;
 use remote_storage::GenericRemoteStorage;
-use tokio::signal::unix::SignalKind;
 use tokio::time::Instant;
 use tokio_util::sync::CancellationToken;
 use tracing::*;
@@ -453,6 +452,24 @@ fn start_pageserver(
     info!("Using auth for http API: {:#?}", conf.http_auth_type);
     info!("Using auth for pg connections: {:#?}", conf.pg_auth_type);
 
+    let tls_server_config = if conf.listen_https_addr.is_some() || conf.enable_tls_page_service_api
+    {
+        let resolver = BACKGROUND_RUNTIME.block_on(ReloadingCertificateResolver::new(
+            "main",
+            &conf.ssl_key_file,
+            &conf.ssl_cert_file,
+            conf.ssl_cert_reload_period,
+        ))?;
+
+        let server_config = rustls::ServerConfig::builder()
+            .with_no_client_auth()
+            .with_cert_resolver(resolver);
+
+        Some(Arc::new(server_config))
+    } else {
+        None
+    };
+
     match var("NEON_AUTH_TOKEN") {
         Ok(v) => {
             info!("Loaded JWT token for authentication with Safekeeper");
@@ -671,17 +688,11 @@ fn start_pageserver(
 
         let https_task = match https_listener {
             Some(https_listener) => {
-                let resolver = MGMT_REQUEST_RUNTIME.block_on(ReloadingCertificateResolver::new(
-                    &conf.ssl_key_file,
-                    &conf.ssl_cert_file,
-                    conf.ssl_cert_reload_period,
-                ))?;
+                let tls_server_config = tls_server_config
+                    .clone()
+                    .expect("tls_server_config is set earlier if https is enabled");
 
-                let server_config = rustls::ServerConfig::builder()
-                    .with_no_client_auth()
-                    .with_cert_resolver(resolver);
-
-                let tls_acceptor = tokio_rustls::TlsAcceptor::from(Arc::new(server_config));
+                let tls_acceptor = tokio_rustls::TlsAcceptor::from(tls_server_config);
 
                 let server =
                     http_utils::server::Server::new(service, https_listener, Some(tls_acceptor))?;
@@ -737,6 +748,11 @@ fn start_pageserver(
             tokio::net::TcpListener::from_std(pageserver_listener)
                 .context("create tokio listener")?
         },
+        if conf.enable_tls_page_service_api {
+            tls_server_config
+        } else {
+            None
+        },
     );
 
     // All started up! Now just sit and wait for shutdown signal.
@@ -744,32 +760,7 @@ fn start_pageserver(
         let signal_token = CancellationToken::new();
         let signal_cancel = signal_token.child_token();
 
-        // Spawn signal handlers. Runs in a loop since we want to be responsive to multiple signals
-        // even after triggering shutdown (e.g. a SIGQUIT after a slow SIGTERM shutdown). See:
-        // https://github.com/neondatabase/neon/issues/9740.
-        tokio::spawn(async move {
-            let mut sigint = tokio::signal::unix::signal(SignalKind::interrupt()).unwrap();
-            let mut sigterm = tokio::signal::unix::signal(SignalKind::terminate()).unwrap();
-            let mut sigquit = tokio::signal::unix::signal(SignalKind::quit()).unwrap();
-
-            loop {
-                let signal = tokio::select! {
-                    _ = sigquit.recv() => {
-                        info!("Got signal SIGQUIT. Terminating in immediate shutdown mode.");
-                        std::process::exit(111);
-                    }
-                    _ = sigint.recv() => "SIGINT",
-                    _ = sigterm.recv() => "SIGTERM",
-                };
-
-                if !signal_token.is_cancelled() {
-                    info!("Got signal {signal}. Terminating gracefully in fast shutdown mode.");
-                    signal_token.cancel();
-                } else {
-                    info!("Got signal {signal}. Already shutting down.");
-                }
-            }
-        });
+        tokio::spawn(utils::signals::signal_handler(signal_token));
 
         // Wait for cancellation signal and shut down the pageserver.
         //

pageserver/src/config.rs

@@ -219,6 +219,11 @@ pub struct PageServerConf {
     pub generate_unarchival_heatmap: bool,
 
     pub tracing: Option<pageserver_api::config::Tracing>,
+
+    /// Enable TLS in page service API.
+    /// Does not force TLS: the client negotiates TLS usage during the handshake.
+    /// Uses key and certificate from ssl_key_file/ssl_cert_file.
+    pub enable_tls_page_service_api: bool,
 }
 
 /// Token for authentication to safekeepers
@@ -391,6 +396,7 @@ impl PageServerConf {
             load_previous_heatmap,
             generate_unarchival_heatmap,
             tracing,
+            enable_tls_page_service_api,
         } = config_toml;
 
         let mut conf = PageServerConf {
@@ -441,6 +447,7 @@ impl PageServerConf {
             page_service_pipelining,
             get_vectored_concurrent_io,
             tracing,
+            enable_tls_page_service_api,
 
             // ------------------------------------------------------------
             // fields that require additional validation or custom handling

pageserver/src/consumption_metrics/metrics.rs

@@ -27,9 +27,6 @@ pub(super) enum Name {
     /// Timeline logical size
     #[serde(rename = "timeline_logical_size")]
     LogicalSize,
-    /// Timeline delta from parent
-    #[serde(rename = "timeline_changed_bytes_from_parent")]
-    ChangesFromParent,
     /// Tenant remote size
     #[serde(rename = "remote_storage_size")]
     RemoteSize,
@@ -178,18 +175,6 @@ impl MetricsKey {
         .absolute_values()
     }
 
-    const fn timeline_changed_bytes_from_parent(
-        tenant_id: TenantId,
-        timeline_id: TimelineId,
-    ) -> AbsoluteValueFactory {
-        MetricsKey {
-            tenant_id,
-            timeline_id: Some(timeline_id),
-            metric: Name::ChangesFromParent,
-        }
-        .absolute_values()
-    }
-
     /// [`Tenant::remote_size`]
     ///
     /// [`Tenant::remote_size`]: crate::tenant::Tenant::remote_size
@@ -379,7 +364,6 @@ struct TimelineSnapshot {
     loaded_at: (Lsn, SystemTime),
     last_record_lsn: Lsn,
     current_exact_logical_size: Option<u64>,
-    changed_bytes_from_parent: Option<u64>,
 }
 
 impl TimelineSnapshot {
@@ -415,23 +399,10 @@ impl TimelineSnapshot {
                 }
             };
 
-            // This is an approximation of how much data has changed on this branch vs. its ancestor: the
-            // number of bytes written to the WAL, clamped to the size of the branch.
-            let changed_bytes_from_parent = current_exact_logical_size.and_then(|size| {
-                if t.get_ancestor_lsn() == Lsn::MAX {
-                    None
-                } else {
-                    t.get_last_record_lsn()
-                        .checked_sub(t.get_ancestor_lsn())
-                        .map(|wal_bytes| std::cmp::min(wal_bytes.0, size))
-                }
-            });
-
             Ok(Some(TimelineSnapshot {
                 loaded_at,
                 last_record_lsn,
                 current_exact_logical_size,
-                changed_bytes_from_parent,
             }))
         }
     }
@@ -509,17 +480,6 @@ impl TimelineSnapshot {
                 metrics.push(factory.at(now, size));
             }
         }
-
-        {
-            let factory = MetricsKey::timeline_changed_bytes_from_parent(tenant_id, timeline_id);
-            let current_or_previous = self
-                .changed_bytes_from_parent
-                .or_else(|| cache.get(factory.key()).map(|item| item.value));
-
-            if let Some(size) = current_or_previous {
-                metrics.push(factory.at(now, size));
-            }
-        }
     }
 }
 

pageserver/src/consumption_metrics/metrics/tests.rs

@@ -18,7 +18,6 @@ fn startup_collected_timeline_metrics_before_advancing() {
         loaded_at: (disk_consistent_lsn, SystemTime::now()),
         last_record_lsn: disk_consistent_lsn,
         current_exact_logical_size: Some(0x42000),
-        changed_bytes_from_parent: Some(0x1000),
     };
 
     let now = DateTime::<Utc>::from(SystemTime::now());
@@ -34,8 +33,7 @@ fn startup_collected_timeline_metrics_before_advancing() {
                 0
             ),
             MetricsKey::written_size(tenant_id, timeline_id).at(now, disk_consistent_lsn.0),
-            MetricsKey::timeline_logical_size(tenant_id, timeline_id).at(now, 0x42000),
-            MetricsKey::timeline_changed_bytes_from_parent(tenant_id, timeline_id).at(now, 0x1000)
+            MetricsKey::timeline_logical_size(tenant_id, timeline_id).at(now, 0x42000)
         ]
     );
 }
@@ -62,7 +60,6 @@ fn startup_collected_timeline_metrics_second_round() {
         loaded_at: (disk_consistent_lsn, init),
         last_record_lsn: disk_consistent_lsn,
         current_exact_logical_size: Some(0x42000),
-        changed_bytes_from_parent: Some(0x1000),
     };
 
     snap.to_metrics(tenant_id, timeline_id, now, &mut metrics, &cache);
@@ -72,8 +69,7 @@ fn startup_collected_timeline_metrics_second_round() {
         &[
             MetricsKey::written_size_delta(tenant_id, timeline_id).from_until(before, now, 0),
             MetricsKey::written_size(tenant_id, timeline_id).at(now, disk_consistent_lsn.0),
-            MetricsKey::timeline_logical_size(tenant_id, timeline_id).at(now, 0x42000),
-            MetricsKey::timeline_changed_bytes_from_parent(tenant_id, timeline_id).at(now, 0x1000)
+            MetricsKey::timeline_logical_size(tenant_id, timeline_id).at(now, 0x42000)
         ]
     );
 }
@@ -108,7 +104,6 @@ fn startup_collected_timeline_metrics_nth_round_at_same_lsn() {
         loaded_at: (disk_consistent_lsn, init),
         last_record_lsn: disk_consistent_lsn,
         current_exact_logical_size: Some(0x42000),
-        changed_bytes_from_parent: Some(0x1000),
     };
 
     snap.to_metrics(tenant_id, timeline_id, now, &mut metrics, &cache);
@@ -118,8 +113,7 @@ fn startup_collected_timeline_metrics_nth_round_at_same_lsn() {
         &[
             MetricsKey::written_size_delta(tenant_id, timeline_id).from_until(just_before, now, 0),
             MetricsKey::written_size(tenant_id, timeline_id).at(now, disk_consistent_lsn.0),
-            MetricsKey::timeline_logical_size(tenant_id, timeline_id).at(now, 0x42000),
-            MetricsKey::timeline_changed_bytes_from_parent(tenant_id, timeline_id).at(now, 0x1000)
+            MetricsKey::timeline_logical_size(tenant_id, timeline_id).at(now, 0x42000)
         ]
     );
 }
@@ -147,7 +141,6 @@ fn post_restart_written_sizes_with_rolled_back_last_record_lsn() {
         loaded_at: (Lsn(50), at_restart),
         last_record_lsn: Lsn(50),
         current_exact_logical_size: None,
-        changed_bytes_from_parent: None,
     };
 
     let mut cache = HashMap::from([
@@ -209,7 +202,6 @@ fn post_restart_current_exact_logical_size_uses_cached() {
         loaded_at: (Lsn(50), at_restart),
         last_record_lsn: Lsn(50),
         current_exact_logical_size: None,
-        changed_bytes_from_parent: Some(0x1000),
     };
 
     let cache = HashMap::from([MetricsKey::timeline_logical_size(tenant_id, timeline_id)

pageserver/src/http/openapi_spec.yml

@@ -212,6 +212,12 @@ paths:
               schema:
                 type: string
                 format: date-time
+        "412":
+          description: No timestamp is found for given LSN, e.g. if there had been no commits till LSN
+          content:
+            application/json:
+              schema:
+                $ref: "#/components/schemas/PreconditionFailedError"
 
   /v1/tenant/{tenant_id}/timeline/{timeline_id}/get_lsn_by_timestamp:
     parameters:

pageserver/src/http/routes.rs

@@ -67,7 +67,7 @@ use crate::tenant::mgr::{
 };
 use crate::tenant::remote_timeline_client::index::GcCompactionState;
 use crate::tenant::remote_timeline_client::{
-    download_index_part, list_remote_tenant_shards, list_remote_timelines,
+    download_index_part, download_tenant_manifest, list_remote_tenant_shards, list_remote_timelines,
 };
 use crate::tenant::secondary::SecondaryController;
 use crate::tenant::size::ModelInputs;
@@ -989,7 +989,7 @@ async fn get_lsn_by_timestamp_handler(
     if !tenant_shard_id.is_shard_zero() {
         // Requires SLRU contents, which are only stored on shard zero
         return Err(ApiError::BadRequest(anyhow!(
-            "Size calculations are only available on shard zero"
+            "Lsn calculations by timestamp are only available on shard zero"
         )));
     }
 
@@ -1064,7 +1064,7 @@ async fn get_timestamp_of_lsn_handler(
     if !tenant_shard_id.is_shard_zero() {
         // Requires SLRU contents, which are only stored on shard zero
         return Err(ApiError::BadRequest(anyhow!(
-            "Size calculations are only available on shard zero"
+            "Timestamp calculations by lsn are only available on shard zero"
         )));
     }
 
@@ -1090,8 +1090,8 @@ async fn get_timestamp_of_lsn_handler(
             .to_string();
             json_response(StatusCode::OK, time)
         }
-        None => Err(ApiError::NotFound(
-            anyhow::anyhow!("Timestamp for lsn {} not found", lsn).into(),
+        None => Err(ApiError::PreconditionFailed(
+            format!("Timestamp for lsn {} not found", lsn).into(),
         )),
     }
 }
@@ -2274,6 +2274,7 @@ async fn timeline_compact_handler(
     if Some(true) == parse_query_param::<_, bool>(&request, "dry_run")? {
         flags |= CompactFlags::DryRun;
     }
+    // Manual compaction does not yield for L0.
 
     let wait_until_uploaded =
         parse_query_param::<_, bool>(&request, "wait_until_uploaded")?.unwrap_or(false);
@@ -2911,9 +2912,22 @@ async fn tenant_scan_remote_handler(
             };
         }
 
+        let result =
+            download_tenant_manifest(&state.remote_storage, &tenant_shard_id, generation, &cancel)
+                .instrument(info_span!("download_tenant_manifest",
+                            tenant_id=%tenant_shard_id.tenant_id,
+                            shard_id=%tenant_shard_id.shard_slug()))
+                .await;
+        let stripe_size = match result {
+            Ok((manifest, _, _)) => manifest.stripe_size,
+            Err(DownloadError::NotFound) => None,
+            Err(err) => return Err(ApiError::InternalServerError(anyhow!(err))),
+        };
+
         response.shards.push(TenantScanRemoteStorageShard {
             tenant_shard_id,
             generation: generation.into(),
+            stripe_size,
         });
     }
 
@@ -3239,7 +3253,7 @@ async fn ingest_aux_files(
         modification
             .put_file(&fname, content.as_bytes(), &ctx)
             .await
-            .map_err(ApiError::InternalServerError)?;
+            .map_err(|e| ApiError::InternalServerError(e.into()))?;
     }
     modification
         .commit(&ctx)
@@ -3368,11 +3382,11 @@ async fn put_tenant_timeline_import_basebackup(
 
         let broker_client = state.broker_client.clone();
 
-        let mut body = StreamReader::new(request.into_body().map(|res| {
-            res.map_err(|error| {
-                std::io::Error::new(std::io::ErrorKind::Other, anyhow::anyhow!(error))
-            })
-        }));
+        let mut body = StreamReader::new(
+            request
+                .into_body()
+                .map(|res| res.map_err(|error| std::io::Error::other(anyhow::anyhow!(error)))),
+        );
 
         tenant.wait_to_become_active(ACTIVE_TENANT_TIMEOUT).await?;
 
@@ -3446,7 +3460,7 @@ async fn put_tenant_timeline_import_wal(
 
         let mut body = StreamReader::new(request.into_body().map(|res| {
             res.map_err(|error| {
-                std::io::Error::new(std::io::ErrorKind::Other, anyhow::anyhow!(error))
+                std::io::Error::other( anyhow::anyhow!(error))
             })
         }));
 

pageserver/src/import_datadir.rs

@@ -27,7 +27,7 @@ use crate::context::RequestContext;
 use crate::metrics::WAL_INGEST;
 use crate::pgdatadir_mapping::*;
 use crate::tenant::Timeline;
-use crate::walingest::WalIngest;
+use crate::walingest::{WalIngest, WalIngestErrorKind};
 
 // Returns checkpoint LSN from controlfile
 pub fn get_lsn_from_controlfile(path: &Utf8Path) -> Result<Lsn> {
@@ -157,9 +157,9 @@ async fn import_rel(
         .put_rel_creation(rel, nblocks as u32, ctx)
         .await
     {
-        match e {
-            RelationError::AlreadyExists => {
-                debug!("Relation {} already exist. We must be extending it.", rel)
+        match e.kind {
+            WalIngestErrorKind::RelationAlreadyExists(rel) => {
+                debug!("Relation {rel} already exists. We must be extending it.")
             }
             _ => return Err(e.into()),
         }

pageserver/src/metrics.rs

@@ -1289,6 +1289,7 @@ pub(crate) enum StorageIoOperation {
     Seek,
     Fsync,
     Metadata,
+    SetLen,
 }
 
 impl StorageIoOperation {
@@ -1303,6 +1304,7 @@ impl StorageIoOperation {
             StorageIoOperation::Seek => "seek",
             StorageIoOperation::Fsync => "fsync",
             StorageIoOperation::Metadata => "metadata",
+            StorageIoOperation::SetLen => "set_len",
         }
     }
 }

pageserver/src/page_service.rs

@@ -105,6 +105,7 @@ pub fn spawn(
     pg_auth: Option<Arc<SwappableJwtAuth>>,
     perf_trace_dispatch: Option<Dispatch>,
     tcp_listener: tokio::net::TcpListener,
+    tls_config: Option<Arc<rustls::ServerConfig>>,
 ) -> Listener {
     let cancel = CancellationToken::new();
     let libpq_ctx = RequestContext::todo_child(
@@ -124,6 +125,7 @@ pub fn spawn(
             perf_trace_dispatch,
             tcp_listener,
             conf.pg_auth_type,
+            tls_config,
             conf.page_service_pipelining.clone(),
             libpq_ctx,
             cancel.clone(),
@@ -181,6 +183,7 @@ pub async fn libpq_listener_main(
     perf_trace_dispatch: Option<Dispatch>,
     listener: tokio::net::TcpListener,
     auth_type: AuthType,
+    tls_config: Option<Arc<rustls::ServerConfig>>,
     pipelining_config: PageServicePipeliningConfig,
     listener_ctx: RequestContext,
     listener_cancel: CancellationToken,
@@ -223,6 +226,7 @@ pub async fn libpq_listener_main(
                     local_auth,
                     socket,
                     auth_type,
+                    tls_config.clone(),
                     pipelining_config.clone(),
                     connection_ctx,
                     connections_cancel.child_token(),
@@ -264,6 +268,7 @@ async fn page_service_conn_main(
     auth: Option<Arc<SwappableJwtAuth>>,
     socket: tokio::net::TcpStream,
     auth_type: AuthType,
+    tls_config: Option<Arc<rustls::ServerConfig>>,
     pipelining_config: PageServicePipeliningConfig,
     connection_ctx: RequestContext,
     cancel: CancellationToken,
@@ -334,7 +339,8 @@ async fn page_service_conn_main(
         cancel.clone(),
         gate_guard,
     );
-    let pgbackend = PostgresBackend::new_from_io(socket_fd, socket, peer_addr, auth_type, None)?;
+    let pgbackend =
+        PostgresBackend::new_from_io(socket_fd, socket, peer_addr, auth_type, tls_config)?;
 
     match pgbackend.run(&mut conn_handler, &cancel).await {
         Ok(()) => {

pageserver/src/pgdatadir_mapping.rs

@@ -9,8 +9,9 @@
 use std::collections::{BTreeMap, HashMap, HashSet, hash_map};
 use std::ops::{ControlFlow, Range};
 
-use crate::PERF_TRACE_TARGET;
-use anyhow::{Context, ensure};
+use crate::walingest::{WalIngestError, WalIngestErrorKind};
+use crate::{PERF_TRACE_TARGET, ensure_walingest};
+use anyhow::Context;
 use bytes::{Buf, Bytes, BytesMut};
 use enum_map::Enum;
 use itertools::Itertools;
@@ -136,12 +137,8 @@ impl From<PageReconstructError> for CalculateLogicalSizeError {
 
 #[derive(Debug, thiserror::Error)]
 pub enum RelationError {
-    #[error("Relation Already Exists")]
-    AlreadyExists,
     #[error("invalid relnode")]
     InvalidRelnode,
-    #[error(transparent)]
-    Other(#[from] anyhow::Error),
 }
 
 ///
@@ -691,7 +688,7 @@ impl Timeline {
         Ok(buf.get_u32_le())
     }
 
-    /// Get size of an SLRU segment
+    /// Does the slru segment exist?
     pub(crate) async fn get_slru_segment_exists(
         &self,
         kind: SlruKind,
@@ -844,9 +841,9 @@ impl Timeline {
         .await
     }
 
-    /// Obtain the possible timestamp range for the given lsn.
+    /// Obtain the timestamp for the given lsn.
     ///
-    /// If the lsn has no timestamps, returns None. returns `(min, max, median)` if it has timestamps.
+    /// If the lsn has no timestamps (e.g. no commits), returns None.
     pub(crate) async fn get_timestamp_for_lsn(
         &self,
         probe_lsn: Lsn,
@@ -1478,8 +1475,8 @@ impl DatadirModification<'_> {
     }
 
     /// Set the current lsn
-    pub(crate) fn set_lsn(&mut self, lsn: Lsn) -> anyhow::Result<()> {
-        ensure!(
+    pub(crate) fn set_lsn(&mut self, lsn: Lsn) -> Result<(), WalIngestError> {
+        ensure_walingest!(
             lsn >= self.lsn,
             "setting an older lsn {} than {} is not allowed",
             lsn,
@@ -1578,7 +1575,7 @@ impl DatadirModification<'_> {
         &mut self,
         rel: RelTag,
         ctx: &RequestContext,
-    ) -> Result<u32, PageReconstructError> {
+    ) -> Result<u32, WalIngestError> {
         // Get current size and put rel creation if rel doesn't exist
         //
         // NOTE: we check the cache first even though get_rel_exists and get_rel_size would
@@ -1593,14 +1590,13 @@ impl DatadirModification<'_> {
             .await?
         {
             // create it with 0 size initially, the logic below will extend it
-            self.put_rel_creation(rel, 0, ctx)
-                .await
-                .context("Relation Error")?;
+            self.put_rel_creation(rel, 0, ctx).await?;
             Ok(0)
         } else {
-            self.tline
+            Ok(self
+                .tline
                 .get_rel_size(rel, Version::Modified(self), ctx)
-                .await
+                .await?)
         }
     }
 
@@ -1637,11 +1633,14 @@ impl DatadirModification<'_> {
         // TODO(vlad): remove this argument and replace the shard check with is_key_local
         shard: &ShardIdentity,
         ctx: &RequestContext,
-    ) -> anyhow::Result<()> {
+    ) -> Result<(), WalIngestError> {
         let mut gaps_at_lsns = Vec::default();
 
         for meta in batch.metadata.iter() {
-            let (rel, blkno) = Key::from_compact(meta.key()).to_rel_block()?;
+            let key = Key::from_compact(meta.key());
+            let (rel, blkno) = key
+                .to_rel_block()
+                .map_err(|_| WalIngestErrorKind::InvalidKey(key, meta.lsn()))?;
             let new_nblocks = blkno + 1;
 
             let old_nblocks = self.create_relation_if_required(rel, ctx).await?;
@@ -1683,8 +1682,8 @@ impl DatadirModification<'_> {
         rel: RelTag,
         blknum: BlockNumber,
         rec: NeonWalRecord,
-    ) -> anyhow::Result<()> {
-        anyhow::ensure!(rel.relnode != 0, RelationError::InvalidRelnode);
+    ) -> Result<(), WalIngestError> {
+        ensure_walingest!(rel.relnode != 0, RelationError::InvalidRelnode);
         self.put(rel_block_to_key(rel, blknum), Value::WalRecord(rec));
         Ok(())
     }
@@ -1696,7 +1695,7 @@ impl DatadirModification<'_> {
         segno: u32,
         blknum: BlockNumber,
         rec: NeonWalRecord,
-    ) -> anyhow::Result<()> {
+    ) -> Result<(), WalIngestError> {
         if !self.tline.tenant_shard_id.is_shard_zero() {
             return Ok(());
         }
@@ -1714,14 +1713,11 @@ impl DatadirModification<'_> {
         rel: RelTag,
         blknum: BlockNumber,
         img: Bytes,
-    ) -> anyhow::Result<()> {
-        anyhow::ensure!(rel.relnode != 0, RelationError::InvalidRelnode);
+    ) -> Result<(), WalIngestError> {
+        ensure_walingest!(rel.relnode != 0, RelationError::InvalidRelnode);
         let key = rel_block_to_key(rel, blknum);
         if !key.is_valid_key_on_write_path() {
-            anyhow::bail!(
-                "the request contains data not supported by pageserver at {}",
-                key
-            );
+            Err(WalIngestErrorKind::InvalidKey(key, self.lsn))?;
         }
         self.put(rel_block_to_key(rel, blknum), Value::Image(img));
         Ok(())
@@ -1733,15 +1729,12 @@ impl DatadirModification<'_> {
         segno: u32,
         blknum: BlockNumber,
         img: Bytes,
-    ) -> anyhow::Result<()> {
+    ) -> Result<(), WalIngestError> {
         assert!(self.tline.tenant_shard_id.is_shard_zero());
 
         let key = slru_block_to_key(kind, segno, blknum);
         if !key.is_valid_key_on_write_path() {
-            anyhow::bail!(
-                "the request contains data not supported by pageserver at {}",
-                key
-            );
+            Err(WalIngestErrorKind::InvalidKey(key, self.lsn))?;
         }
         self.put(key, Value::Image(img));
         Ok(())
@@ -1751,15 +1744,11 @@ impl DatadirModification<'_> {
         &mut self,
         rel: RelTag,
         blknum: BlockNumber,
-    ) -> anyhow::Result<()> {
-        anyhow::ensure!(rel.relnode != 0, RelationError::InvalidRelnode);
+    ) -> Result<(), WalIngestError> {
+        ensure_walingest!(rel.relnode != 0, RelationError::InvalidRelnode);
         let key = rel_block_to_key(rel, blknum);
         if !key.is_valid_key_on_write_path() {
-            anyhow::bail!(
-                "the request contains data not supported by pageserver: {} @ {}",
-                key,
-                self.lsn
-            );
+            Err(WalIngestErrorKind::InvalidKey(key, self.lsn))?;
         }
 
         let batch = self
@@ -1776,15 +1765,11 @@ impl DatadirModification<'_> {
         kind: SlruKind,
         segno: u32,
         blknum: BlockNumber,
-    ) -> anyhow::Result<()> {
+    ) -> Result<(), WalIngestError> {
         assert!(self.tline.tenant_shard_id.is_shard_zero());
         let key = slru_block_to_key(kind, segno, blknum);
         if !key.is_valid_key_on_write_path() {
-            anyhow::bail!(
-                "the request contains data not supported by pageserver: {} @ {}",
-                key,
-                self.lsn
-            );
+            Err(WalIngestErrorKind::InvalidKey(key, self.lsn))?;
         }
 
         let batch = self
@@ -1832,8 +1817,10 @@ impl DatadirModification<'_> {
         dbnode: Oid,
         img: Bytes,
         ctx: &RequestContext,
-    ) -> anyhow::Result<()> {
-        let v2_enabled = self.maybe_enable_rel_size_v2()?;
+    ) -> Result<(), WalIngestError> {
+        let v2_enabled = self
+            .maybe_enable_rel_size_v2()
+            .map_err(WalIngestErrorKind::MaybeRelSizeV2Error)?;
 
         // Add it to the directory (if it doesn't exist already)
         let buf = self.get(DBDIR_KEY, ctx).await?;
@@ -1874,13 +1861,13 @@ impl DatadirModification<'_> {
         xid: u64,
         img: Bytes,
         ctx: &RequestContext,
-    ) -> anyhow::Result<()> {
+    ) -> Result<(), WalIngestError> {
         // Add it to the directory entry
         let dirbuf = self.get(TWOPHASEDIR_KEY, ctx).await?;
         let newdirbuf = if self.tline.pg_version >= 17 {
             let mut dir = TwoPhaseDirectoryV17::des(&dirbuf)?;
             if !dir.xids.insert(xid) {
-                anyhow::bail!("twophase file for xid {} already exists", xid);
+                Err(WalIngestErrorKind::FileAlreadyExists(xid))?;
             }
             self.pending_directory_entries.push((
                 DirectoryKind::TwoPhase,
@@ -1891,7 +1878,7 @@ impl DatadirModification<'_> {
             let xid = xid as u32;
             let mut dir = TwoPhaseDirectory::des(&dirbuf)?;
             if !dir.xids.insert(xid) {
-                anyhow::bail!("twophase file for xid {} already exists", xid);
+                Err(WalIngestErrorKind::FileAlreadyExists(xid.into()))?;
             }
             self.pending_directory_entries.push((
                 DirectoryKind::TwoPhase,
@@ -1909,22 +1896,22 @@ impl DatadirModification<'_> {
         &mut self,
         origin_id: RepOriginId,
         origin_lsn: Lsn,
-    ) -> anyhow::Result<()> {
+    ) -> Result<(), WalIngestError> {
         let key = repl_origin_key(origin_id);
         self.put(key, Value::Image(origin_lsn.ser().unwrap().into()));
         Ok(())
     }
 
-    pub async fn drop_replorigin(&mut self, origin_id: RepOriginId) -> anyhow::Result<()> {
+    pub async fn drop_replorigin(&mut self, origin_id: RepOriginId) -> Result<(), WalIngestError> {
         self.set_replorigin(origin_id, Lsn::INVALID).await
     }
 
-    pub fn put_control_file(&mut self, img: Bytes) -> anyhow::Result<()> {
+    pub fn put_control_file(&mut self, img: Bytes) -> Result<(), WalIngestError> {
         self.put(CONTROLFILE_KEY, Value::Image(img));
         Ok(())
     }
 
-    pub fn put_checkpoint(&mut self, img: Bytes) -> anyhow::Result<()> {
+    pub fn put_checkpoint(&mut self, img: Bytes) -> Result<(), WalIngestError> {
         self.put(CHECKPOINT_KEY, Value::Image(img));
         Ok(())
     }
@@ -1934,7 +1921,7 @@ impl DatadirModification<'_> {
         spcnode: Oid,
         dbnode: Oid,
         ctx: &RequestContext,
-    ) -> anyhow::Result<()> {
+    ) -> Result<(), WalIngestError> {
         let total_blocks = self
             .tline
             .get_db_size(spcnode, dbnode, Version::Modified(self), ctx)
@@ -1973,20 +1960,21 @@ impl DatadirModification<'_> {
         rel: RelTag,
         nblocks: BlockNumber,
         ctx: &RequestContext,
-    ) -> Result<(), RelationError> {
+    ) -> Result<(), WalIngestError> {
         if rel.relnode == 0 {
-            return Err(RelationError::InvalidRelnode);
+            Err(WalIngestErrorKind::LogicalError(anyhow::anyhow!(
+                "invalid relnode"
+            )))?;
         }
         // It's possible that this is the first rel for this db in this
         // tablespace.  Create the reldir entry for it if so.
-        let mut dbdir = DbDirectory::des(&self.get(DBDIR_KEY, ctx).await.context("read db")?)
-            .context("deserialize db")?;
+        let mut dbdir = DbDirectory::des(&self.get(DBDIR_KEY, ctx).await?)?;
 
         let dbdir_exists =
             if let hash_map::Entry::Vacant(e) = dbdir.dbdirs.entry((rel.spcnode, rel.dbnode)) {
                 // Didn't exist. Update dbdir
                 e.insert(false);
-                let buf = DbDirectory::ser(&dbdir).context("serialize db")?;
+                let buf = DbDirectory::ser(&dbdir)?;
                 self.pending_directory_entries.push((
                     DirectoryKind::Db,
                     MetricsUpdate::Set(dbdir.dbdirs.len() as u64),
@@ -2003,27 +1991,25 @@ impl DatadirModification<'_> {
             RelDirectory::default()
         } else {
             // reldir already exists, fetch it
-            RelDirectory::des(&self.get(rel_dir_key, ctx).await.context("read db")?)
-                .context("deserialize db")?
+            RelDirectory::des(&self.get(rel_dir_key, ctx).await?)?
         };
 
-        let v2_enabled = self.maybe_enable_rel_size_v2()?;
+        let v2_enabled = self
+            .maybe_enable_rel_size_v2()
+            .map_err(WalIngestErrorKind::MaybeRelSizeV2Error)?;
 
         if v2_enabled {
             if rel_dir.rels.contains(&(rel.relnode, rel.forknum)) {
-                return Err(RelationError::AlreadyExists);
+                Err(WalIngestErrorKind::RelationAlreadyExists(rel))?;
             }
             let sparse_rel_dir_key =
                 rel_tag_sparse_key(rel.spcnode, rel.dbnode, rel.relnode, rel.forknum);
             // check if the rel_dir_key exists in v2
-            let val = self
-                .sparse_get(sparse_rel_dir_key, ctx)
-                .await
-                .map_err(|e| RelationError::Other(e.into()))?;
+            let val = self.sparse_get(sparse_rel_dir_key, ctx).await?;
             let val = RelDirExists::decode_option(val)
-                .map_err(|_| RelationError::Other(anyhow::anyhow!("invalid reldir key")))?;
+                .map_err(|_| WalIngestErrorKind::InvalidRelDirKey(sparse_rel_dir_key))?;
             if val == RelDirExists::Exists {
-                return Err(RelationError::AlreadyExists);
+                Err(WalIngestErrorKind::RelationAlreadyExists(rel))?;
             }
             self.put(
                 sparse_rel_dir_key,
@@ -2039,9 +2025,7 @@ impl DatadirModification<'_> {
                 // will be key not found errors if we don't create an empty one for rel_size_v2.
                 self.put(
                     rel_dir_key,
-                    Value::Image(Bytes::from(
-                        RelDirectory::ser(&RelDirectory::default()).context("serialize")?,
-                    )),
+                    Value::Image(Bytes::from(RelDirectory::ser(&RelDirectory::default())?)),
                 );
             }
             self.pending_directory_entries
@@ -2049,7 +2033,7 @@ impl DatadirModification<'_> {
         } else {
             // Add the new relation to the rel directory entry, and write it back
             if !rel_dir.rels.insert((rel.relnode, rel.forknum)) {
-                return Err(RelationError::AlreadyExists);
+                Err(WalIngestErrorKind::RelationAlreadyExists(rel))?;
             }
             if !dbdir_exists {
                 self.pending_directory_entries
@@ -2059,9 +2043,7 @@ impl DatadirModification<'_> {
                 .push((DirectoryKind::Rel, MetricsUpdate::Add(1)));
             self.put(
                 rel_dir_key,
-                Value::Image(Bytes::from(
-                    RelDirectory::ser(&rel_dir).context("serialize")?,
-                )),
+                Value::Image(Bytes::from(RelDirectory::ser(&rel_dir)?)),
             );
         }
 
@@ -2086,8 +2068,8 @@ impl DatadirModification<'_> {
         rel: RelTag,
         nblocks: BlockNumber,
         ctx: &RequestContext,
-    ) -> anyhow::Result<()> {
-        anyhow::ensure!(rel.relnode != 0, RelationError::InvalidRelnode);
+    ) -> Result<(), WalIngestError> {
+        ensure_walingest!(rel.relnode != 0, RelationError::InvalidRelnode);
         if self
             .tline
             .get_rel_exists(rel, Version::Modified(self), ctx)
@@ -2117,8 +2099,8 @@ impl DatadirModification<'_> {
         rel: RelTag,
         nblocks: BlockNumber,
         ctx: &RequestContext,
-    ) -> anyhow::Result<()> {
-        anyhow::ensure!(rel.relnode != 0, RelationError::InvalidRelnode);
+    ) -> Result<(), WalIngestError> {
+        ensure_walingest!(rel.relnode != 0, RelationError::InvalidRelnode);
 
         // Put size
         let size_key = rel_size_to_key(rel);
@@ -2142,8 +2124,10 @@ impl DatadirModification<'_> {
         &mut self,
         drop_relations: HashMap<(u32, u32), Vec<RelTag>>,
         ctx: &RequestContext,
-    ) -> anyhow::Result<()> {
-        let v2_enabled = self.maybe_enable_rel_size_v2()?;
+    ) -> Result<(), WalIngestError> {
+        let v2_enabled = self
+            .maybe_enable_rel_size_v2()
+            .map_err(WalIngestErrorKind::MaybeRelSizeV2Error)?;
         for ((spc_node, db_node), rel_tags) in drop_relations {
             let dir_key = rel_dir_to_key(spc_node, db_node);
             let buf = self.get(dir_key, ctx).await?;
@@ -2163,7 +2147,7 @@ impl DatadirModification<'_> {
                     let key =
                         rel_tag_sparse_key(spc_node, db_node, rel_tag.relnode, rel_tag.forknum);
                     let val = RelDirExists::decode_option(self.sparse_get(key, ctx).await?)
-                        .map_err(|_| RelationError::Other(anyhow::anyhow!("invalid reldir key")))?;
+                        .map_err(|_| WalIngestErrorKind::InvalidKey(key, self.lsn))?;
                     if val == RelDirExists::Exists {
                         self.pending_directory_entries
                             .push((DirectoryKind::RelV2, MetricsUpdate::Sub(1)));
@@ -2206,7 +2190,7 @@ impl DatadirModification<'_> {
         segno: u32,
         nblocks: BlockNumber,
         ctx: &RequestContext,
-    ) -> anyhow::Result<()> {
+    ) -> Result<(), WalIngestError> {
         assert!(self.tline.tenant_shard_id.is_shard_zero());
 
         // Add it to the directory entry
@@ -2215,7 +2199,7 @@ impl DatadirModification<'_> {
         let mut dir = SlruSegmentDirectory::des(&buf)?;
 
         if !dir.segments.insert(segno) {
-            anyhow::bail!("slru segment {kind:?}/{segno} already exists");
+            Err(WalIngestErrorKind::SlruAlreadyExists(kind, segno))?;
         }
         self.pending_directory_entries.push((
             DirectoryKind::SlruSegment(kind),
@@ -2242,7 +2226,7 @@ impl DatadirModification<'_> {
         kind: SlruKind,
         segno: u32,
         nblocks: BlockNumber,
-    ) -> anyhow::Result<()> {
+    ) -> Result<(), WalIngestError> {
         assert!(self.tline.tenant_shard_id.is_shard_zero());
 
         // Put size
@@ -2258,7 +2242,7 @@ impl DatadirModification<'_> {
         kind: SlruKind,
         segno: u32,
         ctx: &RequestContext,
-    ) -> anyhow::Result<()> {
+    ) -> Result<(), WalIngestError> {
         // Remove it from the directory entry
         let dir_key = slru_dir_to_key(kind);
         let buf = self.get(dir_key, ctx).await?;
@@ -2283,7 +2267,7 @@ impl DatadirModification<'_> {
     }
 
     /// Drop a relmapper file (pg_filenode.map)
-    pub fn drop_relmap_file(&mut self, _spcnode: Oid, _dbnode: Oid) -> anyhow::Result<()> {
+    pub fn drop_relmap_file(&mut self, _spcnode: Oid, _dbnode: Oid) -> Result<(), WalIngestError> {
         // TODO
         Ok(())
     }
@@ -2293,7 +2277,7 @@ impl DatadirModification<'_> {
         &mut self,
         xid: u64,
         ctx: &RequestContext,
-    ) -> anyhow::Result<()> {
+    ) -> Result<(), WalIngestError> {
         // Remove it from the directory entry
         let buf = self.get(TWOPHASEDIR_KEY, ctx).await?;
         let newdirbuf = if self.tline.pg_version >= 17 {
@@ -2308,7 +2292,8 @@ impl DatadirModification<'_> {
             ));
             Bytes::from(TwoPhaseDirectoryV17::ser(&dir)?)
         } else {
-            let xid: u32 = u32::try_from(xid)?;
+            let xid: u32 = u32::try_from(xid)
+                .map_err(|e| WalIngestErrorKind::LogicalError(anyhow::Error::from(e)))?;
             let mut dir = TwoPhaseDirectory::des(&buf)?;
 
             if !dir.xids.remove(&xid) {
@@ -2333,7 +2318,7 @@ impl DatadirModification<'_> {
         path: &str,
         content: &[u8],
         ctx: &RequestContext,
-    ) -> anyhow::Result<()> {
+    ) -> Result<(), WalIngestError> {
         let key = aux_file::encode_aux_file_key(path);
         // retrieve the key from the engine
         let old_val = match self.get(key, ctx).await {
@@ -2342,7 +2327,7 @@ impl DatadirModification<'_> {
             Err(e) => return Err(e.into()),
         };
         let files: Vec<(&str, &[u8])> = if let Some(ref old_val) = old_val {
-            aux_file::decode_file_value(old_val)?
+            aux_file::decode_file_value(old_val).map_err(WalIngestErrorKind::EncodeAuxFileError)?
         } else {
             Vec::new()
         };
@@ -2387,7 +2372,8 @@ impl DatadirModification<'_> {
             }
             (None, true) => warn!("removing non-existing aux file: {}", path),
         }
-        let new_val = aux_file::encode_file_value(&new_files)?;
+        let new_val = aux_file::encode_file_value(&new_files)
+            .map_err(WalIngestErrorKind::EncodeAuxFileError)?;
         self.put(key, Value::Image(new_val.into()));
 
         Ok(())

pageserver/src/tenant.rs

@@ -45,6 +45,7 @@ use remote_timeline_client::manifest::{
 };
 use remote_timeline_client::{
     FAILED_REMOTE_OP_RETRIES, FAILED_UPLOAD_WARN_THRESHOLD, UploadQueueNotReadyError,
+    download_tenant_manifest,
 };
 use secondary::heatmap::{HeatMapTenant, HeatMapTimeline};
 use storage_broker::BrokerClientChannel;
@@ -99,7 +100,7 @@ use crate::tenant::timeline::delete::DeleteTimelineFlow;
 use crate::tenant::timeline::uninit::cleanup_timeline_directory;
 use crate::virtual_file::VirtualFile;
 use crate::walingest::WalLagCooldown;
-use crate::walredo::PostgresRedoManager;
+use crate::walredo::{PostgresRedoManager, RedoAttemptType};
 use crate::{InitializationOrder, TEMP_FILE_SUFFIX, import_datadir, span, task_mgr, walredo};
 
 static INIT_DB_SEMAPHORE: Lazy<Semaphore> = Lazy::new(|| Semaphore::new(8));
@@ -226,7 +227,8 @@ struct TimelinePreload {
 }
 
 pub(crate) struct TenantPreload {
-    tenant_manifest: TenantManifest,
+    /// The tenant manifest from remote storage, or None if no manifest was found.
+    tenant_manifest: Option<TenantManifest>,
     /// Map from timeline ID to a possible timeline preload. It is None iff the timeline is offloaded according to the manifest.
     timelines: HashMap<TimelineId, Option<TimelinePreload>>,
 }
@@ -282,12 +284,15 @@ pub struct Tenant {
     /// **Lock order**: if acquiring all (or a subset), acquire them in order `timelines`, `timelines_offloaded`, `timelines_creating`
     timelines_offloaded: Mutex<HashMap<TimelineId, Arc<OffloadedTimeline>>>,
 
-    /// Serialize writes of the tenant manifest to remote storage.  If there are concurrent operations
-    /// affecting the manifest, such as timeline deletion and timeline offload, they must wait for
-    /// each other (this could be optimized to coalesce writes if necessary).
+    /// 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.
     ///
-    /// The contents of the Mutex are the last manifest we successfully uploaded
-    tenant_manifest_upload: tokio::sync::Mutex<Option<TenantManifest>>,
+    /// Initially populated during tenant attach, updated via `maybe_upload_tenant_manifest`.
+    ///
+    /// Do not modify this directly. It is used to check whether a new manifest needs to be
+    /// uploaded. The manifest is constructed in `build_tenant_manifest`, and uploaded via
+    /// `maybe_upload_tenant_manifest`.
+    remote_tenant_manifest: tokio::sync::Mutex<Option<TenantManifest>>,
 
     // This mutex prevents creation of new timelines during GC.
     // Adding yet another mutex (in addition to `timelines`) is needed because holding
@@ -468,15 +473,16 @@ impl WalRedoManager {
         base_img: Option<(Lsn, bytes::Bytes)>,
         records: Vec<(Lsn, pageserver_api::record::NeonWalRecord)>,
         pg_version: u32,
+        redo_attempt_type: RedoAttemptType,
     ) -> Result<bytes::Bytes, walredo::Error> {
         match self {
             Self::Prod(_, mgr) => {
-                mgr.request_redo(key, lsn, base_img, records, pg_version)
+                mgr.request_redo(key, lsn, base_img, records, pg_version, redo_attempt_type)
                     .await
             }
             #[cfg(test)]
             Self::Test(mgr) => {
-                mgr.request_redo(key, lsn, base_img, records, pg_version)
+                mgr.request_redo(key, lsn, base_img, records, pg_version, redo_attempt_type)
                     .await
             }
         }
@@ -915,6 +921,7 @@ enum StartCreatingTimelineResult {
     Idempotent(Arc<Timeline>),
 }
 
+#[allow(clippy::large_enum_variant, reason = "TODO")]
 enum TimelineInitAndSyncResult {
     ReadyToActivate(Arc<Timeline>),
     NeedsSpawnImportPgdata(TimelineInitAndSyncNeedsSpawnImportPgdata),
@@ -1001,6 +1008,7 @@ enum CreateTimelineCause {
     Delete,
 }
 
+#[allow(clippy::large_enum_variant, reason = "TODO")]
 enum LoadTimelineCause {
     Attach,
     Unoffload,
@@ -1354,36 +1362,41 @@ impl Tenant {
                     }
                 }
 
-                // Ideally we should use Tenant::set_broken_no_wait, but it is not supposed to be used when tenant is in loading state.
-                enum BrokenVerbosity {
-                    Error,
-                    Info
-                }
-                let make_broken =
-                    |t: &Tenant, err: anyhow::Error, verbosity: BrokenVerbosity| {
-                        match verbosity {
-                            BrokenVerbosity::Info => {
-                                info!("attach cancelled, setting tenant state to Broken: {err}");
-                            },
-                            BrokenVerbosity::Error => {
-                                error!("attach failed, setting tenant state to Broken: {err:?}");
-                            }
+                fn make_broken_or_stopping(t: &Tenant, err: anyhow::Error) {
+                    t.state.send_modify(|state| match state {
+                        // TODO: the old code alluded to DeleteTenantFlow sometimes setting
+                        // TenantState::Stopping before we get here, but this may be outdated.
+                        // Let's find out with a testing assertion. If this doesn't fire, and the
+                        // logs don't show this happening in production, remove the Stopping cases.
+                        TenantState::Stopping{..} if cfg!(any(test, feature = "testing")) => {
+                            panic!("unexpected TenantState::Stopping during attach")
                         }
-                        t.state.send_modify(|state| {
-                            // The Stopping case is for when we have passed control on to DeleteTenantFlow:
-                            // if it errors, we will call make_broken when tenant is already in Stopping.
-                            assert!(
-                                matches!(*state, TenantState::Attaching | TenantState::Stopping { .. }),
-                                "the attach task owns the tenant state until activation is complete"
-                            );
-
-                            *state = TenantState::broken_from_reason(err.to_string());
-                        });
-                    };
+                        // If the tenant is cancelled, assume the error was caused by cancellation.
+                        TenantState::Attaching if t.cancel.is_cancelled() => {
+                            info!("attach cancelled, setting tenant state to Stopping: {err}");
+                            // NB: progress None tells `set_stopping` that attach has cancelled.
+                            *state = TenantState::Stopping { progress: None };
+                        }
+                        // According to the old code, DeleteTenantFlow may already have set this to
+                        // Stopping. Retain its progress.
+                        // TODO: there is no DeleteTenantFlow. Is this still needed? See above.
+                        TenantState::Stopping { progress } if t.cancel.is_cancelled() => {
+                            assert!(progress.is_some(), "concurrent attach cancellation");
+                            info!("attach cancelled, already Stopping: {err}");
+                        }
+                        // Mark the tenant as broken.
+                        TenantState::Attaching | TenantState::Stopping { .. } => {
+                            error!("attach failed, setting tenant state to Broken (was {state}): {err:?}");
+                            *state = TenantState::broken_from_reason(err.to_string())
+                        }
+                        // The attach task owns the tenant state until activated.
+                        state => panic!("invalid tenant state {state} during attach: {err:?}"),
+                    });
+                }
 
                 // TODO: should also be rejecting tenant conf changes that violate this check.
                 if let Err(e) = crate::tenant::storage_layer::inmemory_layer::IndexEntry::validate_checkpoint_distance(tenant_clone.get_checkpoint_distance()) {
-                    make_broken(&tenant_clone, anyhow::anyhow!(e), BrokenVerbosity::Error);
+                    make_broken_or_stopping(&tenant_clone, anyhow::anyhow!(e));
                     return Ok(());
                 }
 
@@ -1435,10 +1448,8 @@ impl Tenant {
                             // stayed in Activating for such a long time that shutdown found it in
                             // that state.
                             tracing::info!(state=%tenant_clone.current_state(), "Tenant shut down before activation");
-                            // Make the tenant broken so that set_stopping will not hang waiting for it to leave
-                            // the Attaching state.  This is an over-reaction (nothing really broke, the tenant is
-                            // just shutting down), but ensures progress.
-                            make_broken(&tenant_clone, anyhow::anyhow!("Shut down while Attaching"), BrokenVerbosity::Info);
+                            // Set the tenant to Stopping to signal `set_stopping` that we're done.
+                            make_broken_or_stopping(&tenant_clone, anyhow::anyhow!("Shut down while Attaching"));
                             return Ok(());
                         },
                     )
@@ -1457,7 +1468,7 @@ impl Tenant {
                         match res {
                             Ok(p) => Some(p),
                             Err(e) => {
-                                make_broken(&tenant_clone, anyhow::anyhow!(e), BrokenVerbosity::Error);
+                                make_broken_or_stopping(&tenant_clone, anyhow::anyhow!(e));
                                 return Ok(());
                             }
                         }
@@ -1483,9 +1494,7 @@ impl Tenant {
                         info!("attach finished, activating");
                         tenant_clone.activate(broker_client, None, &ctx);
                     }
-                    Err(e) => {
-                        make_broken(&tenant_clone, anyhow::anyhow!(e), BrokenVerbosity::Error);
-                    }
+                    Err(e) => make_broken_or_stopping(&tenant_clone, anyhow::anyhow!(e)),
                 }
 
                 // If we are doing an opportunistic warmup attachment at startup, initialize
@@ -1525,28 +1534,27 @@ impl Tenant {
             cancel.clone(),
         )
         .await?;
-        let (offloaded_add, tenant_manifest) =
-            match remote_timeline_client::download_tenant_manifest(
-                remote_storage,
-                &self.tenant_shard_id,
-                self.generation,
-                &cancel,
-            )
-            .await
-            {
-                Ok((tenant_manifest, _generation, _manifest_mtime)) => (
-                    format!("{} offloaded", tenant_manifest.offloaded_timelines.len()),
-                    tenant_manifest,
-                ),
-                Err(DownloadError::NotFound) => {
-                    ("no manifest".to_string(), TenantManifest::empty())
-                }
-                Err(e) => Err(e)?,
-            };
+
+        let tenant_manifest = match download_tenant_manifest(
+            remote_storage,
+            &self.tenant_shard_id,
+            self.generation,
+            &cancel,
+        )
+        .await
+        {
+            Ok((tenant_manifest, _, _)) => Some(tenant_manifest),
+            Err(DownloadError::NotFound) => None,
+            Err(err) => return Err(err.into()),
+        };
 
         info!(
-            "found {} timelines, and {offloaded_add}",
-            remote_timeline_ids.len()
+            "found {} timelines ({} offloaded timelines)",
+            remote_timeline_ids.len(),
+            tenant_manifest
+                .as_ref()
+                .map(|m| m.offloaded_timelines.len())
+                .unwrap_or(0)
         );
 
         for k in other_keys {
@@ -1555,11 +1563,13 @@ impl Tenant {
 
         // Avoid downloading IndexPart of offloaded timelines.
         let mut offloaded_with_prefix = HashSet::new();
-        for offloaded in tenant_manifest.offloaded_timelines.iter() {
-            if remote_timeline_ids.remove(&offloaded.timeline_id) {
-                offloaded_with_prefix.insert(offloaded.timeline_id);
-            } else {
-                // We'll take care later of timelines in the manifest without a prefix
+        if let Some(tenant_manifest) = &tenant_manifest {
+            for offloaded in tenant_manifest.offloaded_timelines.iter() {
+                if remote_timeline_ids.remove(&offloaded.timeline_id) {
+                    offloaded_with_prefix.insert(offloaded.timeline_id);
+                } else {
+                    // We'll take care later of timelines in the manifest without a prefix
+                }
             }
         }
 
@@ -1633,12 +1643,14 @@ impl Tenant {
 
         let mut offloaded_timeline_ids = HashSet::new();
         let mut offloaded_timelines_list = Vec::new();
-        for timeline_manifest in preload.tenant_manifest.offloaded_timelines.iter() {
-            let timeline_id = timeline_manifest.timeline_id;
-            let offloaded_timeline =
-                OffloadedTimeline::from_manifest(self.tenant_shard_id, timeline_manifest);
-            offloaded_timelines_list.push((timeline_id, Arc::new(offloaded_timeline)));
-            offloaded_timeline_ids.insert(timeline_id);
+        if let Some(tenant_manifest) = &preload.tenant_manifest {
+            for timeline_manifest in tenant_manifest.offloaded_timelines.iter() {
+                let timeline_id = timeline_manifest.timeline_id;
+                let offloaded_timeline =
+                    OffloadedTimeline::from_manifest(self.tenant_shard_id, timeline_manifest);
+                offloaded_timelines_list.push((timeline_id, Arc::new(offloaded_timeline)));
+                offloaded_timeline_ids.insert(timeline_id);
+            }
         }
         // Complete deletions for offloaded timeline id's from manifest.
         // The manifest will be uploaded later in this function.
@@ -1796,15 +1808,21 @@ impl Tenant {
             .context("resume_deletion")
             .map_err(LoadLocalTimelineError::ResumeDeletion)?;
         }
-        let needs_manifest_upload =
-            offloaded_timelines_list.len() != preload.tenant_manifest.offloaded_timelines.len();
         {
             let mut offloaded_timelines_accessor = self.timelines_offloaded.lock().unwrap();
             offloaded_timelines_accessor.extend(offloaded_timelines_list.into_iter());
         }
-        if needs_manifest_upload {
-            self.store_tenant_manifest().await?;
+
+        // Stash the preloaded tenant manifest, and upload a new manifest if changed.
+        //
+        // NB: this must happen after the tenant is fully populated above. In particular the
+        // offloaded timelines, which are included in the manifest.
+        {
+            let mut guard = self.remote_tenant_manifest.lock().await;
+            assert!(guard.is_none(), "tenant manifest set before preload"); // first populated here
+            *guard = preload.tenant_manifest;
         }
+        self.maybe_upload_tenant_manifest().await?;
 
         // The local filesystem contents are a cache of what's in the remote IndexPart;
         // IndexPart is the source of truth.
@@ -2218,7 +2236,7 @@ impl Tenant {
         };
 
         // Upload new list of offloaded timelines to S3
-        self.store_tenant_manifest().await?;
+        self.maybe_upload_tenant_manifest().await?;
 
         // Activate the timeline (if it makes sense)
         if !(timeline.is_broken() || timeline.is_stopping()) {
@@ -3429,7 +3447,7 @@ impl Tenant {
             shutdown_mode
         };
 
-        match self.set_stopping(shutdown_progress, false, false).await {
+        match self.set_stopping(shutdown_progress).await {
             Ok(()) => {}
             Err(SetStoppingError::Broken) => {
                 // assume that this is acceptable
@@ -3509,25 +3527,13 @@ impl Tenant {
     /// This function waits for the tenant to become active if it isn't already, before transitioning it into Stopping state.
     ///
     /// This function is not cancel-safe!
-    ///
-    /// `allow_transition_from_loading` is needed for the special case of loading task deleting the tenant.
-    /// `allow_transition_from_attaching` is needed for the special case of attaching deleted tenant.
-    async fn set_stopping(
-        &self,
-        progress: completion::Barrier,
-        _allow_transition_from_loading: bool,
-        allow_transition_from_attaching: bool,
-    ) -> Result<(), SetStoppingError> {
+    async fn set_stopping(&self, progress: completion::Barrier) -> Result<(), SetStoppingError> {
         let mut rx = self.state.subscribe();
 
         // cannot stop before we're done activating, so wait out until we're done activating
         rx.wait_for(|state| match state {
-            TenantState::Attaching if allow_transition_from_attaching => true,
             TenantState::Activating(_) | TenantState::Attaching => {
-                info!(
-                    "waiting for {} to turn Active|Broken|Stopping",
-                    <&'static str>::from(state)
-                );
+                info!("waiting for {state} to turn Active|Broken|Stopping");
                 false
             }
             TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => true,
@@ -3538,25 +3544,24 @@ impl Tenant {
         // we now know we're done activating, let's see whether this task is the winner to transition into Stopping
         let mut err = None;
         let stopping = self.state.send_if_modified(|current_state| match current_state {
-            TenantState::Activating(_) => {
-                unreachable!("1we ensured above that we're done with activation, and, there is no re-activation")
-            }
-            TenantState::Attaching => {
-                if !allow_transition_from_attaching {
-                    unreachable!("2we ensured above that we're done with activation, and, there is no re-activation")
-                };
-                *current_state = TenantState::Stopping { progress };
-                true
+            TenantState::Activating(_) | TenantState::Attaching => {
+                unreachable!("we ensured above that we're done with activation, and, there is no re-activation")
             }
             TenantState::Active => {
                 // FIXME: due to time-of-check vs time-of-use issues, it can happen that new timelines
                 // are created after the transition to Stopping. That's harmless, as the Timelines
                 // won't be accessible to anyone afterwards, because the Tenant is in Stopping state.
-                *current_state = TenantState::Stopping { progress };
+                *current_state = TenantState::Stopping { progress: Some(progress) };
                 // Continue stopping outside the closure. We need to grab timelines.lock()
                 // and we plan to turn it into a tokio::sync::Mutex in a future patch.
                 true
             }
+            TenantState::Stopping { progress: None } => {
+                // An attach was cancelled, and the attach transitioned the tenant from Attaching to
+                // Stopping(None) to let us know it exited. Register our progress and continue.
+                *current_state = TenantState::Stopping { progress: Some(progress) };
+                true
+            }
             TenantState::Broken { reason, .. } => {
                 info!(
                     "Cannot set tenant to Stopping state, it is in Broken state due to: {reason}"
@@ -3564,7 +3569,7 @@ impl Tenant {
                 err = Some(SetStoppingError::Broken);
                 false
             }
-            TenantState::Stopping { progress } => {
+            TenantState::Stopping { progress: Some(progress) } => {
                 info!("Tenant is already in Stopping state");
                 err = Some(SetStoppingError::AlreadyStopping(progress.clone()));
                 false
@@ -4065,18 +4070,20 @@ impl Tenant {
 
     /// Generate an up-to-date TenantManifest based on the state of this Tenant.
     fn build_tenant_manifest(&self) -> TenantManifest {
-        let timelines_offloaded = self.timelines_offloaded.lock().unwrap();
-
-        let mut timeline_manifests = timelines_offloaded
-            .iter()
-            .map(|(_timeline_id, offloaded)| offloaded.manifest())
-            .collect::<Vec<_>>();
-        // Sort the manifests so that our output is deterministic
-        timeline_manifests.sort_by_key(|timeline_manifest| timeline_manifest.timeline_id);
+        // Collect the offloaded timelines, and sort them for deterministic output.
+        let offloaded_timelines = self
+            .timelines_offloaded
+            .lock()
+            .unwrap()
+            .values()
+            .map(|tli| tli.manifest())
+            .sorted_by_key(|m| m.timeline_id)
+            .collect_vec();
 
         TenantManifest {
             version: LATEST_TENANT_MANIFEST_VERSION,
-            offloaded_timelines: timeline_manifests,
+            stripe_size: Some(self.get_shard_stripe_size()),
+            offloaded_timelines,
         }
     }
 
@@ -4299,7 +4306,7 @@ impl Tenant {
             timelines: Mutex::new(HashMap::new()),
             timelines_creating: Mutex::new(HashSet::new()),
             timelines_offloaded: Mutex::new(HashMap::new()),
-            tenant_manifest_upload: Default::default(),
+            remote_tenant_manifest: Default::default(),
             gc_cs: tokio::sync::Mutex::new(()),
             walredo_mgr,
             remote_storage,
@@ -4395,10 +4402,7 @@ impl Tenant {
         .to_string();
 
         fail::fail_point!("tenant-config-before-write", |_| {
-            Err(std::io::Error::new(
-                std::io::ErrorKind::Other,
-                "tenant-config-before-write",
-            ))
+            Err(std::io::Error::other("tenant-config-before-write"))
         });
 
         // Convert the config to a toml file.
@@ -5532,27 +5536,35 @@ impl Tenant {
             .unwrap_or(0)
     }
 
-    /// Serialize and write the latest TenantManifest to remote storage.
-    pub(crate) async fn store_tenant_manifest(&self) -> Result<(), TenantManifestError> {
-        // Only one manifest write may be done at at time, and the contents of the manifest
-        // must be loaded while holding this lock. This makes it safe to call this function
-        // from anywhere without worrying about colliding updates.
+    /// Builds a new tenant manifest, and uploads it if it differs from the last-known tenant
+    /// manifest in `Self::remote_tenant_manifest`.
+    ///
+    /// TODO: instead of requiring callers to remember to call `maybe_upload_tenant_manifest` after
+    /// changing any `Tenant` state that's included in the manifest, consider making the manifest
+    /// the authoritative source of data with an API that automatically uploads on changes. Revisit
+    /// this when the manifest is more widely used and we have a better idea of the data model.
+    pub(crate) async fn maybe_upload_tenant_manifest(&self) -> Result<(), TenantManifestError> {
+        // Multiple tasks may call this function concurrently after mutating the Tenant runtime
+        // state, affecting the manifest generated by `build_tenant_manifest`. We use an async mutex
+        // to serialize these callers. `eq_ignoring_version` acts as a slightly inefficient but
+        // simple coalescing mechanism.
         let mut guard = tokio::select! {
-            g = self.tenant_manifest_upload.lock() => {
-                g
-            },
-            _ = self.cancel.cancelled() => {
-                return Err(TenantManifestError::Cancelled);
-            }
+            guard = self.remote_tenant_manifest.lock() => guard,
+            _ = self.cancel.cancelled() => return Err(TenantManifestError::Cancelled),
         };
 
+        // Build a new manifest.
         let manifest = self.build_tenant_manifest();
-        if Some(&manifest) == (*guard).as_ref() {
-            // Optimisation: skip uploads that don't change anything.
-            return Ok(());
+
+        // Check if the manifest has changed. We ignore the version number here, to avoid
+        // uploading every manifest on version number bumps.
+        if let Some(old) = guard.as_ref() {
+            if manifest.eq_ignoring_version(old) {
+                return Ok(());
+            }
         }
 
-        // Remote storage does no retries internally, so wrap it
+        // Upload the manifest. Remote storage does no retries internally, so retry here.
         match backoff::retry(
             || async {
                 upload_tenant_manifest(
@@ -5564,7 +5576,7 @@ impl Tenant {
                 )
                 .await
             },
-            |_e| self.cancel.is_cancelled(),
+            |_| self.cancel.is_cancelled(),
             FAILED_UPLOAD_WARN_THRESHOLD,
             FAILED_REMOTE_OP_RETRIES,
             "uploading tenant manifest",
@@ -5868,6 +5880,7 @@ pub(crate) mod harness {
             base_img: Option<(Lsn, Bytes)>,
             records: Vec<(Lsn, NeonWalRecord)>,
             _pg_version: u32,
+            _redo_attempt_type: RedoAttemptType,
         ) -> Result<Bytes, walredo::Error> {
             let records_neon = records.iter().all(|r| apply_neon::can_apply_in_neon(&r.1));
             if records_neon {
@@ -8722,6 +8735,21 @@ mod tests {
                 Lsn(0x20),
                 Value::WalRecord(NeonWalRecord::wal_init("i")),
             ),
+            (
+                get_key(4),
+                Lsn(0x30),
+                Value::WalRecord(NeonWalRecord::wal_append_conditional("j", "i")),
+            ),
+            (
+                get_key(5),
+                Lsn(0x20),
+                Value::WalRecord(NeonWalRecord::wal_init("1")),
+            ),
+            (
+                get_key(5),
+                Lsn(0x30),
+                Value::WalRecord(NeonWalRecord::wal_append_conditional("j", "2")),
+            ),
         ];
         let image1 = vec![(get_key(1), "0x10".into())];
 
@@ -8752,8 +8780,18 @@ mod tests {
 
         // Need to remove the limit of "Neon WAL redo requires base image".
 
-        // assert_eq!(tline.get(get_key(3), Lsn(0x50), &ctx).await?, Bytes::new());
-        // assert_eq!(tline.get(get_key(4), Lsn(0x50), &ctx).await?, Bytes::new());
+        assert_eq!(
+            tline.get(get_key(3), Lsn(0x50), &ctx).await?,
+            Bytes::from_static(b"c")
+        );
+        assert_eq!(
+            tline.get(get_key(4), Lsn(0x50), &ctx).await?,
+            Bytes::from_static(b"ij")
+        );
+
+        // Manual testing required: currently, read errors will panic the process in debug mode. So we
+        // cannot enable this assertion in the unit test.
+        // assert!(tline.get(get_key(5), Lsn(0x50), &ctx).await.is_err());
 
         Ok(())
     }
@@ -11533,6 +11571,99 @@ mod tests {
         Ok(())
     }
 
+    #[cfg(feature = "testing")]
+    #[tokio::test]
+    async fn test_bottom_most_compation_redo_failure() -> anyhow::Result<()> {
+        let harness = TenantHarness::create("test_bottom_most_compation_redo_failure").await?;
+        let (tenant, ctx) = harness.load().await;
+
+        fn get_key(id: u32) -> Key {
+            // using aux key here b/c they are guaranteed to be inside `collect_keyspace`.
+            let mut key = Key::from_hex("620000000033333333444444445500000000").unwrap();
+            key.field6 = id;
+            key
+        }
+
+        let img_layer = (0..10)
+            .map(|id| (get_key(id), Bytes::from(format!("value {id}@0x10"))))
+            .collect_vec();
+
+        let delta1 = vec![
+            (
+                get_key(1),
+                Lsn(0x20),
+                Value::WalRecord(NeonWalRecord::wal_append("@0x20")),
+            ),
+            (
+                get_key(1),
+                Lsn(0x24),
+                Value::WalRecord(NeonWalRecord::wal_append("@0x24")),
+            ),
+            (
+                get_key(1),
+                Lsn(0x28),
+                // This record will fail to redo
+                Value::WalRecord(NeonWalRecord::wal_append_conditional("@0x28", "???")),
+            ),
+        ];
+
+        let tline = tenant
+            .create_test_timeline_with_layers(
+                TIMELINE_ID,
+                Lsn(0x10),
+                DEFAULT_PG_VERSION,
+                &ctx,
+                vec![], // in-memory layers
+                vec![DeltaLayerTestDesc::new_with_inferred_key_range(
+                    Lsn(0x20)..Lsn(0x30),
+                    delta1,
+                )], // delta layers
+                vec![(Lsn(0x10), img_layer)], // image layers
+                Lsn(0x50),
+            )
+            .await?;
+        {
+            tline
+                .applied_gc_cutoff_lsn
+                .lock_for_write()
+                .store_and_unlock(Lsn(0x30))
+                .wait()
+                .await;
+            // Update GC info
+            let mut guard = tline.gc_info.write().unwrap();
+            *guard = GcInfo {
+                retain_lsns: vec![],
+                cutoffs: GcCutoffs {
+                    time: Lsn(0x30),
+                    space: Lsn(0x30),
+                },
+                leases: Default::default(),
+                within_ancestor_pitr: false,
+            };
+        }
+
+        let cancel = CancellationToken::new();
+
+        // Compaction will fail, but should not fire any critical error.
+        // Gc-compaction currently cannot figure out what keys are not in the keyspace during the compaction
+        // process. It will always try to redo the logs it reads and if it doesn't work, fail the entire
+        // compaction job. Tracked in <https://github.com/neondatabase/neon/issues/10395>.
+        let res = tline
+            .compact_with_gc(
+                &cancel,
+                CompactOptions {
+                    compact_key_range: None,
+                    compact_lsn_range: None,
+                    ..Default::default()
+                },
+                &ctx,
+            )
+            .await;
+        assert!(res.is_err());
+
+        Ok(())
+    }
+
     #[cfg(feature = "testing")]
     #[tokio::test]
     async fn test_synthetic_size_calculation_with_invisible_branches() -> anyhow::Result<()> {

pageserver/src/tenant/blob_io.rs

@@ -15,20 +15,23 @@
 //! len >= 128: 1CCCXXXX XXXXXXXX XXXXXXXX XXXXXXXX
 //!
 use std::cmp::min;
-use std::io::{Error, ErrorKind};
+use std::sync::Arc;
 
 use async_compression::Level;
 use bytes::{BufMut, BytesMut};
 use pageserver_api::models::ImageCompressionAlgorithm;
 use tokio::io::AsyncWriteExt;
-use tokio_epoll_uring::{BoundedBuf, IoBuf, Slice};
+use tokio_epoll_uring::IoBuf;
+use tokio_util::sync::CancellationToken;
 use tracing::warn;
 
 use crate::context::RequestContext;
 use crate::page_cache::PAGE_SZ;
 use crate::tenant::block_io::BlockCursor;
+use crate::virtual_file::IoBufferMut;
 use crate::virtual_file::VirtualFile;
 use crate::virtual_file::owned_buffers_io::io_buf_ext::{FullSlice, IoBufExt};
+use crate::virtual_file::owned_buffers_io::write::{BufferedWriter, FlushTaskError};
 
 #[derive(Copy, Clone, Debug)]
 pub struct CompressionInfo {
@@ -36,6 +39,14 @@ pub struct CompressionInfo {
     pub compressed_size: Option<usize>,
 }
 
+#[derive(Debug, thiserror::Error)]
+pub enum WriteBlobError {
+    #[error(transparent)]
+    Flush(FlushTaskError),
+    #[error("blob too large ({len} bytes)")]
+    BlobTooLarge { len: usize },
+}
+
 impl BlockCursor<'_> {
     /// Read a blob into a new buffer.
     pub async fn read_blob(
@@ -157,135 +168,62 @@ pub(super) const BYTE_ZSTD: u8 = BYTE_UNCOMPRESSED | 0x10;
 /// A wrapper of `VirtualFile` that allows users to write blobs.
 ///
 /// If a `BlobWriter` is dropped, the internal buffer will be
-/// discarded. You need to call [`flush_buffer`](Self::flush_buffer)
+/// discarded. You need to call [`Self::into_inner`]
 /// manually before dropping.
-pub struct BlobWriter<const BUFFERED: bool> {
-    inner: VirtualFile,
-    offset: u64,
-    /// A buffer to save on write calls, only used if BUFFERED=true
-    buf: Vec<u8>,
+pub struct BlobWriter {
     /// We do tiny writes for the length headers; they need to be in an owned buffer;
     io_buf: Option<BytesMut>,
+    writer: BufferedWriter<IoBufferMut, VirtualFile>,
+    offset: u64,
 }
 
-impl<const BUFFERED: bool> BlobWriter<BUFFERED> {
-    pub fn new(inner: VirtualFile, start_offset: u64) -> Self {
-        Self {
-            inner,
-            offset: start_offset,
-            buf: Vec::with_capacity(Self::CAPACITY),
+impl BlobWriter {
+    pub fn new(
+        file: Arc<VirtualFile>,
+        start_offset: u64,
+        gate: &utils::sync::gate::Gate,
+        cancel: CancellationToken,
+        ctx: &RequestContext,
+        flush_task_span: tracing::Span,
+    ) -> anyhow::Result<Self> {
+        Ok(Self {
             io_buf: Some(BytesMut::new()),
-        }
+            writer: BufferedWriter::new(
+                file,
+                start_offset,
+                || IoBufferMut::with_capacity(Self::CAPACITY),
+                gate.enter()?,
+                cancel,
+                ctx,
+                flush_task_span,
+            ),
+            offset: start_offset,
+        })
     }
 
     pub fn size(&self) -> u64 {
         self.offset
     }
 
-    const CAPACITY: usize = if BUFFERED { 64 * 1024 } else { 0 };
+    const CAPACITY: usize = 64 * 1024;
 
-    /// Writes the given buffer directly to the underlying `VirtualFile`.
-    /// You need to make sure that the internal buffer is empty, otherwise
-    /// data will be written in wrong order.
-    #[inline(always)]
-    async fn write_all_unbuffered<Buf: IoBuf + Send>(
-        &mut self,
-        src_buf: FullSlice<Buf>,
-        ctx: &RequestContext,
-    ) -> (FullSlice<Buf>, Result<(), Error>) {
-        let (src_buf, res) = self.inner.write_all(src_buf, ctx).await;
-        let nbytes = match res {
-            Ok(nbytes) => nbytes,
-            Err(e) => return (src_buf, Err(e)),
-        };
-        self.offset += nbytes as u64;
-        (src_buf, Ok(()))
-    }
-
-    #[inline(always)]
-    /// Flushes the internal buffer to the underlying `VirtualFile`.
-    pub async fn flush_buffer(&mut self, ctx: &RequestContext) -> Result<(), Error> {
-        let buf = std::mem::take(&mut self.buf);
-        let (slice, res) = self.inner.write_all(buf.slice_len(), ctx).await;
-        res?;
-        let mut buf = slice.into_raw_slice().into_inner();
-        buf.clear();
-        self.buf = buf;
-        Ok(())
-    }
-
-    #[inline(always)]
-    /// Writes as much of `src_buf` into the internal buffer as it fits
-    fn write_into_buffer(&mut self, src_buf: &[u8]) -> usize {
-        let remaining = Self::CAPACITY - self.buf.len();
-        let to_copy = src_buf.len().min(remaining);
-        self.buf.extend_from_slice(&src_buf[..to_copy]);
-        self.offset += to_copy as u64;
-        to_copy
-    }
-
-    /// Internal, possibly buffered, write function
+    /// Writes `src_buf` to the file at the current offset.
     async fn write_all<Buf: IoBuf + Send>(
         &mut self,
         src_buf: FullSlice<Buf>,
         ctx: &RequestContext,
-    ) -> (FullSlice<Buf>, Result<(), Error>) {
-        let src_buf = src_buf.into_raw_slice();
-        let src_buf_bounds = src_buf.bounds();
-        let restore = move |src_buf_slice: Slice<_>| {
-            FullSlice::must_new(Slice::from_buf_bounds(
-                src_buf_slice.into_inner(),
-                src_buf_bounds,
-            ))
-        };
+    ) -> (FullSlice<Buf>, Result<(), FlushTaskError>) {
+        let res = self
+            .writer
+            // TODO: why are we taking a FullSlice if we're going to pass a borrow downstack?
+            // Can remove all the complexity around owned buffers upstack
+            .write_buffered_borrowed(&src_buf, ctx)
+            .await
+            .map(|len| {
+                self.offset += len as u64;
+            });
 
-        if !BUFFERED {
-            assert!(self.buf.is_empty());
-            return self
-                .write_all_unbuffered(FullSlice::must_new(src_buf), ctx)
-                .await;
-        }
-        let remaining = Self::CAPACITY - self.buf.len();
-        let src_buf_len = src_buf.bytes_init();
-        if src_buf_len == 0 {
-            return (restore(src_buf), Ok(()));
-        }
-        let mut src_buf = src_buf.slice(0..src_buf_len);
-        // First try to copy as much as we can into the buffer
-        if remaining > 0 {
-            let copied = self.write_into_buffer(&src_buf);
-            src_buf = src_buf.slice(copied..);
-        }
-        // Then, if the buffer is full, flush it out
-        if self.buf.len() == Self::CAPACITY {
-            if let Err(e) = self.flush_buffer(ctx).await {
-                return (restore(src_buf), Err(e));
-            }
-        }
-        // Finally, write the tail of src_buf:
-        // If it wholly fits into the buffer without
-        // completely filling it, then put it there.
-        // If not, write it out directly.
-        let src_buf = if !src_buf.is_empty() {
-            assert_eq!(self.buf.len(), 0);
-            if src_buf.len() < Self::CAPACITY {
-                let copied = self.write_into_buffer(&src_buf);
-                // We just verified above that src_buf fits into our internal buffer.
-                assert_eq!(copied, src_buf.len());
-                restore(src_buf)
-            } else {
-                let (src_buf, res) = self
-                    .write_all_unbuffered(FullSlice::must_new(src_buf), ctx)
-                    .await;
-                if let Err(e) = res {
-                    return (src_buf, Err(e));
-                }
-                src_buf
-            }
-        } else {
-            restore(src_buf)
-        };
-        (src_buf, Ok(()))
+        (src_buf, res)
     }
 
     /// Write a blob of data. Returns the offset that it was written to,
@@ -294,7 +232,7 @@ impl<const BUFFERED: bool> BlobWriter<BUFFERED> {
         &mut self,
         srcbuf: FullSlice<Buf>,
         ctx: &RequestContext,
-    ) -> (FullSlice<Buf>, Result<u64, Error>) {
+    ) -> (FullSlice<Buf>, Result<u64, WriteBlobError>) {
         let (buf, res) = self
             .write_blob_maybe_compressed(srcbuf, ctx, ImageCompressionAlgorithm::Disabled)
             .await;
@@ -308,7 +246,10 @@ impl<const BUFFERED: bool> BlobWriter<BUFFERED> {
         srcbuf: FullSlice<Buf>,
         ctx: &RequestContext,
         algorithm: ImageCompressionAlgorithm,
-    ) -> (FullSlice<Buf>, Result<(u64, CompressionInfo), Error>) {
+    ) -> (
+        FullSlice<Buf>,
+        Result<(u64, CompressionInfo), WriteBlobError>,
+    ) {
         let offset = self.offset;
         let mut compression_info = CompressionInfo {
             written_compressed: false,
@@ -324,17 +265,16 @@ impl<const BUFFERED: bool> BlobWriter<BUFFERED> {
             if len < 128 {
                 // Short blob. Write a 1-byte length header
                 io_buf.put_u8(len as u8);
-                (self.write_all(io_buf.slice_len(), ctx).await, srcbuf)
+                let (slice, res) = self.write_all(io_buf.slice_len(), ctx).await;
+                let res = res.map_err(WriteBlobError::Flush);
+                ((slice, res), srcbuf)
             } else {
                 // Write a 4-byte length header
                 if len > MAX_SUPPORTED_BLOB_LEN {
                     return (
                         (
                             io_buf.slice_len(),
-                            Err(Error::new(
-                                ErrorKind::Other,
-                                format!("blob too large ({len} bytes)"),
-                            )),
+                            Err(WriteBlobError::BlobTooLarge { len }),
                         ),
                         srcbuf,
                     );
@@ -368,7 +308,9 @@ impl<const BUFFERED: bool> BlobWriter<BUFFERED> {
                 assert_eq!(len_buf[0] & 0xf0, 0);
                 len_buf[0] |= high_bit_mask;
                 io_buf.extend_from_slice(&len_buf[..]);
-                (self.write_all(io_buf.slice_len(), ctx).await, srcbuf)
+                let (slice, res) = self.write_all(io_buf.slice_len(), ctx).await;
+                let res = res.map_err(WriteBlobError::Flush);
+                ((slice, res), srcbuf)
             }
         }
         .await;
@@ -383,33 +325,23 @@ impl<const BUFFERED: bool> BlobWriter<BUFFERED> {
         } else {
             self.write_all(srcbuf, ctx).await
         };
+        let res = res.map_err(WriteBlobError::Flush);
         (srcbuf, res.map(|_| (offset, compression_info)))
     }
-}
 
-impl BlobWriter<true> {
     /// Access the underlying `VirtualFile`.
     ///
     /// This function flushes the internal buffer before giving access
     /// to the underlying `VirtualFile`.
-    pub async fn into_inner(mut self, ctx: &RequestContext) -> Result<VirtualFile, Error> {
-        self.flush_buffer(ctx).await?;
-        Ok(self.inner)
-    }
-
-    /// Access the underlying `VirtualFile`.
     ///
-    /// Unlike [`into_inner`](Self::into_inner), this doesn't flush
-    /// the internal buffer before giving access.
-    pub fn into_inner_no_flush(self) -> VirtualFile {
-        self.inner
-    }
-}
-
-impl BlobWriter<false> {
-    /// Access the underlying `VirtualFile`.
-    pub fn into_inner(self) -> VirtualFile {
-        self.inner
+    /// The caller can use the `handle_tail` function to change the tail of the buffer before flushing it to disk.
+    /// The buffer will not be flushed to disk if handle_tail returns `None`.
+    pub async fn into_inner(
+        self,
+        handle_tail: impl FnMut(IoBufferMut) -> Option<IoBufferMut>,
+    ) -> Result<VirtualFile, FlushTaskError> {
+        let (_, file) = self.writer.shutdown(handle_tail).await?;
+        Ok(file)
     }
 }
 
@@ -418,29 +350,33 @@ pub(crate) mod tests {
     use camino::Utf8PathBuf;
     use camino_tempfile::Utf8TempDir;
     use rand::{Rng, SeedableRng};
+    use tracing::info_span;
 
     use super::*;
     use crate::context::DownloadBehavior;
     use crate::task_mgr::TaskKind;
     use crate::tenant::block_io::BlockReaderRef;
 
-    async fn round_trip_test<const BUFFERED: bool>(blobs: &[Vec<u8>]) -> Result<(), Error> {
-        round_trip_test_compressed::<BUFFERED>(blobs, false).await
+    async fn round_trip_test(blobs: &[Vec<u8>]) -> anyhow::Result<()> {
+        round_trip_test_compressed(blobs, false).await
     }
 
-    pub(crate) async fn write_maybe_compressed<const BUFFERED: bool>(
+    pub(crate) async fn write_maybe_compressed(
         blobs: &[Vec<u8>],
         compression: bool,
         ctx: &RequestContext,
-    ) -> Result<(Utf8TempDir, Utf8PathBuf, Vec<u64>), Error> {
+    ) -> anyhow::Result<(Utf8TempDir, Utf8PathBuf, Vec<u64>)> {
         let temp_dir = camino_tempfile::tempdir()?;
         let pathbuf = temp_dir.path().join("file");
+        let gate = utils::sync::gate::Gate::default();
+        let cancel = CancellationToken::new();
 
         // Write part (in block to drop the file)
         let mut offsets = Vec::new();
         {
-            let file = VirtualFile::create(pathbuf.as_path(), ctx).await?;
-            let mut wtr = BlobWriter::<BUFFERED>::new(file, 0);
+            let file = Arc::new(VirtualFile::create_v2(pathbuf.as_path(), ctx).await?);
+            let mut wtr =
+                BlobWriter::new(file, 0, &gate, cancel.clone(), ctx, info_span!("test")).unwrap();
             for blob in blobs.iter() {
                 let (_, res) = if compression {
                     let res = wtr
@@ -457,26 +393,37 @@ pub(crate) mod tests {
                 let offs = res?;
                 offsets.push(offs);
             }
-            // Write out one page worth of zeros so that we can
-            // read again with read_blk
-            let (_, res) = wtr.write_blob(vec![0; PAGE_SZ].slice_len(), ctx).await;
-            let offs = res?;
-            println!("Writing final blob at offs={offs}");
-            wtr.flush_buffer(ctx).await?;
+            wtr.into_inner(|mut buf| {
+                use crate::virtual_file::owned_buffers_io::write::Buffer;
+
+                let len = buf.pending();
+                let cap = buf.cap();
+
+                // pad zeros to the next io alignment requirement.
+                // TODO: this is actually padding to next PAGE_SZ multiple, but only if the buffer capacity is larger than that.
+                // We can't let the fact that we do direct IO, or the buffer capacity, dictate the on-disk format we write here.
+                // Need to find a better API that allows writing the format we intend to.
+                let count = len.next_multiple_of(PAGE_SZ).min(cap) - len;
+                buf.extend_with(0, count);
+
+                Some(buf)
+            })
+            .await?; // TODO: this here is the problem with the tests: we're dropping the tail end
         }
         Ok((temp_dir, pathbuf, offsets))
     }
 
-    async fn round_trip_test_compressed<const BUFFERED: bool>(
+    async fn round_trip_test_compressed(
         blobs: &[Vec<u8>],
         compression: bool,
-    ) -> Result<(), Error> {
+    ) -> anyhow::Result<()> {
         let ctx =
             RequestContext::new(TaskKind::UnitTest, DownloadBehavior::Error).with_scope_unit_test();
         let (_temp_dir, pathbuf, offsets) =
-            write_maybe_compressed::<BUFFERED>(blobs, compression, &ctx).await?;
+            write_maybe_compressed(blobs, compression, &ctx).await?;
 
-        let file = VirtualFile::open(pathbuf, &ctx).await?;
+        println!("Done writing!");
+        let file = VirtualFile::open_v2(pathbuf, &ctx).await?;
         let rdr = BlockReaderRef::VirtualFile(&file);
         let rdr = BlockCursor::new_with_compression(rdr, compression);
         for (idx, (blob, offset)) in blobs.iter().zip(offsets.iter()).enumerate() {
@@ -495,30 +442,27 @@ pub(crate) mod tests {
     }
 
     #[tokio::test]
-    async fn test_one() -> Result<(), Error> {
+    async fn test_one() -> anyhow::Result<()> {
         let blobs = &[vec![12, 21, 22]];
-        round_trip_test::<false>(blobs).await?;
-        round_trip_test::<true>(blobs).await?;
+        round_trip_test(blobs).await?;
         Ok(())
     }
 
     #[tokio::test]
-    async fn test_hello_simple() -> Result<(), Error> {
+    async fn test_hello_simple() -> anyhow::Result<()> {
         let blobs = &[
             vec![0, 1, 2, 3],
             b"Hello, World!".to_vec(),
             Vec::new(),
             b"foobar".to_vec(),
         ];
-        round_trip_test::<false>(blobs).await?;
-        round_trip_test::<true>(blobs).await?;
-        round_trip_test_compressed::<false>(blobs, true).await?;
-        round_trip_test_compressed::<true>(blobs, true).await?;
+        round_trip_test(blobs).await?;
+        round_trip_test_compressed(blobs, true).await?;
         Ok(())
     }
 
     #[tokio::test]
-    async fn test_really_big_array() -> Result<(), Error> {
+    async fn test_really_big_array() -> anyhow::Result<()> {
         let blobs = &[
             b"test".to_vec(),
             random_array(10 * PAGE_SZ),
@@ -527,25 +471,22 @@ pub(crate) mod tests {
             vec![0xf3; 24 * PAGE_SZ],
             b"foobar".to_vec(),
         ];
-        round_trip_test::<false>(blobs).await?;
-        round_trip_test::<true>(blobs).await?;
-        round_trip_test_compressed::<false>(blobs, true).await?;
-        round_trip_test_compressed::<true>(blobs, true).await?;
+        round_trip_test(blobs).await?;
+        round_trip_test_compressed(blobs, true).await?;
         Ok(())
     }
 
     #[tokio::test]
-    async fn test_arrays_inc() -> Result<(), Error> {
+    async fn test_arrays_inc() -> anyhow::Result<()> {
         let blobs = (0..PAGE_SZ / 8)
             .map(|v| random_array(v * 16))
             .collect::<Vec<_>>();
-        round_trip_test::<false>(&blobs).await?;
-        round_trip_test::<true>(&blobs).await?;
+        round_trip_test(&blobs).await?;
         Ok(())
     }
 
     #[tokio::test]
-    async fn test_arrays_random_size() -> Result<(), Error> {
+    async fn test_arrays_random_size() -> anyhow::Result<()> {
         let mut rng = rand::rngs::StdRng::seed_from_u64(42);
         let blobs = (0..1024)
             .map(|_| {
@@ -557,20 +498,18 @@ pub(crate) mod tests {
                 random_array(sz.into())
             })
             .collect::<Vec<_>>();
-        round_trip_test::<false>(&blobs).await?;
-        round_trip_test::<true>(&blobs).await?;
+        round_trip_test(&blobs).await?;
         Ok(())
     }
 
     #[tokio::test]
-    async fn test_arrays_page_boundary() -> Result<(), Error> {
+    async fn test_arrays_page_boundary() -> anyhow::Result<()> {
         let blobs = &[
             random_array(PAGE_SZ - 4),
             random_array(PAGE_SZ - 4),
             random_array(PAGE_SZ - 4),
         ];
-        round_trip_test::<false>(blobs).await?;
-        round_trip_test::<true>(blobs).await?;
+        round_trip_test(blobs).await?;
         Ok(())
     }
 }

pageserver/src/tenant/block_io.rs

@@ -4,14 +4,12 @@
 
 use std::ops::Deref;
 
-use bytes::Bytes;
-
 use super::storage_layer::delta_layer::{Adapter, DeltaLayerInner};
 use crate::context::RequestContext;
 use crate::page_cache::{self, FileId, PAGE_SZ, PageReadGuard, PageWriteGuard, ReadBufResult};
 #[cfg(test)]
 use crate::virtual_file::IoBufferMut;
-use crate::virtual_file::VirtualFile;
+use crate::virtual_file::{IoBuffer, VirtualFile};
 
 /// This is implemented by anything that can read 8 kB (PAGE_SZ)
 /// blocks, using the page cache
@@ -216,12 +214,8 @@ impl<'a> FileBlockReader<'a> {
         match cache
             .read_immutable_buf(self.file_id, blknum, ctx)
             .await
-            .map_err(|e| {
-                std::io::Error::new(
-                    std::io::ErrorKind::Other,
-                    format!("Failed to read immutable buf: {e:#}"),
-                )
-            })? {
+            .map_err(|e| std::io::Error::other(format!("Failed to read immutable buf: {e:#}")))?
+        {
             ReadBufResult::Found(guard) => Ok(guard.into()),
             ReadBufResult::NotFound(write_guard) => {
                 // Read the page from disk into the buffer
@@ -251,17 +245,17 @@ pub trait BlockWriter {
     /// 'buf' must be of size PAGE_SZ. Returns the block number the page was
     /// written to.
     ///
-    fn write_blk(&mut self, buf: Bytes) -> Result<u32, std::io::Error>;
+    fn write_blk(&mut self, buf: IoBuffer) -> Result<u32, std::io::Error>;
 }
 
 ///
 /// A simple in-memory buffer of blocks.
 ///
 pub struct BlockBuf {
-    pub blocks: Vec<Bytes>,
+    pub blocks: Vec<IoBuffer>,
 }
 impl BlockWriter for BlockBuf {
-    fn write_blk(&mut self, buf: Bytes) -> Result<u32, std::io::Error> {
+    fn write_blk(&mut self, buf: IoBuffer) -> Result<u32, std::io::Error> {
         assert!(buf.len() == PAGE_SZ);
         let blknum = self.blocks.len();
         self.blocks.push(buf);

pageserver/src/tenant/disk_btree.rs

@@ -25,7 +25,7 @@ use std::{io, result};
 
 use async_stream::try_stream;
 use byteorder::{BE, ReadBytesExt};
-use bytes::{BufMut, Bytes, BytesMut};
+use bytes::BufMut;
 use either::Either;
 use futures::{Stream, StreamExt};
 use hex;
@@ -34,6 +34,7 @@ use tracing::error;
 
 use crate::context::RequestContext;
 use crate::tenant::block_io::{BlockReader, BlockWriter};
+use crate::virtual_file::{IoBuffer, IoBufferMut, owned_buffers_io::write::Buffer};
 
 // The maximum size of a value stored in the B-tree. 5 bytes is enough currently.
 pub const VALUE_SZ: usize = 5;
@@ -787,12 +788,12 @@ impl<const L: usize> BuildNode<L> {
     ///
     /// Serialize the node to on-disk format.
     ///
-    fn pack(&self) -> Bytes {
+    fn pack(&self) -> IoBuffer {
         assert!(self.keys.len() == self.num_children as usize * self.suffix_len);
         assert!(self.values.len() == self.num_children as usize * VALUE_SZ);
         assert!(self.num_children > 0);
 
-        let mut buf = BytesMut::new();
+        let mut buf = IoBufferMut::with_capacity(PAGE_SZ);
 
         buf.put_u16(self.num_children);
         buf.put_u8(self.level);
@@ -805,7 +806,7 @@ impl<const L: usize> BuildNode<L> {
         assert!(buf.len() == self.size);
 
         assert!(buf.len() <= PAGE_SZ);
-        buf.resize(PAGE_SZ, 0);
+        buf.extend_with(0, PAGE_SZ - buf.len());
         buf.freeze()
     }
 
@@ -839,7 +840,7 @@ pub(crate) mod tests {
 
     #[derive(Clone, Default)]
     pub(crate) struct TestDisk {
-        blocks: Vec<Bytes>,
+        blocks: Vec<IoBuffer>,
     }
     impl TestDisk {
         fn new() -> Self {
@@ -857,7 +858,7 @@ pub(crate) mod tests {
         }
     }
     impl BlockWriter for &mut TestDisk {
-        fn write_blk(&mut self, buf: Bytes) -> io::Result<u32> {
+        fn write_blk(&mut self, buf: IoBuffer) -> io::Result<u32> {
             let blknum = self.blocks.len();
             self.blocks.push(buf);
             Ok(blknum as u32)

pageserver/src/tenant/ephemeral_file.rs

@@ -75,6 +75,7 @@ impl EphemeralFile {
             bytes_written: 0,
             buffered_writer: owned_buffers_io::write::BufferedWriter::new(
                 file,
+                0,
                 || IoBufferMut::with_capacity(TAIL_SZ),
                 gate.enter()?,
                 cancel.child_token(),

pageserver/src/tenant/remote_timeline_client/download.rs

@@ -32,12 +32,14 @@ use super::{
     remote_tenant_manifest_prefix, remote_tenant_path,
 };
 use crate::TEMP_FILE_SUFFIX;
+use crate::assert_u64_eq_usize::UsizeIsU64;
 use crate::config::PageServerConf;
 use crate::context::RequestContext;
 use crate::span::{
     debug_assert_current_span_has_tenant_and_timeline_id, debug_assert_current_span_has_tenant_id,
 };
 use crate::tenant::Generation;
+use crate::tenant::disk_btree::PAGE_SZ;
 use crate::tenant::remote_timeline_client::{remote_layer_path, remote_timelines_path};
 use crate::tenant::storage_layer::LayerName;
 use crate::virtual_file::{MaybeFatalIo, VirtualFile, on_fatal_io_error};
@@ -227,6 +229,7 @@ async fn download_object(
 
                 let mut buffered = owned_buffers_io::write::BufferedWriter::<IoBufferMut, _>::new(
                     destination_file,
+                    0,
                     || IoBufferMut::with_capacity(super::BUFFER_SIZE),
                     gate.enter().map_err(|_| DownloadError::Cancelled)?,
                     cancel.child_token(),
@@ -251,13 +254,41 @@ async fn download_object(
                                 FlushTaskError::Cancelled => DownloadError::Cancelled,
                             })?;
                     }
-                    let inner = buffered
-                        .flush_and_into_inner(ctx)
+                    let mut pad_amount = None;
+                    let (bytes_amount, destination_file) = buffered
+                        .shutdown(|mut buf| {
+                            use crate::virtual_file::owned_buffers_io::write::Buffer;
+
+                            let len = buf.pending();
+                            let cap = buf.cap();
+
+                            // pad zeros to the next io alignment requirement.
+                            // TODO: this is actually padding to next PAGE_SZ multiple, but only if the buffer capacity is larger than that.
+                            // We can't let the fact that we do direct IO, or the buffer capacity, dictate the on-disk format we write here.
+                            // Need to find a better API that allows writing the format we intend to.
+                            let count = len.next_multiple_of(PAGE_SZ).min(cap) - len;
+                            pad_amount = Some(count);
+                            buf.extend_with(0, count);
+
+                            Some(buf)
+                        })
                         .await
                         .map_err(|e| match e {
                             FlushTaskError::Cancelled => DownloadError::Cancelled,
                         })?;
-                    Ok(inner)
+
+                    let pad_amount = pad_amount.expect("shutdown always invokes the closure").into_u64();
+                    let set_len_arg = bytes_amount - pad_amount;
+                    destination_file
+                        .set_len(set_len_arg)
+                        .await
+                        .maybe_fatal_err("download_object set_len")
+                        .with_context(|| {
+                            format!("set len for file at {dst_path}: 0x{set_len_arg:x} = 0x{bytes_amount:x} - 0x{pad_amount:x}")
+                        })
+                        .map_err(DownloadError::Other)?;
+
+                    Ok((set_len_arg, destination_file))
                 }
                 .await?;
 

pageserver/src/tenant/remote_timeline_client/manifest.rs

@@ -1,21 +1,33 @@
 use chrono::NaiveDateTime;
+use pageserver_api::shard::ShardStripeSize;
 use serde::{Deserialize, Serialize};
 use utils::id::TimelineId;
 use utils::lsn::Lsn;
 
-/// Tenant-shard scoped manifest
-#[derive(Clone, Serialize, Deserialize, PartialEq, Eq)]
+/// Tenant shard manifest, stored in remote storage. Contains offloaded timelines and other tenant
+/// shard-wide information that must be persisted in remote storage.
+///
+/// The manifest is always updated on tenant attach, and as needed.
+#[derive(Clone, Debug, Serialize, Deserialize, PartialEq, Eq)]
 pub struct TenantManifest {
-    /// Debugging aid describing the version of this manifest.
-    /// Can also be used for distinguishing breaking changes later on.
+    /// The manifest version. Incremented on manifest format changes, even non-breaking ones.
+    /// Manifests must generally always be backwards and forwards compatible for one release, to
+    /// allow release rollbacks.
     pub version: usize,
 
+    /// This tenant's stripe size. This is only advisory, and used to recover tenant data from
+    /// remote storage. The autoritative source is the storage controller. If None, assume the
+    /// original default value of 32768 blocks (256 MB).
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub stripe_size: Option<ShardStripeSize>,
+
     /// The list of offloaded timelines together with enough information
     /// to not have to actually load them.
     ///
     /// Note: the timelines mentioned in this list might be deleted, i.e.
     /// we don't hold an invariant that the references aren't dangling.
     /// Existence of index-part.json is the actual indicator of timeline existence.
+    #[serde(default)]
     pub offloaded_timelines: Vec<OffloadedTimelineManifest>,
 }
 
@@ -24,7 +36,7 @@ pub struct TenantManifest {
 /// Very similar to [`pageserver_api::models::OffloadedTimelineInfo`],
 /// but the two datastructures serve different needs, this is for a persistent disk format
 /// that must be backwards compatible, while the other is only for informative purposes.
-#[derive(Clone, Serialize, Deserialize, Copy, PartialEq, Eq)]
+#[derive(Clone, Debug, Serialize, Deserialize, Copy, PartialEq, Eq)]
 pub struct OffloadedTimelineManifest {
     pub timeline_id: TimelineId,
     /// Whether the timeline has a parent it has been branched off from or not
@@ -35,20 +47,166 @@ pub struct OffloadedTimelineManifest {
     pub archived_at: NaiveDateTime,
 }
 
-pub const LATEST_TENANT_MANIFEST_VERSION: usize = 1;
+/// The newest manifest version. This should be incremented on changes, even non-breaking ones. We
+/// do not use deny_unknown_fields, so new fields are not breaking.
+///
+/// 1: initial version
+/// 2: +stripe_size
+///
+/// When adding new versions, also add a parse_vX test case below.
+pub const LATEST_TENANT_MANIFEST_VERSION: usize = 2;
 
 impl TenantManifest {
-    pub(crate) fn empty() -> Self {
-        Self {
-            version: LATEST_TENANT_MANIFEST_VERSION,
-            offloaded_timelines: vec![],
+    /// Returns true if the manifests are equal, ignoring the version number. This avoids
+    /// re-uploading all manifests just because the version number is bumped.
+    pub fn eq_ignoring_version(&self, other: &Self) -> bool {
+        // Fast path: if the version is equal, just compare directly.
+        if self.version == other.version {
+            return self == other;
         }
-    }
-    pub fn from_json_bytes(bytes: &[u8]) -> Result<Self, serde_json::Error> {
-        serde_json::from_slice::<Self>(bytes)
+
+        // We could alternatively just clone and modify the version here.
+        let Self {
+            version: _, // ignore version
+            stripe_size,
+            offloaded_timelines,
+        } = self;
+
+        stripe_size == &other.stripe_size && offloaded_timelines == &other.offloaded_timelines
     }
 
-    pub(crate) fn to_json_bytes(&self) -> serde_json::Result<Vec<u8>> {
+    /// Decodes a manifest from JSON.
+    pub fn from_json_bytes(bytes: &[u8]) -> Result<Self, serde_json::Error> {
+        serde_json::from_slice(bytes)
+    }
+
+    /// Encodes a manifest as JSON.
+    pub fn to_json_bytes(&self) -> serde_json::Result<Vec<u8>> {
         serde_json::to_vec(self)
     }
 }
+
+#[cfg(test)]
+mod tests {
+    use std::str::FromStr;
+
+    use utils::id::TimelineId;
+
+    use super::*;
+
+    /// Empty manifests should be parsed. Version is required.
+    #[test]
+    fn parse_empty() -> anyhow::Result<()> {
+        let json = r#"{
+             "version": 0
+         }"#;
+        let expected = TenantManifest {
+            version: 0,
+            stripe_size: None,
+            offloaded_timelines: Vec::new(),
+        };
+        assert_eq!(expected, TenantManifest::from_json_bytes(json.as_bytes())?);
+        Ok(())
+    }
+
+    /// Unknown fields should be ignored, for forwards compatibility.
+    #[test]
+    fn parse_unknown_fields() -> anyhow::Result<()> {
+        let json = r#"{
+             "version": 1,
+             "foo": "bar"
+         }"#;
+        let expected = TenantManifest {
+            version: 1,
+            stripe_size: None,
+            offloaded_timelines: Vec::new(),
+        };
+        assert_eq!(expected, TenantManifest::from_json_bytes(json.as_bytes())?);
+        Ok(())
+    }
+
+    /// v1 manifests should be parsed, for backwards compatibility.
+    #[test]
+    fn parse_v1() -> anyhow::Result<()> {
+        let json = r#"{
+             "version": 1,
+             "offloaded_timelines": [
+                 {
+                     "timeline_id": "5c4df612fd159e63c1b7853fe94d97da",
+                     "archived_at": "2025-03-07T11:07:11.373105434"
+                 },
+                 {
+                     "timeline_id": "f3def5823ad7080d2ea538d8e12163fa",
+                     "ancestor_timeline_id": "5c4df612fd159e63c1b7853fe94d97da",
+                     "ancestor_retain_lsn": "0/1F79038",
+                     "archived_at": "2025-03-05T11:10:22.257901390"
+                 }
+             ]
+         }"#;
+        let expected = TenantManifest {
+            version: 1,
+            stripe_size: None,
+            offloaded_timelines: vec![
+                OffloadedTimelineManifest {
+                    timeline_id: TimelineId::from_str("5c4df612fd159e63c1b7853fe94d97da")?,
+                    ancestor_timeline_id: None,
+                    ancestor_retain_lsn: None,
+                    archived_at: NaiveDateTime::from_str("2025-03-07T11:07:11.373105434")?,
+                },
+                OffloadedTimelineManifest {
+                    timeline_id: TimelineId::from_str("f3def5823ad7080d2ea538d8e12163fa")?,
+                    ancestor_timeline_id: Some(TimelineId::from_str(
+                        "5c4df612fd159e63c1b7853fe94d97da",
+                    )?),
+                    ancestor_retain_lsn: Some(Lsn::from_str("0/1F79038")?),
+                    archived_at: NaiveDateTime::from_str("2025-03-05T11:10:22.257901390")?,
+                },
+            ],
+        };
+        assert_eq!(expected, TenantManifest::from_json_bytes(json.as_bytes())?);
+        Ok(())
+    }
+
+    /// v2 manifests should be parsed, for backwards compatibility.
+    #[test]
+    fn parse_v2() -> anyhow::Result<()> {
+        let json = r#"{
+             "version": 2,
+             "stripe_size": 32768,
+             "offloaded_timelines": [
+                 {
+                     "timeline_id": "5c4df612fd159e63c1b7853fe94d97da",
+                     "archived_at": "2025-03-07T11:07:11.373105434"
+                 },
+                 {
+                     "timeline_id": "f3def5823ad7080d2ea538d8e12163fa",
+                     "ancestor_timeline_id": "5c4df612fd159e63c1b7853fe94d97da",
+                     "ancestor_retain_lsn": "0/1F79038",
+                     "archived_at": "2025-03-05T11:10:22.257901390"
+                 }
+             ]
+         }"#;
+        let expected = TenantManifest {
+            version: 2,
+            stripe_size: Some(ShardStripeSize(32768)),
+            offloaded_timelines: vec![
+                OffloadedTimelineManifest {
+                    timeline_id: TimelineId::from_str("5c4df612fd159e63c1b7853fe94d97da")?,
+                    ancestor_timeline_id: None,
+                    ancestor_retain_lsn: None,
+                    archived_at: NaiveDateTime::from_str("2025-03-07T11:07:11.373105434")?,
+                },
+                OffloadedTimelineManifest {
+                    timeline_id: TimelineId::from_str("f3def5823ad7080d2ea538d8e12163fa")?,
+                    ancestor_timeline_id: Some(TimelineId::from_str(
+                        "5c4df612fd159e63c1b7853fe94d97da",
+                    )?),
+                    ancestor_retain_lsn: Some(Lsn::from_str("0/1F79038")?),
+                    archived_at: NaiveDateTime::from_str("2025-03-05T11:10:22.257901390")?,
+                },
+            ],
+        };
+        assert_eq!(expected, TenantManifest::from_json_bytes(json.as_bytes())?);
+        Ok(())
+    }
+}

pageserver/src/tenant/remote_timeline_client/upload.rs

@@ -61,6 +61,7 @@ pub(crate) async fn upload_index_part(
         .await
         .with_context(|| format!("upload index part for '{tenant_shard_id} / {timeline_id}'"))
 }
+
 /// Serializes and uploads the given tenant manifest data to the remote storage.
 pub(crate) async fn upload_tenant_manifest(
     storage: &GenericRemoteStorage,
@@ -76,16 +77,14 @@ pub(crate) async fn upload_tenant_manifest(
     });
     pausable_failpoint!("before-upload-manifest-pausable");
 
-    let serialized = tenant_manifest.to_json_bytes()?;
-    let serialized = Bytes::from(serialized);
-
-    let tenant_manifest_site = serialized.len();
-
+    let serialized = Bytes::from(tenant_manifest.to_json_bytes()?);
+    let tenant_manifest_size = serialized.len();
     let remote_path = remote_tenant_manifest_path(tenant_shard_id, generation);
+
     storage
         .upload_storage_object(
             futures::stream::once(futures::future::ready(Ok(serialized))),
-            tenant_manifest_site,
+            tenant_manifest_size,
             &remote_path,
             cancel,
         )

pageserver/src/tenant/secondary/downloader.rs

@@ -1521,12 +1521,11 @@ async fn load_heatmap(
     path: &Utf8PathBuf,
     ctx: &RequestContext,
 ) -> Result<Option<HeatMapTenant>, anyhow::Error> {
-    let mut file = match VirtualFile::open(path, ctx).await {
+    let st = match VirtualFile::read_to_string(path, ctx).await {
         Ok(file) => file,
         Err(e) if e.kind() == std::io::ErrorKind::NotFound => return Ok(None),
         Err(e) => Err(e)?,
     };
-    let st = file.read_to_string(ctx).await?;
     let htm = serde_json::from_str(&st)?;
     Ok(Some(htm))
 }

pageserver/src/tenant/storage_layer/batch_split_writer.rs

@@ -5,6 +5,7 @@ use std::sync::Arc;
 use bytes::Bytes;
 use pageserver_api::key::{KEY_SIZE, Key};
 use pageserver_api::value::Value;
+use tokio_util::sync::CancellationToken;
 use utils::id::TimelineId;
 use utils::lsn::Lsn;
 use utils::shard::TenantShardId;
@@ -179,7 +180,7 @@ impl BatchLayerWriter {
 
 /// An image writer that takes images and produces multiple image layers.
 #[must_use]
-pub struct SplitImageLayerWriter {
+pub struct SplitImageLayerWriter<'a> {
     inner: ImageLayerWriter,
     target_layer_size: u64,
     lsn: Lsn,
@@ -188,9 +189,12 @@ pub struct SplitImageLayerWriter {
     tenant_shard_id: TenantShardId,
     batches: BatchLayerWriter,
     start_key: Key,
+    gate: &'a utils::sync::gate::Gate,
+    cancel: CancellationToken,
 }
 
-impl SplitImageLayerWriter {
+impl<'a> SplitImageLayerWriter<'a> {
+    #[allow(clippy::too_many_arguments)]
     pub async fn new(
         conf: &'static PageServerConf,
         timeline_id: TimelineId,
@@ -198,6 +202,8 @@ impl SplitImageLayerWriter {
         start_key: Key,
         lsn: Lsn,
         target_layer_size: u64,
+        gate: &'a utils::sync::gate::Gate,
+        cancel: CancellationToken,
         ctx: &RequestContext,
     ) -> anyhow::Result<Self> {
         Ok(Self {
@@ -208,6 +214,8 @@ impl SplitImageLayerWriter {
                 tenant_shard_id,
                 &(start_key..Key::MAX),
                 lsn,
+                gate,
+                cancel.clone(),
                 ctx,
             )
             .await?,
@@ -217,6 +225,8 @@ impl SplitImageLayerWriter {
             batches: BatchLayerWriter::new(conf).await?,
             lsn,
             start_key,
+            gate,
+            cancel,
         })
     }
 
@@ -239,6 +249,8 @@ impl SplitImageLayerWriter {
                 self.tenant_shard_id,
                 &(key..Key::MAX),
                 self.lsn,
+                self.gate,
+                self.cancel.clone(),
                 ctx,
             )
             .await?;
@@ -291,7 +303,7 @@ impl SplitImageLayerWriter {
 /// into a single file. This behavior might change in the future. For reference, the legacy compaction algorithm
 /// will split them into multiple files based on size.
 #[must_use]
-pub struct SplitDeltaLayerWriter {
+pub struct SplitDeltaLayerWriter<'a> {
     inner: Option<(Key, DeltaLayerWriter)>,
     target_layer_size: u64,
     conf: &'static PageServerConf,
@@ -300,15 +312,19 @@ pub struct SplitDeltaLayerWriter {
     lsn_range: Range<Lsn>,
     last_key_written: Key,
     batches: BatchLayerWriter,
+    gate: &'a utils::sync::gate::Gate,
+    cancel: CancellationToken,
 }
 
-impl SplitDeltaLayerWriter {
+impl<'a> SplitDeltaLayerWriter<'a> {
     pub async fn new(
         conf: &'static PageServerConf,
         timeline_id: TimelineId,
         tenant_shard_id: TenantShardId,
         lsn_range: Range<Lsn>,
         target_layer_size: u64,
+        gate: &'a utils::sync::gate::Gate,
+        cancel: CancellationToken,
     ) -> anyhow::Result<Self> {
         Ok(Self {
             target_layer_size,
@@ -319,6 +335,8 @@ impl SplitDeltaLayerWriter {
             lsn_range,
             last_key_written: Key::MIN,
             batches: BatchLayerWriter::new(conf).await?,
+            gate,
+            cancel,
         })
     }
 
@@ -344,6 +362,8 @@ impl SplitDeltaLayerWriter {
                     self.tenant_shard_id,
                     key,
                     self.lsn_range.clone(),
+                    self.gate,
+                    self.cancel.clone(),
                     ctx,
                 )
                 .await?,
@@ -362,11 +382,13 @@ impl SplitDeltaLayerWriter {
                     self.tenant_shard_id,
                     key,
                     self.lsn_range.clone(),
+                    self.gate,
+                    self.cancel.clone(),
                     ctx,
                 )
                 .await?;
                 let (start_key, prev_delta_writer) =
-                    std::mem::replace(&mut self.inner, Some((key, next_delta_writer))).unwrap();
+                    self.inner.replace((key, next_delta_writer)).unwrap();
                 self.batches.add_unfinished_delta_writer(
                     prev_delta_writer,
                     start_key..key,
@@ -469,6 +491,8 @@ mod tests {
             get_key(0),
             Lsn(0x18),
             4 * 1024 * 1024,
+            &tline.gate,
+            tline.cancel.clone(),
             &ctx,
         )
         .await
@@ -480,6 +504,8 @@ mod tests {
             tenant.tenant_shard_id,
             Lsn(0x18)..Lsn(0x20),
             4 * 1024 * 1024,
+            &tline.gate,
+            tline.cancel.clone(),
         )
         .await
         .unwrap();
@@ -546,6 +572,8 @@ mod tests {
             get_key(0),
             Lsn(0x18),
             4 * 1024 * 1024,
+            &tline.gate,
+            tline.cancel.clone(),
             &ctx,
         )
         .await
@@ -556,6 +584,8 @@ mod tests {
             tenant.tenant_shard_id,
             Lsn(0x18)..Lsn(0x20),
             4 * 1024 * 1024,
+            &tline.gate,
+            tline.cancel.clone(),
         )
         .await
         .unwrap();
@@ -643,6 +673,8 @@ mod tests {
             get_key(0),
             Lsn(0x18),
             4 * 1024,
+            &tline.gate,
+            tline.cancel.clone(),
             &ctx,
         )
         .await
@@ -654,6 +686,8 @@ mod tests {
             tenant.tenant_shard_id,
             Lsn(0x18)..Lsn(0x20),
             4 * 1024,
+            &tline.gate,
+            tline.cancel.clone(),
         )
         .await
         .unwrap();
@@ -730,6 +764,8 @@ mod tests {
             tenant.tenant_shard_id,
             Lsn(0x10)..Lsn(N as u64 * 16 + 0x10),
             4 * 1024 * 1024,
+            &tline.gate,
+            tline.cancel.clone(),
         )
         .await
         .unwrap();

pageserver/src/tenant/storage_layer/delta_layer.rs

@@ -29,11 +29,11 @@
 //!
 use std::collections::{HashMap, VecDeque};
 use std::fs::File;
-use std::io::SeekFrom;
 use std::ops::Range;
 use std::os::unix::fs::FileExt;
 use std::str::FromStr;
 use std::sync::Arc;
+use std::sync::atomic::AtomicU64;
 
 use anyhow::{Context, Result, bail, ensure};
 use camino::{Utf8Path, Utf8PathBuf};
@@ -45,13 +45,13 @@ use pageserver_api::keyspace::KeySpace;
 use pageserver_api::models::ImageCompressionAlgorithm;
 use pageserver_api::shard::TenantShardId;
 use pageserver_api::value::Value;
-use rand::Rng;
-use rand::distributions::Alphanumeric;
 use serde::{Deserialize, Serialize};
 use tokio::sync::OnceCell;
 use tokio_epoll_uring::IoBuf;
+use tokio_util::sync::CancellationToken;
 use tracing::*;
 use utils::bin_ser::BeSer;
+use utils::bin_ser::SerializeError;
 use utils::id::{TenantId, TimelineId};
 use utils::lsn::Lsn;
 
@@ -74,7 +74,8 @@ use crate::tenant::vectored_blob_io::{
     VectoredReadPlanner,
 };
 use crate::virtual_file::owned_buffers_io::io_buf_ext::{FullSlice, IoBufExt};
-use crate::virtual_file::{self, IoBufferMut, MaybeFatalIo, VirtualFile};
+use crate::virtual_file::owned_buffers_io::write::Buffer;
+use crate::virtual_file::{self, IoBuffer, IoBufferMut, MaybeFatalIo, VirtualFile};
 use crate::{DELTA_FILE_MAGIC, STORAGE_FORMAT_VERSION, TEMP_FILE_SUFFIX};
 
 ///
@@ -112,6 +113,15 @@ impl From<&DeltaLayer> for Summary {
 }
 
 impl Summary {
+    /// Serializes the summary header into an aligned buffer of lenth `PAGE_SZ`.
+    pub fn ser_into_page(&self) -> Result<IoBuffer, SerializeError> {
+        let mut buf = IoBufferMut::with_capacity(PAGE_SZ);
+        Self::ser_into(self, &mut buf)?;
+        // Pad zeroes to the buffer so the length is a multiple of the alignment.
+        buf.extend_with(0, buf.capacity() - buf.len());
+        Ok(buf.freeze())
+    }
+
     pub(super) fn expected(
         tenant_id: TenantId,
         timeline_id: TimelineId,
@@ -287,19 +297,19 @@ impl DeltaLayer {
         key_start: Key,
         lsn_range: &Range<Lsn>,
     ) -> Utf8PathBuf {
-        let rand_string: String = rand::thread_rng()
-            .sample_iter(&Alphanumeric)
-            .take(8)
-            .map(char::from)
-            .collect();
+        // Never reuse a filename in the lifetime of a pageserver process so that we need
+        // not worry about laggard Drop impl's async unlink hitting an already reused filename.
+        static NEXT_TEMP_DISAMBIGUATOR: AtomicU64 = AtomicU64::new(1);
+        let filename_disambiguator =
+            NEXT_TEMP_DISAMBIGUATOR.fetch_add(1, std::sync::atomic::Ordering::Relaxed);
 
         conf.timeline_path(tenant_shard_id, timeline_id)
             .join(format!(
-                "{}-XXX__{:016X}-{:016X}.{}.{}",
+                "{}-XXX__{:016X}-{:016X}.{:x}.{}",
                 key_start,
                 u64::from(lsn_range.start),
                 u64::from(lsn_range.end),
-                rand_string,
+                filename_disambiguator,
                 TEMP_FILE_SUFFIX,
             ))
     }
@@ -390,22 +400,27 @@ struct DeltaLayerWriterInner {
 
     tree: DiskBtreeBuilder<BlockBuf, DELTA_KEY_SIZE>,
 
-    blob_writer: BlobWriter<true>,
+    blob_writer: BlobWriter,
 
     // Number of key-lsns in the layer.
     num_keys: usize,
+
+    _gate_guard: utils::sync::gate::GateGuard,
 }
 
 impl DeltaLayerWriterInner {
     ///
     /// Start building a new delta layer.
     ///
+    #[allow(clippy::too_many_arguments)]
     async fn new(
         conf: &'static PageServerConf,
         timeline_id: TimelineId,
         tenant_shard_id: TenantShardId,
         key_start: Key,
         lsn_range: Range<Lsn>,
+        gate: &utils::sync::gate::Gate,
+        cancel: CancellationToken,
         ctx: &RequestContext,
     ) -> anyhow::Result<Self> {
         // Create the file initially with a temporary filename. We don't know
@@ -417,10 +432,17 @@ impl DeltaLayerWriterInner {
         let path =
             DeltaLayer::temp_path_for(conf, &tenant_shard_id, &timeline_id, key_start, &lsn_range);
 
-        let mut file = VirtualFile::create(&path, ctx).await?;
-        // make room for the header block
-        file.seek(SeekFrom::Start(PAGE_SZ as u64)).await?;
-        let blob_writer = BlobWriter::new(file, PAGE_SZ as u64);
+        let file = Arc::new(VirtualFile::create_v2(&path, ctx).await?);
+
+        // Start at PAGE_SZ, make room for the header block
+        let blob_writer = BlobWriter::new(
+            file,
+            PAGE_SZ as u64,
+            gate,
+            cancel,
+            ctx,
+            info_span!(parent: None, "delta_layer_writer_flush_task", tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), timeline_id=%timeline_id, path = %path),
+        )?;
 
         // Initialize the b-tree index builder
         let block_buf = BlockBuf::new();
@@ -435,6 +457,7 @@ impl DeltaLayerWriterInner {
             tree: tree_builder,
             blob_writer,
             num_keys: 0,
+            _gate_guard: gate.enter()?,
         })
     }
 
@@ -530,15 +553,33 @@ impl DeltaLayerWriterInner {
     ) -> anyhow::Result<(PersistentLayerDesc, Utf8PathBuf)> {
         let index_start_blk = self.blob_writer.size().div_ceil(PAGE_SZ as u64) as u32;
 
-        let mut file = self.blob_writer.into_inner(ctx).await?;
+        let file = self
+            .blob_writer
+            .into_inner(|mut buf| {
+                let len = buf.pending();
+                let cap = buf.cap();
+
+                // pad zeros to the next io alignment requirement.
+                // TODO: this is actually padding to next PAGE_SZ multiple, but only if the buffer capacity is larger than that.
+                // We can't let the fact that we do direct IO, or the buffer capacity, dictate the on-disk format we write here.
+                // Need to find a better API that allows writing the format we intend to.
+                let count = len.next_multiple_of(PAGE_SZ).min(cap) - len;
+                buf.extend_with(0, count);
+
+                Some(buf)
+            })
+            .await?;
 
         // Write out the index
         let (index_root_blk, block_buf) = self.tree.finish()?;
-        file.seek(SeekFrom::Start(index_start_blk as u64 * PAGE_SZ as u64))
-            .await?;
+        let mut offset = index_start_blk as u64 * PAGE_SZ as u64;
+
+        // TODO(yuchen): https://github.com/neondatabase/neon/issues/10092
+        // Should we just replace BlockBuf::blocks with one big buffer
         for buf in block_buf.blocks {
-            let (_buf, res) = file.write_all(buf.slice_len(), ctx).await;
+            let (_buf, res) = file.write_all_at(buf.slice_len(), offset, ctx).await;
             res?;
+            offset += PAGE_SZ as u64;
         }
         assert!(self.lsn_range.start < self.lsn_range.end);
         // Fill in the summary on blk 0
@@ -553,11 +594,9 @@ impl DeltaLayerWriterInner {
             index_root_blk,
         };
 
-        let mut buf = Vec::with_capacity(PAGE_SZ);
-        // TODO: could use smallvec here but it's a pain with Slice<T>
-        Summary::ser_into(&summary, &mut buf)?;
-        file.seek(SeekFrom::Start(0)).await?;
-        let (_buf, res) = file.write_all(buf.slice_len(), ctx).await;
+        // Writes summary at the first block (offset 0).
+        let buf = summary.ser_into_page()?;
+        let (_buf, res) = file.write_all_at(buf.slice_len(), 0, ctx).await;
         res?;
 
         let metadata = file
@@ -628,12 +667,15 @@ impl DeltaLayerWriter {
     ///
     /// Start building a new delta layer.
     ///
+    #[allow(clippy::too_many_arguments)]
     pub async fn new(
         conf: &'static PageServerConf,
         timeline_id: TimelineId,
         tenant_shard_id: TenantShardId,
         key_start: Key,
         lsn_range: Range<Lsn>,
+        gate: &utils::sync::gate::Gate,
+        cancel: CancellationToken,
         ctx: &RequestContext,
     ) -> anyhow::Result<Self> {
         Ok(Self {
@@ -644,6 +686,8 @@ impl DeltaLayerWriter {
                     tenant_shard_id,
                     key_start,
                     lsn_range,
+                    gate,
+                    cancel,
                     ctx,
                 )
                 .await?,
@@ -719,12 +763,33 @@ impl DeltaLayerWriter {
 
 impl Drop for DeltaLayerWriter {
     fn drop(&mut self) {
-        if let Some(inner) = self.inner.take() {
-            // We want to remove the virtual file here, so it's fine to not
-            // having completely flushed unwritten data.
-            let vfile = inner.blob_writer.into_inner_no_flush();
-            vfile.remove();
-        }
+        let Some(inner) = self.inner.take() else {
+            return;
+        };
+
+        tokio::spawn(async move {
+            let DeltaLayerWriterInner {
+                blob_writer,
+                _gate_guard,
+                ..
+            } = inner;
+
+            let vfile = match blob_writer.into_inner(|_| None).await {
+                Ok(vfile) => vfile,
+                Err(e) => {
+                    error!(err=%e, "failed to remove delta layer writer file");
+                    drop(_gate_guard);
+                    return;
+                }
+            };
+
+            if let Err(e) = std::fs::remove_file(vfile.path())
+                .maybe_fatal_err("failed to remove the virtual file")
+            {
+                error!(err=%e, path=%vfile.path(), "failed to remove delta layer writer file");
+            }
+            drop(_gate_guard);
+        });
     }
 }
 
@@ -751,7 +816,7 @@ impl DeltaLayer {
     where
         F: Fn(Summary) -> Summary,
     {
-        let mut file = VirtualFile::open_with_options(
+        let file = VirtualFile::open_with_options_v2(
             path,
             virtual_file::OpenOptions::new().read(true).write(true),
             ctx,
@@ -768,11 +833,8 @@ impl DeltaLayer {
 
         let new_summary = rewrite(actual_summary);
 
-        let mut buf = Vec::with_capacity(PAGE_SZ);
-        // TODO: could use smallvec here, but it's a pain with Slice<T>
-        Summary::ser_into(&new_summary, &mut buf).context("serialize")?;
-        file.seek(SeekFrom::Start(0)).await?;
-        let (_buf, res) = file.write_all(buf.slice_len(), ctx).await;
+        let buf = new_summary.ser_into_page().context("serialize")?;
+        let (_buf, res) = file.write_all_at(buf.slice_len(), 0, ctx).await;
         res?;
         Ok(())
     }
@@ -1600,8 +1662,8 @@ pub(crate) mod test {
     use bytes::Bytes;
     use itertools::MinMaxResult;
     use pageserver_api::value::Value;
-    use rand::RngCore;
     use rand::prelude::{SeedableRng, SliceRandom, StdRng};
+    use rand::{Rng, RngCore};
 
     use super::*;
     use crate::DEFAULT_PG_VERSION;
@@ -1885,6 +1947,8 @@ pub(crate) mod test {
             harness.tenant_shard_id,
             entries_meta.key_range.start,
             entries_meta.lsn_range.clone(),
+            &timeline.gate,
+            timeline.cancel.clone(),
             &ctx,
         )
         .await?;
@@ -2079,6 +2143,8 @@ pub(crate) mod test {
                 tenant.tenant_shard_id,
                 Key::MIN,
                 Lsn(0x11)..truncate_at,
+                &branch.gate,
+                branch.cancel.clone(),
                 ctx,
             )
             .await
@@ -2213,6 +2279,8 @@ pub(crate) mod test {
             tenant.tenant_shard_id,
             *key_start,
             (*lsn_min)..lsn_end,
+            &tline.gate,
+            tline.cancel.clone(),
             ctx,
         )
         .await?;

pageserver/src/tenant/storage_layer/image_layer.rs

@@ -27,11 +27,11 @@
 //! actual page images are stored in the "values" part.
 use std::collections::{HashMap, VecDeque};
 use std::fs::File;
-use std::io::SeekFrom;
 use std::ops::Range;
 use std::os::unix::prelude::FileExt;
 use std::str::FromStr;
 use std::sync::Arc;
+use std::sync::atomic::AtomicU64;
 
 use anyhow::{Context, Result, bail, ensure};
 use bytes::Bytes;
@@ -43,13 +43,13 @@ use pageserver_api::key::{DBDIR_KEY, KEY_SIZE, Key};
 use pageserver_api::keyspace::KeySpace;
 use pageserver_api::shard::{ShardIdentity, TenantShardId};
 use pageserver_api::value::Value;
-use rand::Rng;
-use rand::distributions::Alphanumeric;
 use serde::{Deserialize, Serialize};
 use tokio::sync::OnceCell;
 use tokio_stream::StreamExt;
+use tokio_util::sync::CancellationToken;
 use tracing::*;
 use utils::bin_ser::BeSer;
+use utils::bin_ser::SerializeError;
 use utils::id::{TenantId, TimelineId};
 use utils::lsn::Lsn;
 
@@ -72,7 +72,8 @@ use crate::tenant::vectored_blob_io::{
     VectoredReadPlanner,
 };
 use crate::virtual_file::owned_buffers_io::io_buf_ext::IoBufExt;
-use crate::virtual_file::{self, IoBufferMut, MaybeFatalIo, VirtualFile};
+use crate::virtual_file::owned_buffers_io::write::Buffer;
+use crate::virtual_file::{self, IoBuffer, IoBufferMut, MaybeFatalIo, VirtualFile};
 use crate::{IMAGE_FILE_MAGIC, STORAGE_FORMAT_VERSION, TEMP_FILE_SUFFIX};
 
 ///
@@ -111,6 +112,15 @@ impl From<&ImageLayer> for Summary {
 }
 
 impl Summary {
+    /// Serializes the summary header into an aligned buffer of lenth `PAGE_SZ`.
+    pub fn ser_into_page(&self) -> Result<IoBuffer, SerializeError> {
+        let mut buf = IoBufferMut::with_capacity(PAGE_SZ);
+        Self::ser_into(self, &mut buf)?;
+        // Pad zeroes to the buffer so the length is a multiple of the alignment.
+        buf.extend_with(0, buf.capacity() - buf.len());
+        Ok(buf.freeze())
+    }
+
     pub(super) fn expected(
         tenant_id: TenantId,
         timeline_id: TimelineId,
@@ -251,14 +261,17 @@ impl ImageLayer {
         tenant_shard_id: TenantShardId,
         fname: &ImageLayerName,
     ) -> Utf8PathBuf {
-        let rand_string: String = rand::thread_rng()
-            .sample_iter(&Alphanumeric)
-            .take(8)
-            .map(char::from)
-            .collect();
+        // Never reuse a filename in the lifetime of a pageserver process so that we need
+        // not worry about laggard Drop impl's async unlink hitting an already reused filename.
+        static NEXT_TEMP_DISAMBIGUATOR: AtomicU64 = AtomicU64::new(1);
+        let filename_disambiguator =
+            NEXT_TEMP_DISAMBIGUATOR.fetch_add(1, std::sync::atomic::Ordering::Relaxed);
 
         conf.timeline_path(&tenant_shard_id, &timeline_id)
-            .join(format!("{fname}.{rand_string}.{TEMP_FILE_SUFFIX}"))
+            .join(format!(
+                "{fname}.{:x}.{TEMP_FILE_SUFFIX}",
+                filename_disambiguator
+            ))
     }
 
     ///
@@ -348,7 +361,7 @@ impl ImageLayer {
     where
         F: Fn(Summary) -> Summary,
     {
-        let mut file = VirtualFile::open_with_options(
+        let file = VirtualFile::open_with_options_v2(
             path,
             virtual_file::OpenOptions::new().read(true).write(true),
             ctx,
@@ -365,11 +378,8 @@ impl ImageLayer {
 
         let new_summary = rewrite(actual_summary);
 
-        let mut buf = Vec::with_capacity(PAGE_SZ);
-        // TODO: could use smallvec here but it's a pain with Slice<T>
-        Summary::ser_into(&new_summary, &mut buf).context("serialize")?;
-        file.seek(SeekFrom::Start(0)).await?;
-        let (_buf, res) = file.write_all(buf.slice_len(), ctx).await;
+        let buf = new_summary.ser_into_page().context("serialize")?;
+        let (_buf, res) = file.write_all_at(buf.slice_len(), 0, ctx).await;
         res?;
         Ok(())
     }
@@ -737,23 +747,28 @@ struct ImageLayerWriterInner {
     // Number of keys in the layer.
     num_keys: usize,
 
-    blob_writer: BlobWriter<false>,
+    blob_writer: BlobWriter,
     tree: DiskBtreeBuilder<BlockBuf, KEY_SIZE>,
 
     #[cfg(feature = "testing")]
     last_written_key: Key,
+
+    _gate_guard: utils::sync::gate::GateGuard,
 }
 
 impl ImageLayerWriterInner {
     ///
     /// Start building a new image layer.
     ///
+    #[allow(clippy::too_many_arguments)]
     async fn new(
         conf: &'static PageServerConf,
         timeline_id: TimelineId,
         tenant_shard_id: TenantShardId,
         key_range: &Range<Key>,
         lsn: Lsn,
+        gate: &utils::sync::gate::Gate,
+        cancel: CancellationToken,
         ctx: &RequestContext,
     ) -> anyhow::Result<Self> {
         // Create the file initially with a temporary filename.
@@ -768,19 +783,28 @@ impl ImageLayerWriterInner {
             },
         );
         trace!("creating image layer {}", path);
-        let mut file = {
-            VirtualFile::open_with_options(
-                &path,
-                virtual_file::OpenOptions::new()
-                    .write(true)
-                    .create_new(true),
-                ctx,
+        let file = {
+            Arc::new(
+                VirtualFile::open_with_options_v2(
+                    &path,
+                    virtual_file::OpenOptions::new()
+                        .write(true)
+                        .create_new(true),
+                    ctx,
+                )
+                .await?,
             )
-            .await?
         };
-        // make room for the header block
-        file.seek(SeekFrom::Start(PAGE_SZ as u64)).await?;
-        let blob_writer = BlobWriter::new(file, PAGE_SZ as u64);
+
+        // Start at `PAGE_SZ` to make room for the header block.
+        let blob_writer = BlobWriter::new(
+            file,
+            PAGE_SZ as u64,
+            gate,
+            cancel,
+            ctx,
+            info_span!(parent: None, "image_layer_writer_flush_task", tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), timeline_id=%timeline_id, path = %path),
+        )?;
 
         // Initialize the b-tree index builder
         let block_buf = BlockBuf::new();
@@ -801,6 +825,7 @@ impl ImageLayerWriterInner {
             num_keys: 0,
             #[cfg(feature = "testing")]
             last_written_key: Key::MIN,
+            _gate_guard: gate.enter()?,
         };
 
         Ok(writer)
@@ -886,15 +911,30 @@ impl ImageLayerWriterInner {
         crate::metrics::COMPRESSION_IMAGE_INPUT_BYTES_CHOSEN.inc_by(self.uncompressed_bytes_chosen);
         crate::metrics::COMPRESSION_IMAGE_OUTPUT_BYTES.inc_by(compressed_size);
 
-        let mut file = self.blob_writer.into_inner();
+        let file = self
+            .blob_writer
+            .into_inner(|mut buf| {
+                let len = buf.pending();
+                let cap = buf.cap();
+
+                // pad zeros to the next io alignment requirement.
+                let count = len.next_multiple_of(PAGE_SZ).min(cap) - len;
+                buf.extend_with(0, count);
+
+                Some(buf)
+            })
+            .await?;
 
         // Write out the index
-        file.seek(SeekFrom::Start(index_start_blk as u64 * PAGE_SZ as u64))
-            .await?;
+        let mut offset = index_start_blk as u64 * PAGE_SZ as u64;
         let (index_root_blk, block_buf) = self.tree.finish()?;
+
+        // TODO(yuchen): https://github.com/neondatabase/neon/issues/10092
+        // Should we just replace BlockBuf::blocks with one big buffer?
         for buf in block_buf.blocks {
-            let (_buf, res) = file.write_all(buf.slice_len(), ctx).await;
+            let (_buf, res) = file.write_all_at(buf.slice_len(), offset, ctx).await;
             res?;
+            offset += PAGE_SZ as u64;
         }
 
         let final_key_range = if let Some(end_key) = end_key {
@@ -915,11 +955,9 @@ impl ImageLayerWriterInner {
             index_root_blk,
         };
 
-        let mut buf = Vec::with_capacity(PAGE_SZ);
-        // TODO: could use smallvec here but it's a pain with Slice<T>
-        Summary::ser_into(&summary, &mut buf)?;
-        file.seek(SeekFrom::Start(0)).await?;
-        let (_buf, res) = file.write_all(buf.slice_len(), ctx).await;
+        // Writes summary at the first block (offset 0).
+        let buf = summary.ser_into_page()?;
+        let (_buf, res) = file.write_all_at(buf.slice_len(), 0, ctx).await;
         res?;
 
         let metadata = file
@@ -988,18 +1026,30 @@ impl ImageLayerWriter {
     ///
     /// Start building a new image layer.
     ///
+    #[allow(clippy::too_many_arguments)]
     pub async fn new(
         conf: &'static PageServerConf,
         timeline_id: TimelineId,
         tenant_shard_id: TenantShardId,
         key_range: &Range<Key>,
         lsn: Lsn,
+        gate: &utils::sync::gate::Gate,
+        cancel: CancellationToken,
         ctx: &RequestContext,
     ) -> anyhow::Result<ImageLayerWriter> {
         Ok(Self {
             inner: Some(
-                ImageLayerWriterInner::new(conf, timeline_id, tenant_shard_id, key_range, lsn, ctx)
-                    .await?,
+                ImageLayerWriterInner::new(
+                    conf,
+                    timeline_id,
+                    tenant_shard_id,
+                    key_range,
+                    lsn,
+                    gate,
+                    cancel,
+                    ctx,
+                )
+                .await?,
             ),
         })
     }
@@ -1050,9 +1100,33 @@ impl ImageLayerWriter {
 
 impl Drop for ImageLayerWriter {
     fn drop(&mut self) {
-        if let Some(inner) = self.inner.take() {
-            inner.blob_writer.into_inner().remove();
-        }
+        let Some(inner) = self.inner.take() else {
+            return;
+        };
+
+        tokio::spawn(async move {
+            let ImageLayerWriterInner {
+                blob_writer,
+                _gate_guard,
+                ..
+            } = inner;
+
+            let vfile = match blob_writer.into_inner(|_| None).await {
+                Ok(vfile) => vfile,
+                Err(e) => {
+                    error!(err=%e, "failed to remove image layer writer file");
+                    drop(_gate_guard);
+                    return;
+                }
+            };
+
+            if let Err(e) = std::fs::remove_file(vfile.path())
+                .maybe_fatal_err("failed to remove the virtual file")
+            {
+                error!(err=%e, path=%vfile.path(), "failed to remove image layer writer file");
+            }
+            drop(_gate_guard);
+        });
     }
 }
 
@@ -1192,7 +1266,7 @@ mod test {
 
         // This key range contains several 0x8000 page stripes, only one of which belongs to shard zero
         let input_start = Key::from_hex("000000067f00000001000000ae0000000000").unwrap();
-        let input_end = Key::from_hex("000000067f00000001000000ae0000020000").unwrap();
+        let input_end = Key::from_hex("000000067f00000001000000ae0000002000").unwrap();
         let range = input_start..input_end;
 
         // Build an image layer to filter
@@ -1203,6 +1277,8 @@ mod test {
                 harness.tenant_shard_id,
                 &range,
                 lsn,
+                &timeline.gate,
+                timeline.cancel.clone(),
                 &ctx,
             )
             .await
@@ -1235,7 +1311,7 @@ mod test {
             let shard_identity = ShardIdentity::new(
                 ShardNumber(shard_number),
                 shard_count,
-                ShardStripeSize(0x8000),
+                ShardStripeSize(0x800),
             )
             .unwrap();
             let harness = TenantHarness::create_custom(
@@ -1268,6 +1344,8 @@ mod test {
                 harness.tenant_shard_id,
                 &range,
                 lsn,
+                &timeline.gate,
+                timeline.cancel.clone(),
                 &ctx,
             )
             .await
@@ -1287,12 +1365,12 @@ mod test {
 
             // This exact size and those below will need updating as/when the layer encoding changes, but
             // should be deterministic for a given version of the format, as we used no randomness generating the input.
-            assert_eq!(original_size, 1597440);
+            assert_eq!(original_size, 122880);
 
             match shard_number {
                 0 => {
                     // We should have written out just one stripe for our shard identity
-                    assert_eq!(wrote_keys, 0x8000);
+                    assert_eq!(wrote_keys, 0x800);
                     let replacement = replacement.unwrap();
 
                     // We should have dropped some of the data
@@ -1300,7 +1378,7 @@ mod test {
                     assert!(replacement.metadata().file_size > 0);
 
                     // Assert that we dropped ~3/4 of the data.
-                    assert_eq!(replacement.metadata().file_size, 417792);
+                    assert_eq!(replacement.metadata().file_size, 49152);
                 }
                 1 => {
                     // Shard 1 has no keys in our input range
@@ -1309,19 +1387,19 @@ mod test {
                 }
                 2 => {
                     // Shard 2 has one stripes in the input range
-                    assert_eq!(wrote_keys, 0x8000);
+                    assert_eq!(wrote_keys, 0x800);
                     let replacement = replacement.unwrap();
                     assert!(replacement.metadata().file_size < original_size);
                     assert!(replacement.metadata().file_size > 0);
-                    assert_eq!(replacement.metadata().file_size, 417792);
+                    assert_eq!(replacement.metadata().file_size, 49152);
                 }
                 3 => {
                     // Shard 3 has two stripes in the input range
-                    assert_eq!(wrote_keys, 0x10000);
+                    assert_eq!(wrote_keys, 0x1000);
                     let replacement = replacement.unwrap();
                     assert!(replacement.metadata().file_size < original_size);
                     assert!(replacement.metadata().file_size > 0);
-                    assert_eq!(replacement.metadata().file_size, 811008);
+                    assert_eq!(replacement.metadata().file_size, 73728);
                 }
                 _ => unreachable!(),
             }
@@ -1346,6 +1424,8 @@ mod test {
             tenant.tenant_shard_id,
             &key_range,
             lsn,
+            &tline.gate,
+            tline.cancel.clone(),
             ctx,
         )
         .await?;

pageserver/src/tenant/storage_layer/inmemory_layer.rs

@@ -719,6 +719,8 @@ impl InMemoryLayer {
         ctx: &RequestContext,
         key_range: Option<Range<Key>>,
         l0_flush_global_state: &l0_flush::Inner,
+        gate: &utils::sync::gate::Gate,
+        cancel: CancellationToken,
     ) -> Result<Option<(PersistentLayerDesc, Utf8PathBuf)>> {
         // Grab the lock in read-mode. We hold it over the I/O, but because this
         // layer is not writeable anymore, no one should be trying to acquire the
@@ -759,6 +761,8 @@ impl InMemoryLayer {
             self.tenant_shard_id,
             Key::MIN,
             self.start_lsn..end_lsn,
+            gate,
+            cancel,
             ctx,
         )
         .await?;

pageserver/src/tenant/storage_layer/inmemory_layer/vectored_dio_read.rs

@@ -766,7 +766,7 @@ mod tests {
                     rand::Rng::fill(&mut rand::thread_rng(), &mut dst_slice[len..]); // to discover bugs
                     Ok((dst, len))
                 }
-                Err(e) => Err(std::io::Error::new(std::io::ErrorKind::Other, e)),
+                Err(e) => Err(std::io::Error::other(e)),
             }
         }
     }

pageserver/src/tenant/storage_layer/merge_iterator.rs

@@ -59,6 +59,7 @@ impl LayerIterRef<'_> {
 /// 1. Unified iterator for image and delta layers.
 /// 2. `Ord` for use in [`MergeIterator::heap`] (for the k-merge).
 /// 3. Lazy creation of the real delta/image iterator.
+#[allow(clippy::large_enum_variant, reason = "TODO")]
 pub(crate) enum IteratorWrapper<'a> {
     NotLoaded {
         ctx: &'a RequestContext,

pageserver/src/tenant/tasks.rs

@@ -268,7 +268,12 @@ async fn compaction_loop(tenant: Arc<Tenant>, cancel: CancellationToken) {
                 error_run += 1;
                 let backoff =
                     exponential_backoff_duration(error_run, BASE_BACKOFF_SECS, MAX_BACKOFF_SECS);
-                log_compaction_error(&err, Some((error_run, backoff)), cancel.is_cancelled());
+                log_compaction_error(
+                    &err,
+                    Some((error_run, backoff)),
+                    cancel.is_cancelled(),
+                    false,
+                );
                 continue;
             }
         }
@@ -285,6 +290,7 @@ pub(crate) fn log_compaction_error(
     err: &CompactionError,
     retry_info: Option<(u32, Duration)>,
     task_cancelled: bool,
+    degrade_to_warning: bool,
 ) {
     use CompactionError::*;
 
@@ -333,6 +339,7 @@ pub(crate) fn log_compaction_error(
         }
     } else {
         match level {
+            Level::ERROR if degrade_to_warning => warn!("Compaction failed and discarded: {err:#}"),
             Level::ERROR => error!("Compaction failed: {err:#}"),
             Level::INFO => info!("Compaction failed: {err:#}"),
             level => unimplemented!("unexpected level {level:?}"),

pageserver/src/tenant/timeline.rs

@@ -24,6 +24,7 @@ use std::sync::{Arc, Mutex, OnceLock, RwLock, Weak};
 use std::time::{Duration, Instant, SystemTime};
 
 use crate::PERF_TRACE_TARGET;
+use crate::walredo::RedoAttemptType;
 use anyhow::{Context, Result, anyhow, bail, ensure};
 use arc_swap::{ArcSwap, ArcSwapOption};
 use bytes::Bytes;
@@ -115,7 +116,7 @@ use crate::pgdatadir_mapping::{
 use crate::task_mgr::TaskKind;
 use crate::tenant::config::AttachmentMode;
 use crate::tenant::gc_result::GcResult;
-use crate::tenant::layer_map::{LayerMap, SearchResult};
+use crate::tenant::layer_map::LayerMap;
 use crate::tenant::metadata::TimelineMetadata;
 use crate::tenant::storage_layer::delta_layer::DeltaEntry;
 use crate::tenant::storage_layer::inmemory_layer::IndexEntry;
@@ -1039,6 +1040,7 @@ pub(crate) enum ShutdownMode {
     Hard,
 }
 
+#[allow(clippy::large_enum_variant, reason = "TODO")]
 enum ImageLayerCreationOutcome {
     /// We generated an image layer
     Generated {
@@ -1292,6 +1294,12 @@ impl Timeline {
         };
         reconstruct_state.read_path = read_path;
 
+        let redo_attempt_type = if ctx.task_kind() == TaskKind::Compaction {
+            RedoAttemptType::LegacyCompaction
+        } else {
+            RedoAttemptType::ReadPage
+        };
+
         let traversal_res: Result<(), _> = {
             let ctx = RequestContextBuilder::from(ctx)
                 .perf_span(|crnt_perf_span| {
@@ -1379,7 +1387,7 @@ impl Timeline {
 
                     let walredo_deltas = converted.num_deltas();
                     let walredo_res = walredo_self
-                        .reconstruct_value(key, lsn, converted)
+                        .reconstruct_value(key, lsn, converted, redo_attempt_type)
                         .maybe_perf_instrument(&ctx, |crnt_perf_span| {
                             info_span!(
                                 target: PERF_TRACE_TARGET,
@@ -1940,7 +1948,7 @@ impl Timeline {
             )
             .await;
         if let Err(err) = &res {
-            log_compaction_error(err, None, cancel.is_cancelled());
+            log_compaction_error(err, None, cancel.is_cancelled(), false);
         }
         res
     }
@@ -4104,12 +4112,6 @@ impl Timeline {
         cancel: &CancellationToken,
         ctx: &RequestContext,
     ) -> Result<TimelineVisitOutcome, GetVectoredError> {
-        let mut unmapped_keyspace = keyspace.clone();
-        let mut fringe = LayerFringe::new();
-
-        let mut completed_keyspace = KeySpace::default();
-        let mut image_covered_keyspace = KeySpaceRandomAccum::new();
-
         // Prevent GC from progressing while visiting the current timeline.
         // If we are GC-ing because a new image layer was added while traversing
         // the timeline, then it will remove layers that are required for fulfilling
@@ -4120,11 +4122,44 @@ impl Timeline {
         // See `compaction::compact_with_gc` for why we need this.
         let _guard = timeline.gc_compaction_layer_update_lock.read().await;
 
-        loop {
+        // Initialize the fringe
+        let mut fringe = {
+            let mut fringe = LayerFringe::new();
+
+            let guard = timeline.layers.read().await;
+            guard.update_search_fringe(&keyspace, cont_lsn, &mut fringe)?;
+
+            fringe
+        };
+
+        let mut completed_keyspace = KeySpace::default();
+        let mut image_covered_keyspace = KeySpaceRandomAccum::new();
+
+        while let Some((layer_to_read, keyspace_to_read, lsn_range)) = fringe.next_layer() {
             if cancel.is_cancelled() {
                 return Err(GetVectoredError::Cancelled);
             }
 
+            if let Some(ref mut read_path) = reconstruct_state.read_path {
+                read_path.record_layer_visit(&layer_to_read, &keyspace_to_read, &lsn_range);
+            }
+
+            // Visit the layer and plan IOs for it
+            let next_cont_lsn = lsn_range.start;
+            layer_to_read
+                .get_values_reconstruct_data(
+                    keyspace_to_read.clone(),
+                    lsn_range,
+                    reconstruct_state,
+                    ctx,
+                )
+                .await?;
+
+            let mut unmapped_keyspace = keyspace_to_read;
+            cont_lsn = next_cont_lsn;
+
+            reconstruct_state.on_layer_visited(&layer_to_read);
+
             let (keys_done_last_step, keys_with_image_coverage) =
                 reconstruct_state.consume_done_keys();
             unmapped_keyspace.remove_overlapping_with(&keys_done_last_step);
@@ -4135,31 +4170,15 @@ impl Timeline {
                 image_covered_keyspace.add_range(keys_with_image_coverage);
             }
 
+            // Query the layer map for the next layers to read.
+            //
             // Do not descent any further if the last layer we visited
             // completed all keys in the keyspace it inspected. This is not
             // required for correctness, but avoids visiting extra layers
             // which turns out to be a perf bottleneck in some cases.
             if !unmapped_keyspace.is_empty() {
                 let guard = timeline.layers.read().await;
-                let layers = guard.layer_map()?;
-
-                for range in unmapped_keyspace.ranges.iter() {
-                    let results = layers.range_search(range.clone(), cont_lsn);
-
-                    results
-                        .found
-                        .into_iter()
-                        .map(|(SearchResult { layer, lsn_floor }, keyspace_accum)| {
-                            (
-                                guard.upgrade(layer),
-                                keyspace_accum.to_keyspace(),
-                                lsn_floor..cont_lsn,
-                            )
-                        })
-                        .for_each(|(layer, keyspace, lsn_range)| {
-                            fringe.update(layer, keyspace, lsn_range)
-                        });
-                }
+                guard.update_search_fringe(&unmapped_keyspace, cont_lsn, &mut fringe)?;
 
                 // It's safe to drop the layer map lock after planning the next round of reads.
                 // The fringe keeps readable handles for the layers which are safe to read even
@@ -4173,28 +4192,6 @@ impl Timeline {
                 // at two different time points.
                 drop(guard);
             }
-
-            if let Some((layer_to_read, keyspace_to_read, lsn_range)) = fringe.next_layer() {
-                if let Some(ref mut read_path) = reconstruct_state.read_path {
-                    read_path.record_layer_visit(&layer_to_read, &keyspace_to_read, &lsn_range);
-                }
-                let next_cont_lsn = lsn_range.start;
-                layer_to_read
-                    .get_values_reconstruct_data(
-                        keyspace_to_read.clone(),
-                        lsn_range,
-                        reconstruct_state,
-                        ctx,
-                    )
-                    .await?;
-
-                unmapped_keyspace = keyspace_to_read;
-                cont_lsn = next_cont_lsn;
-
-                reconstruct_state.on_layer_visited(&layer_to_read);
-            } else {
-                break;
-            }
         }
 
         Ok(TimelineVisitOutcome {
@@ -4808,7 +4805,13 @@ impl Timeline {
         let ctx = ctx.attached_child();
         let work = async move {
             let Some((desc, path)) = frozen_layer
-                .write_to_disk(&ctx, key_range, self_clone.l0_flush_global_state.inner())
+                .write_to_disk(
+                    &ctx,
+                    key_range,
+                    self_clone.l0_flush_global_state.inner(),
+                    &self_clone.gate,
+                    self_clone.cancel.clone(),
+                )
                 .await?
             else {
                 return Ok(None);
@@ -5346,6 +5349,8 @@ impl Timeline {
                 self.tenant_shard_id,
                 &img_range,
                 lsn,
+                &self.gate,
+                self.cancel.clone(),
                 ctx,
             )
             .await?;
@@ -6357,10 +6362,17 @@ impl Timeline {
         key: Key,
         request_lsn: Lsn,
         mut data: ValueReconstructState,
+        redo_attempt_type: RedoAttemptType,
     ) -> Result<Bytes, PageReconstructError> {
         // Perform WAL redo if needed
         data.records.reverse();
 
+        let fire_critical_error = match redo_attempt_type {
+            RedoAttemptType::ReadPage => true,
+            RedoAttemptType::LegacyCompaction => true,
+            RedoAttemptType::GcCompaction => false,
+        };
+
         // If we have a page image, and no WAL, we're all set
         if data.records.is_empty() {
             if let Some((img_lsn, img)) = &data.img {
@@ -6407,13 +6419,22 @@ impl Timeline {
                     .as_ref()
                     .context("timeline has no walredo manager")
                     .map_err(PageReconstructError::WalRedo)?
-                    .request_redo(key, request_lsn, data.img, data.records, self.pg_version)
+                    .request_redo(
+                        key,
+                        request_lsn,
+                        data.img,
+                        data.records,
+                        self.pg_version,
+                        redo_attempt_type,
+                    )
                     .await;
                 let img = match res {
                     Ok(img) => img,
                     Err(walredo::Error::Cancelled) => return Err(PageReconstructError::Cancelled),
                     Err(walredo::Error::Other(err)) => {
-                        critical!("walredo failure during page reconstruction: {err:?}");
+                        if fire_critical_error {
+                            critical!("walredo failure during page reconstruction: {err:?}");
+                        }
                         return Err(PageReconstructError::WalRedo(
                             err.context("reconstruct a page image"),
                         ));
@@ -6694,6 +6715,8 @@ impl Timeline {
             self.tenant_shard_id,
             &(min_key..end_key),
             lsn,
+            &self.gate,
+            self.cancel.clone(),
             ctx,
         )
         .await?;
@@ -6755,6 +6778,8 @@ impl Timeline {
             self.tenant_shard_id,
             deltas.key_range.start,
             deltas.lsn_range,
+            &self.gate,
+            self.cancel.clone(),
             ctx,
         )
         .await?;

pageserver/src/tenant/timeline/compaction.rs

@@ -7,7 +7,7 @@
 use std::collections::{BinaryHeap, HashMap, HashSet, VecDeque};
 use std::ops::{Deref, Range};
 use std::sync::Arc;
-use std::time::Instant;
+use std::time::{Duration, Instant};
 
 use super::layer_manager::LayerManager;
 use super::{
@@ -16,6 +16,8 @@ use super::{
     Timeline,
 };
 
+use crate::tenant::timeline::DeltaEntry;
+use crate::walredo::RedoAttemptType;
 use anyhow::{Context, anyhow};
 use bytes::Bytes;
 use enumset::EnumSet;
@@ -315,6 +317,9 @@ impl GcCompactionQueue {
                     flags: {
                         let mut flags = EnumSet::new();
                         flags |= CompactFlags::EnhancedGcBottomMostCompaction;
+                        if timeline.get_compaction_l0_first() {
+                            flags |= CompactFlags::YieldForL0;
+                        }
                         flags
                     },
                     sub_compaction: true,
@@ -448,7 +453,7 @@ impl GcCompactionQueue {
     ) -> Result<CompactionOutcome, CompactionError> {
         let res = self.iteration_inner(cancel, ctx, gc_block, timeline).await;
         if let Err(err) = &res {
-            log_compaction_error(err, None, cancel.is_cancelled());
+            log_compaction_error(err, None, cancel.is_cancelled(), true);
         }
         match res {
             Ok(res) => Ok(res),
@@ -742,8 +747,8 @@ impl KeyHistoryRetention {
     async fn pipe_to(
         self,
         key: Key,
-        delta_writer: &mut SplitDeltaLayerWriter,
-        mut image_writer: Option<&mut SplitImageLayerWriter>,
+        delta_writer: &mut SplitDeltaLayerWriter<'_>,
+        mut image_writer: Option<&mut SplitImageLayerWriter<'_>>,
         stat: &mut CompactionStatistics,
         ctx: &RequestContext,
     ) -> anyhow::Result<()> {
@@ -819,15 +824,16 @@ pub struct CompactionStatistics {
     time_acquire_lock_secs: f64,
     time_analyze_secs: f64,
     time_download_layer_secs: f64,
+    time_to_first_kv_pair_secs: f64,
     time_main_loop_secs: f64,
     time_final_phase_secs: f64,
     time_total_secs: f64,
 
     // Summary
-    /// Ratio of the key-value size before/after gc-compaction.
-    uncompressed_size_ratio: f64,
-    /// Ratio of the physical size before/after gc-compaction.
-    physical_size_ratio: f64,
+    /// Ratio of the key-value size after/before gc-compaction.
+    uncompressed_retention_ratio: f64,
+    /// Ratio of the physical size after/before gc-compaction.
+    compressed_retention_ratio: f64,
 }
 
 impl CompactionStatistics {
@@ -896,15 +902,15 @@ impl CompactionStatistics {
     fn finalize(&mut self) {
         let original_key_value_size = self.image_keys_visited.size + self.wal_keys_visited.size;
         let produced_key_value_size = self.image_produced.size + self.wal_produced.size;
-        self.uncompressed_size_ratio =
-            original_key_value_size as f64 / (produced_key_value_size as f64 + 1.0); // avoid div by 0
+        self.uncompressed_retention_ratio =
+            produced_key_value_size as f64 / (original_key_value_size as f64 + 1.0); // avoid div by 0
         let original_physical_size = self.image_layer_visited.size + self.delta_layer_visited.size;
         let produced_physical_size = self.image_layer_produced.size
             + self.delta_layer_produced.size
             + self.image_layer_discarded.size
             + self.delta_layer_discarded.size; // Also include the discarded layers to make the ratio accurate
-        self.physical_size_ratio =
-            original_physical_size as f64 / (produced_physical_size as f64 + 1.0); // avoid div by 0
+        self.compressed_retention_ratio =
+            produced_physical_size as f64 / (original_physical_size as f64 + 1.0); // avoid div by 0
     }
 }
 
@@ -1134,6 +1140,7 @@ impl Timeline {
     ) -> Result<(), CompactionError> {
         let mut drop_layers = Vec::new();
         let mut layers_to_rewrite: Vec<Layer> = Vec::new();
+        let mut rewrite_max_exceeded: bool = false;
 
         // We will use the Lsn cutoff of the last GC as a threshold for rewriting layers: if a
         // layer is behind this Lsn, it indicates that the layer is being retained beyond the
@@ -1142,12 +1149,7 @@ impl Timeline {
         // Holding this read guard also blocks [`Self::gc_timeline`] from entering while we
         // are rewriting layers.
         let latest_gc_cutoff = self.get_applied_gc_cutoff_lsn();
-
-        tracing::info!(
-            "starting shard ancestor compaction, latest_gc_cutoff: {}, pitr cutoff {}",
-            *latest_gc_cutoff,
-            self.gc_info.read().unwrap().cutoffs.time
-        );
+        let pitr_cutoff = self.gc_info.read().unwrap().cutoffs.time;
 
         let layers = self.layers.read().await;
         for layer_desc in layers.layer_map()?.iter_historic_layers() {
@@ -1165,8 +1167,8 @@ impl Timeline {
                 // This ancestral layer only covers keys that belong to other shards.
                 // We include the full metadata in the log: if we had some critical bug that caused
                 // us to incorrectly drop layers, this would simplify manually debugging + reinstating those layers.
-                info!(%layer, old_metadata=?layer.metadata(),
-                    "dropping layer after shard split, contains no keys for this shard.",
+                debug!(%layer, old_metadata=?layer.metadata(),
+                    "dropping layer after shard split, contains no keys for this shard",
                 );
 
                 if cfg!(debug_assertions) {
@@ -1228,9 +1230,10 @@ impl Timeline {
             }
 
             if layers_to_rewrite.len() >= rewrite_max {
-                tracing::info!(%layer, "Will rewrite layer on a future compaction, already rewrote {}",
+                debug!(%layer, "Will rewrite layer on a future compaction, already rewrote {}",
                     layers_to_rewrite.len()
                 );
+                rewrite_max_exceeded = true;
                 continue;
             }
 
@@ -1238,19 +1241,40 @@ impl Timeline {
             layers_to_rewrite.push(layer);
         }
 
-        // Drop read lock on layer map before we start doing time-consuming I/O
+        // Drop read lock on layer map before we start doing time-consuming I/O.
         drop(layers);
 
+        // Drop out early if there's nothing to do.
+        if layers_to_rewrite.is_empty() && drop_layers.is_empty() {
+            return Ok(());
+        }
+
+        info!(
+            "starting shard ancestor compaction, rewriting {} layers and dropping {} layers \
+                (latest_gc_cutoff={} pitr_cutoff={})",
+            layers_to_rewrite.len(),
+            drop_layers.len(),
+            *latest_gc_cutoff,
+            pitr_cutoff,
+        );
+        let started = Instant::now();
+
         let mut replace_image_layers = Vec::new();
 
         for layer in layers_to_rewrite {
-            tracing::info!(layer=%layer, "Rewriting layer after shard split...");
+            if self.cancel.is_cancelled() {
+                return Err(CompactionError::ShuttingDown);
+            }
+
+            info!(layer=%layer, "rewriting layer after shard split");
             let mut image_layer_writer = ImageLayerWriter::new(
                 self.conf,
                 self.timeline_id,
                 self.tenant_shard_id,
                 &layer.layer_desc().key_range,
                 layer.layer_desc().image_layer_lsn(),
+                &self.gate,
+                self.cancel.clone(),
                 ctx,
             )
             .await
@@ -1282,7 +1306,7 @@ impl Timeline {
                     .map_err(CompactionError::Other)?;
                 let new_layer = Layer::finish_creating(self.conf, self, desc, &path)
                     .map_err(CompactionError::Other)?;
-                tracing::info!(layer=%new_layer, "Rewrote layer, {} -> {} bytes",
+                info!(layer=%new_layer, "rewrote layer, {} -> {} bytes",
                     layer.metadata().file_size,
                     new_layer.metadata().file_size);
 
@@ -1294,6 +1318,12 @@ impl Timeline {
             }
         }
 
+        for layer in &drop_layers {
+            info!(%layer, old_metadata=?layer.metadata(),
+                "dropping layer after shard split (no keys for this shard)",
+            );
+        }
+
         // At this point, we have replaced local layer files with their rewritten form, but not yet uploaded
         // metadata to reflect that. If we restart here, the replaced layer files will look invalid (size mismatch
         // to remote index) and be removed. This is inefficient but safe.
@@ -1309,6 +1339,7 @@ impl Timeline {
         // necessary for correctness, but it simplifies testing, and avoids proceeding with another
         // Timeline's compaction while this timeline's uploads may be generating lots of disk I/O
         // load.
+        info!("shard ancestor compaction waiting for uploads");
         match self.remote_client.wait_completion().await {
             Ok(()) => (),
             Err(WaitCompletionError::NotInitialized(ni)) => return Err(CompactionError::from(ni)),
@@ -1317,6 +1348,15 @@ impl Timeline {
             }
         }
 
+        info!(
+            "shard ancestor compaction done in {:.3}s{}",
+            started.elapsed().as_secs_f64(),
+            match rewrite_max_exceeded {
+                true => format!(", more work pending due to rewrite_max={rewrite_max}"),
+                false => String::new(),
+            }
+        );
+
         fail::fail_point!("compact-shard-ancestors-persistent");
 
         Ok(())
@@ -1851,6 +1891,8 @@ impl Timeline {
                                 debug!("Create new layer {}..{}", lsn_range.start, lsn_range.end);
                                 lsn_range.clone()
                             },
+                            &self.gate,
+                            self.cancel.clone(),
                             ctx,
                         )
                         .await
@@ -2406,7 +2448,9 @@ impl Timeline {
                 } else {
                     lsn_split_points[i]
                 };
-                let img = self.reconstruct_value(key, request_lsn, state).await?;
+                let img = self
+                    .reconstruct_value(key, request_lsn, state, RedoAttemptType::GcCompaction)
+                    .await?;
                 Some((request_lsn, img))
             } else {
                 None
@@ -3026,7 +3070,7 @@ impl Timeline {
         .map_err(CompactionError::Other)?;
 
         let time_download_layer = timer.elapsed();
-        let timer = Instant::now();
+        let mut timer = Instant::now();
 
         // Step 2: Produce images+deltas.
         let mut accumulated_values = Vec::new();
@@ -3043,6 +3087,8 @@ impl Timeline {
                     job_desc.compaction_key_range.start,
                     lowest_retain_lsn,
                     self.get_compaction_target_size(),
+                    &self.gate,
+                    self.cancel.clone(),
                     ctx,
                 )
                 .await
@@ -3059,6 +3105,8 @@ impl Timeline {
             self.tenant_shard_id,
             lowest_retain_lsn..end_lsn,
             self.get_compaction_target_size(),
+            &self.gate,
+            self.cancel.clone(),
         )
         .await
         .context("failed to create delta layer writer")
@@ -3101,8 +3149,7 @@ impl Timeline {
         // Actually, we can decide not to write to the image layer at all at this point because
         // the key and LSN range are determined. However, to keep things simple here, we still
         // create this writer, and discard the writer in the end.
-
-        let mut keys_processed = 0;
+        let mut time_to_first_kv_pair = None;
 
         while let Some(((key, lsn, val), desc)) = merge_iter
             .next_with_trace()
@@ -3110,13 +3157,16 @@ impl Timeline {
             .context("failed to get next key-value pair")
             .map_err(CompactionError::Other)?
         {
+            if time_to_first_kv_pair.is_none() {
+                time_to_first_kv_pair = Some(timer.elapsed());
+                timer = Instant::now();
+            }
+
             if cancel.is_cancelled() {
                 return Err(CompactionError::ShuttingDown);
             }
 
-            keys_processed += 1;
             let should_yield = yield_for_l0
-                && keys_processed % 1000 == 0
                 && self
                     .l0_compaction_trigger
                     .notified()
@@ -3153,6 +3203,8 @@ impl Timeline {
                                 self.tenant_shard_id,
                                 desc.key_range.start,
                                 desc.lsn_range.clone(),
+                                &self.gate,
+                                self.cancel.clone(),
                                 ctx,
                             )
                             .await
@@ -3170,6 +3222,8 @@ impl Timeline {
                                 self.tenant_shard_id,
                                 job_desc.compaction_key_range.end,
                                 desc.lsn_range.clone(),
+                                &self.gate,
+                                self.cancel.clone(),
                                 ctx,
                             )
                             .await
@@ -3447,6 +3501,9 @@ impl Timeline {
         let time_final_phase = timer.elapsed();
 
         stat.time_final_phase_secs = time_final_phase.as_secs_f64();
+        stat.time_to_first_kv_pair_secs = time_to_first_kv_pair
+            .unwrap_or(Duration::ZERO)
+            .as_secs_f64();
         stat.time_main_loop_secs = time_main_loop.as_secs_f64();
         stat.time_acquire_lock_secs = time_acquire_lock.as_secs_f64();
         stat.time_download_layer_secs = time_download_layer.as_secs_f64();
@@ -3736,6 +3793,8 @@ impl CompactionJobExecutor for TimelineAdaptor {
             self.timeline.tenant_shard_id,
             key_range.start,
             lsn_range.clone(),
+            &self.timeline.gate,
+            self.timeline.cancel.clone(),
             ctx,
         )
         .await?;
@@ -3811,6 +3870,8 @@ impl TimelineAdaptor {
             self.timeline.tenant_shard_id,
             key_range,
             lsn,
+            &self.timeline.gate,
+            self.timeline.cancel.clone(),
             ctx,
         )
         .await?;
@@ -3907,8 +3968,6 @@ impl CompactionLayer<Key> for OwnArc<DeltaLayer> {
     }
 }
 
-use crate::tenant::timeline::DeltaEntry;
-
 impl CompactionLayer<Key> for ResidentDeltaLayer {
     fn key_range(&self) -> &Range<Key> {
         &self.0.layer_desc().key_range

pageserver/src/tenant/timeline/delete.rs

@@ -410,10 +410,13 @@ impl DeleteTimelineFlow {
         // So indeed, the tenant manifest might refer to an offloaded timeline which has already been deleted.
         // However, we handle this case in tenant loading code so the next time we attach, the issue is
         // resolved.
-        tenant.store_tenant_manifest().await.map_err(|e| match e {
-            TenantManifestError::Cancelled => DeleteTimelineError::Cancelled,
-            _ => DeleteTimelineError::Other(e.into()),
-        })?;
+        tenant
+            .maybe_upload_tenant_manifest()
+            .await
+            .map_err(|err| match err {
+                TenantManifestError::Cancelled => DeleteTimelineError::Cancelled,
+                err => DeleteTimelineError::Other(err.into()),
+            })?;
 
         *guard = Self::Finished;
 

pageserver/src/tenant/timeline/detach_ancestor.rs

@@ -231,6 +231,8 @@ async fn generate_tombstone_image_layer(
             detached.tenant_shard_id,
             &key_range,
             image_lsn,
+            &detached.gate,
+            detached.cancel.clone(),
             ctx,
         )
         .await
@@ -779,6 +781,8 @@ async fn copy_lsn_prefix(
         target_timeline.tenant_shard_id,
         layer.layer_desc().key_range.start,
         layer.layer_desc().lsn_range.start..end_lsn,
+        &target_timeline.gate,
+        target_timeline.cancel.clone(),
         ctx,
     )
     .await

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

@@ -738,6 +738,8 @@ impl ChunkProcessingJob {
             self.timeline.tenant_shard_id,
             &self.range,
             self.pgdata_lsn,
+            &self.timeline.gate,
+            self.timeline.cancel.clone(),
             ctx,
         )
         .await?;

pageserver/src/tenant/timeline/layer_manager.rs

@@ -3,17 +3,18 @@ use std::sync::Arc;
 
 use anyhow::{Context, bail, ensure};
 use itertools::Itertools;
+use pageserver_api::keyspace::KeySpace;
 use pageserver_api::shard::TenantShardId;
 use tokio_util::sync::CancellationToken;
 use tracing::trace;
 use utils::id::TimelineId;
 use utils::lsn::{AtomicLsn, Lsn};
 
-use super::{ReadableLayer, TimelineWriterState};
+use super::{LayerFringe, ReadableLayer, TimelineWriterState};
 use crate::config::PageServerConf;
 use crate::context::RequestContext;
 use crate::metrics::TimelineMetrics;
-use crate::tenant::layer_map::{BatchedUpdates, LayerMap};
+use crate::tenant::layer_map::{BatchedUpdates, LayerMap, SearchResult};
 use crate::tenant::storage_layer::{
     AsLayerDesc, InMemoryLayer, Layer, LayerVisibilityHint, PersistentLayerDesc,
     PersistentLayerKey, ReadableLayerWeak, ResidentLayer,
@@ -38,7 +39,7 @@ impl Default for LayerManager {
 }
 
 impl LayerManager {
-    pub(crate) fn upgrade(&self, weak: ReadableLayerWeak) -> ReadableLayer {
+    fn upgrade(&self, weak: ReadableLayerWeak) -> ReadableLayer {
         match weak {
             ReadableLayerWeak::PersistentLayer(desc) => {
                 ReadableLayer::PersistentLayer(self.get_from_desc(&desc))
@@ -147,6 +148,36 @@ impl LayerManager {
         self.layers().keys().cloned().collect_vec()
     }
 
+    /// Update the [`LayerFringe`] of a read request
+    ///
+    /// Take a key space at a given LSN and query the layer map below each range
+    /// of the key space to find the next layers to visit.
+    pub(crate) fn update_search_fringe(
+        &self,
+        keyspace: &KeySpace,
+        cont_lsn: Lsn,
+        fringe: &mut LayerFringe,
+    ) -> Result<(), Shutdown> {
+        let map = self.layer_map()?;
+
+        for range in keyspace.ranges.iter() {
+            let results = map.range_search(range.clone(), cont_lsn);
+            results
+                .found
+                .into_iter()
+                .map(|(SearchResult { layer, lsn_floor }, keyspace_accum)| {
+                    (
+                        self.upgrade(layer),
+                        keyspace_accum.to_keyspace(),
+                        lsn_floor..cont_lsn,
+                    )
+                })
+                .for_each(|(layer, keyspace, lsn_range)| fringe.update(layer, keyspace, lsn_range));
+        }
+
+        Ok(())
+    }
+
     fn layers(&self) -> &HashMap<PersistentLayerKey, Layer> {
         use LayerManager::*;
         match self {

pageserver/src/tenant/timeline/offload.rs

@@ -111,7 +111,7 @@ pub(crate) async fn offload_timeline(
     // at the next restart attach it again.
     // For that to happen, we'd need to make the manifest reflect our *intended* state,
     // not our actual state of offloaded timelines.
-    tenant.store_tenant_manifest().await?;
+    tenant.maybe_upload_tenant_manifest().await?;
 
     tracing::info!("Timeline offload complete (remaining arc refcount: {remaining_refcount})");
 

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

@@ -580,6 +580,7 @@ impl ConnectionManagerState {
                                 );
                                 Ok(())
                             }
+                            WalReceiverError::Cancelled => Ok(()),
                             WalReceiverError::Other(e) => {
                                 // give out an error to have task_mgr give it a really verbose logging
                                 if cancellation.is_cancelled() {

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

@@ -73,6 +73,7 @@ pub(super) enum WalReceiverError {
     /// Generic error
     Other(anyhow::Error),
     ClosedGate,
+    Cancelled,
 }
 
 impl From<tokio_postgres::Error> for WalReceiverError {
@@ -200,6 +201,9 @@ pub(super) async fn handle_walreceiver_connection(
                                 // with a similar error.
                             },
                             WalReceiverError::SuccessfulCompletion(_) => {}
+                            WalReceiverError::Cancelled => {
+                                debug!("Connection cancelled")
+                            }
                             WalReceiverError::ClosedGate => {
                                 // doesn't happen at runtime
                             }
@@ -273,7 +277,12 @@ pub(super) async fn handle_walreceiver_connection(
 
     let mut waldecoder = WalStreamDecoder::new(startpoint, timeline.pg_version);
 
-    let mut walingest = WalIngest::new(timeline.as_ref(), startpoint, &ctx).await?;
+    let mut walingest = WalIngest::new(timeline.as_ref(), startpoint, &ctx)
+        .await
+        .map_err(|e| match e.kind {
+            crate::walingest::WalIngestErrorKind::Cancelled => WalReceiverError::Cancelled,
+            _ => WalReceiverError::Other(e.into()),
+        })?;
 
     let shard = vec![*timeline.get_shard_identity()];
 
@@ -445,7 +454,7 @@ pub(super) async fn handle_walreceiver_connection(
                         .inspect_err(|err| {
                             // TODO: we can't differentiate cancellation errors with
                             // anyhow::Error, so just ignore it if we're cancelled.
-                            if !cancellation.is_cancelled() {
+                            if !cancellation.is_cancelled() && !timeline.is_stopping() {
                                 critical!("{err:?}")
                             }
                         })?;
@@ -577,7 +586,7 @@ pub(super) async fn handle_walreceiver_connection(
                             .inspect_err(|err| {
                                 // TODO: we can't differentiate cancellation errors with
                                 // anyhow::Error, so just ignore it if we're cancelled.
-                                if !cancellation.is_cancelled() {
+                                if !cancellation.is_cancelled() && !timeline.is_stopping() {
                                     critical!("{err:?}")
                                 }
                             })?;

pageserver/src/tenant/upload_queue.rs

@@ -302,6 +302,7 @@ pub struct UploadQueueStoppedDeletable {
     pub(super) deleted_at: SetDeletedFlagProgress,
 }
 
+#[allow(clippy::large_enum_variant, reason = "TODO")]
 pub enum UploadQueueStopped {
     Deletable(UploadQueueStoppedDeletable),
     Uninitialized,

pageserver/src/tenant/vectored_blob_io.rs

@@ -677,7 +677,6 @@ impl StreamingVectoredReadPlanner {
 
 #[cfg(test)]
 mod tests {
-    use anyhow::Error;
 
     use super::super::blob_io::tests::{random_array, write_maybe_compressed};
     use super::*;
@@ -960,13 +959,16 @@ mod tests {
         }
     }
 
-    async fn round_trip_test_compressed(blobs: &[Vec<u8>], compression: bool) -> Result<(), Error> {
+    async fn round_trip_test_compressed(
+        blobs: &[Vec<u8>],
+        compression: bool,
+    ) -> anyhow::Result<()> {
         let ctx =
             RequestContext::new(TaskKind::UnitTest, DownloadBehavior::Error).with_scope_unit_test();
         let (_temp_dir, pathbuf, offsets) =
-            write_maybe_compressed::<true>(blobs, compression, &ctx).await?;
+            write_maybe_compressed(blobs, compression, &ctx).await?;
 
-        let file = VirtualFile::open(&pathbuf, &ctx).await?;
+        let file = VirtualFile::open_v2(&pathbuf, &ctx).await?;
         let file_len = std::fs::metadata(&pathbuf)?.len();
 
         // Multiply by two (compressed data might need more space), and add a few bytes for the header
@@ -1003,7 +1005,7 @@ mod tests {
     }
 
     #[tokio::test]
-    async fn test_really_big_array() -> Result<(), Error> {
+    async fn test_really_big_array() -> anyhow::Result<()> {
         let blobs = &[
             b"test".to_vec(),
             random_array(10 * PAGE_SZ),
@@ -1018,7 +1020,7 @@ mod tests {
     }
 
     #[tokio::test]
-    async fn test_arrays_inc() -> Result<(), Error> {
+    async fn test_arrays_inc() -> anyhow::Result<()> {
         let blobs = (0..PAGE_SZ / 8)
             .map(|v| random_array(v * 16))
             .collect::<Vec<_>>();

pageserver/src/virtual_file.rs

@@ -12,7 +12,7 @@
 //! src/backend/storage/file/fd.c
 //!
 use std::fs::File;
-use std::io::{Error, ErrorKind, Seek, SeekFrom};
+use std::io::{Error, ErrorKind};
 use std::os::fd::{AsRawFd, FromRawFd, IntoRawFd, OwnedFd, RawFd};
 #[cfg(target_os = "linux")]
 use std::os::unix::fs::OpenOptionsExt;
@@ -185,18 +185,14 @@ impl VirtualFile {
         self.inner.sync_data().await
     }
 
+    pub async fn set_len(&self, len: u64) -> Result<(), Error> {
+        self.inner.set_len(len).await
+    }
+
     pub async fn metadata(&self) -> Result<Metadata, Error> {
         self.inner.metadata().await
     }
 
-    pub fn remove(self) {
-        self.inner.remove();
-    }
-
-    pub async fn seek(&mut self, pos: SeekFrom) -> Result<u64, Error> {
-        self.inner.seek(pos).await
-    }
-
     pub async fn read_exact_at<Buf>(
         &self,
         slice: Slice<Buf>,
@@ -227,25 +223,31 @@ impl VirtualFile {
         self.inner.write_all_at(buf, offset, ctx).await
     }
 
-    pub async fn write_all<Buf: IoBuf + Send>(
-        &mut self,
-        buf: FullSlice<Buf>,
+    pub(crate) async fn read_to_string<P: AsRef<Utf8Path>>(
+        path: P,
         ctx: &RequestContext,
-    ) -> (FullSlice<Buf>, Result<usize, Error>) {
-        self.inner.write_all(buf, ctx).await
-    }
-
-    async fn read_to_end(&mut self, buf: &mut Vec<u8>, ctx: &RequestContext) -> Result<(), Error> {
-        self.inner.read_to_end(buf, ctx).await
-    }
-
-    pub(crate) async fn read_to_string(
-        &mut self,
-        ctx: &RequestContext,
-    ) -> Result<String, anyhow::Error> {
+    ) -> std::io::Result<String> {
+        let file = VirtualFile::open(path, ctx).await?; // TODO: open_v2
         let mut buf = Vec::new();
-        self.read_to_end(&mut buf, ctx).await?;
-        Ok(String::from_utf8(buf)?)
+        let mut tmp = vec![0; 128];
+        let mut pos: u64 = 0;
+        loop {
+            let slice = tmp.slice(..128);
+            let (slice, res) = file.inner.read_at(slice, pos, ctx).await;
+            match res {
+                Ok(0) => break,
+                Ok(n) => {
+                    pos += n as u64;
+                    buf.extend_from_slice(&slice[..n]);
+                }
+                Err(ref e) if e.kind() == std::io::ErrorKind::Interrupted => {}
+                Err(e) => return Err(e),
+            }
+            tmp = slice.into_inner();
+        }
+        String::from_utf8(buf).map_err(|_| {
+            std::io::Error::new(ErrorKind::InvalidData, "file contents are not valid UTF-8")
+        })
     }
 }
 
@@ -292,9 +294,6 @@ pub struct VirtualFileInner {
     /// belongs to a different VirtualFile.
     handle: RwLock<SlotHandle>,
 
-    /// Current file position
-    pos: u64,
-
     /// File path and options to use to open it.
     ///
     /// Note: this only contains the options needed to re-open it. For example,
@@ -608,7 +607,6 @@ impl VirtualFileInner {
 
         let vfile = VirtualFileInner {
             handle: RwLock::new(handle),
-            pos: 0,
             path: path.to_owned(),
             open_options: reopen_options,
         };
@@ -675,6 +673,13 @@ impl VirtualFileInner {
         })
     }
 
+    pub async fn set_len(&self, len: u64) -> Result<(), Error> {
+        with_file!(self, StorageIoOperation::SetLen, |file_guard| {
+            let (_file_guard, res) = io_engine::get().set_len(file_guard, len).await;
+            res.maybe_fatal_err("set_len")
+        })
+    }
+
     /// Helper function internal to `VirtualFile` that looks up the underlying File,
     /// opens it and evicts some other File if necessary. The passed parameter is
     /// assumed to be a function available for the physical `File`.
@@ -742,38 +747,6 @@ impl VirtualFileInner {
         })
     }
 
-    pub fn remove(self) {
-        let path = self.path.clone();
-        drop(self);
-        std::fs::remove_file(path).expect("failed to remove the virtual file");
-    }
-
-    pub async fn seek(&mut self, pos: SeekFrom) -> Result<u64, Error> {
-        match pos {
-            SeekFrom::Start(offset) => {
-                self.pos = offset;
-            }
-            SeekFrom::End(offset) => {
-                self.pos = with_file!(self, StorageIoOperation::Seek, |mut file_guard| file_guard
-                    .with_std_file_mut(|std_file| std_file.seek(SeekFrom::End(offset))))?
-            }
-            SeekFrom::Current(offset) => {
-                let pos = self.pos as i128 + offset as i128;
-                if pos < 0 {
-                    return Err(Error::new(
-                        ErrorKind::InvalidInput,
-                        "offset would be negative",
-                    ));
-                }
-                if pos > u64::MAX as i128 {
-                    return Err(Error::new(ErrorKind::InvalidInput, "offset overflow"));
-                }
-                self.pos = pos as u64;
-            }
-        }
-        Ok(self.pos)
-    }
-
     /// Read the file contents in range `offset..(offset + slice.bytes_total())` into `slice[0..slice.bytes_total()]`.
     ///
     /// The returned `Slice<Buf>` is equivalent to the input `slice`, i.e., it's the same view into the same buffer.
@@ -857,59 +830,7 @@ impl VirtualFileInner {
         (restore(buf), Ok(()))
     }
 
-    /// Writes `buf` to the file at the current offset.
-    ///
-    /// Panics if there is an uninitialized range in `buf`, as that is most likely a bug in the caller.
-    pub async fn write_all<Buf: IoBuf + Send>(
-        &mut self,
-        buf: FullSlice<Buf>,
-        ctx: &RequestContext,
-    ) -> (FullSlice<Buf>, Result<usize, Error>) {
-        let buf = buf.into_raw_slice();
-        let bounds = buf.bounds();
-        let restore =
-            |buf: Slice<_>| FullSlice::must_new(Slice::from_buf_bounds(buf.into_inner(), bounds));
-        let nbytes = buf.len();
-        let mut buf = buf;
-        while !buf.is_empty() {
-            let (tmp, res) = self.write(FullSlice::must_new(buf), ctx).await;
-            buf = tmp.into_raw_slice();
-            match res {
-                Ok(0) => {
-                    return (
-                        restore(buf),
-                        Err(Error::new(
-                            std::io::ErrorKind::WriteZero,
-                            "failed to write whole buffer",
-                        )),
-                    );
-                }
-                Ok(n) => {
-                    buf = buf.slice(n..);
-                }
-                Err(ref e) if e.kind() == std::io::ErrorKind::Interrupted => {}
-                Err(e) => return (restore(buf), Err(e)),
-            }
-        }
-        (restore(buf), Ok(nbytes))
-    }
-
-    async fn write<B: IoBuf + Send>(
-        &mut self,
-        buf: FullSlice<B>,
-        ctx: &RequestContext,
-    ) -> (FullSlice<B>, Result<usize, std::io::Error>) {
-        let pos = self.pos;
-        let (buf, res) = self.write_at(buf, pos, ctx).await;
-        let n = match res {
-            Ok(n) => n,
-            Err(e) => return (buf, Err(e)),
-        };
-        self.pos += n as u64;
-        (buf, Ok(n))
-    }
-
-    pub(crate) async fn read_at<Buf>(
+    pub(super) async fn read_at<Buf>(
         &self,
         buf: tokio_epoll_uring::Slice<Buf>,
         offset: u64,
@@ -937,23 +858,11 @@ impl VirtualFileInner {
         })
     }
 
-    /// The function aborts the process if the error is fatal.
     async fn write_at<B: IoBuf + Send>(
         &self,
         buf: FullSlice<B>,
         offset: u64,
         ctx: &RequestContext,
-    ) -> (FullSlice<B>, Result<usize, Error>) {
-        let (slice, result) = self.write_at_inner(buf, offset, ctx).await;
-        let result = result.maybe_fatal_err("write_at");
-        (slice, result)
-    }
-
-    async fn write_at_inner<B: IoBuf + Send>(
-        &self,
-        buf: FullSlice<B>,
-        offset: u64,
-        ctx: &RequestContext,
     ) -> (FullSlice<B>, Result<usize, Error>) {
         let file_guard = match self.lock_file().await {
             Ok(file_guard) => file_guard,
@@ -962,30 +871,13 @@ impl VirtualFileInner {
         observe_duration!(StorageIoOperation::Write, {
             let ((_file_guard, buf), result) =
                 io_engine::get().write_at(file_guard, offset, buf).await;
+            let result = result.maybe_fatal_err("write_at");
             if let Ok(size) = result {
                 ctx.io_size_metrics().write.add(size.into_u64());
             }
             (buf, result)
         })
     }
-
-    async fn read_to_end(&mut self, buf: &mut Vec<u8>, ctx: &RequestContext) -> Result<(), Error> {
-        let mut tmp = vec![0; 128];
-        loop {
-            let slice = tmp.slice(..128);
-            let (slice, res) = self.read_at(slice, self.pos, ctx).await;
-            match res {
-                Ok(0) => return Ok(()),
-                Ok(n) => {
-                    self.pos += n as u64;
-                    buf.extend_from_slice(&slice[..n]);
-                }
-                Err(ref e) if e.kind() == std::io::ErrorKind::Interrupted => {}
-                Err(e) => return Err(e),
-            }
-            tmp = slice.into_inner();
-        }
-    }
 }
 
 // Adapted from https://doc.rust-lang.org/1.72.0/src/std/os/unix/fs.rs.html#117-135
@@ -1200,19 +1092,6 @@ impl FileGuard {
         let _ = file.into_raw_fd();
         res
     }
-    /// Soft deprecation: we'll move VirtualFile to async APIs and remove this function eventually.
-    fn with_std_file_mut<F, R>(&mut self, with: F) -> R
-    where
-        F: FnOnce(&mut File) -> R,
-    {
-        // SAFETY:
-        // - lifetime of the fd: `file` doesn't outlive the OwnedFd stored in `self`.
-        // - &mut usage below: `self` is `&mut`, hence this call is the only task/thread that has control over the underlying fd
-        let mut file = unsafe { File::from_raw_fd(self.as_ref().as_raw_fd()) };
-        let res = with(&mut file);
-        let _ = file.into_raw_fd();
-        res
-    }
 }
 
 impl tokio_epoll_uring::IoFd for FileGuard {
@@ -1380,7 +1259,6 @@ static SYNC_MODE: AtomicU8 = AtomicU8::new(SyncMode::Sync as u8);
 
 #[cfg(test)]
 mod tests {
-    use std::io::Write;
     use std::os::unix::fs::FileExt;
     use std::sync::Arc;
 
@@ -1433,43 +1311,6 @@ mod tests {
                 MaybeVirtualFile::File(file) => file.write_all_at(&buf[..], offset),
             }
         }
-        async fn seek(&mut self, pos: SeekFrom) -> Result<u64, Error> {
-            match self {
-                MaybeVirtualFile::VirtualFile(file) => file.seek(pos).await,
-                MaybeVirtualFile::File(file) => file.seek(pos),
-            }
-        }
-        async fn write_all<Buf: IoBuf + Send>(
-            &mut self,
-            buf: FullSlice<Buf>,
-            ctx: &RequestContext,
-        ) -> Result<(), Error> {
-            match self {
-                MaybeVirtualFile::VirtualFile(file) => {
-                    let (_buf, res) = file.write_all(buf, ctx).await;
-                    res.map(|_| ())
-                }
-                MaybeVirtualFile::File(file) => file.write_all(&buf[..]),
-            }
-        }
-
-        // Helper function to slurp contents of a file, starting at the current position,
-        // into a string
-        async fn read_string(&mut self, ctx: &RequestContext) -> Result<String, Error> {
-            use std::io::Read;
-            let mut buf = String::new();
-            match self {
-                MaybeVirtualFile::VirtualFile(file) => {
-                    let mut buf = Vec::new();
-                    file.read_to_end(&mut buf, ctx).await?;
-                    return Ok(String::from_utf8(buf).unwrap());
-                }
-                MaybeVirtualFile::File(file) => {
-                    file.read_to_string(&mut buf)?;
-                }
-            }
-            Ok(buf)
-        }
 
         // Helper function to slurp a portion of a file into a string
         async fn read_string_at(
@@ -1565,48 +1406,23 @@ mod tests {
         .await?;
 
         file_a
-            .write_all(b"foobar".to_vec().slice_len(), &ctx)
+            .write_all_at(IoBuffer::from(b"foobar").slice_len(), 0, &ctx)
             .await?;
 
         // cannot read from a file opened in write-only mode
-        let _ = file_a.read_string(&ctx).await.unwrap_err();
+        let _ = file_a.read_string_at(0, 1, &ctx).await.unwrap_err();
 
         // Close the file and re-open for reading
         let mut file_a = A::open(path_a, OpenOptions::new().read(true).to_owned(), &ctx).await?;
 
         // cannot write to a file opened in read-only mode
         let _ = file_a
-            .write_all(b"bar".to_vec().slice_len(), &ctx)
+            .write_all_at(IoBuffer::from(b"bar").slice_len(), 0, &ctx)
             .await
             .unwrap_err();
 
         // Try simple read
-        assert_eq!("foobar", file_a.read_string(&ctx).await?);
-
-        // It's positioned at the EOF now.
-        assert_eq!("", file_a.read_string(&ctx).await?);
-
-        // Test seeks.
-        assert_eq!(file_a.seek(SeekFrom::Start(1)).await?, 1);
-        assert_eq!("oobar", file_a.read_string(&ctx).await?);
-
-        assert_eq!(file_a.seek(SeekFrom::End(-2)).await?, 4);
-        assert_eq!("ar", file_a.read_string(&ctx).await?);
-
-        assert_eq!(file_a.seek(SeekFrom::Start(1)).await?, 1);
-        assert_eq!(file_a.seek(SeekFrom::Current(2)).await?, 3);
-        assert_eq!("bar", file_a.read_string(&ctx).await?);
-
-        assert_eq!(file_a.seek(SeekFrom::Current(-5)).await?, 1);
-        assert_eq!("oobar", file_a.read_string(&ctx).await?);
-
-        // Test erroneous seeks to before byte 0
-        file_a.seek(SeekFrom::End(-7)).await.unwrap_err();
-        assert_eq!(file_a.seek(SeekFrom::Start(1)).await?, 1);
-        file_a.seek(SeekFrom::Current(-2)).await.unwrap_err();
-
-        // the erroneous seek should have left the position unchanged
-        assert_eq!("oobar", file_a.read_string(&ctx).await?);
+        assert_eq!("foobar", file_a.read_string_at(0, 6, &ctx).await?);
 
         // Create another test file, and try FileExt functions on it.
         let path_b = testdir.join("file_b");
@@ -1632,9 +1448,6 @@ mod tests {
 
         // Open a lot of files, enough to cause some evictions. (Or to be precise,
         // open the same file many times. The effect is the same.)
-        //
-        // leave file_a positioned at offset 1 before we start
-        assert_eq!(file_a.seek(SeekFrom::Start(1)).await?, 1);
 
         let mut vfiles = Vec::new();
         for _ in 0..100 {
@@ -1644,7 +1457,7 @@ mod tests {
                 &ctx,
             )
             .await?;
-            assert_eq!("FOOBAR", vfile.read_string(&ctx).await?);
+            assert_eq!("FOOBAR", vfile.read_string_at(0, 6, &ctx).await?);
             vfiles.push(vfile);
         }
 
@@ -1652,8 +1465,8 @@ mod tests {
         assert!(vfiles.len() > TEST_MAX_FILE_DESCRIPTORS * 2);
 
         // The underlying file descriptor for 'file_a' should be closed now. Try to read
-        // from it again. We left the file positioned at offset 1 above.
-        assert_eq!("oobar", file_a.read_string(&ctx).await?);
+        // from it again.
+        assert_eq!("foobar", file_a.read_string_at(0, 6, &ctx).await?);
 
         // Check that all the other FDs still work too. Use them in random order for
         // good measure.
@@ -1747,7 +1560,7 @@ mod tests {
             .await
             .unwrap();
         let mut file = MaybeVirtualFile::from(VirtualFile::open(&path, &ctx).await.unwrap());
-        let post = file.read_string(&ctx).await.unwrap();
+        let post = file.read_string_at(0, 3, &ctx).await.unwrap();
         assert_eq!(post, "foo");
         assert!(!tmp_path.exists());
         drop(file);
@@ -1756,7 +1569,7 @@ mod tests {
             .await
             .unwrap();
         let mut file = MaybeVirtualFile::from(VirtualFile::open(&path, &ctx).await.unwrap());
-        let post = file.read_string(&ctx).await.unwrap();
+        let post = file.read_string_at(0, 3, &ctx).await.unwrap();
         assert_eq!(post, "bar");
         assert!(!tmp_path.exists());
         drop(file);
@@ -1781,7 +1594,7 @@ mod tests {
             .unwrap();
 
         let mut file = MaybeVirtualFile::from(VirtualFile::open(&path, &ctx).await.unwrap());
-        let post = file.read_string(&ctx).await.unwrap();
+        let post = file.read_string_at(0, 3, &ctx).await.unwrap();
         assert_eq!(post, "foo");
         assert!(!tmp_path.exists());
         drop(file);

pageserver/src/virtual_file/io_engine.rs

@@ -209,6 +209,22 @@ impl IoEngine {
             }
         }
     }
+
+    pub(super) async fn set_len(
+        &self,
+        file_guard: FileGuard,
+        len: u64,
+    ) -> (FileGuard, std::io::Result<()>) {
+        match self {
+            IoEngine::NotSet => panic!("not initialized"),
+            // TODO: ftruncate op for tokio-epoll-uring
+            IoEngine::StdFs | IoEngine::TokioEpollUring => {
+                let res = file_guard.with_std_file(|std_file| std_file.set_len(len));
+                (file_guard, res)
+            }
+        }
+    }
+
     pub(super) async fn write_at<B: IoBuf + Send>(
         &self,
         file_guard: FileGuard,

pageserver/src/virtual_file/owned_buffers_io/aligned_buffer/buffer_mut.rs

@@ -282,6 +282,17 @@ unsafe impl<A: Alignment> tokio_epoll_uring::IoBufMut for AlignedBufferMut<A> {
     }
 }
 
+impl<A: Alignment> std::io::Write for AlignedBufferMut<A> {
+    fn write(&mut self, buf: &[u8]) -> std::io::Result<usize> {
+        self.extend_from_slice(buf);
+        Ok(buf.len())
+    }
+
+    fn flush(&mut self) -> std::io::Result<()> {
+        Ok(())
+    }
+}
+
 #[cfg(test)]
 mod tests {
 

pageserver/src/virtual_file/owned_buffers_io/io_buf_aligned.rs

@@ -1,9 +1,11 @@
 use tokio_epoll_uring::{IoBuf, IoBufMut};
 
-use crate::virtual_file::{IoBuffer, IoBufferMut, PageWriteGuardBuf};
+use crate::virtual_file::{self, IoBuffer, IoBufferMut, PageWriteGuardBuf};
 
 /// A marker trait for a mutable aligned buffer type.
-pub trait IoBufAlignedMut: IoBufMut {}
+pub trait IoBufAlignedMut: IoBufMut {
+    const ALIGN: usize = virtual_file::get_io_buffer_alignment();
+}
 
 /// A marker trait for an aligned buffer type.
 pub trait IoBufAligned: IoBuf {}

pageserver/src/virtual_file/owned_buffers_io/write.rs

@@ -1,6 +1,7 @@
 mod flush;
 use std::sync::Arc;
 
+use bytes::BufMut;
 pub(crate) use flush::FlushControl;
 use flush::FlushHandle;
 pub(crate) use flush::FlushTaskError;
@@ -8,6 +9,7 @@ use tokio_epoll_uring::IoBuf;
 use tokio_util::sync::CancellationToken;
 
 use super::io_buf_aligned::IoBufAligned;
+use super::io_buf_aligned::IoBufAlignedMut;
 use super::io_buf_ext::{FullSlice, IoBufExt};
 use crate::context::RequestContext;
 use crate::virtual_file::{IoBuffer, IoBufferMut};
@@ -64,7 +66,7 @@ pub struct BufferedWriter<B: Buffer, W> {
 
 impl<B, Buf, W> BufferedWriter<B, W>
 where
-    B: Buffer<IoBuf = Buf> + Send + 'static,
+    B: IoBufAlignedMut + Buffer<IoBuf = Buf> + Send + 'static,
     Buf: IoBufAligned + Send + Sync + CheapCloneForRead,
     W: OwnedAsyncWriter + Send + Sync + 'static + std::fmt::Debug,
 {
@@ -73,6 +75,7 @@ where
     /// The `buf_new` function provides a way to initialize the owned buffers used by this writer.
     pub fn new(
         writer: Arc<W>,
+        start_offset: u64,
         buf_new: impl Fn() -> B,
         gate_guard: utils::sync::gate::GateGuard,
         cancel: CancellationToken,
@@ -91,7 +94,7 @@ where
                 ctx.attached_child(),
                 flush_task_span,
             ),
-            bytes_submitted: 0,
+            bytes_submitted: start_offset,
         }
     }
 
@@ -116,21 +119,29 @@ where
     }
 
     #[cfg_attr(target_os = "macos", allow(dead_code))]
-    pub async fn flush_and_into_inner(
-        mut self,
-        ctx: &RequestContext,
-    ) -> Result<(u64, Arc<W>), FlushTaskError> {
-        self.flush(ctx).await?;
-
+    pub async fn shutdown(
+        self,
+        mut handle_tail: impl FnMut(B) -> Option<B>,
+    ) -> Result<(u64, W), FlushTaskError> {
         let Self {
-            mutable: buf,
+            mutable: tail,
             maybe_flushed: _,
             writer,
             mut flush_handle,
-            bytes_submitted: bytes_amount,
+            bytes_submitted: submit_offset,
         } = self;
-        flush_handle.shutdown().await?;
-        assert!(buf.is_some());
+
+        let ctx = flush_handle.shutdown().await?;
+        let buf = tail.expect("must not use after an error");
+        let writer = Arc::into_inner(writer).expect("writer is the only strong reference");
+        let mut bytes_amount = submit_offset;
+        if let Some(buf) = handle_tail(buf) {
+            bytes_amount += buf.pending() as u64;
+            // TODO: infinite retries + maybe_fatal_err like we do in the flush loop; can we just send this
+            // as work into the flush loop, as part of flush_handle.shutdown()
+            let (_, res) = writer.write_all_at(buf.flush(), submit_offset, &ctx).await;
+            let _: () = res.unwrap(); // DO NOT MERGE, THIS CAN FAIL E.G. on ENOSPC
+        }
         Ok((bytes_amount, writer))
     }
 
@@ -235,6 +246,10 @@ pub trait Buffer {
     /// panics if `other.len() > self.cap() - self.pending()`.
     fn extend_from_slice(&mut self, other: &[u8]);
 
+    /// Add `count` bytes `val` into `self`.
+    /// Panics if `count > self.cap() - self.pending()`.
+    fn extend_with(&mut self, val: u8, count: usize);
+
     /// Number of bytes in the buffer.
     fn pending(&self) -> usize;
 
@@ -262,6 +277,14 @@ impl Buffer for IoBufferMut {
         IoBufferMut::extend_from_slice(self, other);
     }
 
+    fn extend_with(&mut self, val: u8, count: usize) {
+        if self.len() + count > self.cap() {
+            panic!("Buffer capacity exceeded");
+        }
+
+        IoBufferMut::put_bytes(self, val, count);
+    }
+
     fn pending(&self) -> usize {
         self.len()
     }
@@ -334,6 +357,7 @@ mod tests {
         let cancel = CancellationToken::new();
         let mut writer = BufferedWriter::<_, RecorderWriter>::new(
             recorder,
+            0,
             || IoBufferMut::with_capacity(2),
             gate.enter()?,
             cancel,
@@ -350,7 +374,7 @@ mod tests {
         writer.write_buffered_borrowed(b"j", ctx).await?;
         writer.write_buffered_borrowed(b"klmno", ctx).await?;
 
-        let (_, recorder) = writer.flush_and_into_inner(ctx).await?;
+        let (_, recorder) = writer.shutdown(Some).await?;
         assert_eq!(
             recorder.get_writes(),
             {

pageserver/src/virtual_file/owned_buffers_io/write/flush.rs

@@ -1,5 +1,5 @@
 use std::ops::ControlFlow;
-use std::sync::Arc;
+use std::{marker::PhantomData, sync::Arc};
 
 use tokio_util::sync::CancellationToken;
 use tracing::{Instrument, info, info_span, warn};
@@ -21,7 +21,10 @@ pub struct FlushHandleInner<Buf, W> {
     /// and receives recyled buffer.
     channel: duplex::mpsc::Duplex<FlushRequest<Buf>, FullSlice<Buf>>,
     /// Join handle for the background flush task.
-    join_handle: tokio::task::JoinHandle<Result<Arc<W>, FlushTaskError>>,
+    join_handle: tokio::task::JoinHandle<Result<RequestContext, FlushTaskError>>,
+
+    // TODO: get rit of the type parameter?
+    _phantom: PhantomData<W>,
 }
 
 struct FlushRequest<Buf> {
@@ -144,6 +147,7 @@ where
             inner: Some(FlushHandleInner {
                 channel: front,
                 join_handle,
+                _phantom: PhantomData,
             }),
         }
     }
@@ -179,12 +183,13 @@ where
         Err(self
             .shutdown()
             .await
-            .expect_err("flush task only disconnects duplex if it exits with an error"))
+            .err()
+            .expect("flush task only disconnects duplex if it exits with an error"))
     }
 
     /// Cleans up the channel, join the flush task.
-    pub async fn shutdown(&mut self) -> Result<Arc<W>, FlushTaskError> {
-        let handle = self
+    pub async fn shutdown(&mut self) -> Result<RequestContext, FlushTaskError> {
+        let handle: FlushHandleInner<Buf, W> = self
             .inner
             .take()
             .expect("must not use after we returned an error");
@@ -243,7 +248,7 @@ where
     }
 
     /// Runs the background flush task.
-    async fn run(mut self) -> Result<Arc<W>, FlushTaskError> {
+    async fn run(mut self) -> Result<RequestContext, FlushTaskError> {
         //  Exit condition: channel is closed and there is no remaining buffer to be flushed
         while let Some(request) = self.channel.recv().await {
             #[cfg(test)]
@@ -313,8 +318,7 @@ where
                 continue;
             }
         }
-
-        Ok(self.writer)
+        Ok(self.ctx)
     }
 }
 
@@ -349,7 +353,7 @@ impl FlushNotStarted {
 impl FlushInProgress {
     /// Waits until background flush is done.
     pub async fn wait_until_flush_is_done(self) -> FlushDone {
-        self.done_flush_rx.await.unwrap();
+        let _ = self.done_flush_rx.await;
         FlushDone
     }
 }

pageserver/src/walingest.rs

@@ -21,13 +21,13 @@
 //! redo Postgres process, but some records it can handle directly with
 //! bespoken Rust code.
 
+use std::backtrace::Backtrace;
 use std::collections::HashMap;
 use std::sync::{Arc, OnceLock};
 use std::time::{Duration, Instant, SystemTime};
 
-use anyhow::{Result, bail};
 use bytes::{Buf, Bytes};
-use pageserver_api::key::rel_block_to_key;
+use pageserver_api::key::{Key, rel_block_to_key};
 use pageserver_api::record::NeonWalRecord;
 use pageserver_api::reltag::{BlockNumber, RelTag, SlruKind};
 use pageserver_api::shard::ShardIdentity;
@@ -38,7 +38,7 @@ use postgres_ffi::{
     fsm_logical_to_physical, pg_constants,
 };
 use tracing::*;
-use utils::bin_ser::SerializeError;
+use utils::bin_ser::{DeserializeError, SerializeError};
 use utils::lsn::Lsn;
 use utils::rate_limit::RateLimit;
 use utils::{critical, failpoint_support};
@@ -104,12 +104,101 @@ struct WarnIngestLag {
     timestamp_invalid_msg_ratelimit: RateLimit,
 }
 
+pub struct WalIngestError {
+    pub backtrace: std::backtrace::Backtrace,
+    pub kind: WalIngestErrorKind,
+}
+
+#[derive(thiserror::Error, Debug)]
+pub enum WalIngestErrorKind {
+    #[error(transparent)]
+    #[allow(private_interfaces)]
+    PageReconstructError(#[from] PageReconstructError),
+    #[error(transparent)]
+    DeserializationFailure(#[from] DeserializeError),
+    #[error(transparent)]
+    SerializationFailure(#[from] SerializeError),
+    #[error("the request contains data not supported by pageserver: {0} @ {1}")]
+    InvalidKey(Key, Lsn),
+    #[error("twophase file for xid {0} already exists")]
+    FileAlreadyExists(u64),
+    #[error("slru segment {0:?}/{1} already exists")]
+    SlruAlreadyExists(SlruKind, u32),
+    #[error("relation already exists")]
+    RelationAlreadyExists(RelTag),
+    #[error("invalid reldir key {0}")]
+    InvalidRelDirKey(Key),
+
+    #[error(transparent)]
+    LogicalError(anyhow::Error),
+    #[error(transparent)]
+    EncodeAuxFileError(anyhow::Error),
+    #[error(transparent)]
+    MaybeRelSizeV2Error(anyhow::Error),
+
+    #[error("timeline shutting down")]
+    Cancelled,
+}
+
+impl<T> From<T> for WalIngestError
+where
+    WalIngestErrorKind: From<T>,
+{
+    fn from(value: T) -> Self {
+        WalIngestError {
+            backtrace: Backtrace::capture(),
+            kind: WalIngestErrorKind::from(value),
+        }
+    }
+}
+
+impl std::error::Error for WalIngestError {
+    fn source(&self) -> Option<&(dyn std::error::Error + 'static)> {
+        self.kind.source()
+    }
+}
+
+impl core::fmt::Display for WalIngestError {
+    fn fmt(&self, f: &mut core::fmt::Formatter<'_>) -> core::fmt::Result {
+        self.kind.fmt(f)
+    }
+}
+
+impl core::fmt::Debug for WalIngestError {
+    fn fmt(&self, f: &mut core::fmt::Formatter<'_>) -> core::fmt::Result {
+        if f.alternate() {
+            f.debug_map()
+                .key(&"backtrace")
+                .value(&self.backtrace)
+                .key(&"kind")
+                .value(&self.kind)
+                .finish()
+        } else {
+            writeln!(f, "Error: {:?}", self.kind)?;
+            if self.backtrace.status() == std::backtrace::BacktraceStatus::Captured {
+                writeln!(f, "Stack backtrace: {:?}", self.backtrace)?;
+            }
+            Ok(())
+        }
+    }
+}
+
+#[macro_export]
+macro_rules! ensure_walingest {
+    ($($t:tt)*) => {
+        _ = || -> Result<(), anyhow::Error> {
+            anyhow::ensure!($($t)*);
+            Ok(())
+        }().map_err(WalIngestErrorKind::LogicalError)?;
+    };
+}
+
 impl WalIngest {
     pub async fn new(
         timeline: &Timeline,
         startpoint: Lsn,
         ctx: &RequestContext,
-    ) -> anyhow::Result<WalIngest> {
+    ) -> Result<WalIngest, WalIngestError> {
         // Fetch the latest checkpoint into memory, so that we can compare with it
         // quickly in `ingest_record` and update it when it changes.
         let checkpoint_bytes = timeline.get_checkpoint(startpoint, ctx).await?;
@@ -145,7 +234,7 @@ impl WalIngest {
         interpreted: InterpretedWalRecord,
         modification: &mut DatadirModification<'_>,
         ctx: &RequestContext,
-    ) -> anyhow::Result<bool> {
+    ) -> Result<bool, WalIngestError> {
         WAL_INGEST.records_received.inc();
         let prev_len = modification.len();
 
@@ -288,7 +377,7 @@ impl WalIngest {
     }
 
     /// This is the same as AdjustToFullTransactionId(xid) in PostgreSQL
-    fn adjust_to_full_transaction_id(&self, xid: TransactionId) -> Result<u64> {
+    fn adjust_to_full_transaction_id(&self, xid: TransactionId) -> Result<u64, WalIngestError> {
         let next_full_xid =
             enum_pgversion_dispatch!(&self.checkpoint, CheckPoint, cp, { cp.nextXid.value });
 
@@ -298,9 +387,9 @@ impl WalIngest {
         if xid > next_xid {
             // Wraparound occurred, must be from a prev epoch.
             if epoch == 0 {
-                bail!(
+                Err(WalIngestErrorKind::LogicalError(anyhow::anyhow!(
                     "apparent XID wraparound with prepared transaction XID {xid}, nextXid is {next_full_xid}"
-                );
+                )))?;
             }
             epoch -= 1;
         }
@@ -313,7 +402,7 @@ impl WalIngest {
         clear_vm_bits: ClearVmBits,
         modification: &mut DatadirModification<'_>,
         ctx: &RequestContext,
-    ) -> anyhow::Result<()> {
+    ) -> Result<(), WalIngestError> {
         let ClearVmBits {
             new_heap_blkno,
             old_heap_blkno,
@@ -402,7 +491,7 @@ impl WalIngest {
         create: DbaseCreate,
         modification: &mut DatadirModification<'_>,
         ctx: &RequestContext,
-    ) -> anyhow::Result<()> {
+    ) -> Result<(), WalIngestError> {
         let DbaseCreate {
             db_id,
             tablespace_id,
@@ -505,7 +594,7 @@ impl WalIngest {
         dbase_drop: DbaseDrop,
         modification: &mut DatadirModification<'_>,
         ctx: &RequestContext,
-    ) -> anyhow::Result<()> {
+    ) -> Result<(), WalIngestError> {
         let DbaseDrop {
             db_id,
             tablespace_ids,
@@ -523,7 +612,7 @@ impl WalIngest {
         create: SmgrCreate,
         modification: &mut DatadirModification<'_>,
         ctx: &RequestContext,
-    ) -> anyhow::Result<()> {
+    ) -> Result<(), WalIngestError> {
         let SmgrCreate { rel } = create;
         self.put_rel_creation(modification, rel, ctx).await?;
         Ok(())
@@ -537,7 +626,7 @@ impl WalIngest {
         truncate: XlSmgrTruncate,
         modification: &mut DatadirModification<'_>,
         ctx: &RequestContext,
-    ) -> anyhow::Result<()> {
+    ) -> Result<(), WalIngestError> {
         let XlSmgrTruncate {
             blkno,
             rnode,
@@ -689,7 +778,7 @@ impl WalIngest {
         record: XactRecord,
         modification: &mut DatadirModification<'_>,
         ctx: &RequestContext,
-    ) -> anyhow::Result<()> {
+    ) -> Result<(), WalIngestError> {
         let (xact_common, is_commit, is_prepared) = match record {
             XactRecord::Prepare(XactPrepare { xl_xid, data }) => {
                 let xid: u64 = if modification.tline.pg_version >= 17 {
@@ -813,7 +902,7 @@ impl WalIngest {
         truncate: ClogTruncate,
         modification: &mut DatadirModification<'_>,
         ctx: &RequestContext,
-    ) -> anyhow::Result<()> {
+    ) -> Result<(), WalIngestError> {
         let ClogTruncate {
             pageno,
             oldest_xid,
@@ -889,7 +978,7 @@ impl WalIngest {
         zero_page: ClogZeroPage,
         modification: &mut DatadirModification<'_>,
         ctx: &RequestContext,
-    ) -> anyhow::Result<()> {
+    ) -> Result<(), WalIngestError> {
         let ClogZeroPage { segno, rpageno } = zero_page;
 
         self.put_slru_page_image(
@@ -907,7 +996,7 @@ impl WalIngest {
         &mut self,
         modification: &mut DatadirModification,
         xlrec: &XlMultiXactCreate,
-    ) -> Result<()> {
+    ) -> Result<(), WalIngestError> {
         // Create WAL record for updating the multixact-offsets page
         let pageno = xlrec.mid / pg_constants::MULTIXACT_OFFSETS_PER_PAGE as u32;
         let segno = pageno / pg_constants::SLRU_PAGES_PER_SEGMENT;
@@ -1010,7 +1099,7 @@ impl WalIngest {
         modification: &mut DatadirModification<'_>,
         xlrec: &XlMultiXactTruncate,
         ctx: &RequestContext,
-    ) -> Result<()> {
+    ) -> Result<(), WalIngestError> {
         let (maxsegment, startsegment, endsegment) =
             enum_pgversion_dispatch!(&mut self.checkpoint, CheckPoint, cp, {
                 cp.oldestMulti = xlrec.end_trunc_off;
@@ -1058,7 +1147,7 @@ impl WalIngest {
         zero_page: MultiXactZeroPage,
         modification: &mut DatadirModification<'_>,
         ctx: &RequestContext,
-    ) -> Result<()> {
+    ) -> Result<(), WalIngestError> {
         let MultiXactZeroPage {
             slru_kind,
             segno,
@@ -1080,7 +1169,7 @@ impl WalIngest {
         update: RelmapUpdate,
         modification: &mut DatadirModification<'_>,
         ctx: &RequestContext,
-    ) -> Result<()> {
+    ) -> Result<(), WalIngestError> {
         let RelmapUpdate { update, buf } = update;
 
         modification
@@ -1093,7 +1182,7 @@ impl WalIngest {
         raw_record: RawXlogRecord,
         modification: &mut DatadirModification<'_>,
         ctx: &RequestContext,
-    ) -> Result<()> {
+    ) -> Result<(), WalIngestError> {
         let RawXlogRecord { info, lsn, mut buf } = raw_record;
         let pg_version = modification.tline.pg_version;
 
@@ -1235,12 +1324,12 @@ impl WalIngest {
         put: PutLogicalMessage,
         modification: &mut DatadirModification<'_>,
         ctx: &RequestContext,
-    ) -> Result<()> {
+    ) -> Result<(), WalIngestError> {
         let PutLogicalMessage { path, buf } = put;
         modification.put_file(path.as_str(), &buf, ctx).await
     }
 
-    fn ingest_standby_record(&mut self, record: StandbyRecord) -> Result<()> {
+    fn ingest_standby_record(&mut self, record: StandbyRecord) -> Result<(), WalIngestError> {
         match record {
             StandbyRecord::RunningXacts(running_xacts) => {
                 enum_pgversion_dispatch!(&mut self.checkpoint, CheckPoint, cp, {
@@ -1258,7 +1347,7 @@ impl WalIngest {
         &mut self,
         record: ReploriginRecord,
         modification: &mut DatadirModification<'_>,
-    ) -> Result<()> {
+    ) -> Result<(), WalIngestError> {
         match record {
             ReploriginRecord::Set(set) => {
                 modification
@@ -1278,7 +1367,7 @@ impl WalIngest {
         modification: &mut DatadirModification<'_>,
         rel: RelTag,
         ctx: &RequestContext,
-    ) -> Result<()> {
+    ) -> Result<(), WalIngestError> {
         modification.put_rel_creation(rel, 0, ctx).await?;
         Ok(())
     }
@@ -1291,7 +1380,7 @@ impl WalIngest {
         blknum: BlockNumber,
         img: Bytes,
         ctx: &RequestContext,
-    ) -> Result<(), PageReconstructError> {
+    ) -> Result<(), WalIngestError> {
         self.handle_rel_extend(modification, rel, blknum, ctx)
             .await?;
         modification.put_rel_page_image(rel, blknum, img)?;
@@ -1305,7 +1394,7 @@ impl WalIngest {
         blknum: BlockNumber,
         rec: NeonWalRecord,
         ctx: &RequestContext,
-    ) -> Result<()> {
+    ) -> Result<(), WalIngestError> {
         self.handle_rel_extend(modification, rel, blknum, ctx)
             .await?;
         modification.put_rel_wal_record(rel, blknum, rec)?;
@@ -1318,7 +1407,7 @@ impl WalIngest {
         rel: RelTag,
         nblocks: BlockNumber,
         ctx: &RequestContext,
-    ) -> anyhow::Result<()> {
+    ) -> Result<(), WalIngestError> {
         modification.put_rel_truncation(rel, nblocks, ctx).await?;
         Ok(())
     }
@@ -1329,7 +1418,7 @@ impl WalIngest {
         rel: RelTag,
         blknum: BlockNumber,
         ctx: &RequestContext,
-    ) -> Result<(), PageReconstructError> {
+    ) -> Result<(), WalIngestError> {
         let new_nblocks = blknum + 1;
         // Check if the relation exists. We implicitly create relations on first
         // record.
@@ -1423,7 +1512,7 @@ impl WalIngest {
         blknum: BlockNumber,
         img: Bytes,
         ctx: &RequestContext,
-    ) -> Result<()> {
+    ) -> Result<(), WalIngestError> {
         if !self.shard.is_shard_zero() {
             return Ok(());
         }
@@ -1441,7 +1530,7 @@ impl WalIngest {
         segno: u32,
         blknum: BlockNumber,
         ctx: &RequestContext,
-    ) -> anyhow::Result<()> {
+    ) -> Result<(), WalIngestError> {
         // we don't use a cache for this like we do for relations. SLRUS are explcitly
         // extended with ZEROPAGE records, not with commit records, so it happens
         // a lot less frequently.
@@ -1509,6 +1598,7 @@ async fn get_relsize(
 #[allow(clippy::bool_assert_comparison)]
 #[cfg(test)]
 mod tests {
+    use anyhow::Result;
     use postgres_ffi::RELSEG_SIZE;
 
     use super::*;
@@ -1530,7 +1620,7 @@ mod tests {
     }
 
     #[tokio::test]
-    async fn test_zeroed_checkpoint_decodes_correctly() -> Result<()> {
+    async fn test_zeroed_checkpoint_decodes_correctly() -> Result<(), anyhow::Error> {
         for i in 14..=16 {
             dispatch_pgversion!(i, {
                 pgv::CheckPoint::decode(&pgv::ZERO_CHECKPOINT)?;

pageserver/src/walredo.rs

@@ -136,6 +136,16 @@ macro_rules! bail {
     }
 }
 
+#[derive(Debug, Clone, Copy)]
+pub enum RedoAttemptType {
+    /// Used for the read path. Will fire critical errors and retry twice if failure.
+    ReadPage,
+    // Used for legacy compaction (only used in image compaction). Will fire critical errors and retry once if failure.
+    LegacyCompaction,
+    // Used for gc compaction. Will not fire critical errors and not retry.
+    GcCompaction,
+}
+
 ///
 /// Public interface of WAL redo manager
 ///
@@ -156,11 +166,18 @@ impl PostgresRedoManager {
         base_img: Option<(Lsn, Bytes)>,
         records: Vec<(Lsn, NeonWalRecord)>,
         pg_version: u32,
+        redo_attempt_type: RedoAttemptType,
     ) -> Result<Bytes, Error> {
         if records.is_empty() {
             bail!("invalid WAL redo request with no records");
         }
 
+        let max_retry_attempts = match redo_attempt_type {
+            RedoAttemptType::ReadPage => 2,
+            RedoAttemptType::LegacyCompaction => 1,
+            RedoAttemptType::GcCompaction => 0,
+        };
+
         let base_img_lsn = base_img.as_ref().map(|p| p.0).unwrap_or(Lsn::INVALID);
         let mut img = base_img.map(|p| p.1);
         let mut batch_neon = apply_neon::can_apply_in_neon(&records[0].1);
@@ -180,6 +197,7 @@ impl PostgresRedoManager {
                         &records[batch_start..i],
                         self.conf.wal_redo_timeout,
                         pg_version,
+                        max_retry_attempts,
                     )
                     .await
                 };
@@ -201,6 +219,7 @@ impl PostgresRedoManager {
                 &records[batch_start..],
                 self.conf.wal_redo_timeout,
                 pg_version,
+                max_retry_attempts,
             )
             .await
         }
@@ -424,11 +443,11 @@ impl PostgresRedoManager {
         records: &[(Lsn, NeonWalRecord)],
         wal_redo_timeout: Duration,
         pg_version: u32,
+        max_retry_attempts: u32,
     ) -> Result<Bytes, Error> {
         *(self.last_redo_at.lock().unwrap()) = Some(Instant::now());
 
         let (rel, blknum) = key.to_rel_block().context("invalid record")?;
-        const MAX_RETRY_ATTEMPTS: u32 = 1;
         let mut n_attempts = 0u32;
         loop {
             let base_img = &base_img;
@@ -486,7 +505,7 @@ impl PostgresRedoManager {
                 info!(n_attempts, "retried walredo succeeded");
             }
             n_attempts += 1;
-            if n_attempts > MAX_RETRY_ATTEMPTS || result.is_ok() {
+            if n_attempts > max_retry_attempts || result.is_ok() {
                 return result;
             }
         }
@@ -560,6 +579,7 @@ mod tests {
 
     use super::PostgresRedoManager;
     use crate::config::PageServerConf;
+    use crate::walredo::RedoAttemptType;
 
     #[tokio::test]
     async fn test_ping() {
@@ -593,6 +613,7 @@ mod tests {
                 None,
                 short_records(),
                 14,
+                RedoAttemptType::ReadPage,
             )
             .instrument(h.span())
             .await
@@ -621,6 +642,7 @@ mod tests {
                 None,
                 short_records(),
                 14,
+                RedoAttemptType::ReadPage,
             )
             .instrument(h.span())
             .await
@@ -642,6 +664,7 @@ mod tests {
                 None,
                 short_records(),
                 16, /* 16 currently produces stderr output on startup, which adds a nice extra edge */
+                RedoAttemptType::ReadPage,
             )
             .instrument(h.span())
             .await

pageserver/src/walredo/apply_neon.rs

@@ -276,6 +276,7 @@ pub(crate) fn apply_in_neon(
             append,
             clear,
             will_init,
+            only_if,
         } => {
             use bytes::BufMut;
             if *will_init {
@@ -288,6 +289,13 @@ pub(crate) fn apply_in_neon(
             if *clear {
                 page.clear();
             }
+            if let Some(only_if) = only_if {
+                if page != only_if.as_bytes() {
+                    return Err(anyhow::anyhow!(
+                        "the current image does not match the expected image, cannot append"
+                    ));
+                }
+            }
             page.put_slice(append.as_bytes());
         }
     }