wip

Monomorphization is basically like macro expansion, it just happens at a
later compiler phase.

.dockerignore

@@ -14,6 +14,7 @@
 !pgxn/
 !proxy/
 !safekeeper/
+!s3_scrubber/
 !storage_broker/
 !trace/
 !vendor/postgres-v14/

Cargo.toml

@@ -7,6 +7,7 @@ members = [
     "proxy",
     "safekeeper",
     "storage_broker",
+    "s3_scrubber",
     "workspace_hack",
     "trace",
     "libs/compute_api",
@@ -37,11 +38,11 @@ async-compression = { version = "0.4.0", features = ["tokio", "gzip"] }
 flate2 = "1.0.26"
 async-stream = "0.3"
 async-trait = "0.1"
-aws-config = { version = "0.55", default-features = false, features=["rustls"] }
-aws-sdk-s3 = "0.27"
-aws-smithy-http = "0.55"
-aws-credential-types = "0.55"
-aws-types = "0.55"
+aws-config = { version = "0.56", default-features = false, features=["rustls"] }
+aws-sdk-s3 = "0.29"
+aws-smithy-http = "0.56"
+aws-credential-types = "0.56"
+aws-types = "0.56"
 axum = { version = "0.6.20", features = ["ws"] }
 base64 = "0.13.0"
 bincode = "1.3"
@@ -105,12 +106,12 @@ reqwest-middleware = "0.2.0"
 reqwest-retry = "0.2.2"
 routerify = "3"
 rpds = "0.13"
-rustls = "0.20"
+rustls = "0.21"
 rustls-pemfile = "1"
 rustls-split = "0.3"
 scopeguard = "1.1"
 sysinfo = "0.29.2"
-sentry = { version = "0.30", default-features = false, features = ["backtrace", "contexts", "panic", "rustls", "reqwest" ] }
+sentry = { version = "0.31", default-features = false, features = ["backtrace", "contexts", "panic", "rustls", "reqwest" ] }
 serde = { version = "1.0", features = ["derive"] }
 serde_json = "1"
 serde_with = "2.0"
@@ -125,11 +126,11 @@ sync_wrapper = "0.1.2"
 tar = "0.4"
 test-context = "0.1"
 thiserror = "1.0"
-tls-listener = { version = "0.6", features = ["rustls", "hyper-h1"] }
+tls-listener = { version = "0.7", features = ["rustls", "hyper-h1"] }
 tokio = { version = "1.17", features = ["macros"] }
 tokio-io-timeout = "1.2.0"
-tokio-postgres-rustls = "0.9.0"
-tokio-rustls = "0.23"
+tokio-postgres-rustls = "0.10.0"
+tokio-rustls = "0.24"
 tokio-stream = "0.1"
 tokio-tar = "0.3"
 tokio-util = { version = "0.7", features = ["io"] }
@@ -143,7 +144,7 @@ tracing-subscriber = { version = "0.3", default_features = false, features = ["s
 url = "2.2"
 uuid = { version = "1.2", features = ["v4", "serde"] }
 walkdir = "2.3.2"
-webpki-roots = "0.23"
+webpki-roots = "0.25"
 x509-parser = "0.15"
 
 ## TODO replace this with tracing
@@ -182,8 +183,8 @@ workspace_hack = { version = "0.1", path = "./workspace_hack/" }
 
 ## Build dependencies
 criterion = "0.5.1"
-rcgen = "0.10"
-rstest = "0.17"
+rcgen = "0.11"
+rstest = "0.18"
 tempfile = "3.4"
 tonic-build = "0.9"
 

compute_tools/src/checker.rs

@@ -1,12 +1,39 @@
-use anyhow::{anyhow, Result};
+use anyhow::{anyhow, Ok, Result};
+use postgres::Client;
 use tokio_postgres::NoTls;
 use tracing::{error, instrument};
 
 use crate::compute::ComputeNode;
 
+/// Create a special service table for availability checks
+/// only if it does not exist already.
+pub fn create_availability_check_data(client: &mut Client) -> Result<()> {
+    let query = "
+        DO $$
+        BEGIN
+            IF NOT EXISTS(
+                SELECT 1
+                FROM pg_catalog.pg_tables
+                WHERE tablename = 'health_check'
+            )
+            THEN
+            CREATE TABLE health_check (
+                id serial primary key,
+                updated_at timestamptz default now()
+            );
+            INSERT INTO health_check VALUES (1, now())
+                ON CONFLICT (id) DO UPDATE
+                 SET updated_at = now();
+            END IF;
+        END
+        $$;";
+    client.execute(query, &[])?;
+
+    Ok(())
+}
+
 /// Update timestamp in a row in a special service table to check
 /// that we can actually write some data in this particular timeline.
-/// Create table if it's missing.
 #[instrument(skip_all)]
 pub async fn check_writability(compute: &ComputeNode) -> Result<()> {
     // Connect to the database.
@@ -24,19 +51,15 @@ pub async fn check_writability(compute: &ComputeNode) -> Result<()> {
     });
 
     let query = "
-    CREATE TABLE IF NOT EXISTS health_check (
-        id serial primary key,
-        updated_at timestamptz default now()
-    );
     INSERT INTO health_check VALUES (1, now())
         ON CONFLICT (id) DO UPDATE
          SET updated_at = now();";
 
     let result = client.simple_query(query).await?;
 
-    if result.len() != 2 {
+    if result.len() != 1 {
         return Err(anyhow::format_err!(
-            "expected 2 query results, but got {}",
+            "expected 1 query result, but got {}",
             result.len()
         ));
     }

compute_tools/src/compute.rs

@@ -27,6 +27,7 @@ use utils::measured_stream::MeasuredReader;
 
 use remote_storage::{DownloadError, GenericRemoteStorage, RemotePath};
 
+use crate::checker::create_availability_check_data;
 use crate::pg_helpers::*;
 use crate::spec::*;
 use crate::sync_sk::{check_if_synced, ping_safekeeper};
@@ -696,6 +697,7 @@ impl ComputeNode {
         handle_role_deletions(spec, self.connstr.as_str(), &mut client)?;
         handle_grants(spec, self.connstr.as_str())?;
         handle_extensions(spec, &mut client)?;
+        create_availability_check_data(&mut client)?;
 
         // 'Close' connection
         drop(client);
@@ -1078,7 +1080,8 @@ LIMIT 100",
 
         let mut download_tasks = Vec::new();
         for library in &libs_vec {
-            let (ext_name, ext_path) = remote_extensions.get_ext(library, true)?;
+            let (ext_name, ext_path) =
+                remote_extensions.get_ext(library, true, &self.build_tag, &self.pgversion)?;
             download_tasks.push(self.download_extension(ext_name, ext_path));
         }
         let results = join_all(download_tasks).await;

compute_tools/src/extension_server.rs

@@ -180,7 +180,19 @@ pub async fn download_extension(
 // Create extension control files from spec
 pub fn create_control_files(remote_extensions: &RemoteExtSpec, pgbin: &str) {
     let local_sharedir = Path::new(&get_pg_config("--sharedir", pgbin)).join("extension");
-    for ext_data in remote_extensions.extension_data.values() {
+    for (ext_name, ext_data) in remote_extensions.extension_data.iter() {
+        // Check if extension is present in public or custom.
+        // If not, then it is not allowed to be used by this compute.
+        if let Some(public_extensions) = &remote_extensions.public_extensions {
+            if !public_extensions.contains(ext_name) {
+                if let Some(custom_extensions) = &remote_extensions.custom_extensions {
+                    if !custom_extensions.contains(ext_name) {
+                        continue; // skip this extension, it is not allowed
+                    }
+                }
+            }
+        }
+
         for (control_name, control_content) in &ext_data.control_data {
             let control_path = local_sharedir.join(control_name);
             if !control_path.exists() {

compute_tools/src/http/api.rs

@@ -169,7 +169,12 @@ async fn routes(req: Request<Body>, compute: &Arc<ComputeNode>) -> Response<Body
                     }
                 };
 
-                remote_extensions.get_ext(&filename, is_library)
+                remote_extensions.get_ext(
+                    &filename,
+                    is_library,
+                    &compute.build_tag,
+                    &compute.pgversion,
+                )
             };
 
             match ext {

control_plane/src/endpoint.rs

@@ -138,7 +138,13 @@ impl ComputeControlPlane {
             mode,
             tenant_id,
             pg_version,
-            skip_pg_catalog_updates: false,
+            // We don't setup roles and databases in the spec locally, so we don't need to
+            // do catalog updates. Catalog updates also include check availability
+            // data creation. Yet, we have tests that check that size and db dump
+            // before and after start are the same. So, skip catalog updates,
+            // with this we basically test a case of waking up an idle compute, where
+            // we also skip catalog updates in the cloud.
+            skip_pg_catalog_updates: true,
         });
 
         ep.create_endpoint_dir()?;
@@ -152,7 +158,7 @@ impl ComputeControlPlane {
                 http_port,
                 pg_port,
                 pg_version,
-                skip_pg_catalog_updates: false,
+                skip_pg_catalog_updates: true,
             })?,
         )?;
         std::fs::write(

libs/compute_api/src/spec.rs

@@ -89,6 +89,8 @@ impl RemoteExtSpec {
         &self,
         ext_name: &str,
         is_library: bool,
+        build_tag: &str,
+        pg_major_version: &str,
     ) -> anyhow::Result<(String, RemotePath)> {
         let mut real_ext_name = ext_name;
         if is_library {
@@ -104,11 +106,32 @@ impl RemoteExtSpec {
                 .ok_or(anyhow::anyhow!("library {} is not found", lib_raw_name))?;
         }
 
+        // Check if extension is present in public or custom.
+        // If not, then it is not allowed to be used by this compute.
+        if let Some(public_extensions) = &self.public_extensions {
+            if !public_extensions.contains(&real_ext_name.to_string()) {
+                if let Some(custom_extensions) = &self.custom_extensions {
+                    if !custom_extensions.contains(&real_ext_name.to_string()) {
+                        return Err(anyhow::anyhow!("extension {} is not found", real_ext_name));
+                    }
+                }
+            }
+        }
+
         match self.extension_data.get(real_ext_name) {
-            Some(ext_data) => Ok((
-                real_ext_name.to_string(),
-                RemotePath::from_string(&ext_data.archive_path)?,
-            )),
+            Some(_ext_data) => {
+                // Construct the path to the extension archive
+                // BUILD_TAG/PG_MAJOR_VERSION/extensions/EXTENSION_NAME.tar.zst
+                //
+                // Keep it in sync with path generation in
+                // https://github.com/neondatabase/build-custom-extensions/tree/main
+                let archive_path_str =
+                    format!("{build_tag}/{pg_major_version}/extensions/{real_ext_name}.tar.zst");
+                Ok((
+                    real_ext_name.to_string(),
+                    RemotePath::from_string(&archive_path_str)?,
+                ))
+            }
             None => Err(anyhow::anyhow!(
                 "real_ext_name {} is not found",
                 real_ext_name

libs/remote_storage/src/lib.rs

@@ -13,14 +13,13 @@ use std::{
     collections::HashMap,
     fmt::Debug,
     num::{NonZeroU32, NonZeroUsize},
-    path::{Path, PathBuf, StripPrefixError},
+    path::{Path, PathBuf},
     pin::Pin,
     sync::Arc,
 };
 
 use anyhow::{bail, Context};
 
-use serde::{Deserialize, Serialize};
 use tokio::io;
 use toml_edit::Item;
 use tracing::info;
@@ -45,34 +44,12 @@ pub const DEFAULT_MAX_KEYS_PER_LIST_RESPONSE: Option<i32> = None;
 
 const REMOTE_STORAGE_PREFIX_SEPARATOR: char = '/';
 
-// From the S3 spec
-pub const MAX_KEYS_PER_DELETE: usize = 1000;
-
 /// Path on the remote storage, relative to some inner prefix.
 /// The prefix is an implementation detail, that allows representing local paths
 /// as the remote ones, stripping the local storage prefix away.
 #[derive(Debug, Clone, PartialEq, Eq, PartialOrd, Ord, Hash)]
 pub struct RemotePath(PathBuf);
 
-impl Serialize for RemotePath {
-    fn serialize<S>(&self, serializer: S) -> Result<S::Ok, S::Error>
-    where
-        S: serde::Serializer,
-    {
-        serializer.collect_str(self)
-    }
-}
-
-impl<'de> Deserialize<'de> for RemotePath {
-    fn deserialize<D>(deserializer: D) -> Result<Self, D::Error>
-    where
-        D: serde::Deserializer<'de>,
-    {
-        let str = String::deserialize(deserializer)?;
-        Ok(Self(PathBuf::from(&str)))
-    }
-}
-
 impl std::fmt::Display for RemotePath {
     fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
         write!(f, "{}", self.0.display())
@@ -111,10 +88,6 @@ impl RemotePath {
     pub fn extension(&self) -> Option<&str> {
         self.0.extension()?.to_str()
     }
-
-    pub fn strip_prefix(&self, p: &RemotePath) -> Result<&Path, StripPrefixError> {
-        self.0.strip_prefix(&p.0)
-    }
 }
 
 /// Storage (potentially remote) API to manage its state.
@@ -193,8 +166,6 @@ pub enum DownloadError {
     BadInput(anyhow::Error),
     /// The file was not found in the remote storage.
     NotFound,
-    /// The client was shut down
-    Shutdown,
     /// The file was found in the remote storage, but the download failed.
     Other(anyhow::Error),
 }
@@ -206,7 +177,6 @@ impl std::fmt::Display for DownloadError {
                 write!(f, "Failed to download a remote file due to user input: {e}")
             }
             DownloadError::NotFound => write!(f, "No file found for the remote object id given"),
-            DownloadError::Shutdown => write!(f, "Client shutting down"),
             DownloadError::Other(e) => write!(f, "Failed to download a remote file: {e:?}"),
         }
     }
@@ -271,18 +241,6 @@ impl GenericRemoteStorage {
         }
     }
 
-    /// For small, simple downloads where caller doesn't want to handle the streaming: return the full body
-    pub async fn download_all(&self, from: &RemotePath) -> Result<Vec<u8>, DownloadError> {
-        let mut download = self.download(from).await?;
-
-        let mut bytes = Vec::new();
-        tokio::io::copy(&mut download.download_stream, &mut bytes)
-            .await
-            .with_context(|| format!("Failed to download body from {from}"))
-            .map_err(DownloadError::Other)?;
-        Ok(bytes)
-    }
-
     pub async fn download_byte_range(
         &self,
         from: &RemotePath,

libs/remote_storage/src/s3_bucket.rs

@@ -22,7 +22,7 @@ use aws_sdk_s3::{
     Client,
 };
 use aws_smithy_http::body::SdkBody;
-use hyper::{Body, StatusCode};
+use hyper::Body;
 use scopeguard::ScopeGuard;
 use tokio::{
     io::{self, AsyncRead},
@@ -529,16 +529,7 @@ impl RemoteStorage for S3Bucket {
                     }
                 }
                 Err(e) => {
-                    if let Some(r) = e.raw_response() {
-                        if r.http().status() == StatusCode::NOT_FOUND {
-                            // 404 is acceptable for deletions.  AWS S3 does not return this, but
-                            // some other implementations might (e.g. GCS XML API returns 404 on DeleteObject
-                            // to a missing key)
-                            continue;
-                        } else {
-                            return Err(anyhow::format_err!("DeleteObjects response error: {e}"));
-                        }
-                    }
+                    return Err(e.into());
                 }
             }
         }

libs/utils/src/generation.rs

@@ -0,0 +1,113 @@
+use std::fmt::Debug;
+
+use serde::{Deserialize, Serialize};
+
+/// Tenant generations are used to provide split-brain safety and allow
+/// multiple pageservers to attach the same tenant concurrently.
+///
+/// See docs/rfcs/025-generation-numbers.md for detail on how generation
+/// numbers are used.
+#[derive(Copy, Clone, Eq, PartialEq, PartialOrd, Ord)]
+pub enum Generation {
+    // Generations with this magic value will not add a suffix to S3 keys, and will not
+    // be included in persisted index_part.json.  This value is only to be used
+    // during migration from pre-generation metadata to generation-aware metadata,
+    // and should eventually go away.
+    //
+    // A special Generation is used rather than always wrapping Generation in an Option,
+    // so that code handling generations doesn't have to be aware of the legacy
+    // case everywhere it touches a generation.
+    None,
+    // Generations with this magic value may never be used to construct S3 keys:
+    // we will panic if someone tries to.  This is for Tenants in the "Broken" state,
+    // so that we can satisfy their constructor with a Generation without risking
+    // a code bug using it in an S3 write (broken tenants should never write)
+    Broken,
+    Valid(u32),
+}
+
+/// The Generation type represents a number associated with a Tenant, which
+/// increments every time the tenant is attached to a new pageserver, or
+/// an attached pageserver restarts.
+///
+/// It is included as a suffix in S3 keys, as a protection against split-brain
+/// scenarios where pageservers might otherwise issue conflicting writes to
+/// remote storage
+impl Generation {
+    /// Create a new Generation that represents a legacy key format with
+    /// no generation suffix
+    pub fn none() -> Self {
+        Self::None
+    }
+
+    // Create a new generation that will panic if you try to use get_suffix
+    pub fn broken() -> Self {
+        Self::Broken
+    }
+
+    pub fn new(v: u32) -> Self {
+        Self::Valid(v)
+    }
+
+    pub fn is_none(&self) -> bool {
+        matches!(self, Self::None)
+    }
+
+    pub fn get_suffix(&self) -> String {
+        match self {
+            Self::Valid(v) => {
+                format!("-{:08x}", v)
+            }
+            Self::None => "".into(),
+            Self::Broken => {
+                panic!("Tried to use a broken generation");
+            }
+        }
+    }
+}
+
+impl Serialize for Generation {
+    fn serialize<S>(&self, serializer: S) -> Result<S::Ok, S::Error>
+    where
+        S: serde::Serializer,
+    {
+        if let Self::Valid(v) = self {
+            v.serialize(serializer)
+        } else {
+            // We should never be asked to serialize a None or Broken.  Structures
+            // that include an optional generation should convert None to an
+            // Option<Generation>::None
+            Err(serde::ser::Error::custom(
+                "Tried to serialize invalid generation ({self})",
+            ))
+        }
+    }
+}
+
+impl<'de> Deserialize<'de> for Generation {
+    fn deserialize<D>(deserializer: D) -> Result<Self, D::Error>
+    where
+        D: serde::Deserializer<'de>,
+    {
+        Ok(Self::Valid(u32::deserialize(deserializer)?))
+    }
+}
+
+// We intentionally do not implement Display for Generation, to reduce the
+// risk of a bug where the generation is used in a format!() string directly
+// instead of using get_suffix().
+impl Debug for Generation {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        match self {
+            Self::Valid(v) => {
+                write!(f, "{:08x}", v)
+            }
+            Self::None => {
+                write!(f, "<none>")
+            }
+            Self::Broken => {
+                write!(f, "<broken>")
+            }
+        }
+    }
+}

libs/utils/src/http/error.rs

@@ -24,9 +24,6 @@ pub enum ApiError {
     #[error("Precondition failed: {0}")]
     PreconditionFailed(Box<str>),
 
-    #[error("Shutting down")]
-    ShuttingDown,
-
     #[error(transparent)]
     InternalServerError(anyhow::Error),
 }
@@ -55,10 +52,6 @@ impl ApiError {
                 self.to_string(),
                 StatusCode::PRECONDITION_FAILED,
             ),
-            ApiError::ShuttingDown => HttpErrorBody::response_from_msg_and_status(
-                "Shutting down".to_string(),
-                StatusCode::SERVICE_UNAVAILABLE,
-            ),
             ApiError::InternalServerError(err) => HttpErrorBody::response_from_msg_and_status(
                 err.to_string(),
                 StatusCode::INTERNAL_SERVER_ERROR,

libs/utils/src/id.rs

@@ -244,13 +244,13 @@ id_newtype!(TenantId);
 /// NOTE: It (de)serializes as an array of hex bytes, so the string representation would look
 /// like `[173,80,132,115,129,226,72,254,170,201,135,108,199,26,228,24]`.
 /// See [`Id`] for alternative ways to serialize it.
-#[derive(Clone, Copy, PartialEq, Eq, Hash, PartialOrd, Ord)]
+#[derive(Clone, Copy, PartialEq, Eq, Hash, Serialize, Deserialize, PartialOrd, Ord)]
 pub struct ConnectionId(Id);
 
 id_newtype!(ConnectionId);
 
 // A pair uniquely identifying Neon instance.
-#[derive(Debug, Clone, Copy, PartialOrd, Ord, PartialEq, Eq, Hash)]
+#[derive(Debug, Clone, Copy, PartialOrd, Ord, PartialEq, Eq, Hash, Serialize, Deserialize)]
 pub struct TenantTimelineId {
     pub tenant_id: TenantId,
     pub timeline_id: TimelineId,
@@ -273,36 +273,6 @@ impl TenantTimelineId {
     }
 }
 
-impl Serialize for TenantTimelineId {
-    fn serialize<S>(&self, serializer: S) -> Result<S::Ok, S::Error>
-    where
-        S: serde::Serializer,
-    {
-        serializer.collect_str(self)
-    }
-}
-
-impl<'de> Deserialize<'de> for TenantTimelineId {
-    fn deserialize<D>(deserializer: D) -> Result<Self, D::Error>
-    where
-        D: serde::Deserializer<'de>,
-    {
-        let str = String::deserialize(deserializer)?;
-        if let Some((tenant_part, timeline_part)) = str.split_once('/') {
-            Ok(Self {
-                tenant_id: TenantId(Id::from_hex(tenant_part).map_err(|e| {
-                    serde::de::Error::custom(format!("Malformed tenant in TenantTimelineId: {e}"))
-                })?),
-                timeline_id: TimelineId(Id::from_hex(timeline_part).map_err(|e| {
-                    serde::de::Error::custom(format!("Malformed timeline in TenantTimelineId {e}"))
-                })?),
-            })
-        } else {
-            Err(serde::de::Error::custom("Malformed TenantTimelineId"))
-        }
-    }
-}
-
 impl fmt::Display for TenantTimelineId {
     fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
         write!(f, "{}/{}", self.tenant_id, self.timeline_id)

libs/utils/src/lib.rs

@@ -27,6 +27,9 @@ pub mod id;
 // http endpoint utils
 pub mod http;
 
+// definition of the Generation type for pageserver attachment APIs
+pub mod generation;
+
 // common log initialisation routine
 pub mod logging;
 

pageserver/src/bin/pageserver.rs

@@ -2,14 +2,12 @@
 
 use std::env::{var, VarError};
 use std::sync::Arc;
-use std::time::Duration;
 use std::{env, ops::ControlFlow, path::Path, str::FromStr};
 
 use anyhow::{anyhow, Context};
 use clap::{Arg, ArgAction, Command};
 
 use metrics::launch_timestamp::{set_launch_timestamp_metric, LaunchTimestamp};
-use pageserver::deletion_queue::{DeletionQueue, DeletionQueueError};
 use pageserver::disk_usage_eviction_task::{self, launch_disk_usage_global_eviction_task};
 use pageserver::metrics::{STARTUP_DURATION, STARTUP_IS_LOADING};
 use pageserver::task_mgr::WALRECEIVER_RUNTIME;
@@ -351,35 +349,6 @@ fn start_pageserver(
     // Set up remote storage client
     let remote_storage = create_remote_storage_client(conf)?;
 
-    // Set up deletion queue
-    let deletion_queue_cancel = tokio_util::sync::CancellationToken::new();
-    let (deletion_queue, deletion_frontend, deletion_backend, deletion_executor) =
-        DeletionQueue::new(remote_storage.clone(), conf, deletion_queue_cancel.clone());
-    if let Some(mut deletion_frontend) = deletion_frontend {
-        BACKGROUND_RUNTIME.spawn(async move {
-            deletion_frontend
-                .background()
-                .instrument(info_span!(parent:None, "deletion frontend"))
-                .await
-        });
-    }
-    if let Some(mut deletion_backend) = deletion_backend {
-        BACKGROUND_RUNTIME.spawn(async move {
-            deletion_backend
-                .background()
-                .instrument(info_span!(parent: None, "deletion backend"))
-                .await
-        });
-    }
-    if let Some(mut deletion_executor) = deletion_executor {
-        BACKGROUND_RUNTIME.spawn(async move {
-            deletion_executor
-                .background()
-                .instrument(info_span!(parent: None, "deletion executor"))
-                .await
-        });
-    }
-
     // Up to this point no significant I/O has been done: this should have been fast.  Record
     // duration prior to starting I/O intensive phase of startup.
     startup_checkpoint("initial", "Starting loading tenants");
@@ -417,7 +386,6 @@ fn start_pageserver(
         TenantSharedResources {
             broker_client: broker_client.clone(),
             remote_storage: remote_storage.clone(),
-            deletion_queue_client: deletion_queue.new_client(),
         },
         order,
     ))?;
@@ -514,7 +482,6 @@ fn start_pageserver(
             http_auth,
             broker_client.clone(),
             remote_storage,
-            deletion_queue.clone(),
             disk_usage_eviction_state,
         )?
         .build()
@@ -637,36 +604,6 @@ fn start_pageserver(
             // The plan is to change that over time.
             shutdown_pageserver.take();
             BACKGROUND_RUNTIME.block_on(pageserver::shutdown_pageserver(0));
-
-            // Best effort to persist any outstanding deletions, to avoid leaking objects
-            let dq = deletion_queue.clone();
-            BACKGROUND_RUNTIME.block_on(async move {
-                match tokio::time::timeout(Duration::from_secs(5), dq.new_client().flush()).await {
-                    Ok(flush_r) => {
-                        match flush_r {
-                            Ok(()) => {
-                                info!("Deletion queue flushed successfully on shutdown")
-                            }
-                            Err(e) => {
-                                match e {
-                                    DeletionQueueError::ShuttingDown => {
-                                        // This is not harmful for correctness, but is unexpected: the deletion
-                                        // queue's workers should stay alive as long as there are any client handles instantiated.
-                                        warn!("Deletion queue stopped prematurely");
-                                    }
-                                }
-                            }
-                        }
-                    }
-                    Err(e) => {
-                        warn!("Timed out flushing deletion queue on shutdown ({e})")
-                    }
-                }
-            });
-
-            // Clean shutdown of deletion queue workers
-            deletion_queue_cancel.cancel();
-
             unreachable!()
         }
     })

pageserver/src/config.rs

@@ -566,27 +566,6 @@ impl PageServerConf {
         self.workdir.join("tenants")
     }
 
-    pub fn deletion_prefix(&self) -> PathBuf {
-        self.workdir.join("deletion")
-    }
-
-    pub fn deletion_list_path(&self, sequence: u64) -> PathBuf {
-        // Encode a version in the filename, so that if we ever switch away from JSON we can
-        // increment this.
-        const VERSION: u8 = 1;
-
-        self.deletion_prefix()
-            .join(format!("{sequence:016x}-{VERSION:02x}.list"))
-    }
-
-    pub fn deletion_header_path(&self) -> PathBuf {
-        // Encode a version in the filename, so that if we ever switch away from JSON we can
-        // increment this.
-        const VERSION: u8 = 1;
-
-        self.deletion_prefix().join(format!("header-{VERSION:02x}"))
-    }
-
     pub fn tenant_path(&self, tenant_id: &TenantId) -> PathBuf {
         self.tenants_path().join(tenant_id.to_string())
     }
@@ -664,23 +643,6 @@ impl PageServerConf {
             .join(METADATA_FILE_NAME)
     }
 
-    /// Files on the remote storage are stored with paths, relative to the workdir.
-    /// That path includes in itself both tenant and timeline ids, allowing to have a unique remote storage path.
-    ///
-    /// Errors if the path provided does not start from pageserver's workdir.
-    pub fn remote_path(&self, local_path: &Path) -> anyhow::Result<RemotePath> {
-        local_path
-            .strip_prefix(&self.workdir)
-            .context("Failed to strip workdir prefix")
-            .and_then(RemotePath::new)
-            .with_context(|| {
-                format!(
-                    "Failed to resolve remote part of path {:?} for base {:?}",
-                    local_path, self.workdir
-                )
-            })
-    }
-
     /// Turns storage remote path of a file into its local path.
     pub fn local_path(&self, remote_path: &RemotePath) -> PathBuf {
         remote_path.with_base(&self.workdir)

pageserver/src/deletion_queue.rs

@@ -1,782 +0,0 @@
-mod backend;
-mod executor;
-mod frontend;
-
-use std::collections::HashMap;
-
-use crate::metrics::DELETION_QUEUE_SUBMITTED;
-use remote_storage::{GenericRemoteStorage, RemotePath};
-use serde::Deserialize;
-use serde::Serialize;
-use serde_with::serde_as;
-use thiserror::Error;
-use tokio;
-use tokio_util::sync::CancellationToken;
-use tracing::{self, debug, error};
-use utils::id::{TenantId, TenantTimelineId, TimelineId};
-
-pub(crate) use self::backend::BackendQueueWorker;
-use self::executor::ExecutorWorker;
-use self::frontend::DeletionOp;
-pub(crate) use self::frontend::FrontendQueueWorker;
-use backend::BackendQueueMessage;
-use executor::ExecutorMessage;
-use frontend::FrontendQueueMessage;
-
-use crate::{config::PageServerConf, tenant::storage_layer::LayerFileName};
-
-// TODO: adminstrative "panic button" config property to disable all deletions
-// TODO: configurable for how long to wait before executing deletions
-
-/// We aggregate object deletions from many tenants in one place, for several reasons:
-/// - Coalesce deletions into fewer DeleteObjects calls
-/// - Enable Tenant/Timeline lifetimes to be shorter than the time it takes
-///   to flush any outstanding deletions.
-/// - Globally control throughput of deletions, as these are a low priority task: do
-///   not compete with the same S3 clients/connections used for higher priority uploads.
-/// - Future: enable validating that we may do deletions in a multi-attached scenario,
-///   via generation numbers (see https://github.com/neondatabase/neon/pull/4919)
-///
-/// There are two kinds of deletion: deferred and immediate.  A deferred deletion
-/// may be intentionally delayed to protect passive readers of S3 data, and may
-/// be subject to a generation number validation step.  An immediate deletion is
-/// ready to execute immediately, and is only queued up so that it can be coalesced
-/// with other deletions in flight.
-///
-/// Deferred deletions pass through three steps:
-/// - Frontend: accumulate deletion requests from Timelines, and batch them up into
-///   DeletionLists, which are persisted to S3.
-/// - Backend: accumulate deletion lists, and validate them en-masse prior to passing
-///   the keys in the list onward for actual deletion
-/// - Executor: accumulate object keys that the backend has validated for immediate
-///   deletion, and execute them in batches of 1000 keys via DeleteObjects.
-///
-/// Non-deferred deletions, such as during timeline deletion, bypass the first
-/// two stages and are passed straight into the Executor.
-///
-/// Internally, each stage is joined by a channel to the next.  In S3, there is only
-/// one queue (of DeletionLists), which is written by the frontend and consumed
-/// by the backend.
-#[derive(Clone)]
-pub struct DeletionQueue {
-    client: DeletionQueueClient,
-}
-
-#[derive(Debug)]
-struct FlushOp {
-    tx: tokio::sync::oneshot::Sender<()>,
-}
-
-impl FlushOp {
-    fn fire(self) {
-        if self.tx.send(()).is_err() {
-            // oneshot channel closed. This is legal: a client could be destroyed while waiting for a flush.
-            debug!("deletion queue flush from dropped client");
-        };
-    }
-}
-
-#[derive(Clone)]
-pub struct DeletionQueueClient {
-    tx: tokio::sync::mpsc::Sender<FrontendQueueMessage>,
-    executor_tx: tokio::sync::mpsc::Sender<ExecutorMessage>,
-}
-
-#[derive(Debug, Serialize, Deserialize)]
-struct TimelineDeletionList {
-    objects: Vec<RemotePath>,
-    // TODO: Tenant attachment generation will go here
-    // (see https://github.com/neondatabase/neon/pull/4919)
-    // attach_gen: u32,
-}
-
-#[serde_as]
-#[derive(Debug, Serialize, Deserialize)]
-struct DeletionList {
-    /// Serialization version, for future use
-    version: u8,
-
-    /// Used for constructing a unique key for each deletion list we write out.
-    sequence: u64,
-
-    /// To avoid repeating tenant/timeline IDs in every key, we store keys in
-    /// nested HashMaps by TenantTimelineID
-    objects: HashMap<TenantTimelineId, TimelineDeletionList>,
-    // TODO: Node generation will go here
-    // (see https://github.com/neondatabase/neon/pull/4919)
-    // node_gen: u32,
-}
-
-#[serde_as]
-#[derive(Debug, Serialize, Deserialize)]
-struct DeletionHeader {
-    /// Serialization version, for future use
-    version: u8,
-
-    /// Enable determining the next sequence number even if there are no deletion lists present.
-    /// If there _are_ deletion lists present, then their sequence numbers take precedence over
-    /// this.
-    last_deleted_list_seq: u64,
-    // TODO: this is where we will track a 'clean' sequence number that indicates all deletion
-    // lists <= that sequence have had their generations validated with the control plane
-    // and are OK to execute.
-}
-
-impl DeletionHeader {
-    const VERSION_LATEST: u8 = 1;
-
-    fn new(last_deleted_list_seq: u64) -> Self {
-        Self {
-            version: Self::VERSION_LATEST,
-            last_deleted_list_seq,
-        }
-    }
-}
-
-impl DeletionList {
-    const VERSION_LATEST: u8 = 1;
-    fn new(sequence: u64) -> Self {
-        Self {
-            version: Self::VERSION_LATEST,
-            sequence,
-            objects: HashMap::new(),
-        }
-    }
-
-    fn is_empty(&self) -> bool {
-        self.objects.is_empty()
-    }
-
-    fn len(&self) -> usize {
-        self.objects.values().map(|v| v.objects.len()).sum()
-    }
-
-    fn push(&mut self, tenant: &TenantId, timeline: &TimelineId, mut objects: Vec<RemotePath>) {
-        if objects.is_empty() {
-            // Avoid inserting an empty TimelineDeletionList: this preserves the property
-            // that if we have no keys, then self.objects is empty (used in Self::is_empty)
-            return;
-        }
-
-        let key = TenantTimelineId::new(*tenant, *timeline);
-        let entry = self
-            .objects
-            .entry(key)
-            .or_insert_with(|| TimelineDeletionList {
-                objects: Vec::new(),
-            });
-        entry.objects.append(&mut objects)
-    }
-
-    fn take_paths(&mut self) -> Vec<RemotePath> {
-        self.objects
-            .drain()
-            .flat_map(|(_k, v)| v.objects.into_iter())
-            .collect()
-    }
-}
-
-#[derive(Error, Debug)]
-pub enum DeletionQueueError {
-    #[error("Deletion queue unavailable during shutdown")]
-    ShuttingDown,
-}
-
-impl DeletionQueueClient {
-    async fn do_push(&self, msg: FrontendQueueMessage) -> Result<(), DeletionQueueError> {
-        match self.tx.send(msg).await {
-            Ok(_) => Ok(()),
-            Err(e) => {
-                // This shouldn't happen, we should shut down all tenants before
-                // we shut down the global delete queue.  If we encounter a bug like this,
-                // we may leak objects as deletions won't be processed.
-                error!("Deletion queue closed while pushing, shutting down? ({e})");
-                Err(DeletionQueueError::ShuttingDown)
-            }
-        }
-    }
-
-    /// Submit a list of layers for deletion: this function will return before the deletion is
-    /// persistent, but it may be executed at any time after this function enters: do not push
-    /// layers until you're sure they can be deleted safely (i.e. remote metadata no longer
-    /// references them).
-    pub(crate) async fn push_layers(
-        &self,
-        tenant_id: TenantId,
-        timeline_id: TimelineId,
-        layers: Vec<LayerFileName>,
-    ) -> Result<(), DeletionQueueError> {
-        DELETION_QUEUE_SUBMITTED.inc_by(layers.len() as u64);
-        self.do_push(FrontendQueueMessage::Delete(DeletionOp {
-            tenant_id,
-            timeline_id,
-            layers,
-            objects: Vec::new(),
-        }))
-        .await
-    }
-
-    async fn do_flush(
-        &self,
-        msg: FrontendQueueMessage,
-        rx: tokio::sync::oneshot::Receiver<()>,
-    ) -> Result<(), DeletionQueueError> {
-        self.do_push(msg).await?;
-        if rx.await.is_err() {
-            // This shouldn't happen if tenants are shut down before deletion queue.  If we
-            // encounter a bug like this, then a flusher will incorrectly believe it has flushed
-            // when it hasn't, possibly leading to leaking objects.
-            error!("Deletion queue dropped flush op while client was still waiting");
-            Err(DeletionQueueError::ShuttingDown)
-        } else {
-            Ok(())
-        }
-    }
-
-    /// Wait until all previous deletions are persistent (either executed, or written to a DeletionList)
-    pub async fn flush(&self) -> Result<(), DeletionQueueError> {
-        let (tx, rx) = tokio::sync::oneshot::channel::<()>();
-        self.do_flush(FrontendQueueMessage::Flush(FlushOp { tx }), rx)
-            .await
-    }
-
-    // Wait until all previous deletions are executed
-    pub(crate) async fn flush_execute(&self) -> Result<(), DeletionQueueError> {
-        debug!("flush_execute: flushing to deletion lists...");
-        // Flush any buffered work to deletion lists
-        self.flush().await?;
-
-        // Flush execution of deletion lists
-        let (tx, rx) = tokio::sync::oneshot::channel::<()>();
-        debug!("flush_execute: flushing execution...");
-        self.do_flush(FrontendQueueMessage::FlushExecute(FlushOp { tx }), rx)
-            .await?;
-        debug!("flush_execute: finished flushing execution...");
-        Ok(())
-    }
-
-    /// This interface bypasses the persistent deletion queue, and any validation
-    /// that this pageserver is still elegible to execute the deletions.  It is for
-    /// use in timeline deletions, where the control plane is telling us we may
-    /// delete everything in the timeline.
-    ///
-    /// DO NOT USE THIS FROM GC OR COMPACTION CODE.  Use the regular `push_layers`.
-    pub(crate) async fn push_immediate(
-        &self,
-        objects: Vec<RemotePath>,
-    ) -> Result<(), DeletionQueueError> {
-        self.executor_tx
-            .send(ExecutorMessage::Delete(objects))
-            .await
-            .map_err(|_| DeletionQueueError::ShuttingDown)
-    }
-
-    /// Companion to push_immediate.  When this returns Ok, all prior objects sent
-    /// into push_immediate have been deleted from remote storage.
-    pub(crate) async fn flush_immediate(&self) -> Result<(), DeletionQueueError> {
-        let (tx, rx) = tokio::sync::oneshot::channel::<()>();
-        self.executor_tx
-            .send(ExecutorMessage::Flush(FlushOp { tx }))
-            .await
-            .map_err(|_| DeletionQueueError::ShuttingDown)?;
-
-        rx.await.map_err(|_| DeletionQueueError::ShuttingDown)
-    }
-}
-
-impl DeletionQueue {
-    pub fn new_client(&self) -> DeletionQueueClient {
-        self.client.clone()
-    }
-
-    /// Caller may use the returned object to construct clients with new_client.
-    /// Caller should tokio::spawn the background() members of the two worker objects returned:
-    /// we don't spawn those inside new() so that the caller can use their runtime/spans of choice.
-    ///
-    /// If remote_storage is None, then the returned workers will also be None.
-    pub fn new(
-        remote_storage: Option<GenericRemoteStorage>,
-        conf: &'static PageServerConf,
-        cancel: CancellationToken,
-    ) -> (
-        Self,
-        Option<FrontendQueueWorker>,
-        Option<BackendQueueWorker>,
-        Option<ExecutorWorker>,
-    ) {
-        // Deep channel: it consumes deletions from all timelines and we do not want to block them
-        let (tx, rx) = tokio::sync::mpsc::channel(16384);
-
-        // Shallow channel: it carries DeletionLists which each contain up to thousands of deletions
-        let (backend_tx, backend_rx) = tokio::sync::mpsc::channel(16);
-
-        // Shallow channel: it carries lists of paths, and we expect the main queueing to
-        // happen in the backend (persistent), not in this queue.
-        let (executor_tx, executor_rx) = tokio::sync::mpsc::channel(16);
-
-        let remote_storage = match remote_storage {
-            None => {
-                return (
-                    Self {
-                        client: DeletionQueueClient { tx, executor_tx },
-                    },
-                    None,
-                    None,
-                    None,
-                )
-            }
-            Some(r) => r,
-        };
-
-        (
-            Self {
-                client: DeletionQueueClient {
-                    tx,
-                    executor_tx: executor_tx.clone(),
-                },
-            },
-            Some(FrontendQueueWorker::new(
-                conf,
-                rx,
-                backend_tx,
-                cancel.clone(),
-            )),
-            Some(BackendQueueWorker::new(conf, backend_rx, executor_tx)),
-            Some(ExecutorWorker::new(
-                remote_storage,
-                executor_rx,
-                cancel.clone(),
-            )),
-        )
-    }
-}
-
-#[cfg(test)]
-mod test {
-    use hex_literal::hex;
-    use std::{
-        io::ErrorKind,
-        path::{Path, PathBuf},
-    };
-    use tracing::info;
-
-    use remote_storage::{RemoteStorageConfig, RemoteStorageKind};
-    use tokio::{runtime::EnterGuard, task::JoinHandle};
-
-    use crate::tenant::harness::TenantHarness;
-
-    use super::*;
-    pub const TIMELINE_ID: TimelineId =
-        TimelineId::from_array(hex!("11223344556677881122334455667788"));
-
-    struct TestSetup {
-        runtime: &'static tokio::runtime::Runtime,
-        _entered_runtime: EnterGuard<'static>,
-        harness: TenantHarness,
-        remote_fs_dir: PathBuf,
-        storage: GenericRemoteStorage,
-        deletion_queue: DeletionQueue,
-        fe_worker: JoinHandle<()>,
-        be_worker: JoinHandle<()>,
-        ex_worker: JoinHandle<()>,
-    }
-
-    impl TestSetup {
-        /// Simulate a pageserver restart by destroying and recreating the deletion queue
-        fn restart(&mut self) {
-            let (deletion_queue, fe_worker, be_worker, ex_worker) = DeletionQueue::new(
-                Some(self.storage.clone()),
-                self.harness.conf,
-                CancellationToken::new(),
-            );
-
-            self.deletion_queue = deletion_queue;
-
-            let mut fe_worker = fe_worker.unwrap();
-            let mut be_worker = be_worker.unwrap();
-            let mut ex_worker = ex_worker.unwrap();
-            let mut fe_worker = self
-                .runtime
-                .spawn(async move { fe_worker.background().await });
-            let mut be_worker = self
-                .runtime
-                .spawn(async move { be_worker.background().await });
-            let mut ex_worker = self.runtime.spawn(async move {
-                drop(ex_worker.background().await);
-            });
-            std::mem::swap(&mut self.fe_worker, &mut fe_worker);
-            std::mem::swap(&mut self.be_worker, &mut be_worker);
-            std::mem::swap(&mut self.ex_worker, &mut ex_worker);
-
-            // Join the old workers
-            self.runtime.block_on(fe_worker).unwrap();
-            self.runtime.block_on(be_worker).unwrap();
-            self.runtime.block_on(ex_worker).unwrap();
-        }
-    }
-
-    fn setup(test_name: &str) -> anyhow::Result<TestSetup> {
-        let test_name = Box::leak(Box::new(format!("deletion_queue__{test_name}")));
-        let harness = TenantHarness::create(test_name)?;
-
-        // We do not load() the harness: we only need its config and remote_storage
-
-        // Set up a GenericRemoteStorage targetting a directory
-        let remote_fs_dir = harness.conf.workdir.join("remote_fs");
-        std::fs::create_dir_all(remote_fs_dir)?;
-        let remote_fs_dir = std::fs::canonicalize(harness.conf.workdir.join("remote_fs"))?;
-        let storage_config = RemoteStorageConfig {
-            max_concurrent_syncs: std::num::NonZeroUsize::new(
-                remote_storage::DEFAULT_REMOTE_STORAGE_MAX_CONCURRENT_SYNCS,
-            )
-            .unwrap(),
-            max_sync_errors: std::num::NonZeroU32::new(
-                remote_storage::DEFAULT_REMOTE_STORAGE_MAX_SYNC_ERRORS,
-            )
-            .unwrap(),
-            storage: RemoteStorageKind::LocalFs(remote_fs_dir.clone()),
-        };
-        let storage = GenericRemoteStorage::from_config(&storage_config).unwrap();
-
-        let runtime = Box::leak(Box::new(
-            tokio::runtime::Builder::new_current_thread()
-                .enable_all()
-                .build()?,
-        ));
-        let entered_runtime = runtime.enter();
-
-        let (deletion_queue, fe_worker, be_worker, ex_worker) = DeletionQueue::new(
-            Some(storage.clone()),
-            harness.conf,
-            CancellationToken::new(),
-        );
-
-        let mut fe_worker = fe_worker.unwrap();
-        let mut be_worker = be_worker.unwrap();
-        let mut ex_worker = ex_worker.unwrap();
-        let fe_worker_join = runtime.spawn(async move { fe_worker.background().await });
-        let be_worker_join = runtime.spawn(async move { be_worker.background().await });
-        let ex_worker_join = runtime.spawn(async move {
-            drop(ex_worker.background().await);
-        });
-
-        Ok(TestSetup {
-            runtime,
-            _entered_runtime: entered_runtime,
-            harness,
-            remote_fs_dir,
-            storage,
-            deletion_queue,
-            fe_worker: fe_worker_join,
-            be_worker: be_worker_join,
-            ex_worker: ex_worker_join,
-        })
-    }
-
-    // TODO: put this in a common location so that we can share with remote_timeline_client's tests
-    fn assert_remote_files(expected: &[&str], remote_path: &Path) {
-        let mut expected: Vec<String> = expected.iter().map(|x| String::from(*x)).collect();
-        expected.sort();
-
-        let mut found: Vec<String> = Vec::new();
-        let dir = match std::fs::read_dir(remote_path) {
-            Ok(d) => d,
-            Err(e) => {
-                if e.kind() == ErrorKind::NotFound {
-                    if expected.is_empty() {
-                        // We are asserting prefix is empty: it is expected that the dir is missing
-                        return;
-                    } else {
-                        assert_eq!(expected, Vec::<String>::new());
-                        unreachable!();
-                    }
-                } else {
-                    panic!(
-                        "Unexpected error listing {0}: {e}",
-                        remote_path.to_string_lossy()
-                    );
-                }
-            }
-        };
-
-        for entry in dir.flatten() {
-            let entry_name = entry.file_name();
-            let fname = entry_name.to_str().unwrap();
-            found.push(String::from(fname));
-        }
-        found.sort();
-
-        assert_eq!(expected, found);
-    }
-
-    fn assert_local_files(expected: &[&str], directory: &Path) {
-        let mut dir = match std::fs::read_dir(directory) {
-            Ok(d) => d,
-            Err(_) => {
-                assert_eq!(expected, &Vec::<String>::new());
-                return;
-            }
-        };
-        let mut found = Vec::new();
-        while let Some(dentry) = dir.next() {
-            let dentry = dentry.unwrap();
-            let file_name = dentry.file_name();
-            let file_name_str = file_name.to_string_lossy();
-            found.push(file_name_str.to_string());
-        }
-        found.sort();
-        assert_eq!(expected, found);
-    }
-
-    #[test]
-    fn deletion_queue_smoke() -> anyhow::Result<()> {
-        // Basic test that the deletion queue processes the deletions we pass into it
-        let ctx = setup("deletion_queue_smoke").expect("Failed test setup");
-        let client = ctx.deletion_queue.new_client();
-
-        let layer_file_name_1: LayerFileName = "000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__00000000016B59D8-00000000016B5A51".parse().unwrap();
-        let tenant_id = ctx.harness.tenant_id;
-
-        let content: Vec<u8> = "victim1 contents".into();
-        let relative_remote_path = ctx
-            .harness
-            .conf
-            .remote_path(&ctx.harness.timeline_path(&TIMELINE_ID))
-            .expect("Failed to construct remote path");
-        let remote_timeline_path = ctx.remote_fs_dir.join(relative_remote_path.get_path());
-        let deletion_prefix = ctx.harness.conf.deletion_prefix();
-
-        // Inject a victim file to remote storage
-        info!("Writing");
-        std::fs::create_dir_all(&remote_timeline_path)?;
-        std::fs::write(
-            remote_timeline_path.join(layer_file_name_1.to_string()),
-            content,
-        )?;
-        assert_remote_files(&[&layer_file_name_1.file_name()], &remote_timeline_path);
-
-        // File should still be there after we push it to the queue (we haven't pushed enough to flush anything)
-        info!("Pushing");
-        ctx.runtime.block_on(client.push_layers(
-            tenant_id,
-            TIMELINE_ID,
-            [layer_file_name_1.clone()].to_vec(),
-        ))?;
-        assert_remote_files(&[&layer_file_name_1.file_name()], &remote_timeline_path);
-
-        assert_local_files(&[], &deletion_prefix);
-
-        // File should still be there after we write a deletion list (we haven't pushed enough to execute anything)
-        info!("Flushing");
-        ctx.runtime.block_on(client.flush())?;
-        assert_remote_files(&[&layer_file_name_1.file_name()], &remote_timeline_path);
-        assert_local_files(&["0000000000000001-01.list"], &deletion_prefix);
-
-        // File should go away when we execute
-        info!("Flush-executing");
-        ctx.runtime.block_on(client.flush_execute())?;
-        assert_remote_files(&[], &remote_timeline_path);
-        assert_local_files(&["header-01"], &deletion_prefix);
-
-        // Flushing on an empty queue should succeed immediately, and not write any lists
-        info!("Flush-executing on empty");
-        ctx.runtime.block_on(client.flush_execute())?;
-        assert_local_files(&["header-01"], &deletion_prefix);
-
-        Ok(())
-    }
-
-    #[test]
-    fn deletion_queue_recovery() -> anyhow::Result<()> {
-        // Basic test that the deletion queue processes the deletions we pass into it
-        let mut ctx = setup("deletion_queue_recovery").expect("Failed test setup");
-        let client = ctx.deletion_queue.new_client();
-
-        let layer_file_name_1: LayerFileName = "000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__00000000016B59D8-00000000016B5A51".parse().unwrap();
-        let tenant_id = ctx.harness.tenant_id;
-
-        let content: Vec<u8> = "victim1 contents".into();
-        let relative_remote_path = ctx
-            .harness
-            .conf
-            .remote_path(&ctx.harness.timeline_path(&TIMELINE_ID))
-            .expect("Failed to construct remote path");
-        let remote_timeline_path = ctx.remote_fs_dir.join(relative_remote_path.get_path());
-        let deletion_prefix = ctx.harness.conf.deletion_prefix();
-
-        // Inject a file, delete it, and flush to a deletion list
-        std::fs::create_dir_all(&remote_timeline_path)?;
-        std::fs::write(
-            remote_timeline_path.join(layer_file_name_1.to_string()),
-            content,
-        )?;
-        ctx.runtime.block_on(client.push_layers(
-            tenant_id,
-            TIMELINE_ID,
-            [layer_file_name_1.clone()].to_vec(),
-        ))?;
-        ctx.runtime.block_on(client.flush())?;
-        assert_local_files(&["0000000000000001-01.list"], &deletion_prefix);
-
-        // Restart the deletion queue
-        drop(client);
-        ctx.restart();
-        let client = ctx.deletion_queue.new_client();
-
-        // If we have recovered the deletion list properly, then executing after restart should purge it
-        info!("Flush-executing");
-        ctx.runtime.block_on(client.flush_execute())?;
-        assert_remote_files(&[], &remote_timeline_path);
-        assert_local_files(&["header-01"], &deletion_prefix);
-        Ok(())
-    }
-}
-
-/// A lightweight queue which can issue ordinary DeletionQueueClient objects, but doesn't do any persistence
-/// or coalescing, and doesn't actually execute any deletions unless you call pump() to kick it.
-#[cfg(test)]
-pub mod mock {
-    use tracing::info;
-
-    use super::*;
-    use std::sync::{
-        atomic::{AtomicUsize, Ordering},
-        Arc,
-    };
-
-    pub struct MockDeletionQueue {
-        tx: tokio::sync::mpsc::Sender<FrontendQueueMessage>,
-        executor_tx: tokio::sync::mpsc::Sender<ExecutorMessage>,
-        tx_pump: tokio::sync::mpsc::Sender<FlushOp>,
-        executed: Arc<AtomicUsize>,
-    }
-
-    impl MockDeletionQueue {
-        pub fn new(
-            remote_storage: Option<GenericRemoteStorage>,
-            conf: &'static PageServerConf,
-        ) -> Self {
-            let (tx, mut rx) = tokio::sync::mpsc::channel(16384);
-            let (tx_pump, mut rx_pump) = tokio::sync::mpsc::channel::<FlushOp>(1);
-            let (executor_tx, mut executor_rx) = tokio::sync::mpsc::channel(16384);
-
-            let executed = Arc::new(AtomicUsize::new(0));
-            let executed_bg = executed.clone();
-
-            tokio::spawn(async move {
-                let remote_storage = match &remote_storage {
-                    Some(rs) => rs,
-                    None => {
-                        info!("No remote storage configured, deletion queue will not run");
-                        return;
-                    }
-                };
-                info!("Running mock deletion queue");
-                // Each time we are asked to pump, drain the queue of deletions
-                while let Some(flush_op) = rx_pump.recv().await {
-                    info!("Executing all pending deletions");
-
-                    // Transform all executor messages to generic frontend messages
-                    while let Ok(msg) = executor_rx.try_recv() {
-                        match msg {
-                            ExecutorMessage::Delete(objects) => {
-                                for path in objects {
-                                    match remote_storage.delete(&path).await {
-                                        Ok(_) => {
-                                            debug!("Deleted {path}");
-                                        }
-                                        Err(e) => {
-                                            error!(
-                                                "Failed to delete {path}, leaking object! ({e})"
-                                            );
-                                        }
-                                    }
-                                    executed_bg.fetch_add(1, Ordering::Relaxed);
-                                }
-                            }
-                            ExecutorMessage::Flush(flush_op) => {
-                                flush_op.fire();
-                            }
-                        }
-                    }
-
-                    while let Ok(msg) = rx.try_recv() {
-                        match msg {
-                            FrontendQueueMessage::Delete(op) => {
-                                let timeline_path =
-                                    conf.timeline_path(&op.tenant_id, &op.timeline_id);
-
-                                let mut objects = op.objects;
-                                for layer in op.layers {
-                                    let local_path = timeline_path.join(layer.file_name());
-                                    let path = match conf.remote_path(&local_path) {
-                                        Ok(p) => p,
-                                        Err(e) => {
-                                            panic!("Can't make a timeline path! {e}");
-                                        }
-                                    };
-                                    objects.push(path);
-                                }
-
-                                for path in objects {
-                                    info!("Executing deletion {path}");
-                                    match remote_storage.delete(&path).await {
-                                        Ok(_) => {
-                                            debug!("Deleted {path}");
-                                        }
-                                        Err(e) => {
-                                            error!(
-                                                "Failed to delete {path}, leaking object! ({e})"
-                                            );
-                                        }
-                                    }
-                                    executed_bg.fetch_add(1, Ordering::Relaxed);
-                                }
-                            }
-                            FrontendQueueMessage::Flush(op) => {
-                                op.fire();
-                            }
-                            FrontendQueueMessage::FlushExecute(op) => {
-                                // We have already executed all prior deletions because mock does them inline
-                                op.fire();
-                            }
-                        }
-                        info!("All pending deletions have been executed");
-                    }
-                    flush_op
-                        .tx
-                        .send(())
-                        .expect("Test called flush but dropped before finishing");
-                }
-            });
-
-            Self {
-                tx,
-                tx_pump,
-                executor_tx,
-                executed,
-            }
-        }
-
-        pub fn get_executed(&self) -> usize {
-            self.executed.load(Ordering::Relaxed)
-        }
-
-        pub async fn pump(&self) {
-            let (tx, rx) = tokio::sync::oneshot::channel();
-            self.tx_pump
-                .send(FlushOp { tx })
-                .await
-                .expect("pump called after deletion queue loop stopped");
-            rx.await
-                .expect("Mock delete queue shutdown while waiting to pump");
-        }
-
-        pub(crate) fn new_client(&self) -> DeletionQueueClient {
-            DeletionQueueClient {
-                tx: self.tx.clone(),
-                executor_tx: self.executor_tx.clone(),
-            }
-        }
-    }
-}

pageserver/src/deletion_queue/backend.rs

@@ -1,181 +0,0 @@
-use std::time::Duration;
-
-use tracing::debug;
-use tracing::info;
-use tracing::warn;
-
-use crate::config::PageServerConf;
-use crate::metrics::DELETION_QUEUE_ERRORS;
-
-use super::executor::ExecutorMessage;
-use super::DeletionHeader;
-use super::DeletionList;
-use super::FlushOp;
-
-// After this length of time, execute deletions which are elegible to run,
-// even if we haven't accumulated enough for a full-sized DeleteObjects
-const EXECUTE_IDLE_DEADLINE: Duration = Duration::from_secs(60);
-
-// If we have received this number of keys, proceed with attempting to execute
-const AUTOFLUSH_KEY_COUNT: usize = 16384;
-
-#[derive(Debug)]
-pub(super) enum BackendQueueMessage {
-    Delete(DeletionList),
-    Flush(FlushOp),
-}
-pub struct BackendQueueWorker {
-    conf: &'static PageServerConf,
-    rx: tokio::sync::mpsc::Receiver<BackendQueueMessage>,
-    tx: tokio::sync::mpsc::Sender<ExecutorMessage>,
-
-    // Accumulate some lists to execute in a batch.
-    // The purpose of this accumulation is to implement batched validation of
-    // attachment generations, when split-brain protection is implemented.
-    // (see https://github.com/neondatabase/neon/pull/4919)
-    pending_lists: Vec<DeletionList>,
-
-    // Sum of all the lengths of lists in pending_lists
-    pending_key_count: usize,
-
-    // DeletionLists we have fully executed, which may be deleted
-    // from remote storage.
-    executed_lists: Vec<DeletionList>,
-}
-
-impl BackendQueueWorker {
-    pub(super) fn new(
-        conf: &'static PageServerConf,
-        rx: tokio::sync::mpsc::Receiver<BackendQueueMessage>,
-        tx: tokio::sync::mpsc::Sender<ExecutorMessage>,
-    ) -> Self {
-        Self {
-            conf,
-            rx,
-            tx,
-            pending_lists: Vec::new(),
-            pending_key_count: 0,
-            executed_lists: Vec::new(),
-        }
-    }
-
-    async fn cleanup_lists(&mut self) {
-        debug!(
-            "cleanup_lists: {0} executed lists, {1} pending lists",
-            self.executed_lists.len(),
-            self.pending_lists.len()
-        );
-
-        // Lists are always pushed into the queues + executed list in sequence order, so
-        // no sort is required: can find the highest sequence number by peeking at last element
-        let max_executed_seq = match self.executed_lists.last() {
-            Some(v) => v.sequence,
-            None => {
-                // No executed lists, nothing to clean up.
-                return;
-            }
-        };
-
-        // In case this is the last list, write a header out first so that
-        // we don't risk losing our knowledge of the sequence number (on replay, our
-        // next sequence number is the highest list seen + 1, or read from the header
-        // if there are no lists)
-        let header = DeletionHeader::new(max_executed_seq);
-        debug!("Writing header {:?}", header);
-        let header_bytes =
-            serde_json::to_vec(&header).expect("Failed to serialize deletion header");
-        let header_path = self.conf.deletion_header_path();
-
-        if let Err(e) = tokio::fs::write(&header_path, header_bytes).await {
-            warn!("Failed to upload deletion queue header: {e:#}");
-            DELETION_QUEUE_ERRORS
-                .with_label_values(&["put_header"])
-                .inc();
-            return;
-        }
-
-        while let Some(list) = self.executed_lists.pop() {
-            let list_path = self.conf.deletion_list_path(list.sequence);
-            if let Err(e) = tokio::fs::remove_file(&list_path).await {
-                // Unexpected: we should have permissions and nothing else should
-                // be touching these files
-                tracing::error!("Failed to delete {0}: {e:#}", list_path.display());
-                self.executed_lists.push(list);
-                break;
-            }
-        }
-    }
-
-    pub async fn flush(&mut self) {
-        self.pending_key_count = 0;
-
-        // Submit all keys from pending DeletionLists into the executor
-        for list in &mut self.pending_lists {
-            let objects = list.take_paths();
-            if let Err(_e) = self.tx.send(ExecutorMessage::Delete(objects)).await {
-                warn!("Shutting down");
-                return;
-            };
-        }
-
-        // Flush the executor to ensure all the operations we just submitted have been executed
-        let (tx, rx) = tokio::sync::oneshot::channel::<()>();
-        let flush_op = FlushOp { tx };
-        if let Err(_e) = self.tx.send(ExecutorMessage::Flush(flush_op)).await {
-            warn!("Shutting down");
-            return;
-        };
-        if rx.await.is_err() {
-            warn!("Shutting down");
-            return;
-        }
-
-        // After flush, we are assured that all contents of the pending lists
-        // are executed
-        self.executed_lists.append(&mut self.pending_lists);
-
-        // Erase the lists we executed
-        self.cleanup_lists().await;
-    }
-
-    pub async fn background(&mut self) {
-        // TODO: if we would like to be able to defer deletions while a Layer still has
-        // refs (but it will be elegible for deletion after process ends), then we may
-        // add an ephemeral part to BackendQueueMessage::Delete that tracks which keys
-        // in the deletion list may not be deleted yet, with guards to block on while
-        // we wait to proceed.
-
-        loop {
-            let msg = match tokio::time::timeout(EXECUTE_IDLE_DEADLINE, self.rx.recv()).await {
-                Ok(Some(m)) => m,
-                Ok(None) => {
-                    // All queue senders closed
-                    info!("Shutting down");
-                    break;
-                }
-                Err(_) => {
-                    // Timeout, we hit deadline to execute whatever we have in hand.  These functions will
-                    // return immediately if no work is pending
-                    self.flush().await;
-
-                    continue;
-                }
-            };
-
-            match msg {
-                BackendQueueMessage::Delete(list) => {
-                    self.pending_key_count += list.objects.len();
-                    self.pending_lists.push(list);
-
-                    if self.pending_key_count > AUTOFLUSH_KEY_COUNT {
-                        self.flush().await;
-                    }
-                }
-                BackendQueueMessage::Flush(op) => {
-                    self.flush().await;
-                    op.fire();
-                }
-            }
-        }
-    }
-}

pageserver/src/deletion_queue/executor.rs

@@ -1,143 +0,0 @@
-use remote_storage::GenericRemoteStorage;
-use remote_storage::RemotePath;
-use remote_storage::MAX_KEYS_PER_DELETE;
-use std::time::Duration;
-use tokio_util::sync::CancellationToken;
-use tracing::info;
-use tracing::warn;
-
-use crate::metrics::DELETION_QUEUE_ERRORS;
-use crate::metrics::DELETION_QUEUE_EXECUTED;
-
-use super::DeletionQueueError;
-use super::FlushOp;
-
-const AUTOFLUSH_INTERVAL: Duration = Duration::from_secs(10);
-
-pub(super) enum ExecutorMessage {
-    Delete(Vec<RemotePath>),
-    Flush(FlushOp),
-}
-
-/// Non-persistent deletion queue, for coalescing multiple object deletes into
-/// larger DeleteObjects requests.
-pub struct ExecutorWorker {
-    // Accumulate up to 1000 keys for the next deletion operation
-    accumulator: Vec<RemotePath>,
-
-    rx: tokio::sync::mpsc::Receiver<ExecutorMessage>,
-
-    cancel: CancellationToken,
-    remote_storage: GenericRemoteStorage,
-}
-
-impl ExecutorWorker {
-    pub(super) fn new(
-        remote_storage: GenericRemoteStorage,
-        rx: tokio::sync::mpsc::Receiver<ExecutorMessage>,
-        cancel: CancellationToken,
-    ) -> Self {
-        Self {
-            remote_storage,
-            rx,
-            cancel,
-            accumulator: Vec::new(),
-        }
-    }
-
-    /// Wrap the remote `delete_objects` with a failpoint
-    pub async fn remote_delete(&self) -> Result<(), anyhow::Error> {
-        fail::fail_point!("deletion-queue-before-execute", |_| {
-            info!("Skipping execution, failpoint set");
-            DELETION_QUEUE_ERRORS
-                .with_label_values(&["failpoint"])
-                .inc();
-            Err(anyhow::anyhow!("failpoint hit"))
-        });
-
-        self.remote_storage.delete_objects(&self.accumulator).await
-    }
-
-    /// Block until everything in accumulator has been executed
-    pub async fn flush(&mut self) -> Result<(), DeletionQueueError> {
-        while !self.accumulator.is_empty() && !self.cancel.is_cancelled() {
-            match self.remote_delete().await {
-                Ok(()) => {
-                    // Note: we assume that the remote storage layer returns Ok(()) if some
-                    // or all of the deleted objects were already gone.
-                    DELETION_QUEUE_EXECUTED.inc_by(self.accumulator.len() as u64);
-                    info!(
-                        "Executed deletion batch {}..{}",
-                        self.accumulator
-                            .first()
-                            .expect("accumulator should be non-empty"),
-                        self.accumulator
-                            .last()
-                            .expect("accumulator should be non-empty"),
-                    );
-                    self.accumulator.clear();
-                }
-                Err(e) => {
-                    warn!("DeleteObjects request failed: {e:#}, will retry");
-                    DELETION_QUEUE_ERRORS.with_label_values(&["execute"]).inc();
-                }
-            };
-        }
-        if self.cancel.is_cancelled() {
-            // Expose an error because we may not have actually flushed everything
-            Err(DeletionQueueError::ShuttingDown)
-        } else {
-            Ok(())
-        }
-    }
-
-    pub async fn background(&mut self) -> Result<(), DeletionQueueError> {
-        self.accumulator.reserve(MAX_KEYS_PER_DELETE);
-
-        loop {
-            if self.cancel.is_cancelled() {
-                return Err(DeletionQueueError::ShuttingDown);
-            }
-
-            let msg = match tokio::time::timeout(AUTOFLUSH_INTERVAL, self.rx.recv()).await {
-                Ok(Some(m)) => m,
-                Ok(None) => {
-                    // All queue senders closed
-                    info!("Shutting down");
-                    return Err(DeletionQueueError::ShuttingDown);
-                }
-                Err(_) => {
-                    // Timeout, we hit deadline to execute whatever we have in hand.  These functions will
-                    // return immediately if no work is pending
-                    self.flush().await?;
-
-                    continue;
-                }
-            };
-
-            match msg {
-                ExecutorMessage::Delete(mut list) => {
-                    while !list.is_empty() || self.accumulator.len() == MAX_KEYS_PER_DELETE {
-                        if self.accumulator.len() == MAX_KEYS_PER_DELETE {
-                            self.flush().await?;
-                            // If we have received this number of keys, proceed with attempting to execute
-                            assert_eq!(self.accumulator.len(), 0);
-                        }
-
-                        let available_slots = MAX_KEYS_PER_DELETE - self.accumulator.len();
-                        let take_count = std::cmp::min(available_slots, list.len());
-                        for path in list.drain(list.len() - take_count..) {
-                            self.accumulator.push(path);
-                        }
-                    }
-                }
-                ExecutorMessage::Flush(flush_op) => {
-                    // If flush() errors, we drop the flush_op and the caller will get
-                    // an error recv()'ing their oneshot channel.
-                    self.flush().await?;
-                    flush_op.fire();
-                }
-            }
-        }
-    }
-}

pageserver/src/deletion_queue/frontend.rs

@@ -1,357 +0,0 @@
-use super::BackendQueueMessage;
-use super::DeletionHeader;
-use super::DeletionList;
-use super::FlushOp;
-
-use std::fs::create_dir_all;
-use std::time::Duration;
-
-use regex::Regex;
-use remote_storage::RemotePath;
-use tokio_util::sync::CancellationToken;
-use tracing::debug;
-use tracing::info;
-use tracing::warn;
-use utils::id::TenantId;
-use utils::id::TimelineId;
-
-use crate::config::PageServerConf;
-use crate::metrics::DELETION_QUEUE_ERRORS;
-use crate::metrics::DELETION_QUEUE_SUBMITTED;
-use crate::tenant::storage_layer::LayerFileName;
-
-// The number of keys in a DeletionList before we will proactively persist it
-// (without reaching a flush deadline).  This aims to deliver objects of the order
-// of magnitude 1MB when we are under heavy delete load.
-const DELETION_LIST_TARGET_SIZE: usize = 16384;
-
-// Ordinarily, we only flush to DeletionList periodically, to bound the window during
-// which we might leak objects from not flushing a DeletionList after
-// the objects are already unlinked from timeline metadata.
-const FRONTEND_DEFAULT_TIMEOUT: Duration = Duration::from_millis(10000);
-
-// If someone is waiting for a flush to DeletionList, only delay a little to accumulate
-// more objects before doing the flush.
-const FRONTEND_FLUSHING_TIMEOUT: Duration = Duration::from_millis(100);
-
-#[derive(Debug)]
-pub(super) struct DeletionOp {
-    pub(super) tenant_id: TenantId,
-    pub(super) timeline_id: TimelineId,
-    // `layers` and `objects` are both just lists of objects.  `layers` is used if you do not
-    // have a config object handy to project it to a remote key, and need the consuming worker
-    // to do it for you.
-    pub(super) layers: Vec<LayerFileName>,
-    pub(super) objects: Vec<RemotePath>,
-}
-
-#[derive(Debug)]
-pub(super) enum FrontendQueueMessage {
-    Delete(DeletionOp),
-    // Wait until all prior deletions make it into a persistent DeletionList
-    Flush(FlushOp),
-    // Wait until all prior deletions have been executed (i.e. objects are actually deleted)
-    FlushExecute(FlushOp),
-}
-
-pub struct FrontendQueueWorker {
-    conf: &'static PageServerConf,
-
-    // Incoming frontend requests to delete some keys
-    rx: tokio::sync::mpsc::Receiver<FrontendQueueMessage>,
-
-    // Outbound requests to the backend to execute deletion lists we have composed.
-    tx: tokio::sync::mpsc::Sender<BackendQueueMessage>,
-
-    // The list we are currently building, contains a buffer of keys to delete
-    // and our next sequence number
-    pending: DeletionList,
-
-    // These FlushOps should fire the next time we flush
-    pending_flushes: Vec<FlushOp>,
-
-    // Worker loop is torn down when this fires.
-    cancel: CancellationToken,
-}
-
-impl FrontendQueueWorker {
-    pub(super) fn new(
-        conf: &'static PageServerConf,
-        rx: tokio::sync::mpsc::Receiver<FrontendQueueMessage>,
-        tx: tokio::sync::mpsc::Sender<BackendQueueMessage>,
-        cancel: CancellationToken,
-    ) -> Self {
-        Self {
-            pending: DeletionList::new(1),
-            conf,
-            rx,
-            tx,
-            pending_flushes: Vec::new(),
-            cancel,
-        }
-    }
-    async fn upload_pending_list(&mut self) -> anyhow::Result<()> {
-        let path = self.conf.deletion_list_path(self.pending.sequence);
-
-        let bytes = serde_json::to_vec(&self.pending).expect("Failed to serialize deletion list");
-        tokio::fs::write(&path, &bytes).await?;
-        tokio::fs::File::open(&path).await?.sync_all().await?;
-        Ok(())
-    }
-
-    /// Try to flush `list` to persistent storage
-    ///
-    /// This does not return errors, because on failure to flush we do not lose
-    /// any state: flushing will be retried implicitly on the next deadline
-    async fn flush(&mut self) {
-        if self.pending.is_empty() {
-            for f in self.pending_flushes.drain(..) {
-                f.fire();
-            }
-            return;
-        }
-
-        match self.upload_pending_list().await {
-            Ok(_) => {
-                info!(sequence = self.pending.sequence, "Stored deletion list");
-
-                for f in self.pending_flushes.drain(..) {
-                    f.fire();
-                }
-
-                let mut onward_list = DeletionList::new(self.pending.sequence);
-                std::mem::swap(&mut onward_list.objects, &mut self.pending.objects);
-
-                // We have consumed out of pending: reset it for the next incoming deletions to accumulate there
-                self.pending = DeletionList::new(self.pending.sequence + 1);
-
-                if let Err(e) = self.tx.send(BackendQueueMessage::Delete(onward_list)).await {
-                    // This is allowed to fail: it will only happen if the backend worker is shut down,
-                    // so we can just drop this on the floor.
-                    info!("Deletion list dropped, this is normal during shutdown ({e:#})");
-                }
-            }
-            Err(e) => {
-                DELETION_QUEUE_ERRORS.with_label_values(&["put_list"]).inc();
-                warn!(
-                    sequence = self.pending.sequence,
-                    "Failed to write deletion list to remote storage, will retry later ({e:#})"
-                );
-            }
-        }
-    }
-
-    async fn recover(&mut self) -> Result<(), anyhow::Error> {
-        // Load header: this is not required to be present, e.g. when a pageserver first runs
-        let header_path = self.conf.deletion_header_path();
-
-        // Synchronous, but we only do it once per process lifetime so it's tolerable
-        create_dir_all(&self.conf.deletion_prefix())?;
-
-        let header_bytes = match tokio::fs::read(&header_path).await {
-            Ok(h) => Ok(Some(h)),
-            Err(e) => {
-                if e.kind() == std::io::ErrorKind::NotFound {
-                    debug!(
-                        "Deletion header {0} not found, first start?",
-                        header_path.display()
-                    );
-                    Ok(None)
-                } else {
-                    Err(e)
-                }
-            }
-        }?;
-
-        if let Some(header_bytes) = header_bytes {
-            if let Some(header) = match serde_json::from_slice::<DeletionHeader>(&header_bytes) {
-                Ok(h) => Some(h),
-                Err(e) => {
-                    warn!(
-                        "Failed to deserialize deletion header, ignoring {0}: {e:#}",
-                        header_path.display()
-                    );
-                    // This should never happen unless we make a mistake with our serialization.
-                    // Ignoring a deletion header is not consequential for correctnes because all deletions
-                    // are ultimately allowed to fail: worst case we leak some objects for the scrubber to clean up.
-                    None
-                }
-            } {
-                self.pending.sequence =
-                    std::cmp::max(self.pending.sequence, header.last_deleted_list_seq + 1);
-            };
-        };
-
-        let mut dir = match tokio::fs::read_dir(&self.conf.deletion_prefix()).await {
-            Ok(d) => d,
-            Err(e) => {
-                warn!(
-                    "Failed to open deletion list directory {0}: {e:#}",
-                    header_path.display()
-                );
-
-                // Give up: if we can't read the deletion list directory, we probably can't
-                // write lists into it later, so the queue won't work.
-                return Err(e.into());
-            }
-        };
-
-        let list_name_pattern = Regex::new("([a-zA-Z0-9]{16})-([a-zA-Z0-9]{2}).list").unwrap();
-
-        let mut seqs: Vec<u64> = Vec::new();
-        while let Some(dentry) = dir.next_entry().await? {
-            let file_name = dentry.file_name().to_owned();
-            let basename = file_name.to_string_lossy();
-            let seq_part = if let Some(m) = list_name_pattern.captures(&basename) {
-                m.get(1)
-                    .expect("Non optional group should be present")
-                    .as_str()
-            } else {
-                warn!("Unexpected key in deletion queue: {basename}");
-                continue;
-            };
-
-            let seq: u64 = match u64::from_str_radix(seq_part, 16) {
-                Ok(s) => s,
-                Err(e) => {
-                    warn!("Malformed key '{basename}': {e}");
-                    continue;
-                }
-            };
-            seqs.push(seq);
-        }
-        seqs.sort();
-
-        // Initialize the next sequence number in the frontend based on the maximum of the highest list we see,
-        // and the last list that was deleted according to the header.  Combined with writing out the header
-        // prior to deletions, this guarnatees no re-use of sequence numbers.
-        if let Some(max_list_seq) = seqs.last() {
-            self.pending.sequence = std::cmp::max(self.pending.sequence, max_list_seq + 1);
-        }
-
-        for s in seqs {
-            let list_path = self.conf.deletion_list_path(s);
-            let list_bytes = tokio::fs::read(&list_path).await?;
-
-            let deletion_list = match serde_json::from_slice::<DeletionList>(&list_bytes) {
-                Ok(l) => l,
-                Err(e) => {
-                    // Drop the list on the floor: any objects it referenced will be left behind
-                    // for scrubbing to clean up.  This should never happen unless we have a serialization bug.
-                    warn!(sequence = s, "Failed to deserialize deletion list: {e}");
-                    continue;
-                }
-            };
-
-            // We will drop out of recovery if this fails: it indicates that we are shutting down
-            // or the backend has panicked
-            DELETION_QUEUE_SUBMITTED.inc_by(deletion_list.len() as u64);
-            self.tx
-                .send(BackendQueueMessage::Delete(deletion_list))
-                .await?;
-        }
-
-        info!(next_sequence = self.pending.sequence, "Replay complete");
-
-        Ok(())
-    }
-
-    /// This is the front-end ingest, where we bundle up deletion requests into DeletionList
-    /// and write them out, for later
-    pub async fn background(&mut self) {
-        info!("Started deletion frontend worker");
-
-        let mut recovered: bool = false;
-
-        while !self.cancel.is_cancelled() {
-            let timeout = if self.pending_flushes.is_empty() {
-                FRONTEND_DEFAULT_TIMEOUT
-            } else {
-                FRONTEND_FLUSHING_TIMEOUT
-            };
-
-            let msg = match tokio::time::timeout(timeout, self.rx.recv()).await {
-                Ok(Some(msg)) => msg,
-                Ok(None) => {
-                    // Queue sender destroyed, shutting down
-                    break;
-                }
-                Err(_) => {
-                    // Hit deadline, flush.
-                    self.flush().await;
-                    continue;
-                }
-            };
-
-            // On first message, do recovery.  This avoids unnecessary recovery very
-            // early in startup, and simplifies testing by avoiding a 404 reading the
-            // header on every first pageserver startup.
-            if !recovered {
-                // Before accepting any input from this pageserver lifetime, recover all deletion lists that are in S3
-                if let Err(e) = self.recover().await {
-                    // This should only happen in truly unrecoverable cases, like the recovery finding that the backend
-                    // queue receiver has been dropped.
-                    info!("Deletion queue recover aborted, deletion queue will not proceed ({e})");
-                    return;
-                } else {
-                    recovered = true;
-                }
-            }
-
-            match msg {
-                FrontendQueueMessage::Delete(op) => {
-                    debug!(
-                        "Delete: ingesting {0} layers, {1} other objects",
-                        op.layers.len(),
-                        op.objects.len()
-                    );
-
-                    let timeline_path = self.conf.timeline_path(&op.tenant_id, &op.timeline_id);
-                    let mut layer_paths = Vec::new();
-                    for layer in op.layers {
-                        // TODO go directly to remote path without composing local path
-                        let local_path = timeline_path.join(layer.file_name());
-                        let path = match self.conf.remote_path(&local_path) {
-                            Ok(p) => p,
-                            Err(e) => {
-                                panic!("Can't make a timeline path! {e}");
-                            }
-                        };
-                        layer_paths.push(path);
-                    }
-
-                    self.pending
-                        .push(&op.tenant_id, &op.timeline_id, layer_paths);
-                    self.pending
-                        .push(&op.tenant_id, &op.timeline_id, op.objects);
-                }
-                FrontendQueueMessage::Flush(op) => {
-                    if self.pending.objects.is_empty() {
-                        // Execute immediately
-                        debug!("Flush: No pending objects, flushing immediately");
-                        op.fire()
-                    } else {
-                        // Execute next time we flush
-                        debug!("Flush: adding to pending flush list for next deadline flush");
-                        self.pending_flushes.push(op);
-                    }
-                }
-                FrontendQueueMessage::FlushExecute(op) => {
-                    debug!("FlushExecute: passing through to backend");
-                    // We do not flush to a deletion list here: the client sends a Flush before the FlushExecute
-                    if let Err(e) = self.tx.send(BackendQueueMessage::Flush(op)).await {
-                        info!("Can't flush, shutting down ({e})");
-                        // Caller will get error when their oneshot sender was dropped.
-                    }
-                }
-            }
-
-            if self.pending.objects.len() > DELETION_LIST_TARGET_SIZE
-                || !self.pending_flushes.is_empty()
-            {
-                self.flush().await;
-            }
-        }
-        info!("Deletion queue shut down.");
-    }
-}

pageserver/src/http/openapi_spec.yml

@@ -52,29 +52,6 @@ paths:
               schema:
                 type: object
 
-  /v1/deletion_queue/flush:
-    parameters:
-      - name: execute
-        in: query
-        required: false
-        schema:
-          type: boolean
-        description:
-          If true, attempt to execute deletions.  If false, just flush deletions to persistent deletion lists.
-    put:
-      description: Execute any deletions currently enqueued
-      security: []
-      responses:
-        "200":
-          description: |
-            Flush completed: if execute was true, then enqueued deletions have been completed.  If execute was false,
-            then enqueued deletions have been persisted to deletion lists, and may have been completed.
-          content:
-            application/json:
-              schema:
-                type: object
-
-
   /v1/tenant/{tenant_id}:
     parameters:
       - name: tenant_id

pageserver/src/http/routes.rs

@@ -23,7 +23,6 @@ use super::models::{
     TimelineCreateRequest, TimelineGcRequest, TimelineInfo,
 };
 use crate::context::{DownloadBehavior, RequestContext};
-use crate::deletion_queue::{DeletionQueue, DeletionQueueError};
 use crate::metrics::{StorageTimeOperation, STORAGE_TIME_GLOBAL};
 use crate::pgdatadir_mapping::LsnForTimestamp;
 use crate::task_mgr::TaskKind;
@@ -57,7 +56,6 @@ struct State {
     auth: Option<Arc<JwtAuth>>,
     allowlist_routes: Vec<Uri>,
     remote_storage: Option<GenericRemoteStorage>,
-    deletion_queue: DeletionQueue,
     broker_client: storage_broker::BrokerClientChannel,
     disk_usage_eviction_state: Arc<disk_usage_eviction_task::State>,
 }
@@ -67,7 +65,6 @@ impl State {
         conf: &'static PageServerConf,
         auth: Option<Arc<JwtAuth>>,
         remote_storage: Option<GenericRemoteStorage>,
-        deletion_queue: DeletionQueue,
         broker_client: storage_broker::BrokerClientChannel,
         disk_usage_eviction_state: Arc<disk_usage_eviction_task::State>,
     ) -> anyhow::Result<Self> {
@@ -81,7 +78,6 @@ impl State {
             allowlist_routes,
             remote_storage,
             broker_client,
-            deletion_queue,
             disk_usage_eviction_state,
         })
     }
@@ -494,7 +490,6 @@ async fn tenant_attach_handler(
             tenant_conf,
             state.broker_client.clone(),
             remote_storage.clone(),
-            &state.deletion_queue,
             &ctx,
         )
         .instrument(info_span!("tenant_attach", %tenant_id))
@@ -557,7 +552,6 @@ async fn tenant_load_handler(
         tenant_id,
         state.broker_client.clone(),
         state.remote_storage.clone(),
-        &state.deletion_queue,
         &ctx,
     )
     .instrument(info_span!("load", %tenant_id))
@@ -883,7 +877,6 @@ async fn tenant_create_handler(
         target_tenant_id,
         state.broker_client.clone(),
         state.remote_storage.clone(),
-        &state.deletion_queue,
         &ctx,
     )
     .instrument(info_span!("tenant_create", tenant_id = %target_tenant_id))
@@ -1124,48 +1117,6 @@ async fn always_panic_handler(
     json_response(StatusCode::NO_CONTENT, ())
 }
 
-async fn deletion_queue_flush(
-    r: Request<Body>,
-    cancel: CancellationToken,
-) -> Result<Response<Body>, ApiError> {
-    let state = get_state(&r);
-
-    if state.remote_storage.is_none() {
-        // Nothing to do if remote storage is disabled.
-        return json_response(StatusCode::OK, ());
-    }
-
-    let execute = parse_query_param(&r, "execute")?.unwrap_or(false);
-
-    let queue_client = state.deletion_queue.new_client();
-
-    tokio::select! {
-        flush_result = async {
-            if execute {
-                queue_client.flush_execute().await
-            } else {
-                queue_client.flush().await
-            }
-        } => {
-            match flush_result {
-                Ok(())=> {
-                    json_response(StatusCode::OK, ())
-                },
-                Err(e) => {
-                    match e {
-                        DeletionQueueError::ShuttingDown => {
-            Err(ApiError::ShuttingDown)
-                        }
-                    }
-                }
-            }
-        },
-        _ = cancel.cancelled() => {
-            Err(ApiError::ShuttingDown)
-        }
-    }
-}
-
 async fn disk_usage_eviction_run(
     mut r: Request<Body>,
     _cancel: CancellationToken,
@@ -1375,7 +1326,6 @@ pub fn make_router(
     auth: Option<Arc<JwtAuth>>,
     broker_client: BrokerClientChannel,
     remote_storage: Option<GenericRemoteStorage>,
-    deletion_queue: DeletionQueue,
     disk_usage_eviction_state: Arc<disk_usage_eviction_task::State>,
 ) -> anyhow::Result<RouterBuilder<hyper::Body, ApiError>> {
     let spec = include_bytes!("openapi_spec.yml");
@@ -1405,7 +1355,6 @@ pub fn make_router(
                 conf,
                 auth,
                 remote_storage,
-                deletion_queue,
                 broker_client,
                 disk_usage_eviction_state,
             )
@@ -1490,9 +1439,6 @@ pub fn make_router(
         .put("/v1/disk_usage_eviction/run", |r| {
             api_handler(r, disk_usage_eviction_run)
         })
-        .put("/v1/deletion_queue/flush", |r| {
-            api_handler(r, deletion_queue_flush)
-        })
         .put("/v1/tenant/:tenant_id/break", |r| {
             testing_api_handler("set tenant state to broken", r, handle_tenant_break)
         })

pageserver/src/lib.rs

@@ -3,7 +3,6 @@ pub mod basebackup;
 pub mod config;
 pub mod consumption_metrics;
 pub mod context;
-pub mod deletion_queue;
 pub mod disk_usage_eviction_task;
 pub mod http;
 pub mod import_datadir;

pageserver/src/metrics.rs

@@ -795,31 +795,6 @@ static REMOTE_TIMELINE_CLIENT_CALLS_STARTED_HIST: Lazy<HistogramVec> = Lazy::new
     .expect("failed to define a metric")
 });
 
-pub(crate) static DELETION_QUEUE_SUBMITTED: Lazy<IntCounter> = Lazy::new(|| {
-    register_int_counter!(
-        "pageserver_deletion_queue_submitted_total",
-        "Number of objects submitted for deletion"
-    )
-    .expect("failed to define a metric")
-});
-
-pub(crate) static DELETION_QUEUE_EXECUTED: Lazy<IntCounter> = Lazy::new(|| {
-    register_int_counter!(
-        "pageserver_deletion_queue_executed_total",
-        "Number of objects deleted"
-    )
-    .expect("failed to define a metric")
-});
-
-pub(crate) static DELETION_QUEUE_ERRORS: Lazy<IntCounterVec> = Lazy::new(|| {
-    register_int_counter_vec!(
-        "pageserver_deletion_queue_errors_total",
-        "Incremented on retryable remote I/O errors writing deletion lists or executing deletions.",
-        &["op_kind"],
-    )
-    .expect("failed to define a metric")
-});
-
 static REMOTE_TIMELINE_CLIENT_BYTES_STARTED_COUNTER: Lazy<IntCounterVec> = Lazy::new(|| {
     register_int_counter_vec!(
         "pageserver_remote_timeline_client_bytes_started",

pageserver/src/page_service.rs

@@ -469,7 +469,9 @@ impl PageServerHandler {
         // Create empty timeline
         info!("creating new timeline");
         let tenant = get_active_tenant_with_timeout(tenant_id, &ctx).await?;
-        let timeline = tenant.create_empty_timeline(timeline_id, base_lsn, pg_version, &ctx)?;
+        let timeline = tenant
+            .create_empty_timeline(timeline_id, base_lsn, pg_version, &ctx)
+            .await?;
 
         // TODO mark timeline as not ready until it reaches end_lsn.
         // We might have some wal to import as well, and we should prevent compute

pageserver/src/tenant.rs

@@ -59,7 +59,6 @@ use self::timeline::EvictionTaskTenantState;
 use self::timeline::TimelineResources;
 use crate::config::PageServerConf;
 use crate::context::{DownloadBehavior, RequestContext};
-use crate::deletion_queue::DeletionQueueClient;
 use crate::import_datadir;
 use crate::is_uninit_mark;
 use crate::metrics::TENANT_ACTIVATION;
@@ -69,7 +68,7 @@ use crate::task_mgr;
 use crate::task_mgr::TaskKind;
 use crate::tenant::config::TenantConfOpt;
 use crate::tenant::metadata::load_metadata;
-use crate::tenant::remote_timeline_client::index::IndexPart;
+pub use crate::tenant::remote_timeline_client::index::IndexPart;
 use crate::tenant::remote_timeline_client::MaybeDeletedIndexPart;
 use crate::tenant::storage_layer::DeltaLayer;
 use crate::tenant::storage_layer::ImageLayer;
@@ -86,6 +85,7 @@ pub use pageserver_api::models::TenantState;
 use toml_edit;
 use utils::{
     crashsafe,
+    generation::Generation,
     id::{TenantId, TimelineId},
     lsn::{Lsn, RecordLsn},
 };
@@ -157,7 +157,6 @@ pub const TENANT_DELETED_MARKER_FILE_NAME: &str = "deleted";
 pub struct TenantSharedResources {
     pub broker_client: storage_broker::BrokerClientChannel,
     pub remote_storage: Option<GenericRemoteStorage>,
-    pub deletion_queue_client: DeletionQueueClient,
 }
 
 ///
@@ -180,6 +179,11 @@ pub struct Tenant {
     tenant_conf: Arc<RwLock<TenantConfOpt>>,
 
     tenant_id: TenantId,
+
+    /// The remote storage generation, used to protect S3 objects from split-brain.
+    /// Does not change over the lifetime of the [`Tenant`] object.
+    generation: Generation,
+
     timelines: Mutex<HashMap<TimelineId, Arc<Timeline>>>,
     // This mutex prevents creation of new timelines during GC.
     // Adding yet another mutex (in addition to `timelines`) is needed because holding
@@ -193,9 +197,6 @@ pub struct Tenant {
     // provides access to timeline data sitting in the remote storage
     remote_storage: Option<GenericRemoteStorage>,
 
-    // Access to global deletion queue for when this tenant wants to schedule a deletion
-    deletion_queue_client: Option<DeletionQueueClient>,
-
     /// Cached logical sizes updated updated on each [`Tenant::gather_size_inputs`].
     cached_logical_sizes: tokio::sync::Mutex<HashMap<(TimelineId, Lsn), u64>>,
     cached_synthetic_tenant_size: Arc<AtomicU64>,
@@ -447,6 +448,7 @@ impl Tenant {
                 up_to_date_metadata,
                 first_save,
             )
+            .await
             .context("save_metadata")?;
         }
 
@@ -527,10 +529,10 @@ impl Tenant {
     pub(crate) fn spawn_attach(
         conf: &'static PageServerConf,
         tenant_id: TenantId,
+        generation: Generation,
         broker_client: storage_broker::BrokerClientChannel,
         tenants: &'static tokio::sync::RwLock<TenantsMap>,
         remote_storage: GenericRemoteStorage,
-        deletion_queue_client: DeletionQueueClient,
         ctx: &RequestContext,
     ) -> anyhow::Result<Arc<Tenant>> {
         // TODO dedup with spawn_load
@@ -544,8 +546,8 @@ impl Tenant {
             tenant_conf,
             wal_redo_manager,
             tenant_id,
+            generation,
             Some(remote_storage.clone()),
-            Some(deletion_queue_client),
         ));
 
         // Do all the hard work in the background
@@ -655,12 +657,8 @@ impl Tenant {
             .as_ref()
             .ok_or_else(|| anyhow::anyhow!("cannot attach without remote storage"))?;
 
-        let remote_timeline_ids = remote_timeline_client::list_remote_timelines(
-            remote_storage,
-            self.conf,
-            self.tenant_id,
-        )
-        .await?;
+        let remote_timeline_ids =
+            remote_timeline_client::list_remote_timelines(remote_storage, self.tenant_id).await?;
 
         info!("found {} timelines", remote_timeline_ids.len());
 
@@ -672,6 +670,7 @@ impl Tenant {
                 self.conf,
                 self.tenant_id,
                 timeline_id,
+                self.generation,
             );
             part_downloads.spawn(
                 async move {
@@ -734,7 +733,6 @@ impl Tenant {
                 remote_metadata,
                 TimelineResources {
                     remote_client: Some(remote_client),
-                    deletion_queue_client: self.deletion_queue_client.clone(),
                 },
                 ctx,
             )
@@ -759,7 +757,6 @@ impl Tenant {
                 timeline_id,
                 &index_part.metadata,
                 Some(remote_timeline_client),
-                self.deletion_queue_client.clone(),
                 None,
             )
             .await
@@ -860,7 +857,7 @@ impl Tenant {
             TenantConfOpt::default(),
             wal_redo_manager,
             tenant_id,
-            None,
+            Generation::broken(),
             None,
         ))
     }
@@ -878,6 +875,7 @@ impl Tenant {
     pub(crate) fn spawn_load(
         conf: &'static PageServerConf,
         tenant_id: TenantId,
+        generation: Generation,
         resources: TenantSharedResources,
         init_order: Option<InitializationOrder>,
         tenants: &'static tokio::sync::RwLock<TenantsMap>,
@@ -895,7 +893,6 @@ impl Tenant {
 
         let broker_client = resources.broker_client;
         let remote_storage = resources.remote_storage;
-        let deletion_queue_client = resources.deletion_queue_client;
 
         let wal_redo_manager = Arc::new(PostgresRedoManager::new(conf, tenant_id));
         let tenant = Tenant::new(
@@ -904,8 +901,8 @@ impl Tenant {
             tenant_conf,
             wal_redo_manager,
             tenant_id,
+            generation,
             remote_storage.clone(),
-            Some(deletion_queue_client),
         );
         let tenant = Arc::new(tenant);
 
@@ -1313,7 +1310,6 @@ impl Tenant {
                                 timeline_id,
                                 &local_metadata,
                                 Some(remote_client),
-                                self.deletion_queue_client.clone(),
                                 init_order,
                             )
                             .await
@@ -1363,7 +1359,6 @@ impl Tenant {
                         timeline_id,
                         &local_metadata,
                         None,
-                        None,
                         init_order,
                     )
                     .await
@@ -1456,7 +1451,7 @@ impl Tenant {
     /// For tests, use `DatadirModification::init_empty_test_timeline` + `commit` to setup the
     /// minimum amount of keys required to get a writable timeline.
     /// (Without it, `put` might fail due to `repartition` failing.)
-    pub fn create_empty_timeline(
+    pub async fn create_empty_timeline(
         &self,
         new_timeline_id: TimelineId,
         initdb_lsn: Lsn,
@@ -1468,10 +1463,10 @@ impl Tenant {
             "Cannot create empty timelines on inactive tenant"
         );
 
-        let timelines = self.timelines.lock().unwrap();
-        let timeline_uninit_mark = self.create_timeline_uninit_mark(new_timeline_id, &timelines)?;
-        drop(timelines);
-
+        let timeline_uninit_mark = {
+            let timelines = self.timelines.lock().unwrap();
+            self.create_timeline_uninit_mark(new_timeline_id, &timelines)?
+        };
         let new_metadata = TimelineMetadata::new(
             // Initialize disk_consistent LSN to 0, The caller must import some data to
             // make it valid, before calling finish_creation()
@@ -1490,6 +1485,7 @@ impl Tenant {
             initdb_lsn,
             None,
         )
+        .await
     }
 
     /// Helper for unit tests to create an empty timeline.
@@ -1505,7 +1501,9 @@ impl Tenant {
         pg_version: u32,
         ctx: &RequestContext,
     ) -> anyhow::Result<Arc<Timeline>> {
-        let uninit_tl = self.create_empty_timeline(new_timeline_id, initdb_lsn, pg_version, ctx)?;
+        let uninit_tl = self
+            .create_empty_timeline(new_timeline_id, initdb_lsn, pg_version, ctx)
+            .await?;
         let tline = uninit_tl.raw_timeline().expect("we just created it");
         assert_eq!(tline.get_last_record_lsn(), Lsn(0));
 
@@ -2288,6 +2286,7 @@ impl Tenant {
             ancestor,
             new_timeline_id,
             self.tenant_id,
+            self.generation,
             Arc::clone(&self.walredo_mgr),
             resources,
             pg_version,
@@ -2305,17 +2304,9 @@ impl Tenant {
         tenant_conf: TenantConfOpt,
         walredo_mgr: Arc<dyn WalRedoManager + Send + Sync>,
         tenant_id: TenantId,
+        generation: Generation,
         remote_storage: Option<GenericRemoteStorage>,
-        deletion_queue_client: Option<DeletionQueueClient>,
     ) -> Tenant {
-        #[cfg(not(test))]
-        match state {
-            TenantState::Broken { .. } => {}
-            _ => {
-                // Non-broken tenants must be constructed with a deletion queue
-                assert!(deletion_queue_client.is_some());
-            }
-        }
         let (state, mut rx) = watch::channel(state);
 
         tokio::spawn(async move {
@@ -2372,6 +2363,7 @@ impl Tenant {
 
         Tenant {
             tenant_id,
+            generation,
             conf,
             // using now here is good enough approximation to catch tenants with really long
             // activation times.
@@ -2381,7 +2373,6 @@ impl Tenant {
             gc_cs: tokio::sync::Mutex::new(()),
             walredo_mgr,
             remote_storage,
-            deletion_queue_client,
             state,
             cached_logical_sizes: tokio::sync::Mutex::new(HashMap::new()),
             cached_synthetic_tenant_size: Arc::new(AtomicU64::new(0)),
@@ -2810,13 +2801,15 @@ impl Tenant {
             src_timeline.pg_version,
         );
 
-        let uninitialized_timeline = self.prepare_new_timeline(
-            dst_id,
-            &metadata,
-            timeline_uninit_mark,
-            start_lsn + 1,
-            Some(Arc::clone(src_timeline)),
-        )?;
+        let uninitialized_timeline = self
+            .prepare_new_timeline(
+                dst_id,
+                &metadata,
+                timeline_uninit_mark,
+                start_lsn + 1,
+                Some(Arc::clone(src_timeline)),
+            )
+            .await?;
 
         let new_timeline = uninitialized_timeline.finish_creation()?;
 
@@ -2894,13 +2887,15 @@ impl Tenant {
             pgdata_lsn,
             pg_version,
         );
-        let raw_timeline = self.prepare_new_timeline(
-            timeline_id,
-            &new_metadata,
-            timeline_uninit_mark,
-            pgdata_lsn,
-            None,
-        )?;
+        let raw_timeline = self
+            .prepare_new_timeline(
+                timeline_id,
+                &new_metadata,
+                timeline_uninit_mark,
+                pgdata_lsn,
+                None,
+            )
+            .await?;
 
         let tenant_id = raw_timeline.owning_tenant.tenant_id;
         let unfinished_timeline = raw_timeline.raw_timeline()?;
@@ -2955,16 +2950,14 @@ impl Tenant {
                 self.conf,
                 self.tenant_id,
                 timeline_id,
+                self.generation,
             );
             Some(remote_client)
         } else {
             None
         };
 
-        TimelineResources {
-            remote_client,
-            deletion_queue_client: self.deletion_queue_client.clone(),
-        }
+        TimelineResources { remote_client }
     }
 
     /// Creates intermediate timeline structure and its files.
@@ -2973,7 +2966,7 @@ impl Tenant {
     /// at 'disk_consistent_lsn'. After any initial data has been imported, call
     /// `finish_creation` to insert the Timeline into the timelines map and to remove the
     /// uninit mark file.
-    fn prepare_new_timeline(
+    async fn prepare_new_timeline(
         &self,
         new_timeline_id: TimelineId,
         new_metadata: &TimelineMetadata,
@@ -3001,8 +2994,9 @@ impl Tenant {
 
         timeline_struct.init_empty_layer_map(start_lsn);
 
-        if let Err(e) =
-            self.create_timeline_files(&uninit_mark.timeline_path, &new_timeline_id, new_metadata)
+        if let Err(e) = self
+            .create_timeline_files(&uninit_mark.timeline_path, &new_timeline_id, new_metadata)
+            .await
         {
             error!("Failed to create initial files for timeline {tenant_id}/{new_timeline_id}, cleaning up: {e:?}");
             cleanup_timeline_directory(uninit_mark);
@@ -3018,7 +3012,7 @@ impl Tenant {
         ))
     }
 
-    fn create_timeline_files(
+    async fn create_timeline_files(
         &self,
         timeline_path: &Path,
         new_timeline_id: &TimelineId,
@@ -3037,6 +3031,7 @@ impl Tenant {
             new_metadata,
             true,
         )
+        .await
         .context("Failed to create timeline metadata")?;
         Ok(())
     }
@@ -3481,6 +3476,7 @@ pub mod harness {
         pub conf: &'static PageServerConf,
         pub tenant_conf: TenantConf,
         pub tenant_id: TenantId,
+        pub generation: Generation,
     }
 
     static LOG_HANDLE: OnceCell<()> = OnceCell::new();
@@ -3522,13 +3518,14 @@ pub mod harness {
                 conf,
                 tenant_conf,
                 tenant_id,
+                generation: Generation::new(0xdeadbeef),
             })
         }
 
         pub async fn load(&self) -> (Arc<Tenant>, RequestContext) {
             let ctx = RequestContext::new(TaskKind::UnitTest, DownloadBehavior::Error);
             (
-                self.try_load(&ctx, None, None)
+                self.try_load(&ctx, None)
                     .await
                     .expect("failed to load test tenant"),
                 ctx,
@@ -3539,7 +3536,6 @@ pub mod harness {
             &self,
             ctx: &RequestContext,
             remote_storage: Option<remote_storage::GenericRemoteStorage>,
-            deletion_queue_client: Option<DeletionQueueClient>,
         ) -> anyhow::Result<Arc<Tenant>> {
             let walredo_mgr = Arc::new(TestRedoManager);
 
@@ -3549,8 +3545,8 @@ pub mod harness {
                 TenantConfOpt::from(self.tenant_conf),
                 walredo_mgr,
                 self.tenant_id,
+                self.generation,
                 remote_storage,
-                deletion_queue_client,
             ));
             tenant
                 .load(None, ctx)
@@ -3663,7 +3659,10 @@ mod tests {
             .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
             .await?;
 
-        match tenant.create_empty_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx) {
+        match tenant
+            .create_empty_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
+            .await
+        {
             Ok(_) => panic!("duplicate timeline creation should fail"),
             Err(e) => assert_eq!(
                 e.to_string(),
@@ -4115,7 +4114,7 @@ mod tests {
         std::fs::write(metadata_path, metadata_bytes)?;
 
         let err = harness
-            .try_load(&ctx, None, None)
+            .try_load(&ctx, None)
             .await
             .err()
             .expect("should fail");
@@ -4503,8 +4502,9 @@ mod tests {
             .await;
 
         let initdb_lsn = Lsn(0x20);
-        let utline =
-            tenant.create_empty_timeline(TIMELINE_ID, initdb_lsn, DEFAULT_PG_VERSION, &ctx)?;
+        let utline = tenant
+            .create_empty_timeline(TIMELINE_ID, initdb_lsn, DEFAULT_PG_VERSION, &ctx)
+            .await?;
         let tline = utline.raw_timeline().unwrap();
 
         // Spawn flush loop now so that we can set the `expect_initdb_optimization`
@@ -4569,8 +4569,9 @@ mod tests {
         let harness = TenantHarness::create(name)?;
         {
             let (tenant, ctx) = harness.load().await;
-            let tline =
-                tenant.create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION, &ctx)?;
+            let tline = tenant
+                .create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION, &ctx)
+                .await?;
             // Keeps uninit mark in place
             std::mem::forget(tline);
         }

pageserver/src/tenant/blob_io.rs

@@ -96,18 +96,12 @@ pub trait BlobWriter {
 /// An implementation of BlobWriter to write blobs to anything that
 /// implements std::io::Write.
 ///
-pub struct WriteBlobWriter<W>
-where
-    W: std::io::Write,
-{
+pub struct WriteBlobWriter<W> {
     inner: W,
     offset: u64,
 }
 
-impl<W> WriteBlobWriter<W>
-where
-    W: std::io::Write,
-{
+impl<W> WriteBlobWriter<W> {
     pub fn new(inner: W, start_offset: u64) -> Self {
         WriteBlobWriter {
             inner,
@@ -129,33 +123,38 @@ where
     }
 }
 
-impl<W> BlobWriter for WriteBlobWriter<W>
-where
-    W: std::io::Write,
-{
-    fn write_blob(&mut self, srcbuf: &[u8]) -> Result<u64, Error> {
-        let offset = self.offset;
+macro_rules! write_blob_impl {
+    (WriteBlobWriter<$ty:ty>) => {
+        impl WriteBlobWriter<$ty> {
+            pub async fn write_blob(&mut self, srcbuf: &[u8]) -> Result<u64, Error> {
+                use std::io::Write;
+                let offset = self.offset;
 
-        if srcbuf.len() < 128 {
-            // Short blob. Write a 1-byte length header
-            let len_buf = srcbuf.len() as u8;
-            self.inner.write_all(&[len_buf])?;
-            self.offset += 1;
-        } else {
-            // Write a 4-byte length header
-            if srcbuf.len() > 0x7fff_ffff {
-                return Err(Error::new(
-                    ErrorKind::Other,
-                    format!("blob too large ({} bytes)", srcbuf.len()),
-                ));
+                if srcbuf.len() < 128 {
+                    // Short blob. Write a 1-byte length header
+                    let len_buf = srcbuf.len() as u8;
+                    self.inner.write_all(&[len_buf])?;
+                    self.offset += 1;
+                } else {
+                    // Write a 4-byte length header
+                    if srcbuf.len() > 0x7fff_ffff {
+                        return Err(Error::new(
+                            ErrorKind::Other,
+                            format!("blob too large ({} bytes)", srcbuf.len()),
+                        ));
+                    }
+                    let mut len_buf = ((srcbuf.len()) as u32).to_be_bytes();
+                    len_buf[0] |= 0x80;
+                    self.inner.write_all(&len_buf)?;
+                    self.offset += 4;
+                }
+                self.inner.write_all(srcbuf)?;
+                self.offset += srcbuf.len() as u64;
+                Ok(offset)
             }
-            let mut len_buf = ((srcbuf.len()) as u32).to_be_bytes();
-            len_buf[0] |= 0x80;
-            self.inner.write_all(&len_buf)?;
-            self.offset += 4;
         }
-        self.inner.write_all(srcbuf)?;
-        self.offset += srcbuf.len() as u64;
-        Ok(offset)
-    }
+    };
 }
+
+write_blob_impl!(WriteBlobWriter<crate::tenant::io::BufWriter<crate::virtual_file::VirtualFile>>);
+write_blob_impl!(WriteBlobWriter<crate::virtual_file::VirtualFile>);

pageserver/src/tenant/block_io.rs

@@ -7,9 +7,7 @@ use super::storage_layer::delta_layer::{Adapter, DeltaLayerInner};
 use crate::page_cache::{self, PageReadGuard, ReadBufResult, PAGE_SZ};
 use crate::virtual_file::VirtualFile;
 use bytes::Bytes;
-use std::fs::File;
 use std::ops::{Deref, DerefMut};
-use std::os::unix::fs::FileExt;
 
 /// This is implemented by anything that can read 8 kB (PAGE_SZ)
 /// blocks, using the page cache
@@ -74,7 +72,6 @@ impl<'a> Deref for BlockLease<'a> {
 /// Unlike traits, we also support the read function to be async though.
 pub(crate) enum BlockReaderRef<'a> {
     FileBlockReaderVirtual(&'a FileBlockReader<VirtualFile>),
-    FileBlockReaderFile(&'a FileBlockReader<std::fs::File>),
     EphemeralFile(&'a EphemeralFile),
     Adapter(Adapter<&'a DeltaLayerInner>),
     #[cfg(test)]
@@ -87,7 +84,6 @@ impl<'a> BlockReaderRef<'a> {
         use BlockReaderRef::*;
         match self {
             FileBlockReaderVirtual(r) => r.read_blk(blknum).await,
-            FileBlockReaderFile(r) => r.read_blk(blknum).await,
             EphemeralFile(r) => r.read_blk(blknum).await,
             Adapter(r) => r.read_blk(blknum).await,
             #[cfg(test)]
@@ -150,18 +146,15 @@ pub struct FileBlockReader<F> {
     file_id: page_cache::FileId,
 }
 
-impl<F> FileBlockReader<F>
-where
-    F: FileExt,
-{
-    pub fn new(file: F) -> Self {
+impl FileBlockReader<VirtualFile> {
+    pub fn new(file: VirtualFile) -> Self {
         let file_id = page_cache::next_file_id();
 
         FileBlockReader { file_id, file }
     }
 
     /// Read a page from the underlying file into given buffer.
-    fn fill_buffer(&self, buf: &mut [u8], blkno: u32) -> Result<(), std::io::Error> {
+    async fn fill_buffer(&self, buf: &mut [u8], blkno: u32) -> Result<(), std::io::Error> {
         assert!(buf.len() == PAGE_SZ);
         self.file.read_exact_at(buf, blkno as u64 * PAGE_SZ as u64)
     }
@@ -185,7 +178,7 @@ where
                 ReadBufResult::Found(guard) => break Ok(guard.into()),
                 ReadBufResult::NotFound(mut write_guard) => {
                     // Read the page from disk into the buffer
-                    self.fill_buffer(write_guard.deref_mut(), blknum)?;
+                    self.fill_buffer(write_guard.deref_mut(), blknum).await?;
                     write_guard.mark_valid();
 
                     // Swap for read lock
@@ -196,12 +189,6 @@ where
     }
 }
 
-impl BlockReader for FileBlockReader<File> {
-    fn block_cursor(&self) -> BlockCursor<'_> {
-        BlockCursor::new(BlockReaderRef::FileBlockReaderFile(self))
-    }
-}
-
 impl BlockReader for FileBlockReader<VirtualFile> {
     fn block_cursor(&self) -> BlockCursor<'_> {
         BlockCursor::new(BlockReaderRef::FileBlockReaderVirtual(self))

pageserver/src/tenant/ephemeral_file.rs

@@ -9,7 +9,6 @@ use std::cmp::min;
 use std::fs::OpenOptions;
 use std::io::{self, ErrorKind};
 use std::ops::DerefMut;
-use std::os::unix::prelude::FileExt;
 use std::path::PathBuf;
 use std::sync::atomic::AtomicU64;
 use tracing::*;
@@ -128,10 +127,15 @@ impl EphemeralFile {
                     self.off += n;
                     src_remaining = &src_remaining[n..];
                     if self.off == PAGE_SZ {
-                        match self.ephemeral_file.file.write_all_at(
-                            &self.ephemeral_file.mutable_tail,
-                            self.blknum as u64 * PAGE_SZ as u64,
-                        ) {
+                        match self
+                            .ephemeral_file
+                            .file
+                            .write_all_at(
+                                &self.ephemeral_file.mutable_tail,
+                                self.blknum as u64 * PAGE_SZ as u64,
+                            )
+                            .await
+                        {
                             Ok(_) => {
                                 // Pre-warm the page cache with what we just wrote.
                                 // This isn't necessary for coherency/correctness, but it's how we've always done it.

pageserver/src/tenant/manifest.rs

@@ -26,7 +26,7 @@
 //! recovered from this file. This is tracked in
 //! <https://github.com/neondatabase/neon/issues/4418>
 
-use std::io::{self, Read, Write};
+use std::io::{self, Write};
 
 use crate::virtual_file::VirtualFile;
 use anyhow::Result;
@@ -151,11 +151,12 @@ impl Manifest {
     /// Load a manifest. Returns the manifest and a list of operations. If the manifest is corrupted,
     /// the bool flag will be set to true and the user is responsible to reconstruct a new manifest and
     /// backup the current one.
-    pub fn load(
-        mut file: VirtualFile,
+    pub async fn load(
+        file: VirtualFile,
     ) -> Result<(Self, Vec<Operation>, ManifestPartiallyCorrupted), ManifestLoadError> {
         let mut buf = vec![];
-        file.read_to_end(&mut buf).map_err(ManifestLoadError::Io)?;
+        file.read_exact_at(&mut buf, 0)
+            .map_err(ManifestLoadError::Io)?;
 
         // Read manifest header
         let mut buf = Bytes::from(buf);
@@ -241,8 +242,8 @@ mod tests {
 
     use super::*;
 
-    #[test]
-    fn test_read_manifest() {
+    #[tokio::test]
+    async fn test_read_manifest() {
         let testdir = crate::config::PageServerConf::test_repo_dir("test_read_manifest");
         std::fs::create_dir_all(&testdir).unwrap();
         let file = VirtualFile::create(&testdir.join("MANIFEST")).unwrap();
@@ -274,7 +275,7 @@ mod tests {
                 .truncate(false),
         )
         .unwrap();
-        let (mut manifest, operations, corrupted) = Manifest::load(file).unwrap();
+        let (mut manifest, operations, corrupted) = Manifest::load(file).await.unwrap();
         assert!(!corrupted.0);
         assert_eq!(operations.len(), 2);
         assert_eq!(
@@ -306,7 +307,7 @@ mod tests {
                 .truncate(false),
         )
         .unwrap();
-        let (_manifest, operations, corrupted) = Manifest::load(file).unwrap();
+        let (_manifest, operations, corrupted) = Manifest::load(file).await.unwrap();
         assert!(!corrupted.0);
         assert_eq!(operations.len(), 3);
         assert_eq!(&operations[0], &Operation::Snapshot(snapshot, Lsn::from(0)));

pageserver/src/tenant/metadata.rs

@@ -9,9 +9,9 @@
 //! [`remote_timeline_client`]: super::remote_timeline_client
 
 use std::fs::{File, OpenOptions};
-use std::io::{self, Write};
+use std::io;
 
-use anyhow::{bail, ensure, Context};
+use anyhow::{ensure, Context};
 use serde::{de::Error, Deserialize, Serialize, Serializer};
 use thiserror::Error;
 use tracing::info_span;
@@ -255,7 +255,7 @@ impl Serialize for TimelineMetadata {
 }
 
 /// Save timeline metadata to file
-pub fn save_metadata(
+pub async fn save_metadata(
     conf: &'static PageServerConf,
     tenant_id: &TenantId,
     timeline_id: &TimelineId,
@@ -273,10 +273,7 @@ pub fn save_metadata(
 
     let metadata_bytes = data.to_bytes().context("Failed to get metadata bytes")?;
 
-    if file.write(&metadata_bytes)? != metadata_bytes.len() {
-        bail!("Could not write all the metadata bytes in a single call");
-    }
-    file.sync_all()?;
+    file.write_and_fsync(&metadata_bytes)?;
 
     // fsync the parent directory to ensure the directory entry is durable
     if first_save {

pageserver/src/tenant/mgr.rs

@@ -18,7 +18,6 @@ use utils::crashsafe;
 
 use crate::config::PageServerConf;
 use crate::context::{DownloadBehavior, RequestContext};
-use crate::deletion_queue::DeletionQueue;
 use crate::task_mgr::{self, TaskKind};
 use crate::tenant::config::TenantConfOpt;
 use crate::tenant::delete::DeleteTenantFlow;
@@ -26,6 +25,7 @@ use crate::tenant::{create_tenant_files, CreateTenantFilesMode, Tenant, TenantSt
 use crate::{InitializationOrder, IGNORED_TENANT_FILE_NAME};
 
 use utils::fs_ext::PathExt;
+use utils::generation::Generation;
 use utils::id::{TenantId, TimelineId};
 
 use super::delete::DeleteTenantError;
@@ -203,10 +203,10 @@ pub(crate) fn schedule_local_tenant_processing(
             match Tenant::spawn_attach(
                 conf,
                 tenant_id,
+                Generation::none(),
                 resources.broker_client,
                 tenants,
                 remote_storage,
-                resources.deletion_queue_client,
                 ctx,
             ) {
                 Ok(tenant) => tenant,
@@ -226,7 +226,15 @@ pub(crate) fn schedule_local_tenant_processing(
     } else {
         info!("tenant {tenant_id} is assumed to be loadable, starting load operation");
         // Start loading the tenant into memory. It will initially be in Loading state.
-        Tenant::spawn_load(conf, tenant_id, resources, init_order, tenants, ctx)
+        Tenant::spawn_load(
+            conf,
+            tenant_id,
+            Generation::none(),
+            resources,
+            init_order,
+            tenants,
+            ctx,
+        )
     };
     Ok(tenant)
 }
@@ -351,7 +359,6 @@ pub async fn create_tenant(
     tenant_id: TenantId,
     broker_client: storage_broker::BrokerClientChannel,
     remote_storage: Option<GenericRemoteStorage>,
-    deletion_queue: &DeletionQueue,
     ctx: &RequestContext,
 ) -> Result<Arc<Tenant>, TenantMapInsertError> {
     tenant_map_insert(tenant_id, || {
@@ -365,7 +372,6 @@ pub async fn create_tenant(
         let tenant_resources = TenantSharedResources {
             broker_client,
             remote_storage,
-            deletion_queue_client: deletion_queue.new_client(),
         };
         let created_tenant =
             schedule_local_tenant_processing(conf, &tenant_directory, tenant_resources, None, &TENANTS, ctx)?;
@@ -517,7 +523,6 @@ pub async fn load_tenant(
     tenant_id: TenantId,
     broker_client: storage_broker::BrokerClientChannel,
     remote_storage: Option<GenericRemoteStorage>,
-    deletion_queue: &DeletionQueue,
     ctx: &RequestContext,
 ) -> Result<(), TenantMapInsertError> {
     tenant_map_insert(tenant_id, || {
@@ -531,7 +536,6 @@ pub async fn load_tenant(
         let resources = TenantSharedResources {
             broker_client,
             remote_storage,
-            deletion_queue_client: deletion_queue.new_client(),
         };
         let new_tenant = schedule_local_tenant_processing(conf, &tenant_path,  resources, None,  &TENANTS, ctx)
             .with_context(|| {
@@ -600,7 +604,6 @@ pub async fn attach_tenant(
     tenant_conf: TenantConfOpt,
     broker_client: storage_broker::BrokerClientChannel,
     remote_storage: GenericRemoteStorage,
-    deletion_queue: &DeletionQueue,
     ctx: &RequestContext,
 ) -> Result<(), TenantMapInsertError> {
     tenant_map_insert(tenant_id, || {
@@ -618,7 +621,6 @@ pub async fn attach_tenant(
         let resources = TenantSharedResources {
             broker_client,
             remote_storage: Some(remote_storage),
-            deletion_queue_client: deletion_queue.new_client(),
         };
         let attached_tenant = schedule_local_tenant_processing(conf, &tenant_dir, resources, None, &TENANTS, ctx)?;
         // TODO: tenant object & its background loops remain, untracked in tenant map, if we fail here.

pageserver/src/tenant/remote_timeline_client.rs

@@ -56,11 +56,9 @@
 //! # Consistency
 //!
 //! To have a consistent remote structure, it's important that uploads and
-//! deletions are performed in the right order. For example:
-//! - the index file contains a list of layer files, so it must not be uploaded
-//!    until all the layer files that are in its list have been successfully uploaded.
-//! - objects must be removed from the index before being deleted, and that updated
-//!   index must be written to remote storage before deleting the objects from remote storage.
+//! deletions are performed in the right order. For example, the index file
+//! contains a list of layer files, so it must not be uploaded until all the
+//! layer files that are in its list have been successfully uploaded.
 //!
 //! The contract between client and its user is that the user is responsible of
 //! scheduling operations in an order that keeps the remote consistent as
@@ -72,12 +70,10 @@
 //! correct order, and the client will parallelize the operations in a way that
 //! is safe.
 //!
-//! The caller should be careful with deletion, though:
-//! - they should not delete local files that have been scheduled for upload but
-//!   not yet finished uploading.  Otherwise the upload will fail. To wait for an
-//!   upload to finish, use the 'wait_completion' function (more on that later.)
-//! - they should not to remote deletions via DeletionQueue without waiting for
-//!   the latest metadata to upload via RemoteTimelineClient.
+//! The caller should be careful with deletion, though. They should not delete
+//! local files that have been scheduled for upload but not yet finished uploading.
+//! Otherwise the upload will fail. To wait for an upload to finish, use
+//! the 'wait_completion' function (more on that later.)
 //!
 //! All of this relies on the following invariants:
 //!
@@ -204,11 +200,12 @@
 //! [`Tenant::timeline_init_and_sync`]: super::Tenant::timeline_init_and_sync
 //! [`Timeline::load_layer_map`]: super::Timeline::load_layer_map
 
+mod delete;
 mod download;
 pub mod index;
 mod upload;
 
-use anyhow::{bail, Context};
+use anyhow::Context;
 use chrono::{NaiveDateTime, Utc};
 // re-export these
 pub use download::{is_temp_download_file, list_remote_timelines};
@@ -219,7 +216,7 @@ use utils::backoff::{
 };
 
 use std::collections::{HashMap, VecDeque};
-use std::path::Path;
+use std::path::{Path, PathBuf};
 use std::sync::atomic::{AtomicU32, Ordering};
 use std::sync::{Arc, Mutex};
 
@@ -229,7 +226,6 @@ use tracing::{debug, error, info, instrument, warn};
 use tracing::{info_span, Instrument};
 use utils::lsn::Lsn;
 
-use crate::deletion_queue::DeletionQueueClient;
 use crate::metrics::{
     MeasureRemoteOp, RemoteOpFileKind, RemoteOpKind, RemoteTimelineClientMetrics,
     RemoteTimelineClientMetricsCallTrackSize, REMOTE_ONDEMAND_DOWNLOADED_BYTES,
@@ -238,6 +234,8 @@ use crate::metrics::{
 use crate::task_mgr::shutdown_token;
 use crate::tenant::debug_assert_current_span_has_tenant_and_timeline_id;
 use crate::tenant::remote_timeline_client::index::LayerFileMetadata;
+use crate::tenant::upload_queue::Delete;
+use crate::tenant::TIMELINES_SEGMENT_NAME;
 use crate::{
     config::PageServerConf,
     task_mgr,
@@ -255,6 +253,7 @@ use self::index::IndexPart;
 
 use super::storage_layer::LayerFileName;
 use super::upload_queue::SetDeletedFlagProgress;
+use super::Generation;
 
 // Occasional network issues and such can cause remote operations to fail, and
 // that's expected. If a download fails, we log it at info-level, and retry.
@@ -318,6 +317,7 @@ pub struct RemoteTimelineClient {
 
     tenant_id: TenantId,
     timeline_id: TimelineId,
+    generation: Generation,
 
     upload_queue: Mutex<UploadQueue>,
 
@@ -338,12 +338,14 @@ impl RemoteTimelineClient {
         conf: &'static PageServerConf,
         tenant_id: TenantId,
         timeline_id: TimelineId,
+        generation: Generation,
     ) -> RemoteTimelineClient {
         RemoteTimelineClient {
             conf,
             runtime: BACKGROUND_RUNTIME.handle().to_owned(),
             tenant_id,
             timeline_id,
+            generation,
             storage_impl: remote_storage,
             upload_queue: Mutex::new(UploadQueue::Uninitialized),
             metrics: Arc::new(RemoteTimelineClientMetrics::new(&tenant_id, &timeline_id)),
@@ -452,10 +454,10 @@ impl RemoteTimelineClient {
         );
 
         let index_part = download::download_index_part(
-            self.conf,
             &self.storage_impl,
             &self.tenant_id,
             &self.timeline_id,
+            self.generation,
         )
         .measure_remote_op(
             self.tenant_id,
@@ -634,46 +636,70 @@ impl RemoteTimelineClient {
     /// deletion won't actually be performed, until any previously scheduled
     /// upload operations, and the index file upload, have completed
     /// successfully.
-    pub async fn schedule_layer_file_deletion(
+    pub fn schedule_layer_file_deletion(
         self: &Arc<Self>,
         names: &[LayerFileName],
-        deletion_queue_client: &DeletionQueueClient,
     ) -> anyhow::Result<()> {
-        // Synchronous update of upload queues under mutex
-        {
-            let mut guard = self.upload_queue.lock().unwrap();
-            let upload_queue = guard.initialized_mut()?;
+        let mut guard = self.upload_queue.lock().unwrap();
+        let upload_queue = guard.initialized_mut()?;
 
-            // Deleting layers doesn't affect the values stored in TimelineMetadata,
-            // so we don't need update it. Just serialize it.
-            let metadata = upload_queue.latest_metadata.clone();
+        // Deleting layers doesn't affect the values stored in TimelineMetadata,
+        // so we don't need update it. Just serialize it.
+        let metadata = upload_queue.latest_metadata.clone();
 
-            // Update the remote index file, removing the to-be-deleted files from the index,
-            // before deleting the actual files.
-            for name in names {
-                if upload_queue.latest_files.remove(name).is_some() {
-                    upload_queue.latest_files_changes_since_metadata_upload_scheduled += 1;
-                }
-            }
+        // Update the remote index file, removing the to-be-deleted files from the index,
+        // before deleting the actual files.
+        //
+        // Once we start removing files from upload_queue.latest_files, there's
+        // no going back! Otherwise, some of the files would already be removed
+        // from latest_files, but not yet scheduled for deletion. Use a closure
+        // to syntactically forbid ? or bail! calls here.
+        let no_bail_here = || {
+            // Decorate our list of names with each name's generation, dropping
+            // makes that are unexpectedly missing from our metadata.
+            let with_generations: Vec<_> = names
+                .iter()
+                .filter_map(|name| {
+                    // Remove from latest_files, learning the file's remote generation in the process
+                    let meta = upload_queue.latest_files.remove(name);
+
+                    if let Some(meta) = meta {
+                        upload_queue.latest_files_changes_since_metadata_upload_scheduled += 1;
+                        Some((name, meta.generation))
+                    } else {
+                        // This can only happen if we forgot to to schedule the file upload
+                        // before scheduling the delete. Log it because it is a rare/strange
+                        // situation, and in case something is misbehaving, we'd like to know which
+                        // layers experienced this.
+                        info!(
+                            "Deleting layer {name} not found in latest_files list, never uploaded?"
+                        );
+                        None
+                    }
+                })
+                .collect();
 
             if upload_queue.latest_files_changes_since_metadata_upload_scheduled > 0 {
                 self.schedule_index_upload(upload_queue, metadata);
             }
-        }
 
-        // Barrier: we must ensure all prior uploads and index writes have landed in S3
-        // before emitting deletions.
-        if let Err(e) = self.wait_completion().await {
-            // This can only fail if upload queue is shut down: if this happens, we do
-            // not emit any deletions.  In this condition (remote client is shut down
-            // during compaction or GC) we may leak some objects.
-            bail!("Cannot complete layer file deletions during shutdown ({e})");
-        }
+            // schedule the actual deletions
+            for (name, generation) in with_generations {
+                let op = UploadOp::Delete(Delete {
+                    file_kind: RemoteOpFileKind::Layer,
+                    layer_file_name: name.clone(),
+                    scheduled_from_timeline_delete: false,
+                    generation,
+                });
+                self.calls_unfinished_metric_begin(&op);
+                upload_queue.queued_operations.push_back(op);
+                info!("scheduled layer file deletion {name}");
+            }
 
-        // Enqueue deletions
-        deletion_queue_client
-            .push_layers(self.tenant_id, self.timeline_id, names.to_vec())
-            .await?;
+            // Launch the tasks immediately, if possible
+            self.launch_queued_tasks(upload_queue);
+        };
+        no_bail_here();
         Ok(())
     }
 
@@ -759,10 +785,10 @@ impl RemoteTimelineClient {
         backoff::retry(
             || {
                 upload::upload_index_part(
-                    self.conf,
                     &self.storage_impl,
                     &self.tenant_id,
                     &self.timeline_id,
+                    self.generation,
                     &index_part_with_deleted_at,
                 )
             },
@@ -799,13 +825,12 @@ impl RemoteTimelineClient {
     /// Prerequisites: UploadQueue should be in stopped state and deleted_at should be successfuly set.
     /// The function deletes layer files one by one, then lists the prefix to see if we leaked something
     /// deletes leaked files if any and proceeds with deletion of index file at the end.
-    pub(crate) async fn delete_all(
-        self: &Arc<Self>,
-        deletion_queue: &DeletionQueueClient,
-    ) -> anyhow::Result<()> {
+    pub(crate) async fn delete_all(self: &Arc<Self>) -> anyhow::Result<()> {
         debug_assert_current_span_has_tenant_and_timeline_id();
 
-        let layers: Vec<LayerFileName> = {
+        let (mut receiver, deletions_queued) = {
+            let mut deletions_queued = 0;
+
             let mut locked = self.upload_queue.lock().unwrap();
             let stopped = locked.stopped_mut()?;
 
@@ -817,34 +842,41 @@ impl RemoteTimelineClient {
 
             stopped
                 .upload_queue_for_deletion
-                .latest_files
-                .drain()
-                .map(|kv| kv.0)
-                .collect()
+                .queued_operations
+                .reserve(stopped.upload_queue_for_deletion.latest_files.len());
+
+            // schedule the actual deletions
+            for (name, meta) in &stopped.upload_queue_for_deletion.latest_files {
+                let op = UploadOp::Delete(Delete {
+                    file_kind: RemoteOpFileKind::Layer,
+                    layer_file_name: name.clone(),
+                    scheduled_from_timeline_delete: true,
+                    generation: meta.generation,
+                });
+
+                self.calls_unfinished_metric_begin(&op);
+                stopped
+                    .upload_queue_for_deletion
+                    .queued_operations
+                    .push_back(op);
+
+                info!("scheduled layer file deletion {name}");
+                deletions_queued += 1;
+            }
+
+            self.launch_queued_tasks(&mut stopped.upload_queue_for_deletion);
+
+            (
+                self.schedule_barrier(&mut stopped.upload_queue_for_deletion),
+                deletions_queued,
+            )
         };
 
-        let layer_deletion_count = layers.len();
-
-        let timeline_path = self.conf.timeline_path(&self.tenant_id, &self.timeline_id);
-        let layer_paths = layers
-            .into_iter()
-            .map(|l| {
-                let local_path = timeline_path.join(l.file_name());
-                self.conf
-                    .remote_path(&local_path)
-                    .expect("Timeline path should always convert to remote")
-            })
-            .collect();
-        deletion_queue.push_immediate(layer_paths).await?;
+        receiver.changed().await.context("upload queue shut down")?;
 
         // Do not delete index part yet, it is needed for possible retry. If we remove it first
         // and retry will arrive to different pageserver there wont be any traces of it on remote storage
-        let timeline_path = self.conf.timeline_path(&self.tenant_id, &self.timeline_id);
-        let timeline_storage_path = self.conf.remote_path(&timeline_path)?;
-
-        // Execute all pending deletions, so that when we prroceed to do a list_prefixes below, we aren't
-        // taking the burden of listing all the layers that we already know we should delete.
-        deletion_queue.flush_immediate().await?;
+        let timeline_storage_path = remote_timeline_path(&self.tenant_id, &self.timeline_id);
 
         let remaining = backoff::retry(
             || async {
@@ -873,9 +905,17 @@ impl RemoteTimelineClient {
             })
             .collect();
 
-        let not_referenced_count = remaining.len();
         if !remaining.is_empty() {
-            deletion_queue.push_immediate(remaining).await?;
+            backoff::retry(
+                || async { self.storage_impl.delete_objects(&remaining).await },
+                |_e| false,
+                FAILED_UPLOAD_WARN_THRESHOLD,
+                FAILED_REMOTE_OP_RETRIES,
+                "delete_objects",
+                backoff::Cancel::new(shutdown_token(), || anyhow::anyhow!("Cancelled!")),
+            )
+            .await
+            .context("delete_objects")?;
         }
 
         fail::fail_point!("timeline-delete-before-index-delete", |_| {
@@ -886,14 +926,18 @@ impl RemoteTimelineClient {
 
         let index_file_path = timeline_storage_path.join(Path::new(IndexPart::FILE_NAME));
 
-        debug!("enqueuing index part deletion");
-        deletion_queue
-            .push_immediate([index_file_path].to_vec())
-            .await?;
+        debug!("deleting index part");
 
-        // Timeline deletion is rare and we have probably emitted a reasonably number of objects: wait
-        // for a flush to a persistent deletion list so that we may be sure deletion will occur.
-        deletion_queue.flush_immediate().await?;
+        backoff::retry(
+            || async { self.storage_impl.delete(&index_file_path).await },
+            |_e| false,
+            FAILED_UPLOAD_WARN_THRESHOLD,
+            FAILED_REMOTE_OP_RETRIES,
+            "delete_index",
+            backoff::Cancel::new(shutdown_token(), || anyhow::anyhow!("Cancelled")),
+        )
+        .await
+        .context("delete_index")?;
 
         fail::fail_point!("timeline-delete-after-index-delete", |_| {
             Err(anyhow::anyhow!(
@@ -901,7 +945,7 @@ impl RemoteTimelineClient {
             ))?
         });
 
-        info!(prefix=%timeline_storage_path, referenced=layer_deletion_count, not_referenced=%not_referenced_count, "done deleting in timeline prefix, including index_part.json");
+        info!(prefix=%timeline_storage_path, referenced=deletions_queued, not_referenced=%remaining.len(), "done deleting in timeline prefix, including index_part.json");
 
         Ok(())
     }
@@ -924,6 +968,10 @@ impl RemoteTimelineClient {
                     // have finished.
                     upload_queue.inprogress_tasks.is_empty()
                 }
+                UploadOp::Delete(_) => {
+                    // Wait for preceding uploads to finish. Concurrent deletions are OK, though.
+                    upload_queue.num_inprogress_deletions == upload_queue.inprogress_tasks.len()
+                }
 
                 UploadOp::Barrier(_) => upload_queue.inprogress_tasks.is_empty(),
             };
@@ -951,6 +999,9 @@ impl RemoteTimelineClient {
                 UploadOp::UploadMetadata(_, _) => {
                     upload_queue.num_inprogress_metadata_uploads += 1;
                 }
+                UploadOp::Delete(_) => {
+                    upload_queue.num_inprogress_deletions += 1;
+                }
                 UploadOp::Barrier(sender) => {
                     sender.send_replace(());
                     continue;
@@ -1029,15 +1080,17 @@ impl RemoteTimelineClient {
 
             let upload_result: anyhow::Result<()> = match &task.op {
                 UploadOp::UploadLayer(ref layer_file_name, ref layer_metadata) => {
-                    let path = &self
+                    let path = self
                         .conf
                         .timeline_path(&self.tenant_id, &self.timeline_id)
                         .join(layer_file_name.file_name());
+
                     upload::upload_timeline_layer(
                         self.conf,
                         &self.storage_impl,
-                        path,
+                        &path,
                         layer_metadata,
+                        self.generation,
                     )
                     .measure_remote_op(
                         self.tenant_id,
@@ -1059,10 +1112,10 @@ impl RemoteTimelineClient {
                     };
 
                     let res = upload::upload_index_part(
-                        self.conf,
                         &self.storage_impl,
                         &self.tenant_id,
                         &self.timeline_id,
+                        self.generation,
                         index_part,
                     )
                     .measure_remote_op(
@@ -1082,6 +1135,21 @@ impl RemoteTimelineClient {
                     }
                     res
                 }
+                UploadOp::Delete(delete) => {
+                    let path = &self
+                        .conf
+                        .timeline_path(&self.tenant_id, &self.timeline_id)
+                        .join(delete.layer_file_name.file_name());
+                    delete::delete_layer(self.conf, &self.storage_impl, path, delete.generation)
+                        .measure_remote_op(
+                            self.tenant_id,
+                            self.timeline_id,
+                            delete.file_kind,
+                            RemoteOpKind::Delete,
+                            Arc::clone(&self.metrics),
+                        )
+                        .await
+                }
                 UploadOp::Barrier(_) => {
                     // unreachable. Barrier operations are handled synchronously in
                     // launch_queued_tasks
@@ -1141,7 +1209,15 @@ impl RemoteTimelineClient {
             let mut upload_queue_guard = self.upload_queue.lock().unwrap();
             let upload_queue = match upload_queue_guard.deref_mut() {
                 UploadQueue::Uninitialized => panic!("callers are responsible for ensuring this is only called on an initialized queue"),
-                UploadQueue::Stopped(_) => { None }
+                UploadQueue::Stopped(stopped) => {
+                    // Special care is needed for deletions, if it was an earlier deletion (not scheduled from deletion)
+                    // then stop() took care of it so we just return.
+                    // For deletions that come from delete_all we still want to maintain metrics, launch following tasks, etc.
+                    match &task.op {
+                        UploadOp::Delete(delete) if delete.scheduled_from_timeline_delete => Some(&mut stopped.upload_queue_for_deletion),
+                        _ => None
+                    }
+                },
                 UploadQueue::Initialized(qi) => { Some(qi) }
             };
 
@@ -1163,6 +1239,9 @@ impl RemoteTimelineClient {
                     upload_queue.num_inprogress_metadata_uploads -= 1;
                     upload_queue.last_uploaded_consistent_lsn = lsn; // XXX monotonicity check?
                 }
+                UploadOp::Delete(_) => {
+                    upload_queue.num_inprogress_deletions -= 1;
+                }
                 UploadOp::Barrier(_) => unreachable!(),
             };
 
@@ -1194,6 +1273,13 @@ impl RemoteTimelineClient {
                     reason: "metadata uploads are tiny",
                 },
             ),
+            UploadOp::Delete(delete) => (
+                delete.file_kind,
+                RemoteOpKind::Delete,
+                DontTrackSize {
+                    reason: "should we track deletes? positive or negative sign?",
+                },
+            ),
             UploadOp::Barrier(_) => {
                 // we do not account these
                 return None;
@@ -1253,6 +1339,7 @@ impl RemoteTimelineClient {
                         last_uploaded_consistent_lsn: initialized.last_uploaded_consistent_lsn,
                         num_inprogress_layer_uploads: 0,
                         num_inprogress_metadata_uploads: 0,
+                        num_inprogress_deletions: 0,
                         inprogress_tasks: HashMap::default(),
                         queued_operations: VecDeque::default(),
                     };
@@ -1273,7 +1360,9 @@ impl RemoteTimelineClient {
 
                 // consistency check
                 assert_eq!(
-                    qi.num_inprogress_layer_uploads + qi.num_inprogress_metadata_uploads,
+                    qi.num_inprogress_layer_uploads
+                        + qi.num_inprogress_metadata_uploads
+                        + qi.num_inprogress_deletions,
                     qi.inprogress_tasks.len()
                 );
 
@@ -1298,15 +1387,79 @@ impl RemoteTimelineClient {
     }
 }
 
+pub fn remote_timelines_path(tenant_id: &TenantId) -> RemotePath {
+    let path = format!("tenants/{tenant_id}/{TIMELINES_SEGMENT_NAME}");
+    RemotePath::from_string(&path).expect("Failed to construct path")
+}
+
+pub fn remote_timeline_path(tenant_id: &TenantId, timeline_id: &TimelineId) -> RemotePath {
+    remote_timelines_path(tenant_id).join(&PathBuf::from(timeline_id.to_string()))
+}
+
+pub fn remote_layer_path(
+    tenant_id: &TenantId,
+    timeline_id: &TimelineId,
+    layer_file_name: &LayerFileName,
+    layer_meta: &LayerFileMetadata,
+) -> RemotePath {
+    // Generation-aware key format
+    let path = format!(
+        "tenants/{tenant_id}/{TIMELINES_SEGMENT_NAME}/{timeline_id}/{0}{1}",
+        layer_file_name.file_name(),
+        layer_meta.generation.get_suffix()
+    );
+
+    RemotePath::from_string(&path).expect("Failed to construct path")
+}
+
+pub fn remote_index_path(
+    tenant_id: &TenantId,
+    timeline_id: &TimelineId,
+    generation: Generation,
+) -> RemotePath {
+    RemotePath::from_string(&format!(
+        "tenants/{tenant_id}/{TIMELINES_SEGMENT_NAME}/{timeline_id}/{0}{1}",
+        IndexPart::FILE_NAME,
+        generation.get_suffix()
+    ))
+    .expect("Failed to construct path")
+}
+
+/// Files on the remote storage are stored with paths, relative to the workdir.
+/// That path includes in itself both tenant and timeline ids, allowing to have a unique remote storage path.
+///
+/// Errors if the path provided does not start from pageserver's workdir.
+pub fn remote_path(
+    conf: &PageServerConf,
+    local_path: &Path,
+    generation: Generation,
+) -> anyhow::Result<RemotePath> {
+    let stripped = local_path
+        .strip_prefix(&conf.workdir)
+        .context("Failed to strip workdir prefix")?;
+
+    let suffixed = format!(
+        "{0}{1}",
+        stripped.to_string_lossy(),
+        generation.get_suffix()
+    );
+
+    RemotePath::new(&PathBuf::from(suffixed)).with_context(|| {
+        format!(
+            "to resolve remote part of path {:?} for base {:?}",
+            local_path, conf.workdir
+        )
+    })
+}
+
 #[cfg(test)]
 mod tests {
     use super::*;
     use crate::{
         context::RequestContext,
-        deletion_queue::mock::MockDeletionQueue,
         tenant::{
             harness::{TenantHarness, TIMELINE_ID},
-            Tenant, Timeline,
+            Generation, Tenant, Timeline,
         },
         DEFAULT_PG_VERSION,
     };
@@ -1348,8 +1501,11 @@ mod tests {
         assert_eq!(avec, bvec);
     }
 
-    fn assert_remote_files(expected: &[&str], remote_path: &Path) {
-        let mut expected: Vec<String> = expected.iter().map(|x| String::from(*x)).collect();
+    fn assert_remote_files(expected: &[&str], remote_path: &Path, generation: Generation) {
+        let mut expected: Vec<String> = expected
+            .iter()
+            .map(|x| format!("{}{}", x, generation.get_suffix()))
+            .collect();
         expected.sort();
 
         let mut found: Vec<String> = Vec::new();
@@ -1370,7 +1526,6 @@ mod tests {
         tenant_ctx: RequestContext,
         remote_fs_dir: PathBuf,
         client: Arc<RemoteTimelineClient>,
-        deletion_queue: MockDeletionQueue,
     }
 
     impl TestSetup {
@@ -1401,6 +1556,8 @@ mod tests {
                 storage: RemoteStorageKind::LocalFs(remote_fs_dir.clone()),
             };
 
+            let generation = Generation::new(0xdeadbeef);
+
             let storage = GenericRemoteStorage::from_config(&storage_config).unwrap();
 
             let client = Arc::new(RemoteTimelineClient {
@@ -1408,7 +1565,8 @@ mod tests {
                 runtime: tokio::runtime::Handle::current(),
                 tenant_id: harness.tenant_id,
                 timeline_id: TIMELINE_ID,
-                storage_impl: storage.clone(),
+                generation,
+                storage_impl: storage,
                 upload_queue: Mutex::new(UploadQueue::Uninitialized),
                 metrics: Arc::new(RemoteTimelineClientMetrics::new(
                     &harness.tenant_id,
@@ -1416,8 +1574,6 @@ mod tests {
                 )),
             });
 
-            let deletion_queue = MockDeletionQueue::new(Some(storage), harness.conf);
-
             Ok(Self {
                 harness,
                 tenant,
@@ -1425,7 +1581,6 @@ mod tests {
                 tenant_ctx: ctx,
                 remote_fs_dir,
                 client,
-                deletion_queue,
             })
         }
     }
@@ -1454,7 +1609,6 @@ mod tests {
             tenant_ctx: _tenant_ctx,
             remote_fs_dir,
             client,
-            deletion_queue,
         } = TestSetup::new("upload_scheduling").await.unwrap();
 
         let timeline_path = harness.timeline_path(&TIMELINE_ID);
@@ -1470,6 +1624,8 @@ mod tests {
             .init_upload_queue_for_empty_remote(&metadata)
             .unwrap();
 
+        let generation = Generation::new(0xdeadbeef);
+
         // Create a couple of dummy files,  schedule upload for them
         let layer_file_name_1: LayerFileName = "000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__00000000016B59D8-00000000016B5A51".parse().unwrap();
         let layer_file_name_2: LayerFileName = "000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__00000000016B59D9-00000000016B5A52".parse().unwrap();
@@ -1489,13 +1645,13 @@ mod tests {
         client
             .schedule_layer_file_upload(
                 &layer_file_name_1,
-                &LayerFileMetadata::new(content_1.len() as u64),
+                &LayerFileMetadata::new(content_1.len() as u64, generation),
             )
             .unwrap();
         client
             .schedule_layer_file_upload(
                 &layer_file_name_2,
-                &LayerFileMetadata::new(content_2.len() as u64),
+                &LayerFileMetadata::new(content_2.len() as u64, generation),
             )
             .unwrap();
 
@@ -1559,66 +1715,35 @@ mod tests {
         client
             .schedule_layer_file_upload(
                 &layer_file_name_3,
-                &LayerFileMetadata::new(content_3.len() as u64),
+                &LayerFileMetadata::new(content_3.len() as u64, generation),
             )
             .unwrap();
-
-        {
-            let mut guard = client.upload_queue.lock().unwrap();
-            let upload_queue = guard.initialized_mut().unwrap();
-            assert_eq!(upload_queue.queued_operations.len(), 0);
-            assert_eq!(upload_queue.num_inprogress_layer_uploads, 1);
-        }
-
-        assert_remote_files(
-            &[
-                &layer_file_name_1.file_name(),
-                &layer_file_name_2.file_name(),
-                "index_part.json",
-            ],
-            &remote_timeline_dir,
-        );
-
         client
-            .schedule_layer_file_deletion(
-                &[layer_file_name_1.clone()],
-                &deletion_queue.new_client(),
-            )
-            .await
+            .schedule_layer_file_deletion(&[layer_file_name_1.clone()])
             .unwrap();
-
         {
             let mut guard = client.upload_queue.lock().unwrap();
             let upload_queue = guard.initialized_mut().unwrap();
 
-            // Deletion schedules upload of the index file via RemoteTimelineClient, and
-            // deletion of layer files via DeletionQueue.  The uploads have all been flushed
-            // because schedule_layer_file_deletion does a wait_completion before pushing
-            // to the deletion_queue
-            assert_eq!(upload_queue.queued_operations.len(), 0);
-            assert_eq!(upload_queue.inprogress_tasks.len(), 0);
-            assert_eq!(upload_queue.num_inprogress_layer_uploads, 0);
-            assert_eq!(
-                upload_queue.latest_files_changes_since_metadata_upload_scheduled,
-                0
-            );
+            // Deletion schedules upload of the index file, and the file deletion itself
+            assert!(upload_queue.queued_operations.len() == 2);
+            assert!(upload_queue.inprogress_tasks.len() == 1);
+            assert!(upload_queue.num_inprogress_layer_uploads == 1);
+            assert!(upload_queue.num_inprogress_deletions == 0);
+            assert!(upload_queue.latest_files_changes_since_metadata_upload_scheduled == 0);
         }
         assert_remote_files(
             &[
                 &layer_file_name_1.file_name(),
                 &layer_file_name_2.file_name(),
-                &layer_file_name_3.file_name(),
                 "index_part.json",
             ],
             &remote_timeline_dir,
+            generation,
         );
 
-        // Finish uploads and deletions
+        // Finish them
         client.wait_completion().await.unwrap();
-        deletion_queue.pump().await;
-
-        // 1 layer was deleted
-        assert_eq!(deletion_queue.get_executed(), 1);
 
         assert_remote_files(
             &[
@@ -1627,6 +1752,7 @@ mod tests {
                 "index_part.json",
             ],
             &remote_timeline_dir,
+            generation,
         );
     }
 
@@ -1679,12 +1805,14 @@ mod tests {
 
         // Test
 
+        let generation = Generation::new(0xdeadbeef);
+
         let init = get_bytes_started_stopped();
 
         client
             .schedule_layer_file_upload(
                 &layer_file_name_1,
-                &LayerFileMetadata::new(content_1.len() as u64),
+                &LayerFileMetadata::new(content_1.len() as u64, generation),
             )
             .unwrap();
 

pageserver/src/tenant/remote_timeline_client/delete.rs

@@ -0,0 +1,34 @@
+//! Helper functions to delete files from remote storage with a RemoteStorage
+use anyhow::Context;
+use std::path::Path;
+use tracing::debug;
+
+use remote_storage::GenericRemoteStorage;
+
+use crate::{
+    config::PageServerConf,
+    tenant::{remote_timeline_client::remote_path, Generation},
+};
+
+pub(super) async fn delete_layer<'a>(
+    conf: &'static PageServerConf,
+    storage: &'a GenericRemoteStorage,
+    local_layer_path: &'a Path,
+    generation: Generation,
+) -> anyhow::Result<()> {
+    fail::fail_point!("before-delete-layer", |_| {
+        anyhow::bail!("failpoint before-delete-layer")
+    });
+    debug!("Deleting layer from remote storage: {local_layer_path:?}",);
+
+    let path_to_delete = remote_path(conf, local_layer_path, generation)?;
+
+    // We don't want to print an error if the delete failed if the file has
+    // already been deleted. Thankfully, in this situation S3 already
+    // does not yield an error. While OS-provided local file system APIs do yield
+    // errors, we avoid them in the `LocalFs` wrapper.
+    storage
+        .delete(&path_to_delete)
+        .await
+        .with_context(|| format!("delete remote layer from storage at {path_to_delete:?}"))
+}

pageserver/src/tenant/remote_timeline_client/download.rs

@@ -15,14 +15,16 @@ use tokio_util::sync::CancellationToken;
 use utils::{backoff, crashsafe};
 
 use crate::config::PageServerConf;
+use crate::tenant::remote_timeline_client::{remote_layer_path, remote_timelines_path};
 use crate::tenant::storage_layer::LayerFileName;
 use crate::tenant::timeline::span::debug_assert_current_span_has_tenant_and_timeline_id;
+use crate::tenant::Generation;
 use remote_storage::{DownloadError, GenericRemoteStorage};
 use utils::crashsafe::path_with_suffix_extension;
 use utils::id::{TenantId, TimelineId};
 
 use super::index::{IndexPart, LayerFileMetadata};
-use super::{FAILED_DOWNLOAD_WARN_THRESHOLD, FAILED_REMOTE_OP_RETRIES};
+use super::{remote_index_path, FAILED_DOWNLOAD_WARN_THRESHOLD, FAILED_REMOTE_OP_RETRIES};
 
 static MAX_DOWNLOAD_DURATION: Duration = Duration::from_secs(120);
 
@@ -41,13 +43,11 @@ pub async fn download_layer_file<'a>(
 ) -> Result<u64, DownloadError> {
     debug_assert_current_span_has_tenant_and_timeline_id();
 
-    let timeline_path = conf.timeline_path(&tenant_id, &timeline_id);
+    let local_path = conf
+        .timeline_path(&tenant_id, &timeline_id)
+        .join(layer_file_name.file_name());
 
-    let local_path = timeline_path.join(layer_file_name.file_name());
-
-    let remote_path = conf
-        .remote_path(&local_path)
-        .map_err(DownloadError::Other)?;
+    let remote_path = remote_layer_path(&tenant_id, &timeline_id, layer_file_name, layer_metadata);
 
     // Perform a rename inspired by durable_rename from file_utils.c.
     // The sequence:
@@ -64,33 +64,43 @@ pub async fn download_layer_file<'a>(
     let (mut destination_file, bytes_amount) = download_retry(
         || async {
             // TODO: this doesn't use the cached fd for some reason?
-            let mut destination_file = fs::File::create(&temp_file_path).await.with_context(|| {
-                format!(
-                    "create a destination file for layer '{}'",
-                    temp_file_path.display()
-                )
-            })
-            .map_err(DownloadError::Other)?;
-            let mut download = storage.download(&remote_path).await.with_context(|| {
-                format!(
+            let mut destination_file = fs::File::create(&temp_file_path)
+                .await
+                .with_context(|| {
+                    format!(
+                        "create a destination file for layer '{}'",
+                        temp_file_path.display()
+                    )
+                })
+                .map_err(DownloadError::Other)?;
+            let mut download = storage
+                .download(&remote_path)
+                .await
+                .with_context(|| {
+                    format!(
                     "open a download stream for layer with remote storage path '{remote_path:?}'"
                 )
-            })
-            .map_err(DownloadError::Other)?;
-
-            let bytes_amount = tokio::time::timeout(MAX_DOWNLOAD_DURATION, tokio::io::copy(&mut download.download_stream, &mut destination_file))
-                .await
-                .map_err(|e| DownloadError::Other(anyhow::anyhow!("Timed out  {:?}", e)))?
-                .with_context(|| {
-                    format!("Failed to download layer with remote storage path '{remote_path:?}' into file {temp_file_path:?}")
                 })
                 .map_err(DownloadError::Other)?;
 
-            Ok((destination_file, bytes_amount))
+            let bytes_amount = tokio::time::timeout(
+                MAX_DOWNLOAD_DURATION,
+                tokio::io::copy(&mut download.download_stream, &mut destination_file),
+            )
+            .await
+            .map_err(|e| DownloadError::Other(anyhow::anyhow!("Timed out  {:?}", e)))?
+            .with_context(|| {
+                format!(
+                    "download layer at remote path '{remote_path:?}' into file {temp_file_path:?}"
+                )
+            })
+            .map_err(DownloadError::Other)?;
 
+            Ok((destination_file, bytes_amount))
         },
         &format!("download {remote_path:?}"),
-    ).await?;
+    )
+    .await?;
 
     // Tokio doc here: https://docs.rs/tokio/1.17.0/tokio/fs/struct.File.html states that:
     // A file will not be closed immediately when it goes out of scope if there are any IO operations
@@ -103,12 +113,7 @@ pub async fn download_layer_file<'a>(
     destination_file
         .flush()
         .await
-        .with_context(|| {
-            format!(
-                "failed to flush source file at {}",
-                temp_file_path.display()
-            )
-        })
+        .with_context(|| format!("flush source file at {}", temp_file_path.display()))
         .map_err(DownloadError::Other)?;
 
     let expected = layer_metadata.file_size();
@@ -139,17 +144,12 @@ pub async fn download_layer_file<'a>(
 
     fs::rename(&temp_file_path, &local_path)
         .await
-        .with_context(|| {
-            format!(
-                "Could not rename download layer file to {}",
-                local_path.display(),
-            )
-        })
+        .with_context(|| format!("rename download layer file to {}", local_path.display(),))
         .map_err(DownloadError::Other)?;
 
     crashsafe::fsync_async(&local_path)
         .await
-        .with_context(|| format!("Could not fsync layer file {}", local_path.display(),))
+        .with_context(|| format!("fsync layer file {}", local_path.display(),))
         .map_err(DownloadError::Other)?;
 
     tracing::debug!("download complete: {}", local_path.display());
@@ -173,21 +173,19 @@ pub fn is_temp_download_file(path: &Path) -> bool {
 }
 
 /// List timelines of given tenant in remote storage
-pub async fn list_remote_timelines<'a>(
-    storage: &'a GenericRemoteStorage,
-    conf: &'static PageServerConf,
+pub async fn list_remote_timelines(
+    storage: &GenericRemoteStorage,
     tenant_id: TenantId,
 ) -> anyhow::Result<HashSet<TimelineId>> {
-    let tenant_path = conf.timelines_path(&tenant_id);
-    let tenant_storage_path = conf.remote_path(&tenant_path)?;
+    let remote_path = remote_timelines_path(&tenant_id);
 
     fail::fail_point!("storage-sync-list-remote-timelines", |_| {
         anyhow::bail!("storage-sync-list-remote-timelines");
     });
 
     let timelines = download_retry(
-        || storage.list_prefixes(Some(&tenant_storage_path)),
-        &format!("list prefixes for {tenant_path:?}"),
+        || storage.list_prefixes(Some(&remote_path)),
+        &format!("list prefixes for {tenant_id}"),
     )
     .await?;
 
@@ -202,9 +200,9 @@ pub async fn list_remote_timelines<'a>(
             anyhow::anyhow!("failed to get timeline id for remote tenant {tenant_id}")
         })?;
 
-        let timeline_id: TimelineId = object_name.parse().with_context(|| {
-            format!("failed to parse object name into timeline id '{object_name}'")
-        })?;
+        let timeline_id: TimelineId = object_name
+            .parse()
+            .with_context(|| format!("parse object name into timeline id '{object_name}'"))?;
 
         // list_prefixes is assumed to return unique names. Ensure this here.
         // NB: it's safer to bail out than warn-log this because the pageserver
@@ -222,28 +220,33 @@ pub async fn list_remote_timelines<'a>(
 }
 
 pub(super) async fn download_index_part(
-    conf: &'static PageServerConf,
     storage: &GenericRemoteStorage,
     tenant_id: &TenantId,
     timeline_id: &TimelineId,
+    generation: Generation,
 ) -> Result<IndexPart, DownloadError> {
-    let index_part_path = conf
-        .metadata_path(tenant_id, timeline_id)
-        .with_file_name(IndexPart::FILE_NAME);
-    let part_storage_path = conf
-        .remote_path(&index_part_path)
-        .map_err(DownloadError::BadInput)?;
+    let remote_path = remote_index_path(tenant_id, timeline_id, generation);
 
     let index_part_bytes = download_retry(
-        || storage.download_all(&part_storage_path),
-        &format!("download {part_storage_path:?}"),
+        || async {
+            let mut index_part_download = storage.download(&remote_path).await?;
+
+            let mut index_part_bytes = Vec::new();
+            tokio::io::copy(
+                &mut index_part_download.download_stream,
+                &mut index_part_bytes,
+            )
+            .await
+            .with_context(|| format!("download index part at {remote_path:?}"))
+            .map_err(DownloadError::Other)?;
+            Ok(index_part_bytes)
+        },
+        &format!("download {remote_path:?}"),
     )
     .await?;
 
     let index_part: IndexPart = serde_json::from_slice(&index_part_bytes)
-        .with_context(|| {
-            format!("Failed to deserialize index part file into file {index_part_path:?}")
-        })
+        .with_context(|| format!("download index part file at {remote_path:?}"))
         .map_err(DownloadError::Other)?;
 
     Ok(index_part)

pageserver/src/tenant/remote_timeline_client/index.rs

@@ -2,7 +2,7 @@
 //! Able to restore itself from the storage index parts, that are located in every timeline's remote directory and contain all data about
 //! remote timeline layers and its metadata.
 
-use std::collections::{HashMap, HashSet};
+use std::collections::HashMap;
 
 use chrono::NaiveDateTime;
 use serde::{Deserialize, Serialize};
@@ -12,6 +12,7 @@ use utils::bin_ser::SerializeError;
 use crate::tenant::metadata::TimelineMetadata;
 use crate::tenant::storage_layer::LayerFileName;
 use crate::tenant::upload_queue::UploadQueueInitialized;
+use crate::tenant::Generation;
 
 use utils::lsn::Lsn;
 
@@ -20,22 +21,28 @@ use utils::lsn::Lsn;
 /// Fields have to be `Option`s because remote [`IndexPart`]'s can be from different version, which
 /// might have less or more metadata depending if upgrading or rolling back an upgrade.
 #[derive(Debug, Clone, PartialEq, Eq, PartialOrd, Ord)]
-#[cfg_attr(test, derive(Default))]
+//#[cfg_attr(test, derive(Default))]
 pub struct LayerFileMetadata {
     file_size: u64,
+
+    pub(crate) generation: Generation,
 }
 
 impl From<&'_ IndexLayerMetadata> for LayerFileMetadata {
     fn from(other: &IndexLayerMetadata) -> Self {
         LayerFileMetadata {
             file_size: other.file_size,
+            generation: other.generation,
         }
     }
 }
 
 impl LayerFileMetadata {
-    pub fn new(file_size: u64) -> Self {
-        LayerFileMetadata { file_size }
+    pub fn new(file_size: u64, generation: Generation) -> Self {
+        LayerFileMetadata {
+            file_size,
+            generation,
+        }
     }
 
     pub fn file_size(&self) -> u64 {
@@ -62,10 +69,6 @@ pub struct IndexPart {
     #[serde(skip_serializing_if = "Option::is_none")]
     pub deleted_at: Option<NaiveDateTime>,
 
-    /// Legacy field: equal to the keys of `layer_metadata`, only written out for forward compat
-    #[serde(default, skip_deserializing)]
-    timeline_layers: HashSet<LayerFileName>,
-
     /// Per layer file name metadata, which can be present for a present or missing layer file.
     ///
     /// Older versions of `IndexPart` will not have this property or have only a part of metadata
@@ -91,7 +94,12 @@ impl IndexPart {
     /// - 2: added `deleted_at`
     /// - 3: no longer deserialize `timeline_layers` (serialized format is the same, but timeline_layers
     ///      is always generated from the keys of `layer_metadata`)
-    const LATEST_VERSION: usize = 3;
+    /// - 4: timeline_layers is fully removed.
+    const LATEST_VERSION: usize = 4;
+
+    // Versions we may see when reading from a bucket.
+    pub const KNOWN_VERSIONS: &[usize] = &[1, 2, 3, 4];
+
     pub const FILE_NAME: &'static str = "index_part.json";
 
     pub fn new(
@@ -99,24 +107,30 @@ impl IndexPart {
         disk_consistent_lsn: Lsn,
         metadata: TimelineMetadata,
     ) -> Self {
-        let mut timeline_layers = HashSet::with_capacity(layers_and_metadata.len());
-        let mut layer_metadata = HashMap::with_capacity(layers_and_metadata.len());
-
-        for (remote_name, metadata) in &layers_and_metadata {
-            timeline_layers.insert(remote_name.to_owned());
-            let metadata = IndexLayerMetadata::from(metadata);
-            layer_metadata.insert(remote_name.to_owned(), metadata);
-        }
+        // Transform LayerFileMetadata into IndexLayerMetadata
+        let layer_metadata = layers_and_metadata
+            .into_iter()
+            .map(|(k, v)| (k, IndexLayerMetadata::from(v)))
+            .collect();
 
         Self {
             version: Self::LATEST_VERSION,
-            timeline_layers,
             layer_metadata,
             disk_consistent_lsn,
             metadata,
             deleted_at: None,
         }
     }
+
+    pub fn get_version(&self) -> usize {
+        self.version
+    }
+
+    /// If you want this under normal operations, read it from self.metadata:
+    /// this method is just for the scrubber to use when validating an index.
+    pub fn get_disk_consistent_lsn(&self) -> Lsn {
+        self.disk_consistent_lsn
+    }
 }
 
 impl TryFrom<&UploadQueueInitialized> for IndexPart {
@@ -135,15 +149,20 @@ impl TryFrom<&UploadQueueInitialized> for IndexPart {
 }
 
 /// Serialized form of [`LayerFileMetadata`].
-#[derive(Debug, PartialEq, Eq, Clone, Serialize, Deserialize, Default)]
+#[derive(Debug, PartialEq, Eq, Clone, Serialize, Deserialize)]
 pub struct IndexLayerMetadata {
-    pub(super) file_size: u64,
+    pub file_size: u64,
+
+    #[serde(default = "Generation::none")]
+    #[serde(skip_serializing_if = "Generation::is_none")]
+    pub(super) generation: Generation,
 }
 
-impl From<&'_ LayerFileMetadata> for IndexLayerMetadata {
-    fn from(other: &'_ LayerFileMetadata) -> Self {
+impl From<LayerFileMetadata> for IndexLayerMetadata {
+    fn from(other: LayerFileMetadata) -> Self {
         IndexLayerMetadata {
             file_size: other.file_size,
+            generation: other.generation,
         }
     }
 }
@@ -168,15 +187,16 @@ mod tests {
         let expected = IndexPart {
             // note this is not verified, could be anything, but exists for humans debugging.. could be the git version instead?
             version: 1,
-            timeline_layers: HashSet::new(),
             layer_metadata: HashMap::from([
                 ("000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__0000000001696070-00000000016960E9".parse().unwrap(), IndexLayerMetadata {
                     file_size: 25600000,
+                    generation: Generation::none()
                 }),
                 ("000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__00000000016B59D8-00000000016B5A51".parse().unwrap(), IndexLayerMetadata {
                     // serde_json should always parse this but this might be a double with jq for
                     // example.
                     file_size: 9007199254741001,
+                    generation: Generation::none()
                 })
             ]),
             disk_consistent_lsn: "0/16960E8".parse::<Lsn>().unwrap(),
@@ -205,15 +225,16 @@ mod tests {
         let expected = IndexPart {
             // note this is not verified, could be anything, but exists for humans debugging.. could be the git version instead?
             version: 1,
-            timeline_layers: HashSet::new(),
             layer_metadata: HashMap::from([
                 ("000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__0000000001696070-00000000016960E9".parse().unwrap(), IndexLayerMetadata {
                     file_size: 25600000,
+                    generation: Generation::none()
                 }),
                 ("000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__00000000016B59D8-00000000016B5A51".parse().unwrap(), IndexLayerMetadata {
                     // serde_json should always parse this but this might be a double with jq for
                     // example.
                     file_size: 9007199254741001,
+                    generation: Generation::none()
                 })
             ]),
             disk_consistent_lsn: "0/16960E8".parse::<Lsn>().unwrap(),
@@ -243,15 +264,16 @@ mod tests {
         let expected = IndexPart {
             // note this is not verified, could be anything, but exists for humans debugging.. could be the git version instead?
             version: 2,
-            timeline_layers: HashSet::new(),
             layer_metadata: HashMap::from([
                 ("000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__0000000001696070-00000000016960E9".parse().unwrap(), IndexLayerMetadata {
                     file_size: 25600000,
+                    generation: Generation::none()
                 }),
                 ("000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__00000000016B59D8-00000000016B5A51".parse().unwrap(), IndexLayerMetadata {
                     // serde_json should always parse this but this might be a double with jq for
                     // example.
                     file_size: 9007199254741001,
+                    generation: Generation::none()
                 })
             ]),
             disk_consistent_lsn: "0/16960E8".parse::<Lsn>().unwrap(),
@@ -276,7 +298,6 @@ mod tests {
 
         let expected = IndexPart {
             version: 1,
-            timeline_layers: HashSet::new(),
             layer_metadata: HashMap::new(),
             disk_consistent_lsn: "0/2532648".parse::<Lsn>().unwrap(),
             metadata: TimelineMetadata::from_bytes(&[
@@ -309,4 +330,41 @@ mod tests {
 
         assert_eq!(empty_layers_parsed, expected);
     }
+
+    #[test]
+    fn v4_indexpart_is_parsed() {
+        let example = r#"{
+            "version":4,
+            "layer_metadata":{
+                "000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__0000000001696070-00000000016960E9": { "file_size": 25600000 },
+                "000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__00000000016B59D8-00000000016B5A51": { "file_size": 9007199254741001 }
+            },
+            "disk_consistent_lsn":"0/16960E8",
+            "metadata_bytes":[113,11,159,210,0,54,0,4,0,0,0,0,1,105,96,232,1,0,0,0,0,1,105,96,112,0,0,0,0,0,0,0,0,0,0,0,0,0,1,105,96,112,0,0,0,0,1,105,96,112,0,0,0,14,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],
+            "deleted_at": "2023-07-31T09:00:00.123"
+        }"#;
+
+        let expected = IndexPart {
+            version: 4,
+            layer_metadata: HashMap::from([
+                ("000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__0000000001696070-00000000016960E9".parse().unwrap(), IndexLayerMetadata {
+                    file_size: 25600000,
+                    generation: Generation::none()
+                }),
+                ("000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__00000000016B59D8-00000000016B5A51".parse().unwrap(), IndexLayerMetadata {
+                    // serde_json should always parse this but this might be a double with jq for
+                    // example.
+                    file_size: 9007199254741001,
+                    generation: Generation::none()
+                })
+            ]),
+            disk_consistent_lsn: "0/16960E8".parse::<Lsn>().unwrap(),
+            metadata: TimelineMetadata::from_bytes(&[113,11,159,210,0,54,0,4,0,0,0,0,1,105,96,232,1,0,0,0,0,1,105,96,112,0,0,0,0,0,0,0,0,0,0,0,0,0,1,105,96,112,0,0,0,0,1,105,96,112,0,0,0,14,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0]).unwrap(),
+            deleted_at: Some(chrono::NaiveDateTime::parse_from_str(
+                "2023-07-31T09:00:00.123000000", "%Y-%m-%dT%H:%M:%S.%f").unwrap())
+        };
+
+        let part = serde_json::from_str::<IndexPart>(example).unwrap();
+        assert_eq!(part, expected);
+    }
 }

pageserver/src/tenant/remote_timeline_client/upload.rs

@@ -5,7 +5,11 @@ use fail::fail_point;
 use std::{io::ErrorKind, path::Path};
 use tokio::fs;
 
-use crate::{config::PageServerConf, tenant::remote_timeline_client::index::IndexPart};
+use super::Generation;
+use crate::{
+    config::PageServerConf,
+    tenant::remote_timeline_client::{index::IndexPart, remote_index_path, remote_path},
+};
 use remote_storage::GenericRemoteStorage;
 use utils::id::{TenantId, TimelineId};
 
@@ -15,10 +19,10 @@ use tracing::info;
 
 /// Serializes and uploads the given index part data to the remote storage.
 pub(super) async fn upload_index_part<'a>(
-    conf: &'static PageServerConf,
     storage: &'a GenericRemoteStorage,
     tenant_id: &TenantId,
     timeline_id: &TimelineId,
+    generation: Generation,
     index_part: &'a IndexPart,
 ) -> anyhow::Result<()> {
     tracing::trace!("uploading new index part");
@@ -27,20 +31,16 @@ pub(super) async fn upload_index_part<'a>(
         bail!("failpoint before-upload-index")
     });
 
-    let index_part_bytes = serde_json::to_vec(&index_part)
-        .context("Failed to serialize index part file into bytes")?;
+    let index_part_bytes =
+        serde_json::to_vec(&index_part).context("serialize index part file into bytes")?;
     let index_part_size = index_part_bytes.len();
     let index_part_bytes = tokio::io::BufReader::new(std::io::Cursor::new(index_part_bytes));
 
-    let index_part_path = conf
-        .metadata_path(tenant_id, timeline_id)
-        .with_file_name(IndexPart::FILE_NAME);
-    let storage_path = conf.remote_path(&index_part_path)?;
-
+    let remote_path = remote_index_path(tenant_id, timeline_id, generation);
     storage
-        .upload_storage_object(Box::new(index_part_bytes), index_part_size, &storage_path)
+        .upload_storage_object(Box::new(index_part_bytes), index_part_size, &remote_path)
         .await
-        .with_context(|| format!("Failed to upload index part for '{tenant_id} / {timeline_id}'"))
+        .with_context(|| format!("upload index part for '{tenant_id} / {timeline_id}'"))
 }
 
 /// Attempts to upload given layer files.
@@ -52,12 +52,13 @@ pub(super) async fn upload_timeline_layer<'a>(
     storage: &'a GenericRemoteStorage,
     source_path: &'a Path,
     known_metadata: &'a LayerFileMetadata,
+    generation: Generation,
 ) -> anyhow::Result<()> {
     fail_point!("before-upload-layer", |_| {
         bail!("failpoint before-upload-layer")
     });
-    let storage_path = conf.remote_path(source_path)?;
 
+    let storage_path = remote_path(conf, source_path, generation)?;
     let source_file_res = fs::File::open(&source_path).await;
     let source_file = match source_file_res {
         Ok(source_file) => source_file,
@@ -70,16 +71,15 @@ pub(super) async fn upload_timeline_layer<'a>(
             info!(path = %source_path.display(), "File to upload doesn't exist. Likely the file has been deleted and an upload is not required any more.");
             return Ok(());
         }
-        Err(e) => Err(e)
-            .with_context(|| format!("Failed to open a source file for layer {source_path:?}"))?,
+        Err(e) => {
+            Err(e).with_context(|| format!("open a source file for layer {source_path:?}"))?
+        }
     };
 
     let fs_size = source_file
         .metadata()
         .await
-        .with_context(|| {
-            format!("Failed to get the source file metadata for layer {source_path:?}")
-        })?
+        .with_context(|| format!("get the source file metadata for layer {source_path:?}"))?
         .len();
 
     let metadata_size = known_metadata.file_size();
@@ -87,19 +87,13 @@ pub(super) async fn upload_timeline_layer<'a>(
         bail!("File {source_path:?} has its current FS size {fs_size} diferent from initially determined {metadata_size}");
     }
 
-    let fs_size = usize::try_from(fs_size).with_context(|| {
-        format!("File {source_path:?} size {fs_size} could not be converted to usize")
-    })?;
+    let fs_size = usize::try_from(fs_size)
+        .with_context(|| format!("convert {source_path:?} size {fs_size} usize"))?;
 
     storage
         .upload(source_file, fs_size, &storage_path, None)
         .await
-        .with_context(|| {
-            format!(
-                "Failed to upload a layer from local path '{}'",
-                source_path.display()
-            )
-        })?;
+        .with_context(|| format!("upload layer from local path '{}'", source_path.display()))?;
 
     Ok(())
 }

pageserver/src/tenant/storage_layer/delta_layer.rs

@@ -31,7 +31,7 @@ use crate::config::PageServerConf;
 use crate::context::RequestContext;
 use crate::page_cache::PAGE_SZ;
 use crate::repository::{Key, Value, KEY_SIZE};
-use crate::tenant::blob_io::{BlobWriter, WriteBlobWriter};
+use crate::tenant::blob_io::WriteBlobWriter;
 use crate::tenant::block_io::{BlockBuf, BlockCursor, BlockLease, BlockReader, FileBlockReader};
 use crate::tenant::disk_btree::{DiskBtreeBuilder, DiskBtreeReader, VisitDirection};
 use crate::tenant::storage_layer::{
@@ -45,8 +45,8 @@ use pageserver_api::models::{HistoricLayerInfo, LayerAccessKind};
 use rand::{distributions::Alphanumeric, Rng};
 use serde::{Deserialize, Serialize};
 use std::fs::{self, File};
+use std::io::SeekFrom;
 use std::io::{BufWriter, Write};
-use std::io::{Seek, SeekFrom};
 use std::ops::Range;
 use std::os::unix::fs::FileExt;
 use std::path::{Path, PathBuf};
@@ -632,11 +632,12 @@ impl DeltaLayerWriterInner {
     ///
     /// The values must be appended in key, lsn order.
     ///
-    fn put_value(&mut self, key: Key, lsn: Lsn, val: Value) -> anyhow::Result<()> {
+    async fn put_value(&mut self, key: Key, lsn: Lsn, val: Value) -> anyhow::Result<()> {
         self.put_value_bytes(key, lsn, &Value::ser(&val)?, val.will_init())
+            .await
     }
 
-    fn put_value_bytes(
+    async fn put_value_bytes(
         &mut self,
         key: Key,
         lsn: Lsn,
@@ -645,7 +646,7 @@ impl DeltaLayerWriterInner {
     ) -> anyhow::Result<()> {
         assert!(self.lsn_range.start <= lsn);
 
-        let off = self.blob_writer.write_blob(val)?;
+        let off = self.blob_writer.write_blob(val).await?;
 
         let blob_ref = BlobRef::new(off, will_init);
 
@@ -797,11 +798,11 @@ impl DeltaLayerWriter {
     ///
     /// The values must be appended in key, lsn order.
     ///
-    pub fn put_value(&mut self, key: Key, lsn: Lsn, val: Value) -> anyhow::Result<()> {
-        self.inner.as_mut().unwrap().put_value(key, lsn, val)
+    pub async fn put_value(&mut self, key: Key, lsn: Lsn, val: Value) -> anyhow::Result<()> {
+        self.inner.as_mut().unwrap().put_value(key, lsn, val).await
     }
 
-    pub fn put_value_bytes(
+    pub async fn put_value_bytes(
         &mut self,
         key: Key,
         lsn: Lsn,
@@ -812,6 +813,7 @@ impl DeltaLayerWriter {
             .as_mut()
             .unwrap()
             .put_value_bytes(key, lsn, val, will_init)
+            .await
     }
 
     pub fn size(&self) -> u64 {

pageserver/src/tenant/storage_layer/filename.rs

@@ -212,7 +212,7 @@ pub enum LayerFileName {
 }
 
 impl LayerFileName {
-    pub(crate) fn file_name(&self) -> String {
+    pub fn file_name(&self) -> String {
         self.to_string()
     }
 

pageserver/src/tenant/storage_layer/image_layer.rs

@@ -27,7 +27,7 @@ use crate::config::PageServerConf;
 use crate::context::RequestContext;
 use crate::page_cache::PAGE_SZ;
 use crate::repository::{Key, KEY_SIZE};
-use crate::tenant::blob_io::{BlobWriter, WriteBlobWriter};
+use crate::tenant::blob_io::WriteBlobWriter;
 use crate::tenant::block_io::{BlockBuf, BlockReader, FileBlockReader};
 use crate::tenant::disk_btree::{DiskBtreeBuilder, DiskBtreeReader, VisitDirection};
 use crate::tenant::storage_layer::{
@@ -42,8 +42,8 @@ use pageserver_api::models::{HistoricLayerInfo, LayerAccessKind};
 use rand::{distributions::Alphanumeric, Rng};
 use serde::{Deserialize, Serialize};
 use std::fs::{self, File};
+use std::io::SeekFrom;
 use std::io::Write;
-use std::io::{Seek, SeekFrom};
 use std::ops::Range;
 use std::os::unix::prelude::FileExt;
 use std::path::{Path, PathBuf};
@@ -569,9 +569,9 @@ impl ImageLayerWriterInner {
     ///
     /// The page versions must be appended in blknum order.
     ///
-    fn put_image(&mut self, key: Key, img: &[u8]) -> anyhow::Result<()> {
+    async fn put_image(&mut self, key: Key, img: &[u8]) -> anyhow::Result<()> {
         ensure!(self.key_range.contains(&key));
-        let off = self.blob_writer.write_blob(img)?;
+        let off = self.blob_writer.write_blob(img).await?;
 
         let mut keybuf: [u8; KEY_SIZE] = [0u8; KEY_SIZE];
         key.write_to_byte_slice(&mut keybuf);
@@ -710,8 +710,8 @@ impl ImageLayerWriter {
     ///
     /// The page versions must be appended in blknum order.
     ///
-    pub fn put_image(&mut self, key: Key, img: &[u8]) -> anyhow::Result<()> {
-        self.inner.as_mut().unwrap().put_image(key, img)
+    pub async fn put_image(&mut self, key: Key, img: &[u8]) -> anyhow::Result<()> {
+        self.inner.as_mut().unwrap().put_image(key, img).await
     }
 
     ///

pageserver/src/tenant/storage_layer/inmemory_layer.rs

@@ -348,7 +348,9 @@ impl InMemoryLayer {
             for (lsn, pos) in vec_map.as_slice() {
                 cursor.read_blob_into_buf(*pos, &mut buf).await?;
                 let will_init = Value::des(&buf)?.will_init();
-                delta_layer_writer.put_value_bytes(key, *lsn, &buf, will_init)?;
+                delta_layer_writer
+                    .put_value_bytes(key, *lsn, &buf, will_init)
+                    .await?;
             }
         }
 

pageserver/src/tenant/timeline.rs

@@ -38,7 +38,6 @@ use std::time::{Duration, Instant, SystemTime};
 use crate::context::{
     AccessStatsBehavior, DownloadBehavior, RequestContext, RequestContextBuilder,
 };
-use crate::deletion_queue::DeletionQueueClient;
 use crate::tenant::remote_timeline_client::index::LayerFileMetadata;
 use crate::tenant::storage_layer::delta_layer::DeltaEntry;
 use crate::tenant::storage_layer::{
@@ -68,6 +67,7 @@ use postgres_connection::PgConnectionConfig;
 use postgres_ffi::to_pg_timestamp;
 use utils::{
     completion,
+    generation::Generation,
     id::{TenantId, TimelineId},
     lsn::{AtomicLsn, Lsn, RecordLsn},
     seqwait::SeqWait,
@@ -142,7 +142,6 @@ fn drop_wlock<T>(rlock: tokio::sync::RwLockWriteGuard<'_, T>) {
 /// The outward-facing resources required to build a Timeline
 pub struct TimelineResources {
     pub remote_client: Option<RemoteTimelineClient>,
-    pub deletion_queue_client: Option<DeletionQueueClient>,
 }
 
 pub struct Timeline {
@@ -154,6 +153,10 @@ pub struct Timeline {
     pub tenant_id: TenantId,
     pub timeline_id: TimelineId,
 
+    /// The generation of the tenant that instantiated us: this is used for safety when writing remote objects.
+    /// Never changes for the lifetime of this [`Timeline`] object.
+    generation: Generation,
+
     pub pg_version: u32,
 
     /// The tuple has two elements.
@@ -197,9 +200,6 @@ pub struct Timeline {
     /// See [`remote_timeline_client`](super::remote_timeline_client) module comment for details.
     pub remote_client: Option<Arc<RemoteTimelineClient>>,
 
-    /// Deletion queue: a global queue, separate to the remote storage queue's
-    deletion_queue_client: Option<Arc<DeletionQueueClient>>,
-
     // What page versions do we hold in the repository? If we get a
     // request > last_record_lsn, we need to wait until we receive all
     // the WAL up to the request. The SeqWait provides functions for
@@ -1204,7 +1204,7 @@ impl Timeline {
                 Ok(delta) => Some(delta),
             };
 
-        let layer_metadata = LayerFileMetadata::new(layer_file_size);
+        let layer_metadata = LayerFileMetadata::new(layer_file_size, self.generation);
 
         let new_remote_layer = Arc::new(match local_layer.filename() {
             LayerFileName::Image(image_name) => RemoteLayer::new_img(
@@ -1267,18 +1267,6 @@ impl Timeline {
 
         Ok(())
     }
-
-    async fn delete_all_remote(&self) -> anyhow::Result<()> {
-        if let Some(remote_client) = &self.remote_client {
-            if let Some(deletion_queue_client) = &self.deletion_queue_client {
-                remote_client.delete_all(deletion_queue_client).await
-            } else {
-                Ok(())
-            }
-        } else {
-            Ok(())
-        }
-    }
 }
 
 #[derive(Debug, thiserror::Error)]
@@ -1394,6 +1382,7 @@ impl Timeline {
         ancestor: Option<Arc<Timeline>>,
         timeline_id: TimelineId,
         tenant_id: TenantId,
+        generation: Generation,
         walredo_mgr: Arc<dyn WalRedoManager + Send + Sync>,
         resources: TimelineResources,
         pg_version: u32,
@@ -1423,6 +1412,7 @@ impl Timeline {
                 myself: myself.clone(),
                 timeline_id,
                 tenant_id,
+                generation,
                 pg_version,
                 layers: Arc::new(tokio::sync::RwLock::new(LayerManager::create())),
                 wanted_image_layers: Mutex::new(None),
@@ -1431,7 +1421,6 @@ impl Timeline {
                 walreceiver: Mutex::new(None),
 
                 remote_client: resources.remote_client.map(Arc::new),
-                deletion_queue_client: resources.deletion_queue_client.map(Arc::new),
 
                 // initialize in-memory 'last_record_lsn' from 'disk_consistent_lsn'.
                 last_record_lsn: SeqWait::new(RecordLsn {
@@ -1633,6 +1622,9 @@ impl Timeline {
         let (conf, tenant_id, timeline_id) = (self.conf, self.tenant_id, self.timeline_id);
         let span = tracing::Span::current();
 
+        // Copy to move into the task we're about to spawn
+        let generation = self.generation;
+
         let (loaded_layers, to_sync, total_physical_size) = tokio::task::spawn_blocking({
             move || {
                 let _g = span.entered();
@@ -1674,8 +1666,12 @@ impl Timeline {
                     );
                 }
 
-                let decided =
-                    init::reconcile(discovered_layers, index_part.as_ref(), disk_consistent_lsn);
+                let decided = init::reconcile(
+                    discovered_layers,
+                    index_part.as_ref(),
+                    disk_consistent_lsn,
+                    generation,
+                );
 
                 let mut loaded_layers = Vec::new();
                 let mut needs_upload = Vec::new();
@@ -1768,15 +1764,11 @@ impl Timeline {
         guard.initialize_local_layers(loaded_layers, disk_consistent_lsn + 1);
 
         if let Some(rtc) = self.remote_client.as_ref() {
-            // Deletion queue client is always Some if remote_client is Some
-            let deletion_queue_client = self.deletion_queue_client.as_ref().unwrap();
-
             let (needs_upload, needs_cleanup) = to_sync;
             for (layer, m) in needs_upload {
                 rtc.schedule_layer_file_upload(&layer.layer_desc().filename(), &m)?;
             }
-            rtc.schedule_layer_file_deletion(&needs_cleanup, deletion_queue_client)
-                .await?;
+            rtc.schedule_layer_file_deletion(&needs_cleanup)?;
             rtc.schedule_index_upload_for_file_changes()?;
             // Tenant::create_timeline will wait for these uploads to happen before returning, or
             // on retry.
@@ -2691,7 +2683,7 @@ impl Timeline {
                 (
                     HashMap::from([(
                         layer.filename(),
-                        LayerFileMetadata::new(layer.layer_desc().file_size),
+                        LayerFileMetadata::new(layer.layer_desc().file_size, self.generation),
                     )]),
                     Some(layer),
                 )
@@ -2743,6 +2735,7 @@ impl Timeline {
         if disk_consistent_lsn != old_disk_consistent_lsn {
             assert!(disk_consistent_lsn > old_disk_consistent_lsn);
             self.update_metadata_file(disk_consistent_lsn, layer_paths_to_upload)
+                .await
                 .context("update_metadata_file")?;
             // Also update the in-memory copy
             self.disk_consistent_lsn.store(disk_consistent_lsn);
@@ -2751,7 +2744,7 @@ impl Timeline {
     }
 
     /// Update metadata file
-    fn update_metadata_file(
+    async fn update_metadata_file(
         &self,
         disk_consistent_lsn: Lsn,
         layer_paths_to_upload: HashMap<LayerFileName, LayerFileMetadata>,
@@ -2799,6 +2792,7 @@ impl Timeline {
             &metadata,
             false,
         )
+        .await
         .context("save_metadata")?;
 
         if let Some(remote_client) = &self.remote_client {
@@ -3038,7 +3032,7 @@ impl Timeline {
                                 }
                             }
                         };
-                        image_layer_writer.put_image(key, &img)?;
+                        image_layer_writer.put_image(key, &img).await?;
                         key = key.next();
                     }
                 }
@@ -3087,7 +3081,10 @@ impl Timeline {
                 .metadata()
                 .with_context(|| format!("reading metadata of layer file {}", path.file_name()))?;
 
-            layer_paths_to_upload.insert(path, LayerFileMetadata::new(metadata.len()));
+            layer_paths_to_upload.insert(
+                path,
+                LayerFileMetadata::new(metadata.len(), self.generation),
+            );
 
             self.metrics
                 .resident_physical_size_gauge
@@ -3621,7 +3618,7 @@ impl Timeline {
                 )))
             });
 
-            writer.as_mut().unwrap().put_value(key, lsn, value)?;
+            writer.as_mut().unwrap().put_value(key, lsn, value).await?;
             prev_key = Some(key);
         }
         if let Some(writer) = writer {
@@ -3762,7 +3759,7 @@ impl Timeline {
             if let Some(remote_client) = &self.remote_client {
                 remote_client.schedule_layer_file_upload(
                     &l.filename(),
-                    &LayerFileMetadata::new(metadata.len()),
+                    &LayerFileMetadata::new(metadata.len(), self.generation),
                 )?;
             }
 
@@ -3771,7 +3768,10 @@ impl Timeline {
                 .resident_physical_size_gauge
                 .add(metadata.len());
 
-            new_layer_paths.insert(new_delta_path, LayerFileMetadata::new(metadata.len()));
+            new_layer_paths.insert(
+                new_delta_path,
+                LayerFileMetadata::new(metadata.len(), self.generation),
+            );
             l.access_stats().record_residence_event(
                 LayerResidenceStatus::Resident,
                 LayerResidenceEventReason::LayerCreate,
@@ -3811,13 +3811,7 @@ impl Timeline {
 
         // Also schedule the deletions in remote storage
         if let Some(remote_client) = &self.remote_client {
-            let deletion_queue = self
-                .deletion_queue_client
-                .as_ref()
-                .ok_or_else(|| anyhow::anyhow!("Remote storage enabled without deletion queue"))?;
-            remote_client
-                .schedule_layer_file_deletion(&layer_names_to_delete, deletion_queue)
-                .await?;
+            remote_client.schedule_layer_file_deletion(&layer_names_to_delete)?;
         }
 
         Ok(())
@@ -4130,7 +4124,8 @@ impl Timeline {
         if !layers_to_remove.is_empty() {
             // Persist the new GC cutoff value in the metadata file, before
             // we actually remove anything.
-            self.update_metadata_file(self.disk_consistent_lsn.load(), HashMap::new())?;
+            self.update_metadata_file(self.disk_consistent_lsn.load(), HashMap::new())
+                .await?;
 
             // Actually delete the layers from disk and remove them from the map.
             // (couldn't do this in the loop above, because you cannot modify a collection
@@ -4151,15 +4146,7 @@ impl Timeline {
             }
 
             if let Some(remote_client) = &self.remote_client {
-                // Remote metadata upload was scheduled in `update_metadata_file`: wait
-                // for completion before scheduling any deletions.
-                remote_client.wait_completion().await?;
-                let deletion_queue = self.deletion_queue_client.as_ref().ok_or_else(|| {
-                    anyhow::anyhow!("Remote storage enabled without deletion queue")
-                })?;
-                remote_client
-                    .schedule_layer_file_deletion(&layer_names_to_delete, deletion_queue)
-                    .await?;
+                remote_client.schedule_layer_file_deletion(&layer_names_to_delete)?;
             }
 
             apply.flush();
@@ -4749,7 +4736,6 @@ mod tests {
 
     use utils::{id::TimelineId, lsn::Lsn};
 
-    use crate::deletion_queue::mock::MockDeletionQueue;
     use crate::tenant::{harness::TenantHarness, storage_layer::PersistentLayer};
 
     use super::{EvictionError, Timeline};
@@ -4772,17 +4758,9 @@ mod tests {
             };
             GenericRemoteStorage::from_config(&config).unwrap()
         };
-        let deletion_queue = MockDeletionQueue::new(Some(remote_storage.clone()), harness.conf);
 
         let ctx = any_context();
-        let tenant = harness
-            .try_load(
-                &ctx,
-                Some(remote_storage),
-                Some(deletion_queue.new_client()),
-            )
-            .await
-            .unwrap();
+        let tenant = harness.try_load(&ctx, Some(remote_storage)).await.unwrap();
         let timeline = tenant
             .create_test_timeline(TimelineId::generate(), Lsn(0x10), 14, &ctx)
             .await
@@ -4845,17 +4823,9 @@ mod tests {
             };
             GenericRemoteStorage::from_config(&config).unwrap()
         };
-        let deletion_queue = MockDeletionQueue::new(Some(remote_storage.clone()), harness.conf);
 
         let ctx = any_context();
-        let tenant = harness
-            .try_load(
-                &ctx,
-                Some(remote_storage),
-                Some(deletion_queue.new_client()),
-            )
-            .await
-            .unwrap();
+        let tenant = harness.try_load(&ctx, Some(remote_storage)).await.unwrap();
         let timeline = tenant
             .create_test_timeline(TimelineId::generate(), Lsn(0x10), 14, &ctx)
             .await

pageserver/src/tenant/timeline/delete.rs

@@ -14,7 +14,6 @@ use utils::{
 
 use crate::{
     config::PageServerConf,
-    deletion_queue::DeletionQueueClient,
     task_mgr::{self, TaskKind},
     tenant::{
         metadata::TimelineMetadata,
@@ -239,6 +238,15 @@ async fn delete_local_layer_files(
     Ok(())
 }
 
+/// Removes remote layers and an index file after them.
+async fn delete_remote_layers_and_index(timeline: &Timeline) -> anyhow::Result<()> {
+    if let Some(remote_client) = &timeline.remote_client {
+        remote_client.delete_all().await.context("delete_all")?
+    };
+
+    Ok(())
+}
+
 // This function removs remaining traces of a timeline on disk.
 // Namely: metadata file, timeline directory, delete mark.
 // Note: io::ErrorKind::NotFound are ignored for metadata and timeline dir.
@@ -399,7 +407,6 @@ impl DeleteTimelineFlow {
         timeline_id: TimelineId,
         local_metadata: &TimelineMetadata,
         remote_client: Option<RemoteTimelineClient>,
-        deletion_queue_client: Option<DeletionQueueClient>,
         init_order: Option<&InitializationOrder>,
     ) -> anyhow::Result<()> {
         // Note: here we even skip populating layer map. Timeline is essentially uninitialized.
@@ -409,10 +416,7 @@ impl DeleteTimelineFlow {
                 timeline_id,
                 local_metadata,
                 None, // Ancestor is not needed for deletion.
-                TimelineResources {
-                    remote_client,
-                    deletion_queue_client,
-                },
+                TimelineResources { remote_client },
                 init_order,
                 // Important. We dont pass ancestor above because it can be missing.
                 // Thus we need to skip the validation here.
@@ -555,7 +559,7 @@ impl DeleteTimelineFlow {
     ) -> Result<(), DeleteTimelineError> {
         delete_local_layer_files(conf, tenant.tenant_id, timeline).await?;
 
-        timeline.delete_all_remote().await?;
+        delete_remote_layers_and_index(timeline).await?;
 
         pausable_failpoint!("in_progress_delete");
 

pageserver/src/tenant/timeline/init.rs

@@ -7,6 +7,7 @@ use crate::{
             index::{IndexPart, LayerFileMetadata},
         },
         storage_layer::LayerFileName,
+        Generation,
     },
     METADATA_FILE_NAME,
 };
@@ -104,6 +105,7 @@ pub(super) fn reconcile(
     discovered: Vec<(LayerFileName, u64)>,
     index_part: Option<&IndexPart>,
     disk_consistent_lsn: Lsn,
+    generation: Generation,
 ) -> Vec<(LayerFileName, Result<Decision, FutureLayer>)> {
     use Decision::*;
 
@@ -112,7 +114,15 @@ pub(super) fn reconcile(
 
     let mut discovered = discovered
         .into_iter()
-        .map(|(name, file_size)| (name, (Some(LayerFileMetadata::new(file_size)), None)))
+        .map(|(name, file_size)| {
+            (
+                name,
+                // The generation here will be corrected to match IndexPart in the merge below, unless
+                // it is not in IndexPart, in which case using our current generation makes sense
+                // because it will be uploaded in this generation.
+                (Some(LayerFileMetadata::new(file_size, generation)), None),
+            )
+        })
         .collect::<Collected>();
 
     // merge any index_part information, when available
@@ -137,7 +147,11 @@ pub(super) fn reconcile(
                 Err(FutureLayer { local })
             } else {
                 Ok(match (local, remote) {
-                    (Some(local), Some(remote)) if local != remote => UseRemote { local, remote },
+                    (Some(local), Some(remote)) if local != remote => {
+                        assert_eq!(local.generation, remote.generation);
+
+                        UseRemote { local, remote }
+                    }
                     (Some(x), Some(_)) => UseLocal(x),
                     (None, Some(x)) => Evicted(x),
                     (Some(x), None) => NeedsUpload(x),

pageserver/src/tenant/upload_queue.rs

@@ -1,4 +1,7 @@
+use crate::metrics::RemoteOpFileKind;
+
 use super::storage_layer::LayerFileName;
+use super::Generation;
 use crate::tenant::metadata::TimelineMetadata;
 use crate::tenant::remote_timeline_client::index::IndexPart;
 use crate::tenant::remote_timeline_client::index::LayerFileMetadata;
@@ -60,6 +63,7 @@ pub(crate) struct UploadQueueInitialized {
     // Breakdown of different kinds of tasks currently in-progress
     pub(crate) num_inprogress_layer_uploads: usize,
     pub(crate) num_inprogress_metadata_uploads: usize,
+    pub(crate) num_inprogress_deletions: usize,
 
     /// Tasks that are currently in-progress. In-progress means that a tokio Task
     /// has been launched for it. An in-progress task can be busy uploading, but it can
@@ -117,6 +121,7 @@ impl UploadQueue {
             task_counter: 0,
             num_inprogress_layer_uploads: 0,
             num_inprogress_metadata_uploads: 0,
+            num_inprogress_deletions: 0,
             inprogress_tasks: HashMap::new(),
             queued_operations: VecDeque::new(),
         };
@@ -158,6 +163,7 @@ impl UploadQueue {
             task_counter: 0,
             num_inprogress_layer_uploads: 0,
             num_inprogress_metadata_uploads: 0,
+            num_inprogress_deletions: 0,
             inprogress_tasks: HashMap::new(),
             queued_operations: VecDeque::new(),
         };
@@ -195,6 +201,14 @@ pub(crate) struct UploadTask {
     pub(crate) op: UploadOp,
 }
 
+#[derive(Debug)]
+pub(crate) struct Delete {
+    pub(crate) file_kind: RemoteOpFileKind,
+    pub(crate) layer_file_name: LayerFileName,
+    pub(crate) scheduled_from_timeline_delete: bool,
+    pub(crate) generation: Generation,
+}
+
 #[derive(Debug)]
 pub(crate) enum UploadOp {
     /// Upload a layer file
@@ -203,6 +217,9 @@ pub(crate) enum UploadOp {
     /// Upload the metadata file
     UploadMetadata(IndexPart, Lsn),
 
+    /// Delete a layer file
+    Delete(Delete),
+
     /// Barrier. When the barrier operation is reached,
     Barrier(tokio::sync::watch::Sender<()>),
 }
@@ -213,12 +230,22 @@ impl std::fmt::Display for UploadOp {
             UploadOp::UploadLayer(path, metadata) => {
                 write!(
                     f,
-                    "UploadLayer({}, size={:?})",
+                    "UploadLayer({}, size={:?}, gen={:?})",
                     path.file_name(),
-                    metadata.file_size()
+                    metadata.file_size(),
+                    metadata.generation,
                 )
             }
-            UploadOp::UploadMetadata(_, lsn) => write!(f, "UploadMetadata(lsn: {})", lsn),
+            UploadOp::UploadMetadata(_, lsn) => {
+                write!(f, "UploadMetadata(lsn: {})", lsn)
+            }
+            UploadOp::Delete(delete) => write!(
+                f,
+                "Delete(path: {}, scheduled_from_timeline_delete: {}, gen: {:?})",
+                delete.layer_file_name.file_name(),
+                delete.scheduled_from_timeline_delete,
+                delete.generation
+            ),
             UploadOp::Barrier(_) => write!(f, "Barrier"),
         }
     }

pageserver/src/virtual_file.rs

@@ -13,7 +13,7 @@
 use crate::metrics::{STORAGE_IO_SIZE, STORAGE_IO_TIME};
 use once_cell::sync::OnceCell;
 use std::fs::{self, File, OpenOptions};
-use std::io::{Error, ErrorKind, Read, Seek, SeekFrom, Write};
+use std::io::{Error, ErrorKind, Seek, SeekFrom, Write};
 use std::os::unix::fs::FileExt;
 use std::path::{Path, PathBuf};
 use std::sync::atomic::{AtomicBool, AtomicUsize, Ordering};
@@ -321,54 +321,8 @@ impl VirtualFile {
         drop(self);
         std::fs::remove_file(path).expect("failed to remove the virtual file");
     }
-}
 
-impl Drop for VirtualFile {
-    /// If a VirtualFile is dropped, close the underlying file if it was open.
-    fn drop(&mut self) {
-        let handle = self.handle.get_mut().unwrap();
-
-        // We could check with a read-lock first, to avoid waiting on an
-        // unrelated I/O.
-        let slot = &get_open_files().slots[handle.index];
-        let mut slot_guard = slot.inner.write().unwrap();
-        if slot_guard.tag == handle.tag {
-            slot.recently_used.store(false, Ordering::Relaxed);
-            // there is also operation "close-by-replace" for closes done on eviction for
-            // comparison.
-            STORAGE_IO_TIME
-                .with_label_values(&["close"])
-                .observe_closure_duration(|| drop(slot_guard.file.take()));
-        }
-    }
-}
-
-impl Read for VirtualFile {
-    fn read(&mut self, buf: &mut [u8]) -> Result<usize, Error> {
-        let pos = self.pos;
-        let n = self.read_at(buf, pos)?;
-        self.pos += n as u64;
-        Ok(n)
-    }
-}
-
-impl Write for VirtualFile {
-    fn write(&mut self, buf: &[u8]) -> Result<usize, std::io::Error> {
-        let pos = self.pos;
-        let n = self.write_at(buf, pos)?;
-        self.pos += n as u64;
-        Ok(n)
-    }
-
-    fn flush(&mut self) -> Result<(), std::io::Error> {
-        // flush is no-op for File (at least on unix), so we don't need to do
-        // anything here either.
-        Ok(())
-    }
-}
-
-impl Seek for VirtualFile {
-    fn seek(&mut self, pos: SeekFrom) -> Result<u64, Error> {
+    pub fn seek(&mut self, pos: SeekFrom) -> Result<u64, Error> {
         match pos {
             SeekFrom::Start(offset) => {
                 self.pos = offset;
@@ -392,10 +346,64 @@ impl Seek for VirtualFile {
         }
         Ok(self.pos)
     }
-}
 
-impl FileExt for VirtualFile {
-    fn read_at(&self, buf: &mut [u8], offset: u64) -> Result<usize, Error> {
+    // Copied from https://doc.rust-lang.org/1.72.0/src/std/os/unix/fs.rs.html#117-135
+    pub fn read_exact_at(&self, mut buf: &mut [u8], mut offset: u64) -> Result<(), Error> {
+        while !buf.is_empty() {
+            match self.read_at(buf, offset) {
+                Ok(0) => {
+                    return Err(Error::new(
+                        std::io::ErrorKind::UnexpectedEof,
+                        "failed to fill whole buffer",
+                    ))
+                }
+                Ok(n) => {
+                    buf = &mut buf[n..];
+                    offset += n as u64;
+                }
+                Err(ref e) if e.kind() == std::io::ErrorKind::Interrupted => {}
+                Err(e) => return Err(e),
+            }
+        }
+        Ok(())
+    }
+
+    // Copied from https://doc.rust-lang.org/1.72.0/src/std/os/unix/fs.rs.html#219-235
+    pub async fn write_all_at(&self, mut buf: &[u8], mut offset: u64) -> Result<(), Error> {
+        while !buf.is_empty() {
+            match self.write_at(buf, offset) {
+                Ok(0) => {
+                    return Err(Error::new(
+                        std::io::ErrorKind::WriteZero,
+                        "failed to write whole buffer",
+                    ));
+                }
+                Ok(n) => {
+                    buf = &buf[n..];
+                    offset += n as u64;
+                }
+                Err(ref e) if e.kind() == std::io::ErrorKind::Interrupted => {}
+                Err(e) => return Err(e),
+            }
+        }
+        Ok(())
+    }
+
+    /// Write the given buffer (which has to be below the kernel's internal page size) and fsync
+    ///
+    /// This ensures some level of atomicity (not a good one, but it's the best we have).
+    pub fn write_and_fsync(&mut self, buf: &[u8]) -> Result<(), Error> {
+        if self.write(buf)? != buf.len() {
+            return Err(Error::new(
+                std::io::ErrorKind::Other,
+                "Could not write all the bytes in a single call",
+            ));
+        }
+        self.sync_all()?;
+        Ok(())
+    }
+
+    pub fn read_at(&self, buf: &mut [u8], offset: u64) -> Result<usize, Error> {
         let result = self.with_file("read", |file| file.read_at(buf, offset))?;
         if let Ok(size) = result {
             STORAGE_IO_SIZE
@@ -405,7 +413,7 @@ impl FileExt for VirtualFile {
         result
     }
 
-    fn write_at(&self, buf: &[u8], offset: u64) -> Result<usize, Error> {
+    pub fn write_at(&self, buf: &[u8], offset: u64) -> Result<usize, Error> {
         let result = self.with_file("write", |file| file.write_at(buf, offset))?;
         if let Ok(size) = result {
             STORAGE_IO_SIZE
@@ -416,6 +424,41 @@ impl FileExt for VirtualFile {
     }
 }
 
+impl Drop for VirtualFile {
+    /// If a VirtualFile is dropped, close the underlying file if it was open.
+    fn drop(&mut self) {
+        let handle = self.handle.get_mut().unwrap();
+
+        // We could check with a read-lock first, to avoid waiting on an
+        // unrelated I/O.
+        let slot = &get_open_files().slots[handle.index];
+        let mut slot_guard = slot.inner.write().unwrap();
+        if slot_guard.tag == handle.tag {
+            slot.recently_used.store(false, Ordering::Relaxed);
+            // there is also operation "close-by-replace" for closes done on eviction for
+            // comparison.
+            STORAGE_IO_TIME
+                .with_label_values(&["close"])
+                .observe_closure_duration(|| drop(slot_guard.file.take()));
+        }
+    }
+}
+
+impl Write for VirtualFile {
+    fn write(&mut self, buf: &[u8]) -> Result<usize, std::io::Error> {
+        let pos = self.pos;
+        let n = self.write_at(buf, pos)?;
+        self.pos += n as u64;
+        Ok(n)
+    }
+
+    fn flush(&mut self) -> Result<(), std::io::Error> {
+        // flush is no-op for File (at least on unix), so we don't need to do
+        // anything here either.
+        Ok(())
+    }
+}
+
 impl OpenFiles {
     fn new(num_slots: usize) -> OpenFiles {
         let mut slots = Box::new(Vec::with_capacity(num_slots));
@@ -472,30 +515,60 @@ mod tests {
     use std::sync::Arc;
     use std::thread;
 
-    // Helper function to slurp contents of a file, starting at the current position,
-    // into a string
-    fn read_string<FD>(vfile: &mut FD) -> Result<String, Error>
-    where
-        FD: Read,
-    {
-        let mut buf = String::new();
-        vfile.read_to_string(&mut buf)?;
-        Ok(buf)
+    enum MaybeVirtualFile {
+        VirtualFile(VirtualFile),
+        File(File),
     }
 
-    // Helper function to slurp a portion of a file into a string
-    fn read_string_at<FD>(vfile: &mut FD, pos: u64, len: usize) -> Result<String, Error>
-    where
-        FD: FileExt,
-    {
-        let mut buf = Vec::new();
-        buf.resize(len, 0);
-        vfile.read_exact_at(&mut buf, pos)?;
-        Ok(String::from_utf8(buf).unwrap())
+    impl MaybeVirtualFile {
+        fn read_exact_at(&self, buf: &mut [u8], offset: u64) -> Result<(), Error> {
+            match self {
+                MaybeVirtualFile::VirtualFile(file) => file.read_exact_at(buf, offset),
+                MaybeVirtualFile::File(file) => file.read_exact_at(buf, offset),
+            }
+        }
+        async fn write_all_at(&self, buf: &[u8], offset: u64) -> Result<(), Error> {
+            match self {
+                MaybeVirtualFile::VirtualFile(file) => file.write_all_at(buf, offset).await,
+                MaybeVirtualFile::File(file) => file.write_all_at(buf, offset),
+            }
+        }
+        fn seek(&mut self, pos: SeekFrom) -> Result<u64, Error> {
+            match self {
+                MaybeVirtualFile::VirtualFile(file) => file.seek(pos),
+                MaybeVirtualFile::File(file) => file.seek(pos),
+            }
+        }
+        async fn write_all(&mut self, buf: &[u8]) -> Result<(), Error> {
+            match self {
+                MaybeVirtualFile::VirtualFile(file) => file.write_all(buf),
+                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) -> Result<String, Error> {
+            use std::io::Read;
+            let mut buf = String::new();
+            match self {
+                MaybeVirtualFile::VirtualFile(file) => file.read_to_string(&mut buf)?,
+                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(&mut self, pos: u64, len: usize) -> Result<String, Error> {
+            let mut buf = Vec::new();
+            buf.resize(len, 0);
+            self.read_exact_at(&mut buf, pos)?;
+            Ok(String::from_utf8(buf).unwrap())
+        }
     }
 
-    #[test]
-    fn test_virtual_files() -> Result<(), Error> {
+    #[tokio::test]
+    async fn test_virtual_files() -> Result<(), Error> {
         // The real work is done in the test_files() helper function. This
         // allows us to run the same set of tests against a native File, and
         // VirtualFile. We trust the native Files and wouldn't need to test them,
@@ -504,21 +577,23 @@ mod tests {
         // native files, you will run out of file descriptors if the ulimit
         // is low enough.)
         test_files("virtual_files", |path, open_options| {
-            VirtualFile::open_with_options(path, open_options)
+            let vf = VirtualFile::open_with_options(path, open_options)?;
+            Ok(MaybeVirtualFile::VirtualFile(vf))
         })
+        .await
     }
 
-    #[test]
-    fn test_physical_files() -> Result<(), Error> {
+    #[tokio::test]
+    async fn test_physical_files() -> Result<(), Error> {
         test_files("physical_files", |path, open_options| {
-            open_options.open(path)
+            Ok(MaybeVirtualFile::File(open_options.open(path)?))
         })
+        .await
     }
 
-    fn test_files<OF, FD>(testname: &str, openfunc: OF) -> Result<(), Error>
+    async fn test_files<OF>(testname: &str, openfunc: OF) -> Result<(), Error>
     where
-        FD: Read + Write + Seek + FileExt,
-        OF: Fn(&Path, &OpenOptions) -> Result<FD, std::io::Error>,
+        OF: Fn(&Path, &OpenOptions) -> Result<MaybeVirtualFile, std::io::Error>,
     {
         let testdir = crate::config::PageServerConf::test_repo_dir(testname);
         std::fs::create_dir_all(&testdir)?;
@@ -528,36 +603,36 @@ mod tests {
             &path_a,
             OpenOptions::new().write(true).create(true).truncate(true),
         )?;
-        file_a.write_all(b"foobar")?;
+        file_a.write_all(b"foobar").await?;
 
         // cannot read from a file opened in write-only mode
-        assert!(read_string(&mut file_a).is_err());
+        assert!(file_a.read_string().await.is_err());
 
         // Close the file and re-open for reading
         let mut file_a = openfunc(&path_a, OpenOptions::new().read(true))?;
 
         // cannot write to a file opened in read-only mode
-        assert!(file_a.write(b"bar").is_err());
+        assert!(file_a.write_all(b"bar").await.is_err());
 
         // Try simple read
-        assert_eq!("foobar", read_string(&mut file_a)?);
+        assert_eq!("foobar", file_a.read_string().await?);
 
         // It's positioned at the EOF now.
-        assert_eq!("", read_string(&mut file_a)?);
+        assert_eq!("", file_a.read_string().await?);
 
         // Test seeks.
         assert_eq!(file_a.seek(SeekFrom::Start(1))?, 1);
-        assert_eq!("oobar", read_string(&mut file_a)?);
+        assert_eq!("oobar", file_a.read_string().await?);
 
         assert_eq!(file_a.seek(SeekFrom::End(-2))?, 4);
-        assert_eq!("ar", read_string(&mut file_a)?);
+        assert_eq!("ar", file_a.read_string().await?);
 
         assert_eq!(file_a.seek(SeekFrom::Start(1))?, 1);
         assert_eq!(file_a.seek(SeekFrom::Current(2))?, 3);
-        assert_eq!("bar", read_string(&mut file_a)?);
+        assert_eq!("bar", file_a.read_string().await?);
 
         assert_eq!(file_a.seek(SeekFrom::Current(-5))?, 1);
-        assert_eq!("oobar", read_string(&mut file_a)?);
+        assert_eq!("oobar", file_a.read_string().await?);
 
         // Test erroneous seeks to before byte 0
         assert!(file_a.seek(SeekFrom::End(-7)).is_err());
@@ -565,7 +640,7 @@ mod tests {
         assert!(file_a.seek(SeekFrom::Current(-2)).is_err());
 
         // the erroneous seek should have left the position unchanged
-        assert_eq!("oobar", read_string(&mut file_a)?);
+        assert_eq!("oobar", file_a.read_string().await?);
 
         // Create another test file, and try FileExt functions on it.
         let path_b = testdir.join("file_b");
@@ -577,10 +652,10 @@ mod tests {
                 .create(true)
                 .truncate(true),
         )?;
-        file_b.write_all_at(b"BAR", 3)?;
-        file_b.write_all_at(b"FOO", 0)?;
+        file_b.write_all_at(b"BAR", 3).await?;
+        file_b.write_all_at(b"FOO", 0).await?;
 
-        assert_eq!(read_string_at(&mut file_b, 2, 3)?, "OBA");
+        assert_eq!(file_b.read_string_at(2, 3).await?, "OBA");
 
         // 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.)
@@ -591,7 +666,7 @@ mod tests {
         let mut vfiles = Vec::new();
         for _ in 0..100 {
             let mut vfile = openfunc(&path_b, OpenOptions::new().read(true))?;
-            assert_eq!("FOOBAR", read_string(&mut vfile)?);
+            assert_eq!("FOOBAR", vfile.read_string().await?);
             vfiles.push(vfile);
         }
 
@@ -600,13 +675,13 @@ mod tests {
 
         // 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", read_string(&mut file_a)?);
+        assert_eq!("oobar", file_a.read_string().await?);
 
         // Check that all the other FDs still work too. Use them in random order for
         // good measure.
         vfiles.as_mut_slice().shuffle(&mut thread_rng());
         for vfile in vfiles.iter_mut() {
-            assert_eq!("OOBAR", read_string_at(vfile, 1, 5)?);
+            assert_eq!("OOBAR", vfile.read_string_at(1, 5).await?);
         }
 
         Ok(())

proxy/src/auth/password_hack.rs

@@ -12,13 +12,19 @@ pub struct PasswordHackPayload {
 
 impl PasswordHackPayload {
     pub fn parse(bytes: &[u8]) -> Option<Self> {
-        // The format is `project=<utf-8>;<password-bytes>`.
-        let mut iter = bytes.splitn_str(2, ";");
-        let endpoint = iter.next()?.to_str().ok()?;
-        let endpoint = parse_endpoint_param(endpoint)?.to_owned();
-        let password = iter.next()?.to_owned();
+        // The format is `project=<utf-8>;<password-bytes>` or `project=<utf-8>$<password-bytes>`.
+        let separators = [";", "$"];
+        for sep in separators {
+            if let Some((endpoint, password)) = bytes.split_once_str(sep) {
+                let endpoint = endpoint.to_str().ok()?;
+                return Some(Self {
+                    endpoint: parse_endpoint_param(endpoint)?.to_owned(),
+                    password: password.to_owned(),
+                });
+            }
+        }
 
-        Some(Self { endpoint, password })
+        None
     }
 }
 
@@ -91,4 +97,23 @@ mod tests {
         assert_eq!(payload.endpoint, "foobar");
         assert_eq!(payload.password, b"pass;word");
     }
+
+    #[test]
+    fn parse_password_hack_payload_dollar() {
+        let bytes = b"";
+        assert!(PasswordHackPayload::parse(bytes).is_none());
+
+        let bytes = b"endpoint=";
+        assert!(PasswordHackPayload::parse(bytes).is_none());
+
+        let bytes = b"endpoint=$";
+        let payload = PasswordHackPayload::parse(bytes).expect("parsing failed");
+        assert_eq!(payload.endpoint, "");
+        assert_eq!(payload.password, b"");
+
+        let bytes = b"endpoint=foobar$pass$word";
+        let payload = PasswordHackPayload::parse(bytes).expect("parsing failed");
+        assert_eq!(payload.endpoint, "foobar");
+        assert_eq!(payload.password, b"pass$word");
+    }
 }

proxy/src/console/provider.rs

@@ -63,8 +63,8 @@ pub mod errors {
                         format!("{REQUEST_FAILED}: endpoint is disabled")
                     }
                     http::StatusCode::LOCKED => {
-                        // Status 423: project might be in maintenance mode (or bad state).
-                        format!("{REQUEST_FAILED}: endpoint is temporary unavailable")
+                        // Status 423: project might be in maintenance mode (or bad state), or quotas exceeded.
+                        format!("{REQUEST_FAILED}: endpoint is temporary unavailable. check your quotas and/or contract our support")
                     }
                     _ => REQUEST_FAILED.to_owned(),
                 },
@@ -81,9 +81,15 @@ pub mod errors {
                 // retry some temporary failures because the compute was in a bad state
                 // (bad request can be returned when the endpoint was in transition)
                 Self::Console {
-                    status: http::StatusCode::BAD_REQUEST | http::StatusCode::LOCKED,
+                    status: http::StatusCode::BAD_REQUEST,
                     ..
                 } => true,
+                // locked can be returned when the endpoint was in transition
+                // or when quotas are exceeded. don't retry when quotas are exceeded
+                Self::Console {
+                    status: http::StatusCode::LOCKED,
+                    ref text,
+                } => !text.contains("quota"),
                 // retry server errors
                 Self::Console { status, .. } if status.is_server_error() => true,
                 _ => false,

proxy/src/console/provider/neon.rs

@@ -16,12 +16,21 @@ use tracing::{error, info, info_span, warn, Instrument};
 pub struct Api {
     endpoint: http::Endpoint,
     caches: &'static ApiCaches,
+    jwt: String,
 }
 
 impl Api {
     /// Construct an API object containing the auth parameters.
     pub fn new(endpoint: http::Endpoint, caches: &'static ApiCaches) -> Self {
-        Self { endpoint, caches }
+        let jwt: String = match std::env::var("NEON_PROXY_TO_CONTROLPLANE_TOKEN") {
+            Ok(v) => v,
+            Err(_) => "".to_string(),
+        };
+        Self {
+            endpoint,
+            caches,
+            jwt,
+        }
     }
 
     pub fn url(&self) -> &str {
@@ -39,6 +48,7 @@ impl Api {
                 .endpoint
                 .get("proxy_get_role_secret")
                 .header("X-Request-ID", &request_id)
+                .header("Authorization", &self.jwt)
                 .query(&[("session_id", extra.session_id)])
                 .query(&[
                     ("application_name", extra.application_name),
@@ -83,6 +93,7 @@ impl Api {
                 .endpoint
                 .get("proxy_wake_compute")
                 .header("X-Request-ID", &request_id)
+                .header("Authorization", &self.jwt)
                 .query(&[("session_id", extra.session_id)])
                 .query(&[
                     ("application_name", extra.application_name),

proxy/src/http/websocket.rs

@@ -2,6 +2,7 @@ use crate::{
     cancellation::CancelMap,
     config::ProxyConfig,
     error::io_error,
+    protocol2::{ProxyProtocolAccept, WithClientIp},
     proxy::{handle_client, ClientMode},
 };
 use bytes::{Buf, Bytes};
@@ -292,6 +293,9 @@ pub async fn task_main(
 
     let mut addr_incoming = AddrIncoming::from_listener(ws_listener)?;
     let _ = addr_incoming.set_nodelay(true);
+    let addr_incoming = ProxyProtocolAccept {
+        incoming: addr_incoming,
+    };
 
     let tls_listener = TlsListener::new(tls_acceptor, addr_incoming).filter(|conn| {
         if let Err(err) = conn {
@@ -302,9 +306,11 @@ pub async fn task_main(
         }
     });
 
-    let make_svc =
-        hyper::service::make_service_fn(|stream: &tokio_rustls::server::TlsStream<AddrStream>| {
-            let sni_name = stream.get_ref().1.sni_hostname().map(|s| s.to_string());
+    let make_svc = hyper::service::make_service_fn(
+        |stream: &tokio_rustls::server::TlsStream<WithClientIp<AddrStream>>| {
+            let (io, tls) = stream.get_ref();
+            let peer_addr = io.client_addr().unwrap_or(io.inner.remote_addr());
+            let sni_name = tls.server_name().map(|s| s.to_string());
             let conn_pool = conn_pool.clone();
 
             async move {
@@ -319,13 +325,15 @@ pub async fn task_main(
                         ws_handler(req, config, conn_pool, cancel_map, session_id, sni_name)
                             .instrument(info_span!(
                                 "ws-client",
-                                session = %session_id
+                                session = %session_id,
+                                %peer_addr,
                             ))
                             .await
                     }
                 }))
             }
-        });
+        },
+    );
 
     hyper::Server::builder(accept::from_stream(tls_listener))
         .serve(make_svc)

proxy/src/lib.rs

@@ -16,6 +16,7 @@ pub mod http;
 pub mod logging;
 pub mod metrics;
 pub mod parse;
+pub mod protocol2;
 pub mod proxy;
 pub mod sasl;
 pub mod scram;

proxy/src/protocol2.rs

@@ -0,0 +1,479 @@
+//! Proxy Protocol V2 implementation
+
+use std::{
+    future::poll_fn,
+    future::Future,
+    io,
+    net::SocketAddr,
+    pin::{pin, Pin},
+    task::{ready, Context, Poll},
+};
+
+use bytes::{Buf, BytesMut};
+use hyper::server::conn::{AddrIncoming, AddrStream};
+use pin_project_lite::pin_project;
+use tls_listener::AsyncAccept;
+use tokio::io::{AsyncRead, AsyncReadExt, AsyncWrite, ReadBuf};
+
+pub struct ProxyProtocolAccept {
+    pub incoming: AddrIncoming,
+}
+
+pin_project! {
+    pub struct WithClientIp<T> {
+        #[pin]
+        pub inner: T,
+        buf: BytesMut,
+        tlv_bytes: u16,
+        state: ProxyParse,
+    }
+}
+
+#[derive(Clone, PartialEq, Debug)]
+enum ProxyParse {
+    NotStarted,
+
+    Finished(SocketAddr),
+    None,
+}
+
+impl<T: AsyncWrite> AsyncWrite for WithClientIp<T> {
+    #[inline]
+    fn poll_write(
+        self: Pin<&mut Self>,
+        cx: &mut Context<'_>,
+        buf: &[u8],
+    ) -> Poll<Result<usize, io::Error>> {
+        self.project().inner.poll_write(cx, buf)
+    }
+
+    #[inline]
+    fn poll_flush(self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<Result<(), io::Error>> {
+        self.project().inner.poll_flush(cx)
+    }
+
+    #[inline]
+    fn poll_shutdown(self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<Result<(), io::Error>> {
+        self.project().inner.poll_shutdown(cx)
+    }
+
+    #[inline]
+    fn poll_write_vectored(
+        self: Pin<&mut Self>,
+        cx: &mut Context<'_>,
+        bufs: &[io::IoSlice<'_>],
+    ) -> Poll<Result<usize, io::Error>> {
+        self.project().inner.poll_write_vectored(cx, bufs)
+    }
+
+    #[inline]
+    fn is_write_vectored(&self) -> bool {
+        self.inner.is_write_vectored()
+    }
+}
+
+impl<T> WithClientIp<T> {
+    pub fn new(inner: T) -> Self {
+        WithClientIp {
+            inner,
+            buf: BytesMut::with_capacity(128),
+            tlv_bytes: 0,
+            state: ProxyParse::NotStarted,
+        }
+    }
+
+    pub fn client_addr(&self) -> Option<SocketAddr> {
+        match self.state {
+            ProxyParse::Finished(socket) => Some(socket),
+            _ => None,
+        }
+    }
+}
+
+impl<T: AsyncRead + Unpin> WithClientIp<T> {
+    pub async fn wait_for_addr(&mut self) -> io::Result<Option<SocketAddr>> {
+        match self.state {
+            ProxyParse::NotStarted => {
+                let mut pin = Pin::new(&mut *self);
+                let addr = poll_fn(|cx| pin.as_mut().poll_client_ip(cx)).await?;
+                match addr {
+                    Some(addr) => self.state = ProxyParse::Finished(addr),
+                    None => self.state = ProxyParse::None,
+                }
+                Ok(addr)
+            }
+            ProxyParse::Finished(addr) => Ok(Some(addr)),
+            ProxyParse::None => Ok(None),
+        }
+    }
+}
+
+/// Proxy Protocol Version 2 Header
+const HEADER: [u8; 12] = [
+    0x0D, 0x0A, 0x0D, 0x0A, 0x00, 0x0D, 0x0A, 0x51, 0x55, 0x49, 0x54, 0x0A,
+];
+
+impl<T: AsyncRead> WithClientIp<T> {
+    /// implementation of <https://www.haproxy.org/download/2.4/doc/proxy-protocol.txt>
+    /// Version 2 (Binary Format)
+    fn poll_client_ip(
+        mut self: Pin<&mut Self>,
+        cx: &mut Context<'_>,
+    ) -> Poll<io::Result<Option<SocketAddr>>> {
+        // The binary header format starts with a constant 12 bytes block containing the protocol signature :
+        //    \x0D \x0A \x0D \x0A \x00 \x0D \x0A \x51 \x55 \x49 \x54 \x0A
+        while self.buf.len() < 16 {
+            let mut this = self.as_mut().project();
+            let bytes_read = pin!(this.inner.read_buf(this.buf)).poll(cx)?;
+
+            // exit for bad header
+            let len = usize::min(self.buf.len(), HEADER.len());
+            if self.buf[..len] != HEADER[..len] {
+                return Poll::Ready(Ok(None));
+            }
+
+            // if no more bytes available then exit
+            if ready!(bytes_read) == 0 {
+                return Poll::Ready(Ok(None));
+            };
+        }
+
+        // The next byte (the 13th one) is the protocol version and command.
+        // The highest four bits contains the version. As of this specification, it must
+        // always be sent as \x2 and the receiver must only accept this value.
+        let vc = self.buf[12];
+        let version = vc >> 4;
+        let command = vc & 0b1111;
+        if version != 2 {
+            return Poll::Ready(Err(io::Error::new(
+                io::ErrorKind::Other,
+                "invalid proxy protocol version. expected version 2",
+            )));
+        }
+        match command {
+            // the connection was established on purpose by the proxy
+            // without being relayed. The connection endpoints are the sender and the
+            // receiver. Such connections exist when the proxy sends health-checks to the
+            // server. The receiver must accept this connection as valid and must use the
+            // real connection endpoints and discard the protocol block including the
+            // family which is ignored.
+            0 => {}
+            // the connection was established on behalf of another node,
+            // and reflects the original connection endpoints. The receiver must then use
+            // the information provided in the protocol block to get original the address.
+            1 => {}
+            // other values are unassigned and must not be emitted by senders. Receivers
+            // must drop connections presenting unexpected values here.
+            _ => {
+                return Poll::Ready(Err(io::Error::new(
+                    io::ErrorKind::Other,
+                    "invalid proxy protocol command. expected local (0) or proxy (1)",
+                )))
+            }
+        };
+
+        // The 14th byte contains the transport protocol and address family. The highest 4
+        // bits contain the address family, the lowest 4 bits contain the protocol.
+        let ft = self.buf[13];
+        let address_length = match ft {
+            // - \x11 : TCP over IPv4 : the forwarded connection uses TCP over the AF_INET
+            //   protocol family. Address length is 2*4 + 2*2 = 12 bytes.
+            // - \x12 : UDP over IPv4 : the forwarded connection uses UDP over the AF_INET
+            //   protocol family. Address length is 2*4 + 2*2 = 12 bytes.
+            0x11 | 0x12 => 12,
+            // - \x21 : TCP over IPv6 : the forwarded connection uses TCP over the AF_INET6
+            //   protocol family. Address length is 2*16 + 2*2 = 36 bytes.
+            // - \x22 : UDP over IPv6 : the forwarded connection uses UDP over the AF_INET6
+            //   protocol family. Address length is 2*16 + 2*2 = 36 bytes.
+            0x21 | 0x22 => 36,
+            // unspecified or unix stream. ignore the addresses
+            _ => 0,
+        };
+
+        // The 15th and 16th bytes is the address length in bytes in network endian order.
+        // It is used so that the receiver knows how many address bytes to skip even when
+        // it does not implement the presented protocol. Thus the length of the protocol
+        // header in bytes is always exactly 16 + this value. When a sender presents a
+        // LOCAL connection, it should not present any address so it sets this field to
+        // zero. Receivers MUST always consider this field to skip the appropriate number
+        // of bytes and must not assume zero is presented for LOCAL connections. When a
+        // receiver accepts an incoming connection showing an UNSPEC address family or
+        // protocol, it may or may not decide to log the address information if present.
+        let remaining_length = u16::from_be_bytes(self.buf[14..16].try_into().unwrap());
+        if remaining_length < address_length {
+            return Poll::Ready(Err(io::Error::new(
+                io::ErrorKind::Other,
+                "invalid proxy protocol length. not enough to fit requested IP addresses",
+            )));
+        }
+
+        while self.buf.len() < 16 + address_length as usize {
+            let mut this = self.as_mut().project();
+            if ready!(pin!(this.inner.read_buf(this.buf)).poll(cx)?) == 0 {
+                return Poll::Ready(Err(io::Error::new(
+                    io::ErrorKind::UnexpectedEof,
+                    "stream closed while waiting for proxy protocol addresses",
+                )));
+            }
+        }
+
+        let this = self.as_mut().project();
+
+        // we are sure this is a proxy protocol v2 entry and we have read all the bytes we need
+        // discard the header we have parsed
+        this.buf.advance(16);
+
+        // Starting from the 17th byte, addresses are presented in network byte order.
+        // The address order is always the same :
+        //   - source layer 3 address in network byte order
+        //   - destination layer 3 address in network byte order
+        //   - source layer 4 address if any, in network byte order (port)
+        //   - destination layer 4 address if any, in network byte order (port)
+        let addresses = this.buf.split_to(address_length as usize);
+        let socket = match address_length {
+            12 => {
+                let src_addr: [u8; 4] = addresses[0..4].try_into().unwrap();
+                let src_port = u16::from_be_bytes(addresses[8..10].try_into().unwrap());
+                Some(SocketAddr::from((src_addr, src_port)))
+            }
+            36 => {
+                let src_addr: [u8; 16] = addresses[0..16].try_into().unwrap();
+                let src_port = u16::from_be_bytes(addresses[32..34].try_into().unwrap());
+                Some(SocketAddr::from((src_addr, src_port)))
+            }
+            _ => None,
+        };
+
+        *this.tlv_bytes = remaining_length - address_length;
+        self.as_mut().skip_tlv_inner();
+
+        Poll::Ready(Ok(socket))
+    }
+
+    #[cold]
+    fn read_ip(mut self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<io::Result<()>> {
+        let ip = ready!(self.as_mut().poll_client_ip(cx)?);
+        match ip {
+            Some(x) => *self.as_mut().project().state = ProxyParse::Finished(x),
+            None => *self.as_mut().project().state = ProxyParse::None,
+        }
+        Poll::Ready(Ok(()))
+    }
+
+    #[cold]
+    fn skip_tlv(mut self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<io::Result<()>> {
+        let mut this = self.as_mut().project();
+        // we know that this.buf is empty
+        debug_assert_eq!(this.buf.len(), 0);
+
+        this.buf.reserve((*this.tlv_bytes).clamp(0, 1024) as usize);
+        ready!(pin!(this.inner.read_buf(this.buf)).poll(cx)?);
+        self.skip_tlv_inner();
+
+        Poll::Ready(Ok(()))
+    }
+
+    fn skip_tlv_inner(self: Pin<&mut Self>) {
+        let tlv_bytes_read = match u16::try_from(self.buf.len()) {
+            // we read more than u16::MAX therefore we must have read the full tlv_bytes
+            Err(_) => self.tlv_bytes,
+            // we might not have read the full tlv bytes yet
+            Ok(n) => u16::min(n, self.tlv_bytes),
+        };
+        let this = self.project();
+        *this.tlv_bytes -= tlv_bytes_read;
+        this.buf.advance(tlv_bytes_read as usize);
+    }
+}
+
+impl<T: AsyncRead> AsyncRead for WithClientIp<T> {
+    #[inline]
+    fn poll_read(
+        mut self: Pin<&mut Self>,
+        cx: &mut Context<'_>,
+        buf: &mut ReadBuf<'_>,
+    ) -> Poll<io::Result<()>> {
+        // I'm assuming these 3 comparisons will be easy to branch predict.
+        // especially with the cold attributes
+        // which should make this read wrapper almost invisible
+
+        if let ProxyParse::NotStarted = self.state {
+            ready!(self.as_mut().read_ip(cx)?);
+        }
+
+        while self.tlv_bytes > 0 {
+            ready!(self.as_mut().skip_tlv(cx)?)
+        }
+
+        let this = self.project();
+        if this.buf.is_empty() {
+            this.inner.poll_read(cx, buf)
+        } else {
+            // we know that tlv_bytes is 0
+            debug_assert_eq!(*this.tlv_bytes, 0);
+
+            let write = usize::min(this.buf.len(), buf.remaining());
+            let slice = this.buf.split_to(write).freeze();
+            buf.put_slice(&slice);
+
+            // reset the allocation so it can be freed
+            if this.buf.is_empty() {
+                *this.buf = BytesMut::new();
+            }
+
+            Poll::Ready(Ok(()))
+        }
+    }
+}
+
+impl AsyncAccept for ProxyProtocolAccept {
+    type Connection = WithClientIp<AddrStream>;
+
+    type Error = io::Error;
+
+    fn poll_accept(
+        mut self: Pin<&mut Self>,
+        cx: &mut Context<'_>,
+    ) -> Poll<Option<Result<Self::Connection, Self::Error>>> {
+        let conn = ready!(Pin::new(&mut self.incoming).poll_accept(cx)?);
+        let Some(conn) = conn else {
+            return Poll::Ready(None);
+        };
+
+        Poll::Ready(Some(Ok(WithClientIp::new(conn))))
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::pin::pin;
+
+    use tokio::io::AsyncReadExt;
+
+    use crate::protocol2::{ProxyParse, WithClientIp};
+
+    #[tokio::test]
+    async fn test_ipv4() {
+        let header = super::HEADER
+            // Proxy command, IPV4 | TCP
+            .chain([(2 << 4) | 1, (1 << 4) | 1].as_slice())
+            // 12 + 3 bytes
+            .chain([0, 15].as_slice())
+            // src ip
+            .chain([127, 0, 0, 1].as_slice())
+            // dst ip
+            .chain([192, 168, 0, 1].as_slice())
+            // src port
+            .chain([255, 255].as_slice())
+            // dst port
+            .chain([1, 1].as_slice())
+            // TLV
+            .chain([1, 2, 3].as_slice());
+
+        let extra_data = [0x55; 256];
+
+        let mut read = pin!(WithClientIp::new(header.chain(extra_data.as_slice())));
+
+        let mut bytes = vec![];
+        read.read_to_end(&mut bytes).await.unwrap();
+
+        assert_eq!(bytes, extra_data);
+        assert_eq!(
+            read.state,
+            ProxyParse::Finished(([127, 0, 0, 1], 65535).into())
+        );
+    }
+
+    #[tokio::test]
+    async fn test_ipv6() {
+        let header = super::HEADER
+            // Proxy command, IPV6 | UDP
+            .chain([(2 << 4) | 1, (2 << 4) | 2].as_slice())
+            // 36 + 3 bytes
+            .chain([0, 39].as_slice())
+            // src ip
+            .chain([15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0].as_slice())
+            // dst ip
+            .chain([0, 15, 1, 14, 2, 13, 3, 12, 4, 11, 5, 10, 6, 9, 7, 8].as_slice())
+            // src port
+            .chain([1, 1].as_slice())
+            // dst port
+            .chain([255, 255].as_slice())
+            // TLV
+            .chain([1, 2, 3].as_slice());
+
+        let extra_data = [0x55; 256];
+
+        let mut read = pin!(WithClientIp::new(header.chain(extra_data.as_slice())));
+
+        let mut bytes = vec![];
+        read.read_to_end(&mut bytes).await.unwrap();
+
+        assert_eq!(bytes, extra_data);
+        assert_eq!(
+            read.state,
+            ProxyParse::Finished(
+                ([15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0], 257).into()
+            )
+        );
+    }
+
+    #[tokio::test]
+    async fn test_invalid() {
+        let data = [0x55; 256];
+
+        let mut read = pin!(WithClientIp::new(data.as_slice()));
+
+        let mut bytes = vec![];
+        read.read_to_end(&mut bytes).await.unwrap();
+        assert_eq!(bytes, data);
+        assert_eq!(read.state, ProxyParse::None);
+    }
+
+    #[tokio::test]
+    async fn test_short() {
+        let data = [0x55; 10];
+
+        let mut read = pin!(WithClientIp::new(data.as_slice()));
+
+        let mut bytes = vec![];
+        read.read_to_end(&mut bytes).await.unwrap();
+        assert_eq!(bytes, data);
+        assert_eq!(read.state, ProxyParse::None);
+    }
+
+    #[tokio::test]
+    async fn test_large_tlv() {
+        let tlv = vec![0x55; 32768];
+        let len = (12 + tlv.len() as u16).to_be_bytes();
+
+        let header = super::HEADER
+            // Proxy command, Inet << 4 | Stream
+            .chain([(2 << 4) | 1, (1 << 4) | 1].as_slice())
+            // 12 + 3 bytes
+            .chain(len.as_slice())
+            // src ip
+            .chain([55, 56, 57, 58].as_slice())
+            // dst ip
+            .chain([192, 168, 0, 1].as_slice())
+            // src port
+            .chain([255, 255].as_slice())
+            // dst port
+            .chain([1, 1].as_slice())
+            // TLV
+            .chain(tlv.as_slice());
+
+        let extra_data = [0xaa; 256];
+
+        let mut read = pin!(WithClientIp::new(header.chain(extra_data.as_slice())));
+
+        let mut bytes = vec![];
+        read.read_to_end(&mut bytes).await.unwrap();
+
+        assert_eq!(bytes, extra_data);
+        assert_eq!(
+            read.state,
+            ProxyParse::Finished(([55, 56, 57, 58], 65535).into())
+        );
+    }
+}

proxy/src/proxy.rs

@@ -7,6 +7,7 @@ use crate::{
     compute::{self, PostgresConnection},
     config::{ProxyConfig, TlsConfig},
     console::{self, errors::WakeComputeError, messages::MetricsAuxInfo, Api},
+    protocol2::WithClientIp,
     stream::{PqStream, Stream},
 };
 use anyhow::{bail, Context};
@@ -100,7 +101,7 @@ pub async fn task_main(
     loop {
         tokio::select! {
             accept_result = listener.accept() => {
-                let (socket, peer_addr) = accept_result?;
+                let (socket, _) = accept_result?;
 
                 let session_id = uuid::Uuid::new_v4();
                 let cancel_map = Arc::clone(&cancel_map);
@@ -108,13 +109,19 @@ pub async fn task_main(
                     async move {
                         info!("accepted postgres client connection");
 
+                        let mut socket = WithClientIp::new(socket);
+                        if let Some(ip) = socket.wait_for_addr().await? {
+                            tracing::Span::current().record("peer_addr", &tracing::field::display(ip));
+                        }
+
                         socket
+                            .inner
                             .set_nodelay(true)
                             .context("failed to set socket option")?;
 
                         handle_client(config, &cancel_map, session_id, socket, ClientMode::Tcp).await
                     }
-                    .instrument(info_span!("handle_client", ?session_id, %peer_addr))
+                    .instrument(info_span!("handle_client", ?session_id, peer_addr = tracing::field::Empty))
                     .unwrap_or_else(move |e| {
                         // Acknowledge that the task has finished with an error.
                         error!(?session_id, "per-client task finished with an error: {e:#}");

proxy/src/proxy/tests.rs

@@ -137,6 +137,7 @@ async fn dummy_proxy(
     auth: impl TestAuth + Send,
 ) -> anyhow::Result<()> {
     let cancel_map = CancelMap::default();
+    let client = WithClientIp::new(client);
     let (mut stream, _params) = handshake(client, tls.as_ref(), &cancel_map)
         .await?
         .context("handshake failed")?;

proxy/src/stream.rs

@@ -141,7 +141,7 @@ impl<S> Stream<S> {
     pub fn sni_hostname(&self) -> Option<&str> {
         match self {
             Stream::Raw { .. } => None,
-            Stream::Tls { tls } => tls.get_ref().1.sni_hostname(),
+            Stream::Tls { tls } => tls.get_ref().1.server_name(),
         }
     }
 }

s3_scrubber/Cargo.toml

@@ -0,0 +1,39 @@
+[package]
+name = "s3_scrubber"
+version = "0.1.0"
+edition.workspace = true
+license.workspace = true
+
+[dependencies]
+aws-sdk-s3.workspace = true
+aws-smithy-http.workspace = true
+aws-types.workspace = true
+either.workspace = true
+tokio-rustls.workspace = true
+anyhow.workspace = true
+hex.workspace = true
+thiserror.workspace = true
+rand.workspace = true
+bytes.workspace = true
+bincode.workspace = true
+crc32c.workspace = true
+serde.workspace = true
+serde_json.workspace = true
+serde_with.workspace = true
+workspace_hack.workspace = true
+utils.workspace = true
+
+tokio = { workspace = true, features = ["macros", "rt-multi-thread"] }
+chrono = { workspace = true, default-features = false, features = ["clock", "serde"] }
+reqwest = { workspace = true, default-features = false, features = ["rustls-tls", "json"] }
+aws-config = { workspace = true, default-features = false, features = ["rustls", "credentials-sso"] }
+
+pageserver = {path="../pageserver"}
+
+
+tracing.workspace = true
+tracing-subscriber.workspace = true
+clap.workspace = true
+
+atty = "0.2"
+tracing-appender = "0.2"

s3_scrubber/README.md

@@ -0,0 +1,93 @@
+# Neon S3 scrubber
+
+This tool directly accesses the S3 buckets used by the Neon `pageserver`
+and `safekeeper`, and does housekeeping such as cleaning up objects for tenants & timelines that no longer exist.
+
+## Usage
+
+### Generic Parameters
+
+#### S3
+
+Do `aws sso login --profile dev` to get the SSO access to the bucket to clean, get the SSO_ACCOUNT_ID for your profile (`cat ~/.aws/config` may help).
+
+- `SSO_ACCOUNT_ID`: Credentials id to use for accessing S3 buckets
+- `REGION`: A region where the bucket is located at.
+- `BUCKET`: Bucket name
+
+#### Console API
+
+_This section is only relevant if using a command that requires access to Neon's internal control plane_
+
+- `CLOUD_ADMIN_API_URL`: The URL base to use for checking tenant/timeline for existence via the Cloud API.  e.g. `https://<admin host>/admin`
+
+- `CLOUD_ADMIN_API_TOKEN`: The token to provide when querying the admin API. Get one on the corresponding console page, e.g. `https://<admin host>/app/settings/api-keys`
+
+### Commands
+
+#### `tidy`
+
+Iterate over S3 buckets for storage nodes, checking their contents and removing the data not present in the console. Node S3 data that's not removed is then further checked for discrepancies and, sometimes, validated.
+
+Unless the global `--delete` argument is provided, this command only dry-runs and logs
+what it would have deleted.
+
+```
+tidy --node-kind=<safekeeper|pageserver> [--depth=<tenant|timeline>] [--skip-validation]
+```
+
+- `--node-kind`: whether to inspect safekeeper or pageserver bucket prefix
+- `--depth`: whether to only search for deletable tenants, or also search for
+  deletable timelines within active tenants. Default: `tenant`
+- `--skip-validation`: skip additional post-deletion checks. Default: `false`
+
+For a selected S3 path, the tool lists the S3 bucket given for either tenants or both tenants and timelines — for every found entry, console API is queried: any deleted or missing in the API entity is scheduled for deletion from S3.
+
+If validation is enabled, only the non-deleted tenants' ones are checked.
+For pageserver, timelines' index_part.json on S3 is also checked for various discrepancies: no files are removed, even if there are "extra" S3 files not present in index_part.json: due to the way pageserver updates the remote storage, it's better to do such removals manually, stopping the corresponding tenant first.
+
+Command examples:
+
+`env SSO_ACCOUNT_ID=369495373322 REGION=eu-west-1 BUCKET=neon-dev-storage-eu-west-1 CLOUD_ADMIN_API_TOKEN=${NEON_CLOUD_ADMIN_API_STAGING_KEY} CLOUD_ADMIN_API_URL=[url] cargo run --release -- tidy --node-kind=safekeeper`
+
+`env SSO_ACCOUNT_ID=369495373322 REGION=us-east-2 BUCKET=neon-staging-storage-us-east-2 CLOUD_ADMIN_API_TOKEN=${NEON_CLOUD_ADMIN_API_STAGING_KEY} CLOUD_ADMIN_API_URL=[url] cargo run --release -- tidy --node-kind=pageserver --depth=timeline`
+
+When dry run stats look satisfying, use `-- --delete` before the `tidy` command to
+disable dry run and run the binary with deletion enabled.
+
+See these lines (and lines around) in the logs for the final stats:
+
+- `Finished listing the bucket for tenants`
+- `Finished active tenant and timeline validation`
+- `Total tenant deletion stats`
+- `Total timeline deletion stats`
+
+## Current implementation details
+
+- The tool does not have any peristent state currently: instead, it creates very verbose logs, with every S3 delete request logged, every tenant/timeline id check, etc.
+  Worse, any panic or early errored tasks might force the tool to exit without printing the final summary — all affected ids will still be in the logs though. The tool has retries inside it, so it's error-resistant up to some extent, and recent runs showed no traces of errors/panics.
+
+- Instead of checking non-deleted tenants' timelines instantly, the tool attempts to create separate tasks (futures) for that,
+  complicating the logic and slowing down the process, this should be fixed and done in one "task".
+
+- The tool does uses only publicly available remote resources (S3, console) and does not access pageserver/safekeeper nodes themselves.
+  Yet, its S3 set up should be prepared for running on any pageserver/safekeeper node, using node's S3 credentials, so the node API access logic could be implemented relatively simply on top.
+
+## Cleanup procedure:
+
+### Pageserver preparations
+
+If S3 state is altered first manually, pageserver in-memory state will contain wrong data about S3 state, and tenants/timelines may get recreated on S3 (due to any layer upload due to compaction, pageserver restart, etc.). So before proceeding, for tenants/timelines which are already deleted in the console, we must remove these from pageservers.
+
+First, we need to group pageservers by buckets, `https://<admin host>/admin/pageservers`` can be used for all env nodes, then `cat /storage/pageserver/data/pageserver.toml` on every node will show the bucket names and regions needed.
+
+Per bucket, for every pageserver id related, find deleted tenants:
+
+`curl -X POST "https://<admin_host>/admin/check_pageserver/{id}" -H "Accept: application/json" -H "Authorization: Bearer ${NEON_CLOUD_ADMIN_API_STAGING_KEY}" | jq`
+
+use `?check_timelines=true` to find deleted timelines, but the check runs a separate query on every alive tenant, so that could be long and time out for big pageservers.
+
+Note that some tenants/timelines could be marked as deleted in console, but console might continue querying the node later to fully remove the tenant/timeline: wait for some time before ensuring that the "extra" tenant/timeline is not going away by itself.
+
+When all IDs are collected, manually go to every pageserver and detach/delete the tenant/timeline.
+In future, the cleanup tool may access pageservers directly, but now it's only console and S3 it has access to.

s3_scrubber/src/checks.rs

@@ -0,0 +1,438 @@
+use std::collections::{hash_map, HashMap, HashSet};
+use std::sync::Arc;
+use std::time::Duration;
+
+use anyhow::Context;
+use aws_sdk_s3::Client;
+use tokio::io::AsyncReadExt;
+use tokio::task::JoinSet;
+use tracing::{error, info, info_span, warn, Instrument};
+
+use crate::cloud_admin_api::{BranchData, CloudAdminApiClient, ProjectId};
+use crate::delete_batch_producer::DeleteProducerStats;
+use crate::{list_objects_with_retries, RootTarget, MAX_RETRIES};
+use pageserver::tenant::storage_layer::LayerFileName;
+use pageserver::tenant::IndexPart;
+use utils::id::TenantTimelineId;
+
+pub async fn validate_pageserver_active_tenant_and_timelines(
+    s3_client: Arc<Client>,
+    s3_root: RootTarget,
+    admin_client: Arc<CloudAdminApiClient>,
+    batch_producer_stats: DeleteProducerStats,
+) -> anyhow::Result<BranchCheckStats> {
+    let Some(timeline_stats) = batch_producer_stats.timeline_stats else {
+        info!("No tenant-only checks, exiting");
+        return Ok(BranchCheckStats::default());
+    };
+
+    let s3_active_projects = batch_producer_stats
+        .tenant_stats
+        .active_entries
+        .into_iter()
+        .map(|project| (project.id.clone(), project))
+        .collect::<HashMap<_, _>>();
+    info!("Validating {} active tenants", s3_active_projects.len());
+
+    let mut s3_active_branches_per_project = HashMap::<ProjectId, Vec<BranchData>>::new();
+    let mut s3_blob_data = HashMap::<TenantTimelineId, S3TimelineBlobData>::new();
+    for active_branch in timeline_stats.active_entries {
+        let active_project_id = active_branch.project_id.clone();
+        let active_branch_id = active_branch.id.clone();
+        let active_timeline_id = active_branch.timeline_id;
+
+        s3_active_branches_per_project
+            .entry(active_project_id.clone())
+            .or_default()
+            .push(active_branch);
+
+        let Some(active_project) = s3_active_projects.get(&active_project_id) else {
+            error!("Branch {:?} for project {:?} has no such project in the active projects", active_branch_id, active_project_id);
+            continue;
+        };
+
+        let id = TenantTimelineId::new(active_project.tenant, active_timeline_id);
+        s3_blob_data.insert(
+            id,
+            list_timeline_blobs(&s3_client, id, &s3_root)
+                .await
+                .with_context(|| format!("List timeline {id} blobs"))?,
+        );
+    }
+
+    let mut branch_checks = JoinSet::new();
+    for (_, s3_active_project) in s3_active_projects {
+        let project_id = &s3_active_project.id;
+        let tenant_id = s3_active_project.tenant;
+
+        let mut console_active_branches =
+            branches_for_project_with_retries(&admin_client, project_id)
+                .await
+                .with_context(|| {
+                    format!("Client API branches for project {project_id:?} retrieval")
+                })?
+                .into_iter()
+                .map(|branch| (branch.id.clone(), branch))
+                .collect::<HashMap<_, _>>();
+
+        let active_branches = s3_active_branches_per_project
+            .remove(project_id)
+            .unwrap_or_default();
+        info!(
+            "Spawning tasks for {} tenant {} active timelines",
+            active_branches.len(),
+            tenant_id
+        );
+        for s3_active_branch in active_branches {
+            let console_branch = console_active_branches.remove(&s3_active_branch.id);
+            let timeline_id = s3_active_branch.timeline_id;
+            let id = TenantTimelineId::new(tenant_id, timeline_id);
+            let s3_data = s3_blob_data.remove(&id);
+            let s3_root = s3_root.clone();
+            branch_checks.spawn(
+                async move {
+                    let check_errors = branch_cleanup_and_check_errors(
+                        id,
+                        &s3_root,
+                        &s3_active_branch,
+                        console_branch,
+                        s3_data,
+                    )
+                    .await;
+                    (id, check_errors)
+                }
+                .instrument(info_span!("check_timeline", id = %id)),
+            );
+        }
+    }
+
+    let mut total_stats = BranchCheckStats::default();
+    while let Some((id, branch_check_errors)) = branch_checks
+        .join_next()
+        .await
+        .transpose()
+        .context("branch check task join")?
+    {
+        total_stats.add(id, branch_check_errors);
+    }
+    Ok(total_stats)
+}
+
+async fn branches_for_project_with_retries(
+    admin_client: &CloudAdminApiClient,
+    project_id: &ProjectId,
+) -> anyhow::Result<Vec<BranchData>> {
+    for _ in 0..MAX_RETRIES {
+        match admin_client.branches_for_project(project_id, false).await {
+            Ok(branches) => return Ok(branches),
+            Err(e) => {
+                error!("admin list branches for project {project_id:?} query failed: {e}");
+                tokio::time::sleep(Duration::from_secs(1)).await;
+            }
+        }
+    }
+
+    anyhow::bail!("Failed to list branches for project {project_id:?} {MAX_RETRIES} times")
+}
+
+#[derive(Debug, Default)]
+pub struct BranchCheckStats {
+    pub timelines_with_errors: HashMap<TenantTimelineId, Vec<String>>,
+    pub normal_timelines: HashSet<TenantTimelineId>,
+}
+
+impl BranchCheckStats {
+    pub fn add(&mut self, id: TenantTimelineId, check_errors: Vec<String>) {
+        if check_errors.is_empty() {
+            if !self.normal_timelines.insert(id) {
+                panic!("Checking branch with timeline {id} more than once")
+            }
+        } else {
+            match self.timelines_with_errors.entry(id) {
+                hash_map::Entry::Occupied(_) => {
+                    panic!("Checking branch with timeline {id} more than once")
+                }
+                hash_map::Entry::Vacant(v) => {
+                    v.insert(check_errors);
+                }
+            }
+        }
+    }
+}
+
+async fn branch_cleanup_and_check_errors(
+    id: TenantTimelineId,
+    s3_root: &RootTarget,
+    s3_active_branch: &BranchData,
+    console_branch: Option<BranchData>,
+    s3_data: Option<S3TimelineBlobData>,
+) -> Vec<String> {
+    info!(
+        "Checking timeline for branch branch {:?}/{:?}",
+        s3_active_branch.project_id, s3_active_branch.id
+    );
+    let mut branch_check_errors = Vec::new();
+
+    match console_branch {
+        Some(console_active_branch) => {
+            if console_active_branch.deleted {
+                branch_check_errors.push(format!("Timeline has deleted branch data in the console (id = {:?}, project_id = {:?}), recheck whether if it got removed during the check",
+                    s3_active_branch.id, s3_active_branch.project_id))
+            }
+        },
+        None => branch_check_errors.push(format!("Timeline has no branch data in the console (id = {:?}, project_id = {:?}), recheck whether if it got removed during the check",
+            s3_active_branch.id, s3_active_branch.project_id))
+    }
+
+    let mut keys_to_remove = Vec::new();
+
+    match s3_data {
+        Some(s3_data) => {
+            keys_to_remove.extend(s3_data.keys_to_remove);
+
+            match s3_data.blob_data {
+                BlobDataParseResult::Parsed {
+                    index_part,
+                    mut s3_layers,
+                } => {
+                    if !IndexPart::KNOWN_VERSIONS.contains(&index_part.get_version()) {
+                        branch_check_errors.push(format!(
+                            "index_part.json version: {}",
+                            index_part.get_version()
+                        ))
+                    }
+
+                    if index_part.metadata.disk_consistent_lsn()
+                        != index_part.get_disk_consistent_lsn()
+                    {
+                        branch_check_errors.push(format!(
+                                    "Mismatching disk_consistent_lsn in TimelineMetadata ({}) and in the index_part ({})",
+                                    index_part.metadata.disk_consistent_lsn(),
+                                    index_part.get_disk_consistent_lsn(),
+
+                                ))
+                    }
+
+                    if index_part.layer_metadata.is_empty() {
+                        // not an error, can happen for branches with zero writes, but notice that
+                        info!("index_part.json has no layers");
+                    }
+
+                    for (layer, metadata) in index_part.layer_metadata {
+                        if metadata.file_size == 0 {
+                            branch_check_errors.push(format!(
+                                            "index_part.json contains a layer {} that has 0 size in its layer metadata", layer.file_name(),
+                                        ))
+                        }
+
+                        if !s3_layers.remove(&layer) {
+                            branch_check_errors.push(format!(
+                                "index_part.json contains a layer {} that is not present in S3",
+                                layer.file_name(),
+                            ))
+                        }
+                    }
+
+                    if !s3_layers.is_empty() {
+                        branch_check_errors.push(format!(
+                            "index_part.json does not contain layers from S3: {:?}",
+                            s3_layers
+                                .iter()
+                                .map(|layer_name| layer_name.file_name())
+                                .collect::<Vec<_>>(),
+                        ));
+                        keys_to_remove.extend(s3_layers.iter().map(|layer_name| {
+                            let mut key = s3_root.timeline_root(id).prefix_in_bucket;
+                            let delimiter = s3_root.delimiter();
+                            if !key.ends_with(delimiter) {
+                                key.push_str(delimiter);
+                            }
+                            key.push_str(&layer_name.file_name());
+                            key
+                        }));
+                    }
+                }
+                BlobDataParseResult::Incorrect(parse_errors) => branch_check_errors.extend(
+                    parse_errors
+                        .into_iter()
+                        .map(|error| format!("parse error: {error}")),
+                ),
+            }
+        }
+        None => branch_check_errors.push("Timeline has no data on S3 at all".to_string()),
+    }
+
+    if branch_check_errors.is_empty() {
+        info!("No check errors found");
+    } else {
+        warn!("Found check errors: {branch_check_errors:?}");
+    }
+
+    if !keys_to_remove.is_empty() {
+        error!("The following keys should be removed from S3: {keys_to_remove:?}")
+    }
+
+    branch_check_errors
+}
+
+#[derive(Debug)]
+struct S3TimelineBlobData {
+    blob_data: BlobDataParseResult,
+    keys_to_remove: Vec<String>,
+}
+
+#[derive(Debug)]
+enum BlobDataParseResult {
+    Parsed {
+        index_part: IndexPart,
+        s3_layers: HashSet<LayerFileName>,
+    },
+    Incorrect(Vec<String>),
+}
+
+async fn list_timeline_blobs(
+    s3_client: &Client,
+    id: TenantTimelineId,
+    s3_root: &RootTarget,
+) -> anyhow::Result<S3TimelineBlobData> {
+    let mut s3_layers = HashSet::new();
+    let mut index_part_object = None;
+
+    let timeline_dir_target = s3_root.timeline_root(id);
+    let mut continuation_token = None;
+
+    let mut errors = Vec::new();
+    let mut keys_to_remove = Vec::new();
+
+    loop {
+        let fetch_response =
+            list_objects_with_retries(s3_client, &timeline_dir_target, continuation_token.clone())
+                .await?;
+
+        let subdirectories = fetch_response.common_prefixes().unwrap_or_default();
+        if !subdirectories.is_empty() {
+            errors.push(format!(
+                "S3 list response should not contain any subdirectories, but got {subdirectories:?}"
+            ));
+        }
+
+        for (object, key) in fetch_response
+            .contents()
+            .unwrap_or_default()
+            .iter()
+            .filter_map(|object| Some((object, object.key()?)))
+        {
+            let blob_name = key.strip_prefix(&timeline_dir_target.prefix_in_bucket);
+            match blob_name {
+                Some("index_part.json") => index_part_object = Some(object.clone()),
+                Some(maybe_layer_name) => match maybe_layer_name.parse::<LayerFileName>() {
+                    Ok(new_layer) => {
+                        s3_layers.insert(new_layer);
+                    }
+                    Err(e) => {
+                        errors.push(
+                            format!("S3 list response got an object with key {key} that is not a layer name: {e}"),
+                        );
+                        keys_to_remove.push(key.to_string());
+                    }
+                },
+                None => {
+                    errors.push(format!("S3 list response got an object with odd key {key}"));
+                    keys_to_remove.push(key.to_string());
+                }
+            }
+        }
+
+        match fetch_response.next_continuation_token {
+            Some(new_token) => continuation_token = Some(new_token),
+            None => break,
+        }
+    }
+
+    if index_part_object.is_none() {
+        errors.push("S3 list response got no index_part.json file".to_string());
+    }
+
+    if let Some(index_part_object_key) = index_part_object.as_ref().and_then(|object| object.key())
+    {
+        let index_part_bytes = download_object_with_retries(
+            s3_client,
+            &timeline_dir_target.bucket_name,
+            index_part_object_key,
+        )
+        .await
+        .context("index_part.json download")?;
+
+        match serde_json::from_slice(&index_part_bytes) {
+            Ok(index_part) => {
+                return Ok(S3TimelineBlobData {
+                    blob_data: BlobDataParseResult::Parsed {
+                        index_part,
+                        s3_layers,
+                    },
+                    keys_to_remove,
+                })
+            }
+            Err(index_parse_error) => errors.push(format!(
+                "index_part.json body parsing error: {index_parse_error}"
+            )),
+        }
+    } else {
+        errors.push(format!(
+            "Index part object {index_part_object:?} has no key"
+        ));
+    }
+
+    if errors.is_empty() {
+        errors.push(
+            "Unexpected: no errors did not lead to a successfully parsed blob return".to_string(),
+        );
+    }
+
+    Ok(S3TimelineBlobData {
+        blob_data: BlobDataParseResult::Incorrect(errors),
+        keys_to_remove,
+    })
+}
+
+async fn download_object_with_retries(
+    s3_client: &Client,
+    bucket_name: &str,
+    key: &str,
+) -> anyhow::Result<Vec<u8>> {
+    for _ in 0..MAX_RETRIES {
+        let mut body_buf = Vec::new();
+        let response_stream = match s3_client
+            .get_object()
+            .bucket(bucket_name)
+            .key(key)
+            .send()
+            .await
+        {
+            Ok(response) => response,
+            Err(e) => {
+                error!("Failed to download object for key {key}: {e}");
+                tokio::time::sleep(Duration::from_secs(1)).await;
+                continue;
+            }
+        };
+
+        match response_stream
+            .body
+            .into_async_read()
+            .read_to_end(&mut body_buf)
+            .await
+        {
+            Ok(bytes_read) => {
+                info!("Downloaded {bytes_read} bytes for object object with key {key}");
+                return Ok(body_buf);
+            }
+            Err(e) => {
+                error!("Failed to stream object body for key {key}: {e}");
+                tokio::time::sleep(Duration::from_secs(1)).await;
+            }
+        }
+    }
+
+    anyhow::bail!("Failed to download objects with key {key} {MAX_RETRIES} times")
+}

s3_scrubber/src/cloud_admin_api.rs

@@ -0,0 +1,418 @@
+#![allow(unused)]
+
+use chrono::{DateTime, Utc};
+use reqwest::{header, Client, Url};
+use tokio::sync::Semaphore;
+
+use utils::id::{TenantId, TimelineId};
+use utils::lsn::Lsn;
+
+#[derive(Debug)]
+pub struct Error {
+    context: String,
+    kind: ErrorKind,
+}
+
+impl Error {
+    fn new(context: String, kind: ErrorKind) -> Self {
+        Self { context, kind }
+    }
+}
+
+impl std::fmt::Display for Error {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        match &self.kind {
+            ErrorKind::RequestSend(e) => write!(
+                f,
+                "Failed to send a request. Context: {}, error: {}",
+                self.context, e
+            ),
+            ErrorKind::BodyRead(e) => {
+                write!(
+                    f,
+                    "Failed to read a request body. Context: {}, error: {}",
+                    self.context, e
+                )
+            }
+            ErrorKind::UnexpectedState => write!(f, "Unexpected state: {}", self.context),
+        }
+    }
+}
+
+#[derive(Debug, Clone, serde::Deserialize, Hash, PartialEq, Eq)]
+#[serde(transparent)]
+pub struct ProjectId(pub String);
+
+#[derive(Clone, Debug, serde::Deserialize, Hash, PartialEq, Eq)]
+#[serde(transparent)]
+pub struct BranchId(pub String);
+
+impl std::error::Error for Error {}
+
+#[derive(Debug)]
+pub enum ErrorKind {
+    RequestSend(reqwest::Error),
+    BodyRead(reqwest::Error),
+    UnexpectedState,
+}
+
+pub struct CloudAdminApiClient {
+    request_limiter: Semaphore,
+    token: String,
+    base_url: Url,
+    http_client: Client,
+}
+
+#[derive(Debug, serde::Deserialize)]
+struct AdminApiResponse<T> {
+    data: T,
+    total: Option<usize>,
+}
+
+#[derive(Debug, serde::Deserialize)]
+pub struct PageserverData {
+    pub id: u64,
+    pub created_at: DateTime<Utc>,
+    pub updated_at: DateTime<Utc>,
+    pub region_id: String,
+    pub version: i64,
+    pub instance_id: String,
+    pub port: u16,
+    pub http_host: String,
+    pub http_port: u16,
+    pub active: bool,
+    pub projects_count: usize,
+    pub availability_zone_id: String,
+}
+
+#[derive(Debug, Clone, serde::Deserialize)]
+pub struct SafekeeperData {
+    pub id: u64,
+    pub created_at: DateTime<Utc>,
+    pub updated_at: DateTime<Utc>,
+    pub region_id: String,
+    pub version: i64,
+    pub instance_id: String,
+    pub active: bool,
+    pub host: String,
+    pub port: u16,
+    pub projects_count: usize,
+    pub availability_zone_id: String,
+}
+
+#[serde_with::serde_as]
+#[derive(Debug, Clone, serde::Deserialize)]
+pub struct ProjectData {
+    pub id: ProjectId,
+    pub name: String,
+    pub region_id: String,
+    pub platform_id: String,
+    pub user_id: String,
+    pub pageserver_id: u64,
+    #[serde_as(as = "serde_with::DisplayFromStr")]
+    pub tenant: TenantId,
+    pub safekeepers: Vec<SafekeeperData>,
+    pub deleted: bool,
+    pub created_at: DateTime<Utc>,
+    pub updated_at: DateTime<Utc>,
+    pub pg_version: u32,
+    pub max_project_size: u64,
+    pub remote_storage_size: u64,
+    pub resident_size: u64,
+    pub synthetic_storage_size: u64,
+    pub compute_time: u64,
+    pub data_transfer: u64,
+    pub data_storage: u64,
+    pub maintenance_set: Option<String>,
+}
+
+#[serde_with::serde_as]
+#[derive(Debug, serde::Deserialize)]
+pub struct BranchData {
+    pub id: BranchId,
+    pub created_at: DateTime<Utc>,
+    pub updated_at: DateTime<Utc>,
+    pub name: String,
+    pub project_id: ProjectId,
+    #[serde_as(as = "serde_with::DisplayFromStr")]
+    pub timeline_id: TimelineId,
+    #[serde(default)]
+    pub parent_id: Option<BranchId>,
+    #[serde(default)]
+    #[serde_as(as = "Option<serde_with::DisplayFromStr>")]
+    pub parent_lsn: Option<Lsn>,
+    pub default: bool,
+    pub deleted: bool,
+    pub logical_size: Option<u64>,
+    pub physical_size: Option<u64>,
+    pub written_size: Option<u64>,
+}
+
+impl CloudAdminApiClient {
+    pub fn new(token: String, base_url: Url) -> Self {
+        Self {
+            token,
+            base_url,
+            request_limiter: Semaphore::new(200),
+            http_client: Client::new(), // TODO timeout configs at least
+        }
+    }
+
+    pub async fn find_tenant_project(
+        &self,
+        tenant_id: TenantId,
+    ) -> Result<Option<ProjectData>, Error> {
+        let _permit = self
+            .request_limiter
+            .acquire()
+            .await
+            .expect("Semaphore is not closed");
+
+        let response = self
+            .http_client
+            .get(self.append_url("/projects"))
+            .query(&[
+                ("tenant_id", tenant_id.to_string()),
+                ("show_deleted", "true".to_string()),
+            ])
+            .header(header::ACCEPT, "application/json")
+            .bearer_auth(&self.token)
+            .send()
+            .await
+            .map_err(|e| {
+                Error::new(
+                    "Find project for tenant".to_string(),
+                    ErrorKind::RequestSend(e),
+                )
+            })?;
+
+        let response: AdminApiResponse<Vec<ProjectData>> = response.json().await.map_err(|e| {
+            Error::new(
+                "Find project for tenant".to_string(),
+                ErrorKind::BodyRead(e),
+            )
+        })?;
+        match response.data.len() {
+            0 => Ok(None),
+            1 => Ok(Some(
+                response
+                    .data
+                    .into_iter()
+                    .next()
+                    .expect("Should have exactly one element"),
+            )),
+            too_many => Err(Error::new(
+                format!("Find project for tenant returned {too_many} projects instead of 0 or 1"),
+                ErrorKind::UnexpectedState,
+            )),
+        }
+    }
+
+    pub async fn find_timeline_branch(
+        &self,
+        timeline_id: TimelineId,
+    ) -> Result<Option<BranchData>, Error> {
+        let _permit = self
+            .request_limiter
+            .acquire()
+            .await
+            .expect("Semaphore is not closed");
+
+        let response = self
+            .http_client
+            .get(self.append_url("/branches"))
+            .query(&[
+                ("timeline_id", timeline_id.to_string()),
+                ("show_deleted", "true".to_string()),
+            ])
+            .header(header::ACCEPT, "application/json")
+            .bearer_auth(&self.token)
+            .send()
+            .await
+            .map_err(|e| {
+                Error::new(
+                    "Find branch for timeline".to_string(),
+                    ErrorKind::RequestSend(e),
+                )
+            })?;
+
+        let response: AdminApiResponse<Vec<BranchData>> = response.json().await.map_err(|e| {
+            Error::new(
+                "Find branch for timeline".to_string(),
+                ErrorKind::BodyRead(e),
+            )
+        })?;
+        match response.data.len() {
+            0 => Ok(None),
+            1 => Ok(Some(
+                response
+                    .data
+                    .into_iter()
+                    .next()
+                    .expect("Should have exactly one element"),
+            )),
+            too_many => Err(Error::new(
+                format!("Find branch for timeline returned {too_many} branches instead of 0 or 1"),
+                ErrorKind::UnexpectedState,
+            )),
+        }
+    }
+
+    pub async fn list_pageservers(&self) -> Result<Vec<PageserverData>, Error> {
+        let _permit = self
+            .request_limiter
+            .acquire()
+            .await
+            .expect("Semaphore is not closed");
+
+        let response = self
+            .http_client
+            .get(self.append_url("/pageservers"))
+            .header(header::ACCEPT, "application/json")
+            .bearer_auth(&self.token)
+            .send()
+            .await
+            .map_err(|e| Error::new("List pageservers".to_string(), ErrorKind::RequestSend(e)))?;
+
+        let response: AdminApiResponse<Vec<PageserverData>> = response
+            .json()
+            .await
+            .map_err(|e| Error::new("List pageservers".to_string(), ErrorKind::BodyRead(e)))?;
+
+        Ok(response.data)
+    }
+
+    pub async fn list_safekeepers(&self) -> Result<Vec<SafekeeperData>, Error> {
+        let _permit = self
+            .request_limiter
+            .acquire()
+            .await
+            .expect("Semaphore is not closed");
+
+        let response = self
+            .http_client
+            .get(self.append_url("/safekeepers"))
+            .header(header::ACCEPT, "application/json")
+            .bearer_auth(&self.token)
+            .send()
+            .await
+            .map_err(|e| Error::new("List safekeepers".to_string(), ErrorKind::RequestSend(e)))?;
+
+        let response: AdminApiResponse<Vec<SafekeeperData>> = response
+            .json()
+            .await
+            .map_err(|e| Error::new("List safekeepers".to_string(), ErrorKind::BodyRead(e)))?;
+
+        Ok(response.data)
+    }
+
+    pub async fn projects_for_pageserver(
+        &self,
+        pageserver_id: u64,
+        show_deleted: bool,
+    ) -> Result<Vec<ProjectData>, Error> {
+        let _permit = self
+            .request_limiter
+            .acquire()
+            .await
+            .expect("Semaphore is not closed");
+
+        let response = self
+            .http_client
+            .get(self.append_url("/projects"))
+            .query(&[
+                ("pageserver_id", &pageserver_id.to_string()),
+                ("show_deleted", &show_deleted.to_string()),
+            ])
+            .header(header::ACCEPT, "application/json")
+            .bearer_auth(&self.token)
+            .send()
+            .await
+            .map_err(|e| Error::new("Project for tenant".to_string(), ErrorKind::RequestSend(e)))?;
+
+        let response: AdminApiResponse<Vec<ProjectData>> = response
+            .json()
+            .await
+            .map_err(|e| Error::new("Project for tenant".to_string(), ErrorKind::BodyRead(e)))?;
+
+        Ok(response.data)
+    }
+
+    pub async fn project_for_tenant(
+        &self,
+        tenant_id: TenantId,
+        show_deleted: bool,
+    ) -> Result<Option<ProjectData>, Error> {
+        let _permit = self
+            .request_limiter
+            .acquire()
+            .await
+            .expect("Semaphore is not closed");
+
+        let response = self
+            .http_client
+            .get(self.append_url("/projects"))
+            .query(&[
+                ("search", &tenant_id.to_string()),
+                ("show_deleted", &show_deleted.to_string()),
+            ])
+            .header(header::ACCEPT, "application/json")
+            .bearer_auth(&self.token)
+            .send()
+            .await
+            .map_err(|e| Error::new("Project for tenant".to_string(), ErrorKind::RequestSend(e)))?;
+
+        let response: AdminApiResponse<Vec<ProjectData>> = response
+            .json()
+            .await
+            .map_err(|e| Error::new("Project for tenant".to_string(), ErrorKind::BodyRead(e)))?;
+
+        match response.data.as_slice() {
+            [] => Ok(None),
+            [_single] => Ok(Some(response.data.into_iter().next().unwrap())),
+            multiple => Err(Error::new(
+                format!("Got more than one project for tenant {tenant_id} : {multiple:?}"),
+                ErrorKind::UnexpectedState,
+            )),
+        }
+    }
+
+    pub async fn branches_for_project(
+        &self,
+        project_id: &ProjectId,
+        show_deleted: bool,
+    ) -> Result<Vec<BranchData>, Error> {
+        let _permit = self
+            .request_limiter
+            .acquire()
+            .await
+            .expect("Semaphore is not closed");
+
+        let response = self
+            .http_client
+            .get(self.append_url("/branches"))
+            .query(&[
+                ("project_id", &project_id.0),
+                ("show_deleted", &show_deleted.to_string()),
+            ])
+            .header(header::ACCEPT, "application/json")
+            .bearer_auth(&self.token)
+            .send()
+            .await
+            .map_err(|e| Error::new("Project for tenant".to_string(), ErrorKind::RequestSend(e)))?;
+
+        let response: AdminApiResponse<Vec<BranchData>> = response
+            .json()
+            .await
+            .map_err(|e| Error::new("Project for tenant".to_string(), ErrorKind::BodyRead(e)))?;
+
+        Ok(response.data)
+    }
+
+    fn append_url(&self, subpath: &str) -> Url {
+        // TODO fugly, but `.join` does not work when called
+        (self.base_url.to_string() + subpath)
+            .parse()
+            .unwrap_or_else(|e| panic!("Could not append {subpath} to base url: {e}"))
+    }
+}

s3_scrubber/src/delete_batch_producer.rs

@@ -0,0 +1,354 @@
+mod tenant_batch;
+mod timeline_batch;
+
+use std::future::Future;
+use std::str::FromStr;
+use std::sync::Arc;
+use std::time::Duration;
+
+use anyhow::Context;
+use aws_sdk_s3::Client;
+use either::Either;
+use tokio::sync::mpsc::UnboundedReceiver;
+use tokio::sync::Mutex;
+use tokio::task::{JoinHandle, JoinSet};
+use tracing::{error, info, info_span, Instrument};
+
+use crate::cloud_admin_api::{BranchData, CloudAdminApiClient, ProjectData};
+use crate::{list_objects_with_retries, RootTarget, S3Target, TraversingDepth, MAX_RETRIES};
+use utils::id::{TenantId, TenantTimelineId};
+
+/// Typical tenant to remove contains 1 layer and 1 index_part.json blobs
+/// Also, there are some non-standard tenants to remove, having more layers.
+/// delete_objects request allows up to 1000 keys, so be on a safe side and allow most
+/// batch processing tasks to do 1 delete objects request only.
+///
+/// Every batch item will be additionally S3 LS'ed later, so keep the batch size
+/// even lower to allow multiple concurrent tasks do the LS requests.
+const BATCH_SIZE: usize = 100;
+
+pub struct DeleteBatchProducer {
+    delete_tenants_sender_task: JoinHandle<anyhow::Result<ProcessedS3List<TenantId, ProjectData>>>,
+    delete_timelines_sender_task:
+        JoinHandle<anyhow::Result<ProcessedS3List<TenantTimelineId, BranchData>>>,
+    delete_batch_creator_task: JoinHandle<()>,
+    delete_batch_receiver: Arc<Mutex<UnboundedReceiver<DeleteBatch>>>,
+}
+
+pub struct DeleteProducerStats {
+    pub tenant_stats: ProcessedS3List<TenantId, ProjectData>,
+    pub timeline_stats: Option<ProcessedS3List<TenantTimelineId, BranchData>>,
+}
+
+impl DeleteProducerStats {
+    pub fn tenants_checked(&self) -> usize {
+        self.tenant_stats.entries_total
+    }
+
+    pub fn active_tenants(&self) -> usize {
+        self.tenant_stats.active_entries.len()
+    }
+
+    pub fn timelines_checked(&self) -> usize {
+        self.timeline_stats
+            .as_ref()
+            .map(|stats| stats.entries_total)
+            .unwrap_or(0)
+    }
+}
+
+#[derive(Debug, Default, Clone)]
+pub struct DeleteBatch {
+    pub tenants: Vec<TenantId>,
+    pub timelines: Vec<TenantTimelineId>,
+}
+
+impl DeleteBatch {
+    pub fn merge(&mut self, other: Self) {
+        self.tenants.extend(other.tenants);
+        self.timelines.extend(other.timelines);
+    }
+
+    pub fn len(&self) -> usize {
+        self.tenants.len() + self.timelines.len()
+    }
+
+    pub fn is_empty(&self) -> bool {
+        self.len() == 0
+    }
+}
+
+impl DeleteBatchProducer {
+    pub fn start(
+        admin_client: Arc<CloudAdminApiClient>,
+        s3_client: Arc<Client>,
+        s3_root_target: RootTarget,
+        traversing_depth: TraversingDepth,
+    ) -> Self {
+        let (delete_elements_sender, mut delete_elements_receiver) =
+            tokio::sync::mpsc::unbounded_channel();
+        let delete_elements_sender = Arc::new(delete_elements_sender);
+        let admin_client = Arc::new(admin_client);
+
+        let (projects_to_check_sender, mut projects_to_check_receiver) =
+            tokio::sync::mpsc::unbounded_channel();
+        let delete_tenants_root_target = s3_root_target.clone();
+        let delete_tenants_client = Arc::clone(&s3_client);
+        let delete_tenants_admin_client = Arc::clone(&admin_client);
+        let delete_sender = Arc::clone(&delete_elements_sender);
+        let delete_tenants_sender_task = tokio::spawn(
+            async move {
+                tenant_batch::schedule_cleanup_deleted_tenants(
+                    &delete_tenants_root_target,
+                    &delete_tenants_client,
+                    &delete_tenants_admin_client,
+                    projects_to_check_sender,
+                    delete_sender,
+                    traversing_depth,
+                )
+                .await
+            }
+            .instrument(info_span!("delete_tenants_sender")),
+        );
+        let delete_timelines_sender_task = tokio::spawn(async move {
+            timeline_batch::schedule_cleanup_deleted_timelines(
+                &s3_root_target,
+                &s3_client,
+                &admin_client,
+                &mut projects_to_check_receiver,
+                delete_elements_sender,
+            )
+            .in_current_span()
+            .await
+        });
+
+        let (delete_batch_sender, delete_batch_receiver) = tokio::sync::mpsc::unbounded_channel();
+        let delete_batch_creator_task = tokio::spawn(
+            async move {
+                'outer: loop {
+                    let mut delete_batch = DeleteBatch::default();
+                    while delete_batch.len() < BATCH_SIZE {
+                        match delete_elements_receiver.recv().await {
+                            Some(new_task) => match new_task {
+                                Either::Left(tenant_id) => delete_batch.tenants.push(tenant_id),
+                                Either::Right(timeline_id) => {
+                                    delete_batch.timelines.push(timeline_id)
+                                }
+                            },
+                            None => {
+                                info!("Task finished: sender dropped");
+                                delete_batch_sender.send(delete_batch).ok();
+                                break 'outer;
+                            }
+                        }
+                    }
+
+                    if !delete_batch.is_empty() {
+                        delete_batch_sender.send(delete_batch).ok();
+                    }
+                }
+            }
+            .instrument(info_span!("delete batch creator")),
+        );
+
+        Self {
+            delete_tenants_sender_task,
+            delete_timelines_sender_task,
+            delete_batch_creator_task,
+            delete_batch_receiver: Arc::new(Mutex::new(delete_batch_receiver)),
+        }
+    }
+
+    pub fn subscribe(&self) -> Arc<Mutex<UnboundedReceiver<DeleteBatch>>> {
+        self.delete_batch_receiver.clone()
+    }
+
+    pub async fn join(self) -> anyhow::Result<DeleteProducerStats> {
+        let (delete_tenants_task_result, delete_timelines_task_result, batch_task_result) = tokio::join!(
+            self.delete_tenants_sender_task,
+            self.delete_timelines_sender_task,
+            self.delete_batch_creator_task,
+        );
+
+        let tenant_stats = match delete_tenants_task_result {
+            Ok(Ok(stats)) => stats,
+            Ok(Err(tenant_deletion_error)) => return Err(tenant_deletion_error),
+            Err(join_error) => {
+                anyhow::bail!("Failed to join the delete tenant producing task: {join_error}")
+            }
+        };
+
+        let timeline_stats = match delete_timelines_task_result {
+            Ok(Ok(stats)) => Some(stats),
+            Ok(Err(timeline_deletion_error)) => return Err(timeline_deletion_error),
+            Err(join_error) => {
+                anyhow::bail!("Failed to join the delete timeline producing task: {join_error}")
+            }
+        };
+
+        match batch_task_result {
+            Ok(()) => (),
+            Err(join_error) => anyhow::bail!("Failed to join the batch forming task: {join_error}"),
+        };
+
+        Ok(DeleteProducerStats {
+            tenant_stats,
+            timeline_stats,
+        })
+    }
+}
+
+pub struct ProcessedS3List<I, A> {
+    pub entries_total: usize,
+    pub entries_to_delete: Vec<I>,
+    pub active_entries: Vec<A>,
+}
+
+impl<I, A> Default for ProcessedS3List<I, A> {
+    fn default() -> Self {
+        Self {
+            entries_total: 0,
+            entries_to_delete: Vec::new(),
+            active_entries: Vec::new(),
+        }
+    }
+}
+
+impl<I, A> ProcessedS3List<I, A> {
+    fn merge(&mut self, other: Self) {
+        self.entries_total += other.entries_total;
+        self.entries_to_delete.extend(other.entries_to_delete);
+        self.active_entries.extend(other.active_entries);
+    }
+
+    fn change_ids<NewI>(self, transform: impl Fn(I) -> NewI) -> ProcessedS3List<NewI, A> {
+        ProcessedS3List {
+            entries_total: self.entries_total,
+            entries_to_delete: self.entries_to_delete.into_iter().map(transform).collect(),
+            active_entries: self.active_entries,
+        }
+    }
+}
+
+async fn process_s3_target_recursively<F, Fut, I, E, A>(
+    s3_client: &Client,
+    target: &S3Target,
+    find_active_and_deleted_entries: F,
+) -> anyhow::Result<ProcessedS3List<I, A>>
+where
+    I: FromStr<Err = E> + Send + Sync,
+    E: Send + Sync + std::error::Error + 'static,
+    F: FnOnce(Vec<I>) -> Fut + Clone,
+    Fut: Future<Output = anyhow::Result<ProcessedS3List<I, A>>>,
+{
+    let mut continuation_token = None;
+    let mut total_entries = ProcessedS3List::default();
+
+    loop {
+        let fetch_response =
+            list_objects_with_retries(s3_client, target, continuation_token.clone()).await?;
+
+        let new_entry_ids = fetch_response
+            .common_prefixes()
+            .unwrap_or_default()
+            .iter()
+            .filter_map(|prefix| prefix.prefix())
+            .filter_map(|prefix| -> Option<&str> {
+                prefix
+                    .strip_prefix(&target.prefix_in_bucket)?
+                    .strip_suffix('/')
+            })
+            .map(|entry_id_str| {
+                entry_id_str
+                    .parse()
+                    .with_context(|| format!("Incorrect entry id str: {entry_id_str}"))
+            })
+            .collect::<anyhow::Result<Vec<I>>>()
+            .context("list and parse bucket's entry ids")?;
+
+        total_entries.merge(
+            (find_active_and_deleted_entries.clone())(new_entry_ids)
+                .await
+                .context("filter active and deleted entry ids")?,
+        );
+
+        match fetch_response.next_continuation_token {
+            Some(new_token) => continuation_token = Some(new_token),
+            None => break,
+        }
+    }
+
+    Ok(total_entries)
+}
+
+enum FetchResult<A> {
+    Found(A),
+    Deleted,
+    Absent,
+}
+
+async fn split_to_active_and_deleted_entries<I, A, F, Fut>(
+    new_entry_ids: Vec<I>,
+    find_active_entry: F,
+) -> anyhow::Result<ProcessedS3List<I, A>>
+where
+    I: std::fmt::Display + Send + Sync + 'static + Copy,
+    A: Send + 'static,
+    F: FnOnce(I) -> Fut + Send + Sync + 'static + Clone,
+    Fut: Future<Output = anyhow::Result<FetchResult<A>>> + Send,
+{
+    let entries_total = new_entry_ids.len();
+    let mut check_tasks = JoinSet::new();
+    let mut active_entries = Vec::with_capacity(entries_total);
+    let mut entries_to_delete = Vec::with_capacity(entries_total);
+
+    for new_entry_id in new_entry_ids {
+        let check_closure = find_active_entry.clone();
+        check_tasks.spawn(
+            async move {
+                (
+                    new_entry_id,
+                    async {
+                        for _ in 0..MAX_RETRIES {
+                            let closure_clone = check_closure.clone();
+                            match closure_clone(new_entry_id).await {
+                                Ok(active_entry) => return Ok(active_entry),
+                                Err(e) => {
+                                    error!("find active entry admin API call failed: {e}");
+                                    tokio::time::sleep(Duration::from_secs(1)).await;
+                                }
+                            }
+                        }
+
+                        anyhow::bail!("Failed to check entry {new_entry_id} {MAX_RETRIES} times")
+                    }
+                    .await,
+                )
+            }
+            .instrument(info_span!("filter_active_entries")),
+        );
+    }
+
+    while let Some(task_result) = check_tasks.join_next().await {
+        let (entry_id, entry_data_fetch_result) = task_result.context("task join")?;
+        match entry_data_fetch_result.context("entry data fetch")? {
+            FetchResult::Found(active_entry) => {
+                info!("Entry {entry_id} is alive, cannot delete");
+                active_entries.push(active_entry);
+            }
+            FetchResult::Deleted => {
+                info!("Entry {entry_id} deleted in the admin data, can safely delete");
+                entries_to_delete.push(entry_id);
+            }
+            FetchResult::Absent => {
+                info!("Entry {entry_id} absent in the admin data, can safely delete");
+                entries_to_delete.push(entry_id);
+            }
+        }
+    }
+    Ok(ProcessedS3List {
+        entries_total,
+        entries_to_delete,
+        active_entries,
+    })
+}

s3_scrubber/src/delete_batch_producer/tenant_batch.rs

@@ -0,0 +1,87 @@
+use std::sync::Arc;
+
+use anyhow::Context;
+use aws_sdk_s3::Client;
+use either::Either;
+use tokio::sync::mpsc::UnboundedSender;
+use tracing::info;
+
+use crate::cloud_admin_api::{CloudAdminApiClient, ProjectData};
+use crate::delete_batch_producer::FetchResult;
+use crate::{RootTarget, TraversingDepth};
+use utils::id::{TenantId, TenantTimelineId};
+
+use super::ProcessedS3List;
+
+pub async fn schedule_cleanup_deleted_tenants(
+    s3_root_target: &RootTarget,
+    s3_client: &Arc<Client>,
+    admin_client: &Arc<CloudAdminApiClient>,
+    projects_to_check_sender: UnboundedSender<ProjectData>,
+    delete_sender: Arc<UnboundedSender<Either<TenantId, TenantTimelineId>>>,
+    traversing_depth: TraversingDepth,
+) -> anyhow::Result<ProcessedS3List<TenantId, ProjectData>> {
+    info!(
+        "Starting to list the bucket from root {}",
+        s3_root_target.bucket_name()
+    );
+    s3_client
+        .head_bucket()
+        .bucket(s3_root_target.bucket_name())
+        .send()
+        .await
+        .with_context(|| format!("bucket {} was not found", s3_root_target.bucket_name()))?;
+
+    let check_client = Arc::clone(admin_client);
+    let tenant_stats = super::process_s3_target_recursively(
+        s3_client,
+        s3_root_target.tenants_root(),
+        |s3_tenants| async move {
+            let another_client = Arc::clone(&check_client);
+            super::split_to_active_and_deleted_entries(s3_tenants, move |tenant_id| async move {
+                let project_data = another_client
+                    .find_tenant_project(tenant_id)
+                    .await
+                    .with_context(|| format!("Tenant {tenant_id} project admin check"))?;
+
+                Ok(if let Some(console_project) = project_data {
+                    if console_project.deleted {
+                        delete_sender.send(Either::Left(tenant_id)).ok();
+                        FetchResult::Deleted
+                    } else {
+                        if traversing_depth == TraversingDepth::Timeline {
+                            projects_to_check_sender.send(console_project.clone()).ok();
+                        }
+                        FetchResult::Found(console_project)
+                    }
+                } else {
+                    delete_sender.send(Either::Left(tenant_id)).ok();
+                    FetchResult::Absent
+                })
+            })
+            .await
+        },
+    )
+    .await
+    .context("tenant batch processing")?;
+
+    info!(
+        "Among {} tenants, found {} tenants to delete and {} active ones",
+        tenant_stats.entries_total,
+        tenant_stats.entries_to_delete.len(),
+        tenant_stats.active_entries.len(),
+    );
+
+    let tenant_stats = match traversing_depth {
+        TraversingDepth::Tenant => {
+            info!("Finished listing the bucket for tenants only");
+            tenant_stats
+        }
+        TraversingDepth::Timeline => {
+            info!("Finished listing the bucket for tenants and sent {} active tenants to check for timelines", tenant_stats.active_entries.len());
+            tenant_stats
+        }
+    };
+
+    Ok(tenant_stats)
+}

s3_scrubber/src/delete_batch_producer/timeline_batch.rs

@@ -0,0 +1,102 @@
+use std::sync::Arc;
+
+use anyhow::Context;
+use aws_sdk_s3::Client;
+use either::Either;
+use tokio::sync::mpsc::{UnboundedReceiver, UnboundedSender};
+use tracing::{info, info_span, Instrument};
+
+use crate::cloud_admin_api::{BranchData, CloudAdminApiClient, ProjectData};
+use crate::delete_batch_producer::{FetchResult, ProcessedS3List};
+use crate::RootTarget;
+use utils::id::{TenantId, TenantTimelineId};
+
+pub async fn schedule_cleanup_deleted_timelines(
+    s3_root_target: &RootTarget,
+    s3_client: &Arc<Client>,
+    admin_client: &Arc<CloudAdminApiClient>,
+    projects_to_check_receiver: &mut UnboundedReceiver<ProjectData>,
+    delete_elements_sender: Arc<UnboundedSender<Either<TenantId, TenantTimelineId>>>,
+) -> anyhow::Result<ProcessedS3List<TenantTimelineId, BranchData>> {
+    info!(
+        "Starting to list the bucket from root {}",
+        s3_root_target.bucket_name()
+    );
+    s3_client
+        .head_bucket()
+        .bucket(s3_root_target.bucket_name())
+        .send()
+        .await
+        .with_context(|| format!("bucket {} was not found", s3_root_target.bucket_name()))?;
+
+    let mut timeline_stats = ProcessedS3List::default();
+    while let Some(project_to_check) = projects_to_check_receiver.recv().await {
+        let check_client = Arc::clone(admin_client);
+
+        let check_s3_client = Arc::clone(s3_client);
+
+        let check_delete_sender = Arc::clone(&delete_elements_sender);
+
+        let check_root = s3_root_target.clone();
+
+        let new_stats = async move {
+            let tenant_id_to_check = project_to_check.tenant;
+            let check_target = check_root.timelines_root(tenant_id_to_check);
+            let stats = super::process_s3_target_recursively(
+                &check_s3_client,
+                &check_target,
+                |s3_timelines| async move {
+                    let another_client = check_client.clone();
+                    super::split_to_active_and_deleted_entries(
+                        s3_timelines,
+                        move |timeline_id| async move {
+                            let console_branch = another_client
+                                .find_timeline_branch(timeline_id)
+                                .await
+                                .map_err(|e| {
+                                    anyhow::anyhow!(
+                                        "Timeline {timeline_id} branch admin check: {e}"
+                                    )
+                                })?;
+
+                            let id = TenantTimelineId::new(tenant_id_to_check, timeline_id);
+                            Ok(match console_branch {
+                                Some(console_branch) => {
+                                    if console_branch.deleted {
+                                        check_delete_sender.send(Either::Right(id)).ok();
+                                        FetchResult::Deleted
+                                    } else {
+                                        FetchResult::Found(console_branch)
+                                    }
+                                }
+                                None => {
+                                    check_delete_sender.send(Either::Right(id)).ok();
+                                    FetchResult::Absent
+                                }
+                            })
+                        },
+                    )
+                    .await
+                },
+            )
+            .await
+            .with_context(|| format!("tenant {tenant_id_to_check} timeline batch processing"))?
+            .change_ids(|timeline_id| TenantTimelineId::new(tenant_id_to_check, timeline_id));
+
+            Ok::<_, anyhow::Error>(stats)
+        }
+        .instrument(info_span!("delete_timelines_sender", tenant = %project_to_check.tenant))
+        .await?;
+
+        timeline_stats.merge(new_stats);
+    }
+
+    info!(
+        "Among {} timelines, found {} timelines to delete and {} active ones",
+        timeline_stats.entries_total,
+        timeline_stats.entries_to_delete.len(),
+        timeline_stats.active_entries.len(),
+    );
+
+    Ok(timeline_stats)
+}

s3_scrubber/src/lib.rs

@@ -0,0 +1,204 @@
+pub mod checks;
+pub mod cloud_admin_api;
+pub mod delete_batch_producer;
+mod s3_deletion;
+
+use std::env;
+use std::fmt::Display;
+use std::time::Duration;
+
+use aws_config::environment::EnvironmentVariableCredentialsProvider;
+use aws_config::imds::credentials::ImdsCredentialsProvider;
+use aws_config::meta::credentials::CredentialsProviderChain;
+use aws_config::sso::SsoCredentialsProvider;
+use aws_sdk_s3::config::Region;
+use aws_sdk_s3::{Client, Config};
+
+pub use s3_deletion::S3Deleter;
+use tracing::error;
+use tracing_appender::non_blocking::WorkerGuard;
+use tracing_subscriber::{fmt, prelude::*, EnvFilter};
+use utils::id::{TenantId, TenantTimelineId};
+
+const MAX_RETRIES: usize = 20;
+const CLOUD_ADMIN_API_TOKEN_ENV_VAR: &str = "CLOUD_ADMIN_API_TOKEN";
+
+#[derive(Debug, Clone)]
+pub struct S3Target {
+    pub bucket_name: String,
+    pub prefix_in_bucket: String,
+    pub delimiter: String,
+}
+
+#[derive(clap::ValueEnum, Debug, Clone, Copy, PartialEq, Eq)]
+pub enum TraversingDepth {
+    Tenant,
+    Timeline,
+}
+
+impl Display for TraversingDepth {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        f.write_str(match self {
+            Self::Tenant => "tenant",
+            Self::Timeline => "timeline",
+        })
+    }
+}
+
+impl S3Target {
+    pub fn with_sub_segment(&self, new_segment: &str) -> Self {
+        let mut new_self = self.clone();
+        let _ = new_self.prefix_in_bucket.pop();
+        new_self.prefix_in_bucket =
+            [&new_self.prefix_in_bucket, new_segment, ""].join(&new_self.delimiter);
+        new_self
+    }
+}
+
+#[derive(Clone)]
+pub enum RootTarget {
+    Pageserver(S3Target),
+    Safekeeper(S3Target),
+}
+
+impl RootTarget {
+    pub fn tenants_root(&self) -> &S3Target {
+        match self {
+            Self::Pageserver(root) => root,
+            Self::Safekeeper(root) => root,
+        }
+    }
+
+    pub fn tenant_root(&self, tenant_id: TenantId) -> S3Target {
+        self.tenants_root().with_sub_segment(&tenant_id.to_string())
+    }
+
+    pub fn timelines_root(&self, tenant_id: TenantId) -> S3Target {
+        match self {
+            Self::Pageserver(_) => self.tenant_root(tenant_id).with_sub_segment("timelines"),
+            Self::Safekeeper(_) => self.tenant_root(tenant_id),
+        }
+    }
+
+    pub fn timeline_root(&self, id: TenantTimelineId) -> S3Target {
+        self.timelines_root(id.tenant_id)
+            .with_sub_segment(&id.timeline_id.to_string())
+    }
+
+    pub fn bucket_name(&self) -> &str {
+        match self {
+            Self::Pageserver(root) => &root.bucket_name,
+            Self::Safekeeper(root) => &root.bucket_name,
+        }
+    }
+
+    pub fn delimiter(&self) -> &str {
+        match self {
+            Self::Pageserver(root) => &root.delimiter,
+            Self::Safekeeper(root) => &root.delimiter,
+        }
+    }
+}
+
+pub fn get_cloud_admin_api_token_or_exit() -> String {
+    match env::var(CLOUD_ADMIN_API_TOKEN_ENV_VAR) {
+        Ok(token) => token,
+        Err(env::VarError::NotPresent) => {
+            error!("{CLOUD_ADMIN_API_TOKEN_ENV_VAR} env variable is not present");
+            std::process::exit(1);
+        }
+        Err(env::VarError::NotUnicode(not_unicode_string)) => {
+            error!("{CLOUD_ADMIN_API_TOKEN_ENV_VAR} env variable's value is not a valid unicode string: {not_unicode_string:?}");
+            std::process::exit(1);
+        }
+    }
+}
+
+pub fn init_logging(binary_name: &str, dry_run: bool, node_kind: &str) -> WorkerGuard {
+    let file_name = if dry_run {
+        format!(
+            "{}_{}_{}__dry.log",
+            binary_name,
+            node_kind,
+            chrono::Utc::now().format("%Y_%m_%d__%H_%M_%S")
+        )
+    } else {
+        format!(
+            "{}_{}_{}.log",
+            binary_name,
+            node_kind,
+            chrono::Utc::now().format("%Y_%m_%d__%H_%M_%S")
+        )
+    };
+
+    let (file_writer, guard) =
+        tracing_appender::non_blocking(tracing_appender::rolling::never("./logs/", file_name));
+
+    let file_logs = fmt::Layer::new()
+        .with_target(false)
+        .with_ansi(false)
+        .with_writer(file_writer);
+    let stdout_logs = fmt::Layer::new()
+        .with_target(false)
+        .with_ansi(atty::is(atty::Stream::Stdout))
+        .with_writer(std::io::stdout);
+    tracing_subscriber::registry()
+        .with(EnvFilter::try_from_default_env().unwrap_or_else(|_| EnvFilter::new("info")))
+        .with(file_logs)
+        .with(stdout_logs)
+        .init();
+
+    guard
+}
+
+pub fn init_s3_client(account_id: String, bucket_region: Region) -> Client {
+    let credentials_provider = {
+        // uses "AWS_ACCESS_KEY_ID", "AWS_SECRET_ACCESS_KEY"
+        CredentialsProviderChain::first_try("env", EnvironmentVariableCredentialsProvider::new())
+            // uses sso
+            .or_else(
+                "sso",
+                SsoCredentialsProvider::builder()
+                    .account_id(account_id)
+                    .role_name("PowerUserAccess")
+                    .start_url("https://neondb.awsapps.com/start")
+                    .region(Region::from_static("eu-central-1"))
+                    .build(),
+            )
+            // uses imds v2
+            .or_else("imds", ImdsCredentialsProvider::builder().build())
+    };
+
+    let config = Config::builder()
+        .region(bucket_region)
+        .credentials_provider(credentials_provider)
+        .build();
+
+    Client::from_conf(config)
+}
+
+async fn list_objects_with_retries(
+    s3_client: &Client,
+    s3_target: &S3Target,
+    continuation_token: Option<String>,
+) -> anyhow::Result<aws_sdk_s3::operation::list_objects_v2::ListObjectsV2Output> {
+    for _ in 0..MAX_RETRIES {
+        match s3_client
+            .list_objects_v2()
+            .bucket(&s3_target.bucket_name)
+            .prefix(&s3_target.prefix_in_bucket)
+            .delimiter(&s3_target.delimiter)
+            .set_continuation_token(continuation_token.clone())
+            .send()
+            .await
+        {
+            Ok(response) => return Ok(response),
+            Err(e) => {
+                error!("list_objects_v2 query failed: {e}");
+                tokio::time::sleep(Duration::from_secs(1)).await;
+            }
+        }
+    }
+
+    anyhow::bail!("Failed to list objects {MAX_RETRIES} times")
+}

s3_scrubber/src/main.rs

@@ -0,0 +1,268 @@
+use std::collections::HashMap;
+use std::env;
+use std::fmt::Display;
+use std::num::NonZeroUsize;
+use std::sync::Arc;
+
+use anyhow::Context;
+use aws_sdk_s3::config::Region;
+use reqwest::Url;
+use s3_scrubber::cloud_admin_api::CloudAdminApiClient;
+use s3_scrubber::delete_batch_producer::DeleteBatchProducer;
+use s3_scrubber::{
+    checks, get_cloud_admin_api_token_or_exit, init_logging, init_s3_client, RootTarget, S3Deleter,
+    S3Target, TraversingDepth,
+};
+use tracing::{info, info_span, warn};
+
+use clap::{Parser, Subcommand, ValueEnum};
+
+#[derive(Parser)]
+#[command(author, version, about, long_about = None)]
+#[command(arg_required_else_help(true))]
+struct Cli {
+    #[command(subcommand)]
+    command: Command,
+
+    #[arg(short, long, default_value_t = false)]
+    delete: bool,
+}
+
+#[derive(ValueEnum, Clone, Copy, Eq, PartialEq)]
+enum NodeKind {
+    Safekeeper,
+    Pageserver,
+}
+
+impl NodeKind {
+    fn as_str(&self) -> &'static str {
+        match self {
+            Self::Safekeeper => "safekeeper",
+            Self::Pageserver => "pageserver",
+        }
+    }
+}
+
+impl Display for NodeKind {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        f.write_str(self.as_str())
+    }
+}
+
+#[derive(Subcommand)]
+enum Command {
+    Tidy {
+        #[arg(short, long)]
+        node_kind: NodeKind,
+        #[arg(short, long, default_value_t=TraversingDepth::Tenant)]
+        depth: TraversingDepth,
+        #[arg(short, long, default_value_t = false)]
+        skip_validation: bool,
+    },
+}
+
+struct BucketConfig {
+    region: String,
+    bucket: String,
+    sso_account_id: String,
+}
+
+impl Display for BucketConfig {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        write!(f, "{}/{}/{}", self.sso_account_id, self.region, self.bucket)
+    }
+}
+
+impl BucketConfig {
+    fn from_env() -> anyhow::Result<Self> {
+        let sso_account_id =
+            env::var("SSO_ACCOUNT_ID").context("'SSO_ACCOUNT_ID' param retrieval")?;
+        let region = env::var("REGION").context("'REGION' param retrieval")?;
+        let bucket = env::var("BUCKET").context("'BUCKET' param retrieval")?;
+
+        Ok(Self {
+            region,
+            bucket,
+            sso_account_id,
+        })
+    }
+}
+
+struct ConsoleConfig {
+    admin_api_url: Url,
+}
+
+impl ConsoleConfig {
+    fn from_env() -> anyhow::Result<Self> {
+        let admin_api_url: Url = env::var("CLOUD_ADMIN_API_URL")
+            .context("'CLOUD_ADMIN_API_URL' param retrieval")?
+            .parse()
+            .context("'CLOUD_ADMIN_API_URL' param parsing")?;
+
+        Ok(Self { admin_api_url })
+    }
+}
+
+async fn tidy(
+    cli: &Cli,
+    bucket_config: BucketConfig,
+    console_config: ConsoleConfig,
+    node_kind: NodeKind,
+    depth: TraversingDepth,
+    skip_validation: bool,
+) -> anyhow::Result<()> {
+    let binary_name = env::args()
+        .next()
+        .context("binary name in not the first argument")?;
+
+    let dry_run = !cli.delete;
+    let _guard = init_logging(&binary_name, dry_run, node_kind.as_str());
+    let _main_span = info_span!("tidy", binary = %binary_name, %dry_run).entered();
+
+    if dry_run {
+        info!("Dry run, not removing items for real");
+    } else {
+        warn!("Dry run disabled, removing bucket items for real");
+    }
+
+    info!("skip_validation={skip_validation}");
+
+    info!("Starting extra S3 removal in {bucket_config} for node kind '{node_kind}', traversing depth: {depth:?}");
+
+    info!("Starting extra tenant S3 removal in {bucket_config} for node kind '{node_kind}'");
+    let cloud_admin_api_client = Arc::new(CloudAdminApiClient::new(
+        get_cloud_admin_api_token_or_exit(),
+        console_config.admin_api_url,
+    ));
+
+    let bucket_region = Region::new(bucket_config.region);
+    let delimiter = "/".to_string();
+    let s3_client = Arc::new(init_s3_client(bucket_config.sso_account_id, bucket_region));
+    let s3_root = match node_kind {
+        NodeKind::Pageserver => RootTarget::Pageserver(S3Target {
+            bucket_name: bucket_config.bucket,
+            prefix_in_bucket: ["pageserver", "v1", "tenants", ""].join(&delimiter),
+            delimiter,
+        }),
+        NodeKind::Safekeeper => RootTarget::Safekeeper(S3Target {
+            bucket_name: bucket_config.bucket,
+            prefix_in_bucket: ["safekeeper", "v1", "wal", ""].join(&delimiter),
+            delimiter,
+        }),
+    };
+
+    let delete_batch_producer = DeleteBatchProducer::start(
+        Arc::clone(&cloud_admin_api_client),
+        Arc::clone(&s3_client),
+        s3_root.clone(),
+        depth,
+    );
+
+    let s3_deleter = S3Deleter::new(
+        dry_run,
+        NonZeroUsize::new(15).unwrap(),
+        Arc::clone(&s3_client),
+        delete_batch_producer.subscribe(),
+        s3_root.clone(),
+    );
+
+    let (deleter_task_result, batch_producer_task_result) =
+        tokio::join!(s3_deleter.remove_all(), delete_batch_producer.join());
+
+    let deletion_stats = deleter_task_result.context("s3 deletion")?;
+    info!(
+        "Deleted {} tenants ({} keys) and {} timelines ({} keys) total. Dry run: {}",
+        deletion_stats.deleted_tenant_keys.len(),
+        deletion_stats.deleted_tenant_keys.values().sum::<usize>(),
+        deletion_stats.deleted_timeline_keys.len(),
+        deletion_stats.deleted_timeline_keys.values().sum::<usize>(),
+        dry_run,
+    );
+    info!(
+        "Total tenant deletion stats: {:?}",
+        deletion_stats
+            .deleted_tenant_keys
+            .into_iter()
+            .map(|(id, key)| (id.to_string(), key))
+            .collect::<HashMap<_, _>>()
+    );
+    info!(
+        "Total timeline deletion stats: {:?}",
+        deletion_stats
+            .deleted_timeline_keys
+            .into_iter()
+            .map(|(id, key)| (id.to_string(), key))
+            .collect::<HashMap<_, _>>()
+    );
+
+    let batch_producer_stats = batch_producer_task_result.context("delete batch producer join")?;
+    info!(
+        "Total bucket tenants listed: {}; for {} active tenants, timelines checked: {}",
+        batch_producer_stats.tenants_checked(),
+        batch_producer_stats.active_tenants(),
+        batch_producer_stats.timelines_checked()
+    );
+
+    if node_kind == NodeKind::Pageserver {
+        info!("node_kind != pageserver, finish without performing validation step");
+        return Ok(());
+    }
+
+    if skip_validation {
+        info!("--skip-validation is set, exiting");
+        return Ok(());
+    }
+
+    info!("validating active tenants and timelines for pageserver S3 data");
+
+    // TODO kb real stats for validation + better stats for every place: add and print `min`, `max`, `mean` values at least
+    let validation_stats = checks::validate_pageserver_active_tenant_and_timelines(
+        s3_client,
+        s3_root,
+        cloud_admin_api_client,
+        batch_producer_stats,
+    )
+    .await
+    .context("active tenant and timeline validation")?;
+    info!("Finished active tenant and timeline validation, correct timelines: {}, timeline validation errors: {}",
+        validation_stats.normal_timelines.len(), validation_stats.timelines_with_errors.len());
+    if !validation_stats.timelines_with_errors.is_empty() {
+        warn!(
+            "Validation errors: {:#?}",
+            validation_stats
+                .timelines_with_errors
+                .into_iter()
+                .map(|(id, errors)| (id.to_string(), format!("{errors:?}")))
+                .collect::<HashMap<_, _>>()
+        );
+    }
+
+    info!("Done");
+    Ok(())
+}
+
+#[tokio::main]
+async fn main() -> anyhow::Result<()> {
+    let cli = Cli::parse();
+
+    let bucket_config = BucketConfig::from_env()?;
+
+    match &cli.command {
+        Command::Tidy {
+            node_kind,
+            depth,
+            skip_validation,
+        } => {
+            let console_config = ConsoleConfig::from_env()?;
+            tidy(
+                &cli,
+                bucket_config,
+                console_config,
+                *node_kind,
+                *depth,
+                *skip_validation,
+            )
+            .await
+        }
+    }
+}

s3_scrubber/src/s3_deletion.rs

@@ -0,0 +1,434 @@
+use std::collections::BTreeMap;
+use std::num::NonZeroUsize;
+use std::sync::Arc;
+use std::time::Duration;
+
+use anyhow::Context;
+use aws_sdk_s3::types::{Delete, ObjectIdentifier};
+use aws_sdk_s3::Client;
+use tokio::sync::mpsc::error::TryRecvError;
+use tokio::sync::mpsc::UnboundedReceiver;
+use tokio::sync::Mutex;
+use tokio::task::JoinSet;
+use tracing::{debug, error, info, info_span, Instrument};
+
+use crate::delete_batch_producer::DeleteBatch;
+use crate::{list_objects_with_retries, RootTarget, S3Target, TenantId, MAX_RETRIES};
+use utils::id::TenantTimelineId;
+
+pub struct S3Deleter {
+    dry_run: bool,
+    concurrent_tasks_count: NonZeroUsize,
+    delete_batch_receiver: Arc<Mutex<UnboundedReceiver<DeleteBatch>>>,
+    s3_client: Arc<Client>,
+    s3_target: RootTarget,
+}
+
+impl S3Deleter {
+    pub fn new(
+        dry_run: bool,
+        concurrent_tasks_count: NonZeroUsize,
+        s3_client: Arc<Client>,
+        delete_batch_receiver: Arc<Mutex<UnboundedReceiver<DeleteBatch>>>,
+        s3_target: RootTarget,
+    ) -> Self {
+        Self {
+            dry_run,
+            concurrent_tasks_count,
+            delete_batch_receiver,
+            s3_client,
+            s3_target,
+        }
+    }
+
+    pub async fn remove_all(self) -> anyhow::Result<DeletionStats> {
+        let mut deletion_tasks = JoinSet::new();
+        for id in 0..self.concurrent_tasks_count.get() {
+            let closure_client = Arc::clone(&self.s3_client);
+            let closure_s3_target = self.s3_target.clone();
+            let closure_batch_receiver = Arc::clone(&self.delete_batch_receiver);
+            let dry_run = self.dry_run;
+            deletion_tasks.spawn(
+                async move {
+                    info!("Task started");
+                    (
+                        id,
+                        async move {
+                            let mut task_stats = DeletionStats::default();
+                            loop {
+                                let mut guard = closure_batch_receiver.lock().await;
+                                let receiver_result = guard.try_recv();
+                                drop(guard);
+                                match receiver_result {
+                                    Ok(batch) => {
+                                        let stats = delete_batch(
+                                            &closure_client,
+                                            &closure_s3_target,
+                                            batch,
+                                            dry_run,
+                                        )
+                                        .await
+                                        .context("batch deletion")?;
+                                        debug!(
+                                            "Batch processed, number of objects deleted per tenant in the batch is: {}, per timeline — {}",
+                                            stats.deleted_tenant_keys.len(),
+                                            stats.deleted_timeline_keys.len(),
+                                        );
+                                        task_stats.merge(stats);
+                                    }
+                                    Err(TryRecvError::Empty) => {
+                                        debug!("No tasks yet, waiting");
+                                        tokio::time::sleep(Duration::from_secs(1)).await;
+                                        continue;
+                                    }
+                                    Err(TryRecvError::Disconnected) => {
+                                        info!("Task finished: sender dropped");
+                                        return Ok(task_stats);
+                                    }
+                                }
+                            }
+                        }
+                        .in_current_span()
+                        .await,
+                    )
+                }
+                .instrument(info_span!("deletion_task", %id)),
+            );
+        }
+
+        let mut total_stats = DeletionStats::default();
+        while let Some(task_result) = deletion_tasks.join_next().await {
+            match task_result {
+                Ok((id, Ok(task_stats))) => {
+                    info!("Task {id} completed");
+                    total_stats.merge(task_stats);
+                }
+                Ok((id, Err(e))) => {
+                    error!("Task {id} failed: {e:#}");
+                    return Err(e);
+                }
+                Err(join_error) => anyhow::bail!("Failed to join on a task: {join_error:?}"),
+            }
+        }
+
+        Ok(total_stats)
+    }
+}
+
+/// S3 delete_objects allows up to 1000 keys to be passed in a single request.
+/// Yet if you pass too many key requests, apparently S3 could return with OK and
+/// actually delete nothing, so keep the number lower.
+const MAX_ITEMS_TO_DELETE: usize = 200;
+
+#[derive(Debug, Default)]
+pub struct DeletionStats {
+    pub deleted_tenant_keys: BTreeMap<TenantId, usize>,
+    pub deleted_timeline_keys: BTreeMap<TenantTimelineId, usize>,
+}
+
+impl DeletionStats {
+    fn merge(&mut self, other: Self) {
+        self.deleted_tenant_keys.extend(other.deleted_tenant_keys);
+        self.deleted_timeline_keys
+            .extend(other.deleted_timeline_keys);
+    }
+}
+
+async fn delete_batch(
+    s3_client: &Client,
+    s3_target: &RootTarget,
+    batch: DeleteBatch,
+    dry_run: bool,
+) -> anyhow::Result<DeletionStats> {
+    let (deleted_tenant_keys, deleted_timeline_keys) = tokio::join!(
+        delete_tenants_batch(batch.tenants, s3_target, s3_client, dry_run),
+        delete_timelines_batch(batch.timelines, s3_target, s3_client, dry_run),
+    );
+
+    Ok(DeletionStats {
+        deleted_tenant_keys: deleted_tenant_keys.context("tenant batch deletion")?,
+        deleted_timeline_keys: deleted_timeline_keys.context("timeline batch deletion")?,
+    })
+}
+
+async fn delete_tenants_batch(
+    batched_tenants: Vec<TenantId>,
+    s3_target: &RootTarget,
+    s3_client: &Client,
+    dry_run: bool,
+) -> Result<BTreeMap<TenantId, usize>, anyhow::Error> {
+    info!("Deleting tenants batch of size {}", batched_tenants.len());
+    info!("Tenant ids to remove: {batched_tenants:?}");
+    let deleted_keys = delete_elements(
+        &batched_tenants,
+        s3_target,
+        s3_client,
+        dry_run,
+        |root_target, tenant_to_delete| root_target.tenant_root(tenant_to_delete),
+    )
+    .await?;
+
+    if !dry_run {
+        let mut last_err = None;
+        for _ in 0..MAX_RETRIES {
+            match ensure_tenant_batch_deleted(s3_client, s3_target, &batched_tenants).await {
+                Ok(()) => {
+                    last_err = None;
+                    break;
+                }
+                Err(e) => {
+                    error!("Failed to ensure the tenant batch is deleted: {e}");
+                    last_err = Some(e);
+                }
+            }
+        }
+
+        if let Some(e) = last_err {
+            anyhow::bail!(
+                "Failed to ensure that tenant batch is deleted {MAX_RETRIES} times: {e:?}"
+            );
+        }
+    }
+
+    Ok(deleted_keys)
+}
+
+async fn delete_timelines_batch(
+    batched_timelines: Vec<TenantTimelineId>,
+    s3_target: &RootTarget,
+    s3_client: &Client,
+    dry_run: bool,
+) -> Result<BTreeMap<TenantTimelineId, usize>, anyhow::Error> {
+    info!(
+        "Deleting timelines batch of size {}",
+        batched_timelines.len()
+    );
+    info!(
+        "Timeline ids to remove: {:?}",
+        batched_timelines
+            .iter()
+            .map(|id| id.to_string())
+            .collect::<Vec<_>>()
+    );
+    let deleted_keys = delete_elements(
+        &batched_timelines,
+        s3_target,
+        s3_client,
+        dry_run,
+        |root_target, timeline_to_delete| root_target.timeline_root(timeline_to_delete),
+    )
+    .await?;
+
+    if !dry_run {
+        let mut last_err = None;
+        for _ in 0..MAX_RETRIES {
+            match ensure_timeline_batch_deleted(s3_client, s3_target, &batched_timelines).await {
+                Ok(()) => {
+                    last_err = None;
+                    break;
+                }
+                Err(e) => {
+                    error!("Failed to ensure the timelines batch is deleted: {e}");
+                    last_err = Some(e);
+                }
+            }
+        }
+
+        if let Some(e) = last_err {
+            anyhow::bail!(
+                "Failed to ensure that timeline batch is deleted {MAX_RETRIES} times: {e:?}"
+            );
+        }
+    }
+    Ok(deleted_keys)
+}
+
+async fn delete_elements<I>(
+    batched_ids: &Vec<I>,
+    s3_target: &RootTarget,
+    s3_client: &Client,
+    dry_run: bool,
+    target_producer: impl Fn(&RootTarget, I) -> S3Target,
+) -> Result<BTreeMap<I, usize>, anyhow::Error>
+where
+    I: Ord + PartialOrd + Copy,
+{
+    let mut deleted_keys = BTreeMap::new();
+    let mut object_ids_to_delete = Vec::with_capacity(MAX_ITEMS_TO_DELETE);
+    for &id_to_delete in batched_ids {
+        let mut continuation_token = None;
+        let mut subtargets = vec![target_producer(s3_target, id_to_delete)];
+        while let Some(current_target) = subtargets.pop() {
+            loop {
+                let fetch_response = list_objects_with_retries(
+                    s3_client,
+                    &current_target,
+                    continuation_token.clone(),
+                )
+                .await?;
+
+                for object_id in fetch_response
+                    .contents()
+                    .unwrap_or_default()
+                    .iter()
+                    .filter_map(|object| object.key())
+                    .map(|key| ObjectIdentifier::builder().key(key).build())
+                {
+                    if object_ids_to_delete.len() >= MAX_ITEMS_TO_DELETE {
+                        let object_ids_for_request = std::mem::replace(
+                            &mut object_ids_to_delete,
+                            Vec::with_capacity(MAX_ITEMS_TO_DELETE),
+                        );
+                        send_delete_request(
+                            s3_client,
+                            s3_target.bucket_name(),
+                            object_ids_for_request,
+                            dry_run,
+                        )
+                        .await
+                        .context("object ids deletion")?;
+                    }
+
+                    object_ids_to_delete.push(object_id);
+                    *deleted_keys.entry(id_to_delete).or_default() += 1;
+                }
+
+                subtargets.extend(
+                    fetch_response
+                        .common_prefixes()
+                        .unwrap_or_default()
+                        .iter()
+                        .filter_map(|common_prefix| common_prefix.prefix())
+                        .map(|prefix| {
+                            let mut new_target = current_target.clone();
+                            new_target.prefix_in_bucket = prefix.to_string();
+                            new_target
+                        }),
+                );
+
+                match fetch_response.next_continuation_token {
+                    Some(new_token) => continuation_token = Some(new_token),
+                    None => break,
+                }
+            }
+        }
+    }
+    if !object_ids_to_delete.is_empty() {
+        info!("Removing last objects of the batch");
+        send_delete_request(
+            s3_client,
+            s3_target.bucket_name(),
+            object_ids_to_delete,
+            dry_run,
+        )
+        .await
+        .context("Last object ids deletion")?;
+    }
+    Ok(deleted_keys)
+}
+
+pub async fn send_delete_request(
+    s3_client: &Client,
+    bucket_name: &str,
+    ids: Vec<ObjectIdentifier>,
+    dry_run: bool,
+) -> anyhow::Result<()> {
+    info!("Removing {} object ids from S3", ids.len());
+    info!("Object ids to remove: {ids:?}");
+    let delete_request = s3_client
+        .delete_objects()
+        .bucket(bucket_name)
+        .delete(Delete::builder().set_objects(Some(ids)).build());
+    if dry_run {
+        info!("Dry run, skipping the actual removal");
+        Ok(())
+    } else {
+        let original_request = delete_request.clone();
+
+        for _ in 0..MAX_RETRIES {
+            match delete_request
+                .clone()
+                .send()
+                .await
+                .context("delete request processing")
+            {
+                Ok(delete_response) => {
+                    info!("Delete response: {delete_response:?}");
+                    match delete_response.errors() {
+                        Some(delete_errors) => {
+                            error!("Delete request returned errors: {delete_errors:?}");
+                            tokio::time::sleep(Duration::from_secs(1)).await;
+                        }
+                        None => {
+                            info!("Successfully removed an object batch from S3");
+                            return Ok(());
+                        }
+                    }
+                }
+                Err(e) => {
+                    error!("Failed to send a delete request: {e:#}");
+                    tokio::time::sleep(Duration::from_secs(1)).await;
+                }
+            }
+        }
+
+        error!("Failed to do deletion, request: {original_request:?}");
+        anyhow::bail!("Failed to run deletion request {MAX_RETRIES} times");
+    }
+}
+
+async fn ensure_tenant_batch_deleted(
+    s3_client: &Client,
+    s3_target: &RootTarget,
+    batch: &[TenantId],
+) -> anyhow::Result<()> {
+    let mut not_deleted_tenants = Vec::with_capacity(batch.len());
+
+    for &tenant_id in batch {
+        let fetch_response =
+            list_objects_with_retries(s3_client, &s3_target.tenant_root(tenant_id), None).await?;
+
+        if fetch_response.is_truncated()
+            || fetch_response.contents().is_some()
+            || fetch_response.common_prefixes().is_some()
+        {
+            error!(
+                "Tenant {tenant_id} should be deleted, but its list response is {fetch_response:?}"
+            );
+            not_deleted_tenants.push(tenant_id);
+        }
+    }
+
+    anyhow::ensure!(
+        not_deleted_tenants.is_empty(),
+        "Failed to delete all tenants in a batch. Tenants {not_deleted_tenants:?} should be deleted."
+    );
+    Ok(())
+}
+
+async fn ensure_timeline_batch_deleted(
+    s3_client: &Client,
+    s3_target: &RootTarget,
+    batch: &[TenantTimelineId],
+) -> anyhow::Result<()> {
+    let mut not_deleted_timelines = Vec::with_capacity(batch.len());
+
+    for &id in batch {
+        let fetch_response =
+            list_objects_with_retries(s3_client, &s3_target.timeline_root(id), None).await?;
+
+        if fetch_response.is_truncated()
+            || fetch_response.contents().is_some()
+            || fetch_response.common_prefixes().is_some()
+        {
+            error!("Timeline {id} should be deleted, but its list response is {fetch_response:?}");
+            not_deleted_timelines.push(id);
+        }
+    }
+
+    anyhow::ensure!(
+        not_deleted_timelines.is_empty(),
+        "Failed to delete all timelines in a batch"
+    );
+    Ok(())
+}

test_runner/fixtures/pageserver/http.py

@@ -613,8 +613,3 @@ class PageserverHttpClient(requests.Session):
             },
         )
         self.verbose_error(res)
-
-    def deletion_queue_flush(self, execute: bool = False):
-        self.put(
-            f"http://localhost:{self.port}/v1/deletion_queue/flush?execute={'true' if execute else 'false'}"
-        ).raise_for_status()

test_runner/regress/test_remote_storage.py

@@ -12,10 +12,7 @@ from typing import Dict, List, Optional, Tuple
 import pytest
 from fixtures.log_helper import log
 from fixtures.neon_fixtures import (
-    NeonEnv,
     NeonEnvBuilder,
-    PgBin,
-    last_flush_lsn_upload,
     wait_for_last_flush_lsn,
 )
 from fixtures.pageserver.http import PageserverApiException, PageserverHttpClient
@@ -253,20 +250,35 @@ def test_remote_storage_upload_queue_retries(
 
     client = env.pageserver.http_client()
 
-    def configure_storage_write_failpoints(action):
+    endpoint = env.endpoints.create_start("main", tenant_id=tenant_id)
+
+    endpoint.safe_psql("CREATE TABLE foo (id INTEGER PRIMARY KEY, val text)")
+
+    def configure_storage_sync_failpoints(action):
         client.configure_failpoints(
             [
                 ("before-upload-layer", action),
                 ("before-upload-index", action),
+                ("before-delete-layer", action),
             ]
         )
 
-    def configure_storage_delete_failpoints(action):
-        client.configure_failpoints(
+    def overwrite_data_and_wait_for_it_to_arrive_at_pageserver(data):
+        # create initial set of layers & upload them with failpoints configured
+        endpoint.safe_psql_many(
             [
-                ("deletion-queue-before-execute", action),
+                f"""
+               INSERT INTO foo (id, val)
+               SELECT g, '{data}'
+               FROM generate_series(1, 20000) g
+               ON CONFLICT (id) DO UPDATE
+               SET val = EXCLUDED.val
+               """,
+                # to ensure that GC can actually remove some layers
+                "VACUUM foo",
             ]
         )
+        wait_for_last_flush_lsn(env, endpoint, tenant_id, timeline_id)
 
     def get_queued_count(file_kind, op_kind):
         val = client.get_remote_timeline_client_metric(
@@ -279,52 +291,23 @@ def test_remote_storage_upload_queue_retries(
         assert val is not None, "expecting metric to be present"
         return int(val)
 
-    def get_deletions_executed() -> int:
-        executed = client.get_metric_value("pageserver_deletion_queue_executed_total")
-        if executed is None:
-            return 0
-        else:
-            return int(executed)
+    # create some layers & wait for uploads to finish
+    overwrite_data_and_wait_for_it_to_arrive_at_pageserver("a")
+    client.timeline_checkpoint(tenant_id, timeline_id)
+    client.timeline_compact(tenant_id, timeline_id)
+    overwrite_data_and_wait_for_it_to_arrive_at_pageserver("b")
+    client.timeline_checkpoint(tenant_id, timeline_id)
+    client.timeline_compact(tenant_id, timeline_id)
+    gc_result = client.timeline_gc(tenant_id, timeline_id, 0)
+    print_gc_result(gc_result)
+    assert gc_result["layers_removed"] > 0
 
-    def get_deletion_errors(op_type) -> int:
-        executed = client.get_metric_value(
-            "pageserver_deletion_queue_errors_total", {"op_kind": op_type}
-        )
-        if executed is None:
-            return 0
-        else:
-            return int(executed)
-
-    def assert_queued_count(file_kind: str, op_kind: str, fn):
-        v = get_queued_count(file_kind=file_kind, op_kind=op_kind)
-        log.info(f"queue count: {file_kind} {op_kind} {v}")
-        assert fn(v)
-
-    # Push some uploads into the remote_timeline_client queues, before failpoints
-    # are enabled: these should execute and the queue should revert to zero depth
-    generate_uploads_and_deletions(env, tenant_id=tenant_id, timeline_id=timeline_id)
-
-    wait_until(2, 1, lambda: assert_queued_count("layer", "upload", lambda v: v == 0))
-    wait_until(2, 1, lambda: assert_queued_count("index", "upload", lambda v: v == 0))
-
-    # Wait for some deletions to happen in the above compactions, assert that
-    # our metrics of interest exist
-    wait_until(2, 1, lambda: assert_deletion_queue(client, lambda v: v is not None))
-
-    # Before enabling failpoints, flushing deletions through should work
-    client.deletion_queue_flush(execute=True)
-    executed = client.get_metric_value("pageserver_deletion_queue_executed_total")
-    assert executed is not None
-    assert executed > 0
+    wait_until(2, 1, lambda: get_queued_count(file_kind="layer", op_kind="upload") == 0)
+    wait_until(2, 1, lambda: get_queued_count(file_kind="index", op_kind="upload") == 0)
+    wait_until(2, 1, lambda: get_queued_count(file_kind="layer", op_kind="delete") == 0)
 
     # let all future operations queue up
-    configure_storage_write_failpoints("return")
-    configure_storage_delete_failpoints("return")
-
-    # Snapshot of executed deletions: should not increment while failpoint is enabled
-    deletions_executed_pre_failpoint = client.get_metric_value(
-        "pageserver_deletion_queue_executed_total"
-    )
+    configure_storage_sync_failpoints("return")
 
     # Create more churn to generate all upload ops.
     # The checkpoint / compact / gc ops will block because they call remote_client.wait_completion().
@@ -332,77 +315,38 @@ def test_remote_storage_upload_queue_retries(
     churn_thread_result = [False]
 
     def churn_while_failpoints_active(result):
-        generate_uploads_and_deletions(
-            env, init=False, tenant_id=tenant_id, timeline_id=timeline_id, data="d"
-        )
+        overwrite_data_and_wait_for_it_to_arrive_at_pageserver("c")
+        client.timeline_checkpoint(tenant_id, timeline_id)
+        client.timeline_compact(tenant_id, timeline_id)
+        overwrite_data_and_wait_for_it_to_arrive_at_pageserver("d")
+        client.timeline_checkpoint(tenant_id, timeline_id)
+        client.timeline_compact(tenant_id, timeline_id)
+        gc_result = client.timeline_gc(tenant_id, timeline_id, 0)
+        print_gc_result(gc_result)
+        assert gc_result["layers_removed"] > 0
         result[0] = True
 
     churn_while_failpoints_active_thread = threading.Thread(
         target=churn_while_failpoints_active, args=[churn_thread_result]
     )
-    log.info("Entered churn phase")
     churn_while_failpoints_active_thread.start()
 
-    try:
-        # wait for churn thread's data to get stuck in the upload queue
-        wait_until(10, 0.1, lambda: assert_queued_count("layer", "upload", lambda v: v > 0))
-        wait_until(10, 0.1, lambda: assert_queued_count("index", "upload", lambda v: v >= 2))
+    # wait for churn thread's data to get stuck in the upload queue
+    wait_until(10, 0.1, lambda: get_queued_count(file_kind="layer", op_kind="upload") > 0)
+    wait_until(10, 0.1, lambda: get_queued_count(file_kind="index", op_kind="upload") >= 2)
+    wait_until(10, 0.1, lambda: get_queued_count(file_kind="layer", op_kind="delete") > 0)
 
-        # Deletion queue should not grow, because deletions wait for upload of
-        # metadata, and we blocked that upload.
-        wait_until(10, 0.5, lambda: assert_deletion_queue(client, lambda v: v == 0))
+    # unblock churn operations
+    configure_storage_sync_failpoints("off")
 
-        # No more deletions should have executed
-        assert get_deletions_executed() == deletions_executed_pre_failpoint
-
-        # unblock write operations
-        log.info("Unblocking remote writes")
-        configure_storage_write_failpoints("off")
-
-        # ... and wait for them to finish. Exponential back-off in upload queue, so, gracious timeouts.
-        wait_until(30, 1, lambda: assert_queued_count("layer", "upload", lambda v: v == 0))
-        wait_until(30, 1, lambda: assert_queued_count("index", "upload", lambda v: v == 0))
-
-        # Deletions should have been enqueued now that index uploads proceeded
-        log.info("Waiting to see deletions enqueued")
-        wait_until(10, 1, lambda: assert_deletion_queue(client, lambda v: v > 0))
-
-        # Run flush in the backgrorund because it will block on the failpoint
-        class background_flush(threading.Thread):
-            def run(self):
-                client.deletion_queue_flush(execute=True)
-
-        flusher = background_flush()
-        flusher.start()
-
-        def assert_failpoint_hit():
-            assert get_deletion_errors("failpoint") > 0
-
-        # Our background flush thread should induce us to hit the failpoint
-        wait_until(20, 0.25, assert_failpoint_hit)
-
-        # Deletions should not have been executed while failpoint is still active.
-        assert get_deletion_queue_depth(client) is not None
-        assert get_deletion_queue_depth(client) > 0
-        assert get_deletions_executed() == deletions_executed_pre_failpoint
-
-        log.info("Unblocking remote deletes")
-        configure_storage_delete_failpoints("off")
-
-        # An API flush should now complete
-        flusher.join()
-
-        # Queue should drain, which should involve executing some deletions
-        wait_until(2, 1, lambda: assert_deletion_queue(client, lambda v: v == 0))
-        assert get_deletions_executed() > deletions_executed_pre_failpoint
-
-    finally:
-        # The churn thread doesn't make progress once it blocks on the first wait_completion() call,
-        # so, give it some time to wrap up.
-        log.info("Joining churn workload")
-        churn_while_failpoints_active_thread.join(30)
-        log.info("Joined churn workload")
+    # ... and wait for them to finish. Exponential back-off in upload queue, so, gracious timeouts.
+    wait_until(30, 1, lambda: get_queued_count(file_kind="layer", op_kind="upload") == 0)
+    wait_until(30, 1, lambda: get_queued_count(file_kind="index", op_kind="upload") == 0)
+    wait_until(30, 1, lambda: get_queued_count(file_kind="layer", op_kind="delete") == 0)
 
+    # The churn thread doesn't make progress once it blocks on the first wait_completion() call,
+    # so, give it some time to wrap up.
+    churn_while_failpoints_active_thread.join(30)
     assert not churn_while_failpoints_active_thread.is_alive()
     assert churn_thread_result[0]
 
@@ -488,6 +432,7 @@ def test_remote_timeline_client_calls_started_metric(
     calls_started: Dict[Tuple[str, str], List[int]] = {
         ("layer", "upload"): [0],
         ("index", "upload"): [0],
+        ("layer", "delete"): [0],
     }
 
     def fetch_calls_started():
@@ -659,6 +604,7 @@ def test_timeline_deletion_with_files_stuck_in_upload_queue(
     checkpoint_allowed_to_fail.set()
     env.pageserver.allowed_errors.append(
         ".* ERROR .*Error processing HTTP request: InternalServerError\\(timeline is Stopping"
+        ".* ERROR .*[Cc]ould not flush frozen layer.*"
     )
 
     # Generous timeout, because currently deletions can get blocked waiting for compaction
@@ -985,154 +931,4 @@ def assert_nothing_to_upload(
     assert Lsn(detail["last_record_lsn"]) == Lsn(detail["remote_consistent_lsn"])
 
 
-def get_deletion_queue_depth(ps_http) -> int:
-    """
-    Queue depth if at least one deletion has been submitted, else None
-    """
-    submitted = ps_http.get_metric_value("pageserver_deletion_queue_submitted_total")
-
-    if submitted is None:
-        return 0
-
-    executed = ps_http.get_metric_value("pageserver_deletion_queue_executed_total")
-    executed = 0 if executed is None else executed
-
-    depth = submitted - executed
-    assert depth >= 0
-
-    log.info(f"get_deletion_queue_depth: {depth} ({submitted} - {executed})")
-    return int(depth)
-
-
-def assert_deletion_queue(ps_http, size_fn) -> None:
-    v = get_deletion_queue_depth(ps_http)
-    assert v is not None
-    assert size_fn(v) is True
-
-
 # TODO Test that we correctly handle GC of files that are stuck in upload queue.
-
-
-def generate_uploads_and_deletions(
-    env: NeonEnv,
-    *,
-    init: bool = True,
-    tenant_id: Optional[TenantId] = None,
-    timeline_id: Optional[TimelineId] = None,
-    data: Optional[str] = None,
-):
-    """
-    Using the environment's default tenant + timeline, generate a load pattern
-    that results in some uploads and some deletions to remote storage.
-    """
-
-    if tenant_id is None:
-        tenant_id = env.initial_tenant
-    assert tenant_id is not None
-
-    if timeline_id is None:
-        timeline_id = env.initial_timeline
-    assert timeline_id is not None
-
-    ps_http = env.pageserver.http_client()
-
-    with env.endpoints.create_start("main", tenant_id=tenant_id) as endpoint:
-        if init:
-            endpoint.safe_psql("CREATE TABLE foo (id INTEGER PRIMARY KEY, val text)")
-            last_flush_lsn_upload(env, endpoint, tenant_id, timeline_id)
-
-        def churn(data):
-            endpoint.safe_psql_many(
-                [
-                    f"""
-                INSERT INTO foo (id, val)
-                SELECT g, '{data}'
-                FROM generate_series(1, 20000) g
-                ON CONFLICT (id) DO UPDATE
-                SET val = EXCLUDED.val
-                """,
-                    # to ensure that GC can actually remove some layers
-                    "VACUUM foo",
-                ]
-            )
-            assert tenant_id is not None
-            assert timeline_id is not None
-            wait_for_last_flush_lsn(env, endpoint, tenant_id, timeline_id)
-            ps_http.timeline_checkpoint(tenant_id, timeline_id)
-
-        # Compaction should generate some GC-elegible layers
-        for i in range(0, 2):
-            churn(f"{i if data is None else data}")
-
-        gc_result = ps_http.timeline_gc(tenant_id, timeline_id, 0)
-        print_gc_result(gc_result)
-        assert gc_result["layers_removed"] > 0
-
-
-@pytest.mark.parametrize("remote_storage_kind", [RemoteStorageKind.LOCAL_FS])
-def test_deletion_queue_recovery(
-    neon_env_builder: NeonEnvBuilder,
-    remote_storage_kind: RemoteStorageKind,
-    pg_bin: PgBin,
-):
-    neon_env_builder.enable_remote_storage(
-        remote_storage_kind=remote_storage_kind,
-        test_name="test_deletion_queue_recovery",
-    )
-
-    env = neon_env_builder.init_start(
-        initial_tenant_conf={
-            # small checkpointing and compaction targets to ensure we generate many upload operations
-            "checkpoint_distance": f"{128 * 1024}",
-            "compaction_threshold": "1",
-            "compaction_target_size": f"{128 * 1024}",
-            # no PITR horizon, we specify the horizon when we request on-demand GC
-            "pitr_interval": "0s",
-            # disable background compaction and GC. We invoke it manually when we want it to happen.
-            "gc_period": "0s",
-            "compaction_period": "0s",
-            # create image layers eagerly, so that GC can remove some layers
-            "image_creation_threshold": "1",
-        }
-    )
-
-    ps_http = env.pageserver.http_client()
-
-    # Prevent deletion lists from being executed, to build up some backlog of deletions
-    ps_http.configure_failpoints(
-        [
-            ("deletion-queue-before-execute", "return"),
-        ]
-    )
-
-    generate_uploads_and_deletions(env)
-
-    # There should be entries in the deletion queue
-    assert_deletion_queue(ps_http, lambda n: n > 0)
-    ps_http.deletion_queue_flush()
-    before_restart_depth = get_deletion_queue_depth(ps_http)
-
-    log.info(f"Restarting pageserver with {before_restart_depth} deletions enqueued")
-    env.pageserver.stop(immediate=True)
-    env.pageserver.start()
-
-    def assert_deletions_submitted(n: int):
-        assert ps_http.get_metric_value("pageserver_deletion_queue_submitted_total") == n
-
-    # After restart, issue a flush to kick the deletion frorntend to do recovery.
-    # It should recover all the operations we submitted before the restart.
-    ps_http.deletion_queue_flush(execute=False)
-    wait_until(20, 0.25, lambda: assert_deletions_submitted(before_restart_depth))
-
-    # The queue should drain through completely if we flush it
-    ps_http.deletion_queue_flush(execute=True)
-    wait_until(10, 1, lambda: assert_deletion_queue(ps_http, lambda n: n == 0))
-
-    # Restart again
-    env.pageserver.stop(immediate=True)
-    env.pageserver.start()
-
-    # No deletion lists should be recovered: this demonstrates that deletion lists
-    # were cleaned up after being executed.
-    time.sleep(1)
-    assert_deletion_queue(ps_http, lambda n: n == 0)

test_runner/regress/test_tenant_delete.py

@@ -47,15 +47,6 @@ def test_tenant_delete_smoke(
     )
 
     env = neon_env_builder.init_start()
-    env.pageserver.allowed_errors.extend(
-        [
-            # The deletion queue will complain when it encounters simulated S3 errors
-            ".*deletion frontend: Failed to write deletion list.*",
-            ".*deletion backend: Failed to delete deletion list.*",
-            ".*deletion executor: DeleteObjects request failed.*",
-            ".*deletion backend: Failed to upload deletion queue header.*",
-        ]
-    )
 
     # lucky race with stopping from flushing a layer we fail to schedule any uploads
     env.pageserver.allowed_errors.append(
@@ -100,9 +91,7 @@ def test_tenant_delete_smoke(
 
     iterations = poll_for_remote_storage_iterations(remote_storage_kind)
 
-    # We are running with failures enabled, so this may take some time to make
-    # it through all the remote storage operations required to complete
-    tenant_delete_wait_completed(ps_http, tenant_id, iterations * 10)
+    tenant_delete_wait_completed(ps_http, tenant_id, iterations)
 
     tenant_path = env.tenant_dir(tenant_id=tenant_id)
     assert not tenant_path.exists()
@@ -212,17 +201,6 @@ def test_delete_tenant_exercise_crash_safety_failpoints(
         ]
     )
 
-    if simulate_failures:
-        env.pageserver.allowed_errors.extend(
-            [
-                # The deletion queue will complain when it encounters simulated S3 errors
-                ".*deletion frontend: Failed to write deletion list.*",
-                ".*deletion backend: Failed to delete deletion list.*",
-                ".*deletion executor: DeleteObjects request failed.*",
-                ".*deletion backend: Failed to upload deletion queue header.*",
-            ]
-        )
-
     ps_http = env.pageserver.http_client()
 
     timeline_id = env.neon_cli.create_timeline("delete", tenant_id=tenant_id)

test_runner/regress/test_timeline_delete.py

@@ -488,14 +488,7 @@ def test_timeline_delete_fail_before_local_delete(neon_env_builder: NeonEnvBuild
     # Wait for tenant to finish loading.
     wait_until_tenant_active(ps_http, tenant_id=env.initial_tenant, iterations=10, period=1)
 
-    # Timeline deletion takes some finite time after startup
-    wait_timeline_detail_404(
-        ps_http,
-        tenant_id=env.initial_tenant,
-        timeline_id=leaf_timeline_id,
-        iterations=20,
-        interval=0.5,
-    )
+    wait_timeline_detail_404(ps_http, env.initial_tenant, leaf_timeline_id, iterations=4)
 
     assert (
         not leaf_timeline_path.exists()
@@ -541,7 +534,7 @@ def test_timeline_delete_fail_before_local_delete(neon_env_builder: NeonEnvBuild
     wait_until(
         2,
         0.5,
-        lambda: assert_prefix_empty(neon_env_builder, prefix="/tenants"),
+        lambda: assert_prefix_empty(neon_env_builder),
     )
 
 
@@ -695,7 +688,7 @@ def test_delete_timeline_client_hangup(neon_env_builder: NeonEnvBuilder):
     wait_until(50, 0.1, first_request_finished)
 
     # check that the timeline is gone
-    wait_timeline_detail_404(ps_http, env.initial_tenant, child_timeline_id, iterations=4)
+    wait_timeline_detail_404(ps_http, env.initial_tenant, child_timeline_id, iterations=2)
 
 
 @pytest.mark.parametrize(
@@ -779,11 +772,7 @@ def test_timeline_delete_works_for_remote_smoke(
 
     # for some reason the check above doesnt immediately take effect for the below.
     # Assume it is mock server inconsistency and check twice.
-    wait_until(
-        2,
-        0.5,
-        lambda: assert_prefix_empty(neon_env_builder, "/tenants"),
-    )
+    wait_until(2, 0.5, lambda: assert_prefix_empty(neon_env_builder))
 
 
 def test_delete_orphaned_objects(
@@ -838,8 +827,6 @@ def test_delete_orphaned_objects(
     reason = timeline_info["state"]["Broken"]["reason"]
     assert reason.endswith(f"failpoint: {failpoint}"), reason
 
-    ps_http.deletion_queue_flush(execute=True)
-
     for orphan in orphans:
         assert not orphan.exists()
         assert env.pageserver.log_contains(

workspace_hack/Cargo.toml

@@ -14,13 +14,17 @@ publish = false
 ### BEGIN HAKARI SECTION
 [dependencies]
 anyhow = { version = "1", features = ["backtrace"] }
+aws-config = { version = "0.56", default-features = false, features = ["credentials-sso", "rustls"] }
+aws-runtime = { version = "0.56", default-features = false, features = ["event-stream"] }
+aws-sigv4 = { version = "0.56", features = ["sign-eventstream"] }
+aws-smithy-http = { version = "0.56", default-features = false, features = ["event-stream", "rt-tokio"] }
 axum = { version = "0.6", features = ["ws"] }
+base64 = { version = "0.21", features = ["alloc"] }
 bytes = { version = "1", features = ["serde"] }
 chrono = { version = "0.4", default-features = false, features = ["clock", "serde"] }
 clap = { version = "4", features = ["derive", "string"] }
 clap_builder = { version = "4", default-features = false, features = ["color", "help", "std", "string", "suggestions", "usage"] }
 crossbeam-utils = { version = "0.8" }
-digest = { version = "0.10", features = ["mac", "std"] }
 either = { version = "1" }
 fail = { version = "0.5", default-features = false, features = ["failpoints"] }
 futures = { version = "0.3" }
@@ -29,6 +33,7 @@ futures-core = { version = "0.3" }
 futures-executor = { version = "0.3" }
 futures-sink = { version = "0.3" }
 futures-util = { version = "0.3", features = ["channel", "io", "sink"] }
+hex = { version = "0.4", features = ["serde"] }
 hyper = { version = "0.14", features = ["full"] }
 itertools = { version = "0.10" }
 libc = { version = "0.2", features = ["extra_traits"] }
@@ -44,14 +49,15 @@ regex = { version = "1" }
 regex-syntax = { version = "0.7" }
 reqwest = { version = "0.11", default-features = false, features = ["blocking", "json", "multipart", "rustls-tls"] }
 ring = { version = "0.16", features = ["std"] }
-rustls = { version = "0.20", features = ["dangerous_configuration"] }
+rustls = { version = "0.21", features = ["dangerous_configuration"] }
 scopeguard = { version = "1" }
 serde = { version = "1", features = ["alloc", "derive"] }
 serde_json = { version = "1", features = ["raw_value"] }
 smallvec = { version = "1", default-features = false, features = ["write"] }
 socket2 = { version = "0.4", default-features = false, features = ["all"] }
+time = { version = "0.3", features = ["formatting", "macros", "parsing"] }
 tokio = { version = "1", features = ["fs", "io-std", "io-util", "macros", "net", "process", "rt-multi-thread", "signal", "test-util"] }
-tokio-rustls = { version = "0.23" }
+tokio-rustls = { version = "0.24" }
 tokio-util = { version = "0.7", features = ["codec", "io"] }
 toml_datetime = { version = "0.6", default-features = false, features = ["serde"] }
 toml_edit = { version = "0.19", features = ["serde"] }
@@ -59,6 +65,7 @@ tower = { version = "0.4", features = ["balance", "buffer", "limit", "retry", "t
 tracing = { version = "0.1", features = ["log"] }
 tracing-core = { version = "0.1" }
 url = { version = "2", features = ["serde"] }
+uuid = { version = "1", features = ["serde", "v4"] }
 
 [build-dependencies]
 anyhow = { version = "1", features = ["backtrace"] }
@@ -74,7 +81,7 @@ prost = { version = "0.11" }
 regex = { version = "1" }
 regex-syntax = { version = "0.7" }
 serde = { version = "1", features = ["alloc", "derive"] }
-syn-dff4ba8e3ae991db = { package = "syn", version = "1", features = ["extra-traits", "full", "visit", "visit-mut"] }
-syn-f595c2ba2a3f28df = { package = "syn", version = "2", features = ["extra-traits", "full", "visit-mut"] }
+syn-dff4ba8e3ae991db = { package = "syn", version = "1", features = ["extra-traits", "full", "visit"] }
+syn-f595c2ba2a3f28df = { package = "syn", version = "2", features = ["extra-traits", "full", "visit", "visit-mut"] }
 
 ### END HAKARI SECTION