Implement validation of generations before delete

This reverts commit 8c2ff87f1a.

.github/actions/run-python-test-set/action.yml

@@ -145,7 +145,11 @@ runs:
 
         if [ "${RERUN_FLAKY}" == "true" ]; then
           mkdir -p $TEST_OUTPUT
-          poetry run ./scripts/flaky_tests.py "${TEST_RESULT_CONNSTR}" --days 10 --output "$TEST_OUTPUT/flaky.json"
+          poetry run ./scripts/flaky_tests.py "${TEST_RESULT_CONNSTR}" \
+                                              --days 7 \
+                                              --output "$TEST_OUTPUT/flaky.json" \
+                                              --pg-version "${DEFAULT_PG_VERSION}" \
+                                              --build-type "${BUILD_TYPE}"
 
           EXTRA_PARAMS="--flaky-tests-json $TEST_OUTPUT/flaky.json $EXTRA_PARAMS"
         fi

.github/workflows/build_and_test.yml

@@ -752,7 +752,7 @@ jobs:
       run:
         shell: sh -eu {0}
     env:
-      VM_BUILDER_VERSION: v0.16.3
+      VM_BUILDER_VERSION: v0.17.5
 
     steps:
       - name: Checkout
@@ -775,6 +775,7 @@ jobs:
         run: |
           ./vm-builder \
             -enable-file-cache \
+            -cgroup-uid=postgres \
             -src=369495373322.dkr.ecr.eu-central-1.amazonaws.com/compute-node-${{ matrix.version }}:${{needs.tag.outputs.build-tag}} \
             -dst=369495373322.dkr.ecr.eu-central-1.amazonaws.com/vm-compute-node-${{ matrix.version }}:${{needs.tag.outputs.build-tag}}
 
@@ -903,7 +904,7 @@ jobs:
     container:
       image: 369495373322.dkr.ecr.eu-central-1.amazonaws.com/base:pinned
       options: --init
-    needs: [ promote-images, tag ]
+    needs: [ tag ]
     steps:
       - name: Set PR's status to pending and request a remote CI test
         run: |

CODEOWNERS

@@ -1,11 +1,12 @@
-/compute_tools/ @neondatabase/control-plane
+/compute_tools/ @neondatabase/control-plane @neondatabase/compute
 /control_plane/ @neondatabase/compute @neondatabase/storage
 /libs/pageserver_api/ @neondatabase/compute @neondatabase/storage
-/libs/postgres_ffi/ @neondatabase/compute 
-/libs/remote_storage/ @neondatabase/storage 
-/libs/safekeeper_api/ @neondatabase/safekeepers  
-/pageserver/ @neondatabase/compute @neondatabase/storage 
+/libs/postgres_ffi/ @neondatabase/compute
+/libs/remote_storage/ @neondatabase/storage
+/libs/safekeeper_api/ @neondatabase/safekeepers
+/libs/vm_monitor/ @neondatabase/autoscaling @neondatabase/compute
+/pageserver/ @neondatabase/compute @neondatabase/storage
 /pgxn/ @neondatabase/compute
-/proxy/ @neondatabase/control-plane 
+/proxy/ @neondatabase/proxy
 /safekeeper/ @neondatabase/safekeepers
 /vendor/ @neondatabase/compute

Cargo.lock

@@ -1001,6 +1001,7 @@ dependencies = [
  "comfy-table",
  "compute_api",
  "git-version",
+ "hyper",
  "nix 0.26.2",
  "once_cell",
  "pageserver_api",
@@ -1016,6 +1017,7 @@ dependencies = [
  "storage_broker",
  "tar",
  "thiserror",
+ "tokio",
  "toml",
  "tracing",
  "url",
@@ -2684,6 +2686,7 @@ dependencies = [
  "bytes",
  "const_format",
  "enum-map",
+ "hex",
  "postgres_ffi",
  "serde",
  "serde_json",
@@ -5014,6 +5017,7 @@ dependencies = [
  "nix 0.26.2",
  "once_cell",
  "pin-project-lite",
+ "postgres_connection",
  "pq_proto",
  "rand",
  "regex",

Dockerfile.compute-node

@@ -211,8 +211,8 @@ RUN wget https://github.com/df7cb/postgresql-unit/archive/refs/tags/7.7.tar.gz -
 FROM build-deps AS vector-pg-build
 COPY --from=pg-build /usr/local/pgsql/ /usr/local/pgsql/
 
-RUN wget https://github.com/pgvector/pgvector/archive/refs/tags/v0.4.4.tar.gz -O pgvector.tar.gz && \
-    echo "1cb70a63f8928e396474796c22a20be9f7285a8a013009deb8152445b61b72e6 pgvector.tar.gz" | sha256sum --check && \
+RUN wget https://github.com/pgvector/pgvector/archive/refs/tags/v0.5.0.tar.gz -O pgvector.tar.gz && \
+    echo "d8aa3504b215467ca528525a6de12c3f85f9891b091ce0e5864dd8a9b757f77b pgvector.tar.gz" | sha256sum --check && \
     mkdir pgvector-src && cd pgvector-src && tar xvzf ../pgvector.tar.gz --strip-components=1 -C . && \
     make -j $(getconf _NPROCESSORS_ONLN) PG_CONFIG=/usr/local/pgsql/bin/pg_config && \
     make -j $(getconf _NPROCESSORS_ONLN) install PG_CONFIG=/usr/local/pgsql/bin/pg_config && \

compute_tools/README.md

@@ -19,9 +19,10 @@ Also `compute_ctl` spawns two separate service threads:
 - `http-endpoint` runs a Hyper HTTP API server, which serves readiness and the
   last activity requests.
 
-If the `vm-informant` binary is present at `/bin/vm-informant`, it will also be started. For VM
-compute nodes, `vm-informant` communicates with the VM autoscaling system. It coordinates
-downscaling and (eventually) will request immediate upscaling under resource pressure.
+If `AUTOSCALING` environment variable is set, `compute_ctl` will start the
+`vm-monitor` located in [`neon/libs/vm_monitor`]. For VM compute nodes,
+`vm-monitor` communicates with the VM autoscaling system. It coordinates
+downscaling and requests immediate upscaling under resource pressure.
 
 Usage example:
 ```sh

compute_tools/src/bin/compute_ctl.rs

@@ -20,9 +20,10 @@
 //! - `http-endpoint` runs a Hyper HTTP API server, which serves readiness and the
 //!   last activity requests.
 //!
-//! If the `vm-informant` binary is present at `/bin/vm-informant`, it will also be started. For VM
-//! compute nodes, `vm-informant` communicates with the VM autoscaling system. It coordinates
-//! downscaling and (eventually) will request immediate upscaling under resource pressure.
+//! If `AUTOSCALING` environment variable is set, `compute_ctl` will start the
+//! `vm-monitor` located in [`neon/libs/vm_monitor`]. For VM compute nodes,
+//! `vm-monitor` communicates with the VM autoscaling system. It coordinates
+//! downscaling and requests immediate upscaling under resource pressure.
 //!
 //! Usage example:
 //! ```sh
@@ -278,8 +279,9 @@ fn main() -> Result<()> {
             use tokio_util::sync::CancellationToken;
             use tracing::warn;
             let vm_monitor_addr = matches.get_one::<String>("vm-monitor-addr");
-            let cgroup = matches.get_one::<String>("filecache-connstr");
-            let file_cache_connstr = matches.get_one::<String>("cgroup");
+            let file_cache_connstr = matches.get_one::<String>("filecache-connstr");
+            let cgroup = matches.get_one::<String>("cgroup");
+            let file_cache_on_disk = matches.get_flag("file-cache-on-disk");
 
             // Only make a runtime if we need to.
             // Note: it seems like you can make a runtime in an inner scope and
@@ -312,6 +314,7 @@ fn main() -> Result<()> {
                         cgroup: cgroup.cloned(),
                         pgconnstr: file_cache_connstr.cloned(),
                         addr: vm_monitor_addr.cloned().unwrap(),
+                        file_cache_on_disk,
                     })),
                     token.clone(),
                 ))
@@ -482,6 +485,11 @@ fn cli() -> clap::Command {
                 )
                 .value_name("FILECACHE_CONNSTR"),
         )
+        .arg(
+            Arg::new("file-cache-on-disk")
+                .long("file-cache-on-disk")
+                .action(clap::ArgAction::SetTrue),
+        )
 }
 
 #[test]

compute_tools/src/compute.rs

@@ -1,4 +1,5 @@
 use std::collections::HashMap;
+use std::env;
 use std::fs;
 use std::io::BufRead;
 use std::os::unix::fs::PermissionsExt;
@@ -175,6 +176,27 @@ impl TryFrom<ComputeSpec> for ParsedSpec {
     }
 }
 
+/// If we are a VM, returns a [`Command`] that will run in the `neon-postgres`
+/// cgroup. Otherwise returns the default `Command::new(cmd)`
+///
+/// This function should be used to start postgres, as it will start it in the
+/// neon-postgres cgroup if we are a VM. This allows autoscaling to control
+/// postgres' resource usage. The cgroup will exist in VMs because vm-builder
+/// creates it during the sysinit phase of its inittab.
+fn maybe_cgexec(cmd: &str) -> Command {
+    // The cplane sets this env var for autoscaling computes.
+    // use `var_os` so we don't have to worry about the variable being valid
+    // unicode. Should never be an concern . . . but just in case
+    if env::var_os("AUTOSCALING").is_some() {
+        let mut command = Command::new("cgexec");
+        command.args(["-g", "memory:neon-postgres"]);
+        command.arg(cmd);
+        command
+    } else {
+        Command::new(cmd)
+    }
+}
+
 /// Create special neon_superuser role, that's a slightly nerfed version of a real superuser
 /// that we give to customers
 fn create_neon_superuser(spec: &ComputeSpec, client: &mut Client) -> Result<()> {
@@ -451,7 +473,7 @@ impl ComputeNode {
     pub fn sync_safekeepers(&self, storage_auth_token: Option<String>) -> Result<Lsn> {
         let start_time = Utc::now();
 
-        let sync_handle = Command::new(&self.pgbin)
+        let sync_handle = maybe_cgexec(&self.pgbin)
             .args(["--sync-safekeepers"])
             .env("PGDATA", &self.pgdata) // we cannot use -D in this mode
             .envs(if let Some(storage_auth_token) = &storage_auth_token {
@@ -586,7 +608,7 @@ impl ComputeNode {
 
         // Start postgres
         info!("starting postgres");
-        let mut pg = Command::new(&self.pgbin)
+        let mut pg = maybe_cgexec(&self.pgbin)
             .args(["-D", pgdata])
             .spawn()
             .expect("cannot start postgres process");
@@ -614,7 +636,7 @@ impl ComputeNode {
         let pgdata_path = Path::new(&self.pgdata);
 
         // Run postgres as a child process.
-        let mut pg = Command::new(&self.pgbin)
+        let mut pg = maybe_cgexec(&self.pgbin)
             .args(["-D", &self.pgdata])
             .envs(if let Some(storage_auth_token) = &storage_auth_token {
                 vec![("NEON_AUTH_TOKEN", storage_auth_token)]

control_plane/Cargo.toml

@@ -12,6 +12,7 @@ git-version.workspace = true
 nix.workspace = true
 once_cell.workspace = true
 postgres.workspace = true
+hyper.workspace = true
 regex.workspace = true
 reqwest = { workspace = true, features = ["blocking", "json"] }
 serde.workspace = true
@@ -20,6 +21,7 @@ serde_with.workspace = true
 tar.workspace = true
 thiserror.workspace = true
 toml.workspace = true
+tokio.workspace = true
 url.workspace = true
 # Note: Do not directly depend on pageserver or safekeeper; use pageserver_api or safekeeper_api
 # instead, so that recompile times are better.

control_plane/src/attachment_service.rs

@@ -0,0 +1,104 @@
+use crate::{background_process, local_env::LocalEnv};
+use anyhow::anyhow;
+use pageserver_api::control_api::HexTenantId;
+use serde::{Deserialize, Serialize};
+use std::{path::PathBuf, process::Child};
+use utils::id::{NodeId, TenantId};
+
+pub struct AttachmentService {
+    env: LocalEnv,
+    listen: String,
+    path: PathBuf,
+}
+
+const COMMAND: &str = "attachment_service";
+
+#[derive(Serialize, Deserialize)]
+pub struct AttachHookRequest {
+    pub tenant_id: HexTenantId,
+    pub pageserver_id: Option<NodeId>,
+}
+
+#[derive(Serialize, Deserialize)]
+pub struct AttachHookResponse {
+    pub gen: Option<u32>,
+}
+
+impl AttachmentService {
+    pub fn from_env(env: &LocalEnv) -> Self {
+        let path = env.base_data_dir.join("attachments.json");
+
+        // Makes no sense to construct this if pageservers aren't going to use it: assume
+        // pageservers have control plane API set
+        let listen_url = env.pageserver.control_plane_api.clone().unwrap();
+
+        let listen = format!(
+            "{}:{}",
+            listen_url.host_str().unwrap(),
+            listen_url.port().unwrap()
+        );
+
+        Self {
+            env: env.clone(),
+            path,
+            listen,
+        }
+    }
+
+    fn pid_file(&self) -> PathBuf {
+        self.env.base_data_dir.join("attachment_service.pid")
+    }
+
+    pub fn start(&self) -> anyhow::Result<Child> {
+        let path_str = self.path.to_string_lossy();
+
+        background_process::start_process(
+            COMMAND,
+            &self.env.base_data_dir,
+            &self.env.attachment_service_bin(),
+            ["-l", &self.listen, "-p", &path_str],
+            [],
+            background_process::InitialPidFile::Create(&self.pid_file()),
+            // TODO: a real status check
+            || Ok(true),
+        )
+    }
+
+    pub fn stop(&self, immediate: bool) -> anyhow::Result<()> {
+        background_process::stop_process(immediate, COMMAND, &self.pid_file())
+    }
+
+    /// Call into the attach_hook API, for use before handing out attachments to pageservers
+    pub fn attach_hook(
+        &self,
+        tenant_id: TenantId,
+        pageserver_id: NodeId,
+    ) -> anyhow::Result<Option<u32>> {
+        use hyper::StatusCode;
+
+        let url = self
+            .env
+            .pageserver
+            .control_plane_api
+            .clone()
+            .unwrap()
+            .join("attach_hook")
+            .unwrap();
+        let client = reqwest::blocking::ClientBuilder::new()
+            .build()
+            .expect("Failed to construct http client");
+
+        let request = AttachHookRequest {
+            tenant_id: HexTenantId::new(tenant_id),
+            pageserver_id: Some(pageserver_id),
+        };
+
+        let response = client.post(url).json(&request).send()?;
+        if response.status() != StatusCode::OK {
+            return Err(anyhow!("Unexpected status {0}", response.status()));
+        }
+
+        let response = response.json::<AttachHookResponse>()?;
+        Ok(response.gen)
+    }
+}

control_plane/src/bin/attachment_service.rs

@@ -0,0 +1,264 @@
+/// The attachment service mimics the aspects of the control plane API
+/// that are required for a pageserver to operate.
+///
+/// This enables running & testing pageservers without a full-blown
+/// deployment of the Neon cloud platform.
+///
+use anyhow::anyhow;
+use clap::Parser;
+use hyper::StatusCode;
+use hyper::{Body, Request, Response};
+use pageserver_api::control_api::*;
+use serde::{Deserialize, Serialize};
+use std::path::{Path, PathBuf};
+use std::{collections::HashMap, sync::Arc};
+use utils::logging::{self, LogFormat};
+
+use utils::{
+    http::{
+        endpoint::{self},
+        error::ApiError,
+        json::{json_request, json_response},
+        RequestExt, RouterBuilder,
+    },
+    id::{NodeId, TenantId},
+    tcp_listener,
+};
+
+use control_plane::attachment_service::{AttachHookRequest, AttachHookResponse};
+
+#[derive(Parser)]
+#[command(author, version, about, long_about = None)]
+#[command(arg_required_else_help(true))]
+struct Cli {
+    #[arg(short, long)]
+    listen: String,
+
+    #[arg(short, long)]
+    path: PathBuf,
+}
+
+// The persistent state of each Tenant
+#[derive(Serialize, Deserialize, Clone)]
+struct TenantState {
+    // Currently attached pageserver
+    pageserver: Option<NodeId>,
+
+    // Latest generation number: next time we attach, increment this
+    // and use the incremented number when attaching
+    generation: u32,
+}
+
+fn to_hex_map<S, V>(input: &HashMap<TenantId, V>, serializer: S) -> Result<S::Ok, S::Error>
+where
+    S: serde::Serializer,
+    V: Clone + Serialize,
+{
+    eprintln!("to_hex_map");
+    let transformed = input
+        .iter()
+        .map(|(k, v)| (HexTenantId::new(k.clone()), v.clone()));
+
+    transformed
+        .collect::<HashMap<HexTenantId, V>>()
+        .serialize(serializer)
+}
+
+fn from_hex_map<'de, D, V>(deserializer: D) -> Result<HashMap<TenantId, V>, D::Error>
+where
+    D: serde::de::Deserializer<'de>,
+    V: Deserialize<'de>,
+{
+    eprintln!("from_hex_map");
+    let hex_map = HashMap::<HexTenantId, V>::deserialize(deserializer)?;
+
+    Ok(hex_map.into_iter().map(|(k, v)| (k.take(), v)).collect())
+}
+
+// Top level state available to all HTTP handlers
+#[derive(Serialize, Deserialize)]
+struct PersistentState {
+    #[serde(serialize_with = "to_hex_map", deserialize_with = "from_hex_map")]
+    tenants: HashMap<TenantId, TenantState>,
+
+    #[serde(skip)]
+    path: PathBuf,
+}
+
+impl PersistentState {
+    async fn save(&self) -> anyhow::Result<()> {
+        let bytes = serde_json::to_vec(self)?;
+        tokio::fs::write(&self.path, &bytes).await?;
+
+        Ok(())
+    }
+
+    async fn load(path: &Path) -> anyhow::Result<Self> {
+        let bytes = tokio::fs::read(path).await?;
+        let mut decoded = serde_json::from_slice::<Self>(&bytes)?;
+        decoded.path = path.to_owned();
+        Ok(decoded)
+    }
+
+    async fn load_or_new(path: &Path) -> Self {
+        match Self::load(path).await {
+            Ok(s) => s,
+            Err(e) => {
+                tracing::info!(
+                    "Creating new state file at {0} (load returned {e})",
+                    path.to_string_lossy()
+                );
+                Self {
+                    tenants: HashMap::new(),
+                    path: path.to_owned(),
+                }
+            }
+        }
+    }
+}
+
+/// State available to HTTP request handlers
+#[derive(Clone)]
+struct State {
+    inner: Arc<tokio::sync::RwLock<PersistentState>>,
+}
+
+impl State {
+    fn new(persistent_state: PersistentState) -> State {
+        Self {
+            inner: Arc::new(tokio::sync::RwLock::new(persistent_state)),
+        }
+    }
+}
+
+#[inline(always)]
+fn get_state(request: &Request<Body>) -> &State {
+    request
+        .data::<Arc<State>>()
+        .expect("unknown state type")
+        .as_ref()
+}
+
+/// Pageserver calls into this on startup, to learn which tenants it should attach
+async fn handle_re_attach(mut req: Request<Body>) -> Result<Response<Body>, ApiError> {
+    let reattach_req = json_request::<ReAttachRequest>(&mut req).await?;
+
+    let state = get_state(&req).inner.clone();
+    let mut locked = state.write().await;
+
+    let mut response = ReAttachResponse {
+        tenants: Vec::new(),
+    };
+    for (t, state) in &mut locked.tenants {
+        if state.pageserver == Some(reattach_req.node_id) {
+            state.generation += 1;
+            response.tenants.push(ReAttachResponseTenant {
+                id: HexTenantId::new(t.clone()),
+                generation: state.generation,
+            });
+        }
+    }
+
+    locked
+        .save()
+        .await
+        .map_err(|e| ApiError::InternalServerError(e))?;
+
+    json_response(StatusCode::OK, response)
+}
+
+/// Pageserver calls into this before doing deletions, to confirm that it still
+/// holds the latest generation for the tenants with deletions enqueued
+async fn handle_validate(mut req: Request<Body>) -> Result<Response<Body>, ApiError> {
+    let validate_req = json_request::<ValidateRequest>(&mut req).await?;
+
+    let state = get_state(&req).inner.clone();
+    let locked = state.read().await;
+
+    let mut response = ValidateResponse {
+        tenants: Vec::new(),
+    };
+
+    for req_tenant in validate_req.tenants {
+        if let Some(tenant_state) = locked.tenants.get(req_tenant.id.as_ref()) {
+            let valid = tenant_state.generation == req_tenant.gen;
+            response.tenants.push(ValidateResponseTenant {
+                id: req_tenant.id,
+                valid,
+            });
+        }
+    }
+
+    json_response(StatusCode::OK, response)
+}
+/// Call into this before attaching a tenant to a pageserver, to acquire a generation number
+/// (in the real control plane this is unnecessary, because the same program is managing
+///  generation numbers and doing attachments).
+async fn handle_attach_hook(mut req: Request<Body>) -> Result<Response<Body>, ApiError> {
+    let attach_req = json_request::<AttachHookRequest>(&mut req).await?;
+
+    let state = get_state(&req).inner.clone();
+    let mut locked = state.write().await;
+
+    let tenant_state = locked
+        .tenants
+        .entry(attach_req.tenant_id.take())
+        .or_insert_with(|| TenantState {
+            pageserver: attach_req.pageserver_id,
+            generation: 0,
+        });
+
+    if attach_req.pageserver_id.is_some() {
+        tenant_state.generation += 1;
+    }
+    let generation = tenant_state.generation;
+
+    locked
+        .save()
+        .await
+        .map_err(|e| ApiError::InternalServerError(e))?;
+
+    json_response(
+        StatusCode::OK,
+        AttachHookResponse {
+            gen: attach_req.pageserver_id.map(|_| generation),
+        },
+    )
+}
+
+fn make_router(persistent_state: PersistentState) -> RouterBuilder<hyper::Body, ApiError> {
+    endpoint::make_router()
+        .data(Arc::new(State::new(persistent_state)))
+        .post("/re-attach", |r| handle_re_attach(r))
+        .post("/validate", |r| handle_validate(r))
+        .post("/attach_hook", |r| handle_attach_hook(r))
+}
+
+#[tokio::main]
+async fn main() -> anyhow::Result<()> {
+    logging::init(
+        LogFormat::Plain,
+        logging::TracingErrorLayerEnablement::Disabled,
+    )?;
+
+    let args = Cli::parse();
+    tracing::info!(
+        "Starting, state at {}, listening on {}",
+        args.path.to_string_lossy(),
+        args.listen
+    );
+
+    let persistent_state = PersistentState::load_or_new(&args.path).await;
+
+    let http_listener = tcp_listener::bind(&args.listen)?;
+    let router = make_router(persistent_state)
+        .build()
+        .map_err(|err| anyhow!(err))?;
+    let service = utils::http::RouterService::new(router).unwrap();
+    let server = hyper::Server::from_tcp(http_listener)?.serve(service);
+
+    tracing::info!("Serving on {0}", args.listen.as_str());
+    server.await?;
+
+    Ok(())
+}

control_plane/src/bin/neon_local.rs

@@ -8,6 +8,7 @@
 use anyhow::{anyhow, bail, Context, Result};
 use clap::{value_parser, Arg, ArgAction, ArgMatches, Command};
 use compute_api::spec::ComputeMode;
+use control_plane::attachment_service::AttachmentService;
 use control_plane::endpoint::ComputeControlPlane;
 use control_plane::local_env::LocalEnv;
 use control_plane::pageserver::PageServerNode;
@@ -43,6 +44,8 @@ project_git_version!(GIT_VERSION);
 
 const DEFAULT_PG_VERSION: &str = "15";
 
+const DEFAULT_PAGESERVER_CONTROL_PLANE_API: &str = "http://127.0.0.1:1234/";
+
 fn default_conf() -> String {
     format!(
         r#"
@@ -56,11 +59,13 @@ listen_pg_addr = '{DEFAULT_PAGESERVER_PG_ADDR}'
 listen_http_addr = '{DEFAULT_PAGESERVER_HTTP_ADDR}'
 pg_auth_type = '{trust_auth}'
 http_auth_type = '{trust_auth}'
+control_plane_api = '{DEFAULT_PAGESERVER_CONTROL_PLANE_API}'
 
 [[safekeepers]]
 id = {DEFAULT_SAFEKEEPER_ID}
 pg_port = {DEFAULT_SAFEKEEPER_PG_PORT}
 http_port = {DEFAULT_SAFEKEEPER_HTTP_PORT}
+
 "#,
         trust_auth = AuthType::Trust,
     )
@@ -107,6 +112,7 @@ fn main() -> Result<()> {
             "start" => handle_start_all(sub_args, &env),
             "stop" => handle_stop_all(sub_args, &env),
             "pageserver" => handle_pageserver(sub_args, &env),
+            "attachment_service" => handle_attachment_service(sub_args, &env),
             "safekeeper" => handle_safekeeper(sub_args, &env),
             "endpoint" => handle_endpoint(sub_args, &env),
             "pg" => bail!("'pg' subcommand has been renamed to 'endpoint'"),
@@ -342,13 +348,25 @@ fn handle_tenant(tenant_match: &ArgMatches, env: &mut local_env::LocalEnv) -> an
             }
         }
         Some(("create", create_match)) => {
-            let initial_tenant_id = parse_tenant_id(create_match)?;
             let tenant_conf: HashMap<_, _> = create_match
                 .get_many::<String>("config")
                 .map(|vals| vals.flat_map(|c| c.split_once(':')).collect())
                 .unwrap_or_default();
-            let new_tenant_id = pageserver.tenant_create(initial_tenant_id, tenant_conf)?;
-            println!("tenant {new_tenant_id} successfully created on the pageserver");
+
+            // If tenant ID was not specified, generate one
+            let tenant_id = parse_tenant_id(create_match)?.unwrap_or(TenantId::generate());
+
+            let generation = if env.pageserver.control_plane_api.is_some() {
+                // We must register the tenant with the attachment service, so
+                // that when the pageserver restarts, it will be re-attached.
+                let attachment_service = AttachmentService::from_env(env);
+                attachment_service.attach_hook(tenant_id, env.pageserver.id)?
+            } else {
+                None
+            };
+
+            pageserver.tenant_create(tenant_id, generation, tenant_conf)?;
+            println!("tenant {tenant_id} successfully created on the pageserver");
 
             // Create an initial timeline for the new tenant
             let new_timeline_id = parse_timeline_id(create_match)?;
@@ -358,7 +376,7 @@ fn handle_tenant(tenant_match: &ArgMatches, env: &mut local_env::LocalEnv) -> an
                 .context("Failed to parse postgres version from the argument string")?;
 
             let timeline_info = pageserver.timeline_create(
-                new_tenant_id,
+                tenant_id,
                 new_timeline_id,
                 None,
                 None,
@@ -369,17 +387,17 @@ fn handle_tenant(tenant_match: &ArgMatches, env: &mut local_env::LocalEnv) -> an
 
             env.register_branch_mapping(
                 DEFAULT_BRANCH_NAME.to_string(),
-                new_tenant_id,
+                tenant_id,
                 new_timeline_id,
             )?;
 
             println!(
-                "Created an initial timeline '{new_timeline_id}' at Lsn {last_record_lsn} for tenant: {new_tenant_id}",
+                "Created an initial timeline '{new_timeline_id}' at Lsn {last_record_lsn} for tenant: {tenant_id}",
             );
 
             if create_match.get_flag("set-default") {
-                println!("Setting tenant {new_tenant_id} as a default one");
-                env.default_tenant_id = Some(new_tenant_id);
+                println!("Setting tenant {tenant_id} as a default one");
+                env.default_tenant_id = Some(tenant_id);
             }
         }
         Some(("set-default", set_default_match)) => {
@@ -817,6 +835,33 @@ fn handle_pageserver(sub_match: &ArgMatches, env: &local_env::LocalEnv) -> Resul
     Ok(())
 }
 
+fn handle_attachment_service(sub_match: &ArgMatches, env: &local_env::LocalEnv) -> Result<()> {
+    let svc = AttachmentService::from_env(env);
+    match sub_match.subcommand() {
+        Some(("start", _start_match)) => {
+            if let Err(e) = svc.start() {
+                eprintln!("start failed: {e}");
+                exit(1);
+            }
+        }
+
+        Some(("stop", stop_match)) => {
+            let immediate = stop_match
+                .get_one::<String>("stop-mode")
+                .map(|s| s.as_str())
+                == Some("immediate");
+
+            if let Err(e) = svc.stop(immediate) {
+                eprintln!("stop failed: {}", e);
+                exit(1);
+            }
+        }
+        Some((sub_name, _)) => bail!("Unexpected attachment_service subcommand '{}'", sub_name),
+        None => bail!("no attachment_service subcommand provided"),
+    }
+    Ok(())
+}
+
 fn get_safekeeper(env: &local_env::LocalEnv, id: NodeId) -> Result<SafekeeperNode> {
     if let Some(node) = env.safekeepers.iter().find(|node| node.id == id) {
         Ok(SafekeeperNode::from_env(env, node))
@@ -897,6 +942,16 @@ fn handle_start_all(sub_match: &ArgMatches, env: &local_env::LocalEnv) -> anyhow
 
     broker::start_broker_process(env)?;
 
+    // Only start the attachment service if the pageserver is configured to need it
+    if env.pageserver.control_plane_api.is_some() {
+        let attachment_service = AttachmentService::from_env(env);
+        if let Err(e) = attachment_service.start() {
+            eprintln!("attachment_service start failed: {:#}", e);
+            try_stop_all(env, true);
+            exit(1);
+        }
+    }
+
     let pageserver = PageServerNode::from_env(env);
     if let Err(e) = pageserver.start(&pageserver_config_overrides(sub_match)) {
         eprintln!("pageserver {} start failed: {:#}", env.pageserver.id, e);
@@ -955,6 +1010,13 @@ fn try_stop_all(env: &local_env::LocalEnv, immediate: bool) {
     if let Err(e) = broker::stop_broker_process(env) {
         eprintln!("neon broker stop failed: {e:#}");
     }
+
+    if env.pageserver.control_plane_api.is_some() {
+        let attachment_service = AttachmentService::from_env(env);
+        if let Err(e) = attachment_service.stop(immediate) {
+            eprintln!("attachment service stop failed: {e:#}");
+        }
+    }
 }
 
 fn cli() -> Command {
@@ -1138,6 +1200,14 @@ fn cli() -> Command {
                             .arg(stop_mode_arg.clone()))
                 .subcommand(Command::new("restart").about("Restart local pageserver").arg(pageserver_config_args.clone()))
         )
+        .subcommand(
+            Command::new("attachment_service")
+                .arg_required_else_help(true)
+                .about("Manage attachment_service")
+                .subcommand(Command::new("start").about("Start local pageserver").arg(pageserver_config_args.clone()))
+                .subcommand(Command::new("stop").about("Stop local pageserver")
+                            .arg(stop_mode_arg.clone()))
+        )
         .subcommand(
             Command::new("safekeeper")
                 .arg_required_else_help(true)

control_plane/src/lib.rs

@@ -7,6 +7,7 @@
 // local installations.
 //
 
+pub mod attachment_service;
 mod background_process;
 pub mod broker;
 pub mod endpoint;

control_plane/src/local_env.rs

@@ -118,6 +118,9 @@ pub struct PageServerConf {
     // auth type used for the PG and HTTP ports
     pub pg_auth_type: AuthType,
     pub http_auth_type: AuthType,
+
+    // Control plane location
+    pub control_plane_api: Option<Url>,
 }
 
 impl Default for PageServerConf {
@@ -128,6 +131,7 @@ impl Default for PageServerConf {
             listen_http_addr: String::new(),
             pg_auth_type: AuthType::Trust,
             http_auth_type: AuthType::Trust,
+            control_plane_api: None,
         }
     }
 }
@@ -202,6 +206,10 @@ impl LocalEnv {
         self.neon_distrib_dir.join("pageserver")
     }
 
+    pub fn attachment_service_bin(&self) -> PathBuf {
+        self.neon_distrib_dir.join("attachment_service")
+    }
+
     pub fn safekeeper_bin(&self) -> PathBuf {
         self.neon_distrib_dir.join("safekeeper")
     }

control_plane/src/pageserver.rs

@@ -126,6 +126,13 @@ impl PageServerNode {
             broker_endpoint_param,
         ];
 
+        if let Some(control_plane_api) = &self.env.pageserver.control_plane_api {
+            overrides.push(format!(
+                "control_plane_api='{}'",
+                control_plane_api.as_str()
+            ));
+        }
+
         if self.env.pageserver.http_auth_type != AuthType::Trust
             || self.env.pageserver.pg_auth_type != AuthType::Trust
         {
@@ -316,7 +323,8 @@ impl PageServerNode {
 
     pub fn tenant_create(
         &self,
-        new_tenant_id: Option<TenantId>,
+        new_tenant_id: TenantId,
+        generation: Option<u32>,
         settings: HashMap<&str, &str>,
     ) -> anyhow::Result<TenantId> {
         let mut settings = settings.clone();
@@ -382,11 +390,9 @@ impl PageServerNode {
                 .context("Failed to parse 'gc_feedback' as bool")?,
         };
 
-        // If tenant ID was not specified, generate one
-        let new_tenant_id = new_tenant_id.unwrap_or(TenantId::generate());
-
         let request = models::TenantCreateRequest {
             new_tenant_id,
+            generation,
             config,
         };
         if !settings.is_empty() {

docs/rfcs/025-generation-numbers.md

@@ -0,0 +1,957 @@
+# Pageserver: split-brain safety for remote storage through generation numbers
+
+## Summary
+
+A scheme of logical "generation numbers" for tenant attachment to pageservers is proposed, along with
+changes to the remote storage format to include these generation numbers in S3 keys.
+
+Using the control plane as the issuer of these generation numbers enables strong anti-split-brain
+properties in the pageserver cluster without implementing a consensus mechanism directly
+in the pageservers.
+
+## Motivation
+
+Currently, the pageserver's remote storage format does not provide a mechanism for addressing
+split brain conditions that may happen when replacing a node or when migrating
+a tenant from one pageserver to another.
+
+From a remote storage perspective, a split brain condition occurs whenever two nodes both think
+they have the same tenant attached, and both can write to S3. This can happen in the case of a
+network partition, pathologically long delays (e.g. suspended VM), or software bugs.
+
+In the current deployment model, control plane guarantees that a tenant is attached to one
+pageserver at a time, thereby ruling out split-brain conditions resulting from dual
+attachment (however, there is always the risk of a control plane bug). This control
+plane guarantee prevents robust response to failures, as if a pageserver is unresponsive
+we may not detach from it. The mechanism in this RFC fixes this, by making it safe to
+attach to a new, different pageserver even if an unresponsive pageserver may be running.
+
+Futher, lack of safety during split-brain conditions blocks two important features where occasional
+split-brain conditions are part of the design assumptions:
+
+- seamless tenant migration ([RFC PR](https://github.com/neondatabase/neon/pull/5029))
+- automatic pageserver instance failure handling (aka "failover") (RFC TBD)
+
+### Prior art
+
+- 020-pageserver-s3-coordination.md
+- 023-the-state-of-pageserver-tenant-relocation.md
+- 026-pageserver-s3-mvcc.md
+
+This RFC has broad similarities to the proposal to implement a MVCC scheme in
+S3 object names, but this RFC avoids a general purpose transaction scheme in
+favour of more specialized "generations" that work like a transaction ID that
+always has the same lifetime as a pageserver process or tenant attachment, whichever
+is shorter.
+
+## Requirements
+
+- Accommodate storage backends with no atomic or fencing capability (i.e. work within
+  S3's limitation that there are no atomics and clients can't be fenced)
+- Don't depend on any STONITH or node fencing in the compute layer (i.e. we will not
+  assume that we can reliably kill and EC2 instance and have it die)
+- Scoped per-tenant, not per-pageserver; for _seamless tenant migration_, we need
+  per-tenant granularity, and for _failover_, we likely want to spread the workload
+  of the failed pageserver instance to a number of peers, rather than monolithically
+  moving the entire workload to another machine.
+  We do not rule out the latter case, but should not constrain ourselves to it.
+
+## Design Tenets
+
+These are not requirements, but are ideas that guide the following design:
+
+- Avoid implementing another consensus system: we already have a strongly consistent
+  database in the control plane that can do atomic operations where needed, and we also
+  have a Paxos implementation in the safekeeper.
+- Avoiding locking in to specific models of how failover will work (e.g. do not assume that
+  all the tenants on a pageserver will fail over as a unit).
+- Be strictly correct when it comes to data integrity. Occasional failures of availability
+  are tolerable, occasional data loss is not.
+
+## Non Goals
+
+The changes in this RFC intentionally isolate the design decision of how to define
+logical generations numbers and object storage format in a way that is somewhat flexible with
+respect to how actual orchestration of failover works.
+
+This RFC intentionally does not cover:
+
+- Failure detection
+- Orchestration of failover
+- Standby modes to keep data ready for fast migration
+- Intentional multi-writer operation on tenants (multi-writer scenarios are assumed to be transient split-brain situations).
+- Sharding.
+
+The interaction between this RFC and those features is discussed in [Appendix B](#appendix-b-interoperability-with-other-features)
+
+## Impacted Components
+
+pageserver, control plane, safekeeper (optional)
+
+## Implementation Part 1: Correctness
+
+### Summary
+
+- A per-tenant **generation number** is introduced to uniquely identifying tenant attachments to pageserver processes.
+
+  - This generation number increments each time the control plane modifies a tenant (`Project`)'s assigned pageserver, or when the assigned pageserver restarts.
+  - the control plane is the authority for generation numbers: only it may
+    increment a generation number.
+
+- **Object keys are suffixed** with the generation number
+- **Safety for multiply-attached tenants** is provided by the
+  generation number in the object key: the competing pageservers will not
+  try to write to the same keys.
+- **Safety in split brain for multiple nodes running with
+  the same node ID** is provided by the pageserver calling out to the control plane
+  on startup, to re-attach and thereby increment the generations of any attached tenants
+- **Safety for deletions** is achieved by deferring the DELETE from S3 to a point in time where the deleting node has validated with control plane that no attachment with a higher generation has a reference to the to-be-DELETEd key.
+- **The control plane is used to issue generation numbers** to avoid the need for
+  a built-in consensus system in the pageserver, although this could in principle
+  be changed without changing the storage format.
+
+### Generation numbers
+
+A generation number is associated with each tenant in the control plane,
+and each time the attachment status of the tenant changes, this is incremented.
+Changes in attachment status include:
+
+- Attaching the tenant to a different pageserver
+- A pageserver restarting, and "re-attaching" its tenants on startup
+
+These increments of attachment generation provide invariants we need to avoid
+split-brain issues in storage:
+
+- If two pageservers have the same tenant attached, the attachments are guaranteed to have different generation numbers, because the generation would increment
+  while attaching the second one.
+- If there are multiple pageservers running with the same node ID, all the attachments on all pageservers are guaranteed to have different generation numbers, because the generation would increment
+  when the second node started and re-attached its tenants.
+
+As long as the infrastructure does not transparently replace an underlying
+physical machine, we are totally safe. See the later [unsafe case](#unsafe-case-on-badly-behaved-infrastructure) section for details.
+
+### Object Key Changes
+
+#### Generation suffix
+
+All object keys (layer objects and index objects) will contain the attachment
+generation as a [suffix](#why-a-generation-suffix-rather-than-prefix).
+This suffix is the primary mechanism for protecting against split-brain situations, and
+enabling safe multi-attachment of tenants:
+
+- Two pageservers running with the same node ID (e.g. after a failure, where there is
+  some rogue pageserver still running) will not try to write to the same objects, because at startup they will have re-attached tenants and thereby incremented
+  generation numbers.
+- Multiple attachments (to different pageservers) of the same tenant will not try to write to the same objects, as each attachment would have a distinct generation.
+
+The generation is appended in hex format (8 byte string representing
+u32), to all our existing key names. A u32's range limit would permit
+27 restarts _per second_ over a 5 year system lifetime: orders of magnitude more than
+is realistic.
+
+The exact meaning of the generation suffix can evolve over time if necessary, for
+example if we chose to implement a failover mechanism internally to the pageservers
+rather than going via the control plane. The storage format just sees it as a number,
+with the only semantic property being that the highest numbered index is the latest.
+
+#### Index changes
+
+Since object keys now include a generation suffix, the index of these keys must also be updated. IndexPart currently stores keys and LSNs sufficient to reconstruct key names: this would be extended to store the generation as well.
+
+This will increase the size of the file, but only modestly: layers are already encoded as
+their string-ized form, so the overhead is about 10 bytes per layer. This will be less if/when
+the index storage format is migrated to a binary format from JSON.
+
+#### Visibility
+
+_This section doesn't describe code changes, but extends on the consequences of the
+object key changes given above_
+
+##### Visibility of objects to pageservers
+
+Pageservers can of course list objects in S3 at any time, but in practice their
+visible set is based on the contents of their LayerMap, which is initialized
+from the `index_part.json.???` that they load.
+
+Starting with the `index_part` from the most recent previous generation
+(see [loading index_part](#finding-the-remote-indices-for-timelines)), a pageserver
+initially has visibility of all the objects that were referenced in the loaded index.
+These objects are guaranteed to remain visible until the current generation is
+superseded, via pageservers in older generations avoiding deletions (see [deletion](#deletion)).
+
+The "most recent previous generation" is _not_ necessarily the most recent
+in terms of walltime, it is the one that is readable at the time a new generation
+starts. Consider the following sequence of a tenant being re-attached to different
+pageserver nodes:
+
+- Create + attach on PS1 in generation 1
+- PS1 Do some work, write out index_part.json-0001
+- Attach to PS2 in generation 2
+- Read index_part.json-0001
+- PS2 starts doing some work...
+- Attach to PS3 in generation 3
+- Read index_part.json-0001
+- **...PS2 finishes its work: now it writes index_part.json-0002**
+- PS3 writes out index_part.json-0003
+
+In the above sequence, the ancestry of indices is:
+
+```
+0001 -> 0002
+     |
+     -> 0003
+```
+
+This is not an issue for safety: if the 0002 references some object that is
+not in 0001, then 0003 simply does not see it, and will re-do whatever
+work was required (e.g. ingesting WAL or doing compaction). Objects referenced
+by only the 0002 index will never be read by future attachment generations, and
+will eventually be cleaned up by a scrub (see [scrubbing](#cleaning-up-orphan-objects-scrubbing)).
+
+##### Visibility of LSNs to clients
+
+Because index_part.json is now written with a generation suffix, which data
+is visible depends on which generation the reader is operating in:
+
+- If one was passively reading from S3 from outside of a pageserver, the
+  visibility of data would depend on which index_part.json-<generation> file
+  one had chosen to read from.
+- If two pageservers have the same tenant attached, they may have different
+  data visible as they're independently replaying the WAL, and maintaining
+  independent LayerMaps that are written to independent index_part.json files.
+  Data does not have to be remotely committed to be visible.
+- For a pageserver writing with a stale generation, historic LSNs
+  remain readable until another pageserver (with a higher generation suffix)
+  decides to execute GC deletions. At this point, we may think of the stale
+  attachment's generation as having logically ended: during its existence
+  the generation had a consistent view of the world.
+- For a newly attached pageserver, its highest visible LSN may appears to
+  go backwards with respect to an earlier attachment, if that earlier
+  attachment had not uploaded all data to S3 before the new attachment.
+
+### Deletion
+
+#### Generation number validation
+
+While writes are de-conflicted by writers always using their own generation number in the key,
+deletions are slightly more challenging: if a pageserver A is isolated, and the true active node is
+pageserver B, then it is dangerous for A to do any object deletions, even of objects that it wrote
+itself, because pageserver's B metadata might reference those objects.
+
+We solve this by inserting a "generation validation" step between the write of a remote index
+that un-links a particular object from the index, and the actual deletion of the object, such
+that deletions strictly obey the following ordering:
+
+1. Write out index_part.json: this guarantees that any subsequent reader of the metadata will
+   not try and read the object we unlinked.
+2. Call out to control plane to validate that the generation which we use for our attachment is still the latest.
+3. If step 2 passes, it is safe to delete the object. Why? The check-in with control plane
+   together with our visibility rules guarantees that any later generation
+   will use either the exact `index_part.json` that we uploaded in step 1, or a successor
+   of it; not an earlier one. In both cases, the `index_part.json` doesn't reference the
+   key we are deleting anymore, so, the key is invisible to any later attachment generation.
+   Hence it's safe to delete it.
+
+Note that at step 2 we are only confirming that deletions of objects _no longer referenced
+by the specific `index_part.json` written in step 1_ are safe. If we were attempting other deletions concurrently,
+these would need their own generation validation step.
+
+If step 2 fails, we may leak the object. This is safe, but has a cost: see [scrubbing](#cleaning-up-orphan-objects-scrubbing). We may avoid this entirely outside of node
+failures, if we do proper flushing of deletions on clean shutdown and clean migration.
+
+To avoid doing a huge number of control plane requests to perform generation validation,
+validation of many tenants will be done in a single request, and deletions will be queued up
+prior to validation: see [Persistent deletion queue](#persistent-deletion-queue) for more.
+
+#### `remote_consistent_lsn` updates
+
+Remote objects are not the only kind of deletion the pageserver does: it also indirectly deletes
+WAL data, by feeding back remote_consistent_lsn to safekeepers, as a signal to the safekeepers that
+they may drop data below this LSN.
+
+For the same reasons that deletion of objects must be guarded by an attachment generation number
+validation step, updates to `remote_consistent_lsn` are subject to the same rules, using
+an ordering as follows:
+
+1. upload the index_part that covers data up to LSN `L0` to S3
+2. Call out to control plane to validate that the generation which we use for our attachment is still the latest.
+3. advance the `remote_consistent_lsn` that we advertise to the safekeepers to `L0`
+
+If step 2 fails, then the `remote_consistent_lsn` advertised
+to safekeepers will not advance again until a pageserver
+with the latest generation is ready to do so.
+
+**Note:** at step 3 we are not advertising the _latest_ remote_consistent_lsn, we are
+advertising the value in the index_part that we uploaded in step 1. This provides
+a strong ordering guarantee.
+
+Internally to the pageserver, each timeline will have two remote_consistent_lsn values: the one that
+reflects its latest write to remote storage, and the one that reflects the most
+recent validation of generation number. It is only the latter value that may
+be advertised to the outside world (i.e. to the safekeeper).
+
+The control plane remains unaware of `remote_consistent_lsn`: it only has to validate
+the freshness of generation numbers, thereby granting the pageserver permission to
+share the information with the safekeeper.
+
+For convenience, in subsequent sections and RFCs we will use "deletion" to mean both deletion
+of objects in S3, and updates to the `remote_consistent_lsn`, as updates to the remote consistent
+LSN are de-facto deletions done via the safekeeper, and both kinds of deletion are subject to
+the same generation validation requirement.
+
+### Pageserver attach/startup changes
+
+#### Attachment
+
+Calls to `/v1/tenant/{tenant_id}/attach` are augmented with an additional
+`generation` field in the body.
+
+The pageserver does not persist this: a generation is only good for the lifetime
+of a process.
+
+#### Finding the remote indices for timelines
+
+Because index files are now suffixed with generation numbers, the pageserver
+cannot always GET the remote index in one request, because it can't always
+know a-priori what the latest remote index is.
+
+Typically, the most recent generation to write an index would be our own
+generation minus 1. However, this might not be the case: the previous
+node might have started and acquired a generation number, and then crashed
+before writing out a remote index.
+
+In the general case and as a fallback, the pageserver may list all the `index_part.json`
+files for a timeline, sort them by generation, and pick the highest that is `<=`
+its current generation for this attachment. The tenant should never load an index
+with an attachment generation _newer_ than its own.
+These two rules combined ensure that objects written by later generations are never visible to earlier generations.
+
+Note that if a given attachment picks an index part from an earlier generation (say n-2), but crashes & restarts before it writes its own generation's index part, next time it tries to pick an index part there may be an index part from generation n-1.
+It would pick the n-1 index part in that case, because it's sorted higher than the previous one from generation n-2.
+So, above rules guarantee no determinism in selecting the index part.
+are allowed to be attached with stale attachment generations during a multiply-attached
+phase in a migration, and in this instance if the old location's pageserver restarts,
+it should not try and load the newer generation's index.
+
+To summarize, on starting a timeline, the pageserver will:
+
+1. Issue a GET for index_part.json-<my generation - 1>
+2. If 1 failed, issue a ListObjectsv2 request for index_part.json\* and
+   pick the newest.
+
+One could optimize this further by using the control plane to record specifically
+which generation most recently wrote an index_part.json, if necessary, to increase
+the probability of finding the index_part.json in one GET. One could also improve
+the chances by having pageservers proactively write out index_part.json after they
+get a new generation ID.
+
+#### Re-attachment on startup
+
+On startup, the pageserver will call out to an new control plane `/re-attach`
+API (see [Generation API](#generation-api)). This returns a list of
+tenants that should be attached to the pageserver, and their generation numbers, which
+the control plane will increment before returning.
+
+The pageserver should still scan its local disk on startup, but should _delete_
+any local content for tenants not indicated in the `/re-attach` response: their
+absence is an implicit detach operation.
+
+**Note** if a tenant is omitted from the re-attach response, its local disk content
+will be deleted. This will change in subsequent work, when the control plane gains
+the concept of a secondary/standby location: a node with local content may revert
+to this status and retain some local content.
+
+#### Cleaning up previous generations' remote indices
+
+Deletion of old indices is not necessary for correctness, although it is necessary
+to avoid the ListObjects fallback in the previous section becoming ever more expensive.
+
+Once the new attachment has written out its index_part.json, it may asynchronously clean up historic index_part.json
+objects that were found.
+
+We may choose to implement this deletion either as an explicit step after we
+write out index_part for the first time in a pageserver's lifetime, or for
+simplicity just do it periodically as part of the background scrub (see [scrubbing](#cleaning-up-orphan-objects-scrubbing));
+
+### Control Plane Changes
+
+#### Store generations for attaching tenants
+
+- The `Project` table must store the generation number for use when
+  attaching the tenant to a new pageserver.
+- The `/v1/tenant/:tenant_id/attach` pageserver API will require the generation number,
+  which the control plane can supply by simply incrementing the `Project`'s
+  generation number each time the tenant is attached to a different server: the same database
+  transaction that changes the assigned pageserver should also change the generation number.
+
+#### Generation API
+
+This section describes an API that could be provided directly by the control plane,
+or built as a separate microservice. In earlier parts of the RFC, when we
+discuss the control plane providing generation numbers, we are referring to this API.
+
+The API endpoints used by the pageserver to acquire and validate generation
+numbers are quite simple, and only require access to some persistent and
+linerizable storage (such as a database).
+
+Building this into the control plane is proposed as a least-effort option to exploit existing infrastructure and implement generation number issuance in the same transaction that mandates it (i.e., the transaction that updates the `Project` assignment to another pageserver).
+However, this is not mandatory: this "Generation Number Issuer" could
+be built as a microservice. In practice, we will write such a miniature service
+anyway, to enable E2E pageserver/compute testing without control plane.
+
+The endpoints required by pageservers are:
+
+##### `/re-attach`
+
+- Request: `{node_id: <u32>}`
+- Response:
+  - 200 `{tenants: [{id: <TenantId>, gen: <u32>}]}`
+  - 404: unknown node_id
+  - (Future: 429: flapping detected, perhaps nodes are fighting for the same node ID,
+    or perhaps this node was in a retry loop)
+  - (On unknown tenants, omit tenant from `tenants` array)
+- Server behavior: query database for which tenants should be attached to this pageserver.
+  - for each tenant that should be attached, increment the attachment generation and
+    include the new generation in the response
+- Client behavior:
+  - for all tenants in the response, activate with the new generation number
+  - for any local disk content _not_ referenced in the response, act as if we
+    had been asked to detach it (i.e. delete local files)
+
+**Note** the `node_id` in this request will change in future if we move to ephemeral
+node IDs, to be replaced with some correlation ID that helps the control plane realize
+if a process is running with the same storage as a previous pageserver process (e.g.
+we might use EC instance ID, or we might just write some UUID to the disk the first
+time we use it)
+
+##### `/validate`
+
+- Request: `{'tenants': [{tenant: <tenant id>, attach_gen: <gen>}, ...]}'`
+- Response:
+  - 200 `{'tenants': [{tenant: <tenant id>, status: <bool>}...]}`
+  - (On unknown tenants, omit tenant from `tenants` array)
+- Purpose: enable the pageserver to discover for the given attachments whether they are still the latest.
+- Server behavior: this is a read-only operation: simply compare the generations in the request with
+  the generations known to the server, and set status to `true` if they match.
+- Client behavior: clients must not do deletions within a tenant's remote data until they have
+  received a response indicating the generation they hold for the attachment is current.
+
+#### Use of `/load` and `/ignore` APIs
+
+Because the pageserver will be changed to only attach tenants on startup
+based on the control plane's response to a `/re-attach` request, the load/ignore
+APIs no longer make sense in their current form.
+
+The `/load` API becomes functionally equivalent to attach, and will be removed:
+any location that used `/load` before should just attach instead.
+
+The `/ignore` API is equivalent to detaching, but without deleting local files.
+
+### Timeline/Branch creation & deletion
+
+All of the previous arguments for safety have described operations within
+a timeline, where we may describe a sequence that includes updates to
+index_part.json, and where reads and writes are coming from a postgres
+endpoint (writes via the safekeeper).
+
+Creating or destroying timeline is a bit different, because writes
+are coming from the control plane.
+
+We must be safe against scenarios such as:
+
+- A tenant is attached to pageserver B while pageserver A is
+  in the middle of servicing an RPC from the control plane to
+  create or delete a tenant.
+- A pageserver A has been sent a timeline creation request
+  but becomes unresponsive. The tenant is attached to a
+  different pageserver B, and the timeline creation request
+  is sent there too.
+
+#### Timeline Creation
+
+If some very slow node tries to do a timeline creation _after_
+a more recent generation node has already created the timeline
+and written some data into it, that must not cause harm. This
+is provided in timeline creations by the way all the objects
+within the timeline's remote path include a generation suffix:
+a slow node in an old generation that attempts to "create" a timeline
+that already exists will just emit an index_part.json with
+an old generation suffix.
+
+Timeline IDs are never reused, so we don't have
+to worry about the case of create/delete/create cycles. If they
+were re-used during a disaster recovery "un-delete" of a timeline,
+that special case can be handled by calling out to all available pageservers
+to check that they return 404 for the timeline, and to flush their
+deletion queues in case they had any deletions pending from the
+timeline.
+
+The above makes it safe for control plane to change the assignment of
+tenant to pageserver in control plane while a timeline creation is ongoing.
+The reason is that the creation request against the new assigned pageserver
+uses a new generation number. However, care must be taken by control plane
+to ensure that a "timeline creation successul" response from some pageserver
+is checked for the pageserver's generation for that timeline's tenant still being the latest.
+If it is not the latest, the response does not constitute a successful timeline creation.
+It is acceptable to discard such responses, the scrubber will clean up the S3 state.
+It is better to issue a timelien deletion request to the stale attachment.
+
+#### Timeline Deletion
+
+Tenant/timeline deletion operations are exempt from generation validation
+on deletes, and therefore don't have to go through the same deletion
+queue as GC/compaction layer deletions. This is because once a
+delete is issued by the control plane, it is a promise that the
+control plane will keep trying until the deletion is done, so even stale
+pageservers are permitted to go ahead and delete the objects.
+
+The implications of this for control plane are:
+
+- During timeline/tenant deletion, the control plane must wait for the deletion to
+  be truly complete (status 404) and also handle the case where the pageserver
+  becomes unavailable, either by waiting for a replacement with the same node_id,
+  or by *re-attaching the tenant elsewhere.
+
+- The control plane must persist its intent to delete
+  a timeline/tenant before issuing any RPCs, and then once it starts, it must
+  keep retrying until the tenant/timeline is gone. This is already handled
+  by using a persistent `Operation` record that is retried indefinitely.
+
+Timeline deletion may result in a special kind of object leak, where
+the latest generation attachment completes a deletion (including erasing
+all objects in the timeline path), but some slow/partitioned node is
+writing into the timeline path with a stale generation number. This would
+not be caught by any per-timeline scrubbing (see [scrubbing](#cleaning-up-orphan-objects-scrubbing)), since scrubbing happens on the
+attached pageserver, and once the timeline is deleted it isn't attached anywhere.
+This scenario should be pretty rare, and the control plane can make it even
+rarer by ensuring that if a tenant is in a multi-attached state (e.g. during
+migration), we wait for that to complete before processing the deletion. Beyond
+that, we may implement some other top-level scrub of timelines in
+an external tool, to identify any tenant/timeline paths that are not found
+in the control plane database.
+
+#### Examples
+
+- Deletion, node restarts partway through:
+  - By the time we returned 202, we have written a remote delete marker
+  - Any subsequent incarnation of the same node_id will see the remote
+    delete marker and continue to process the deletion
+  - If the original pageserver is lost permanently and no replacement
+    with the same node_id is available, then the control plane must recover
+    by re-attaching the tenant to a different node.
+- Creation, node becomes unresponsive partway through.
+  - Control plane will see HTTP request timeout, keep re-issuing
+    request to whoever is the latest attachment point for the tenant
+    until it succeeds.
+  - Stale nodes may be trying to execute timeline creation: they will
+    write out index_part.json files with
+    stale attachment generation: these will be eventually cleaned up
+    by the same mechanism as other old indices.
+
+### Unsafe case on badly behaved infrastructure
+
+This section is only relevant if running on a different environment
+than EC2 machines with ephemeral disks.
+
+If we ever run pageservers on infrastructure that might transparently restart
+a pageserver while leaving an old process running (e.g. a VM gets rescheduled
+without the old one being fenced), then there is a risk of corruption, when
+the control plane attaches the tenant, as follows:
+
+- If the control plane sends an `/attach` request to node A, then node A dies
+  and is replaced, and the control plane's retries the request without
+  incrementing that attachment ID, then it could end up with two physical nodes
+  both using the same generation number.
+- This is not an issue when using EC2 instances with ephemeral storage, as long
+  as the control plane never re-uses a node ID, but it would need re-examining
+  if running on different infrastructure.
+- To robustly protect against this class of issue, we would either:
+  - add a "node generation" to distinguish between different processes holding the
+    same node_id.
+  - or, dispense with static node_id entirely and issue an ephemeral ID to each
+    pageserver process when it starts.
+
+## Implementation Part 2: Optimizations
+
+### Persistent deletion queue
+
+Between writing our a new index_part.json that doesn't reference an object,
+and executing the deletion, an object passes through a window where it is
+only referenced in memory, and could be leaked if the pageserver is stopped
+uncleanly. That introduces conflicting incentives: on the one hand, we would
+like to delay and batch deletions to
+1. minimize the cost of the mandatory validations calls to control plane, and
+2. minimize cost for DeleteObjects requests.
+On the other hand we would also like to minimize leakage by executing
+deletions promptly.
+
+To resolve this, we may make the deletion queue persistent
+and then executing these in the background at a later time.
+
+_Note: The deletion queue's reason for existence is optimization rather than correctness,
+so there is a lot of flexibility in exactly how the it should work,
+as long as it obeys the rule to validate generations before executing deletions,
+so the following details are not essential to the overall RFC._
+
+#### Scope
+
+The deletion queue will be global per pageserver, not per-tenant. There
+are several reasons for this choice:
+
+- Use the queue as a central point to coalesce validation requests to the
+  control plane: this avoids individual `Timeline` objects ever touching
+  the control plane API, and avoids them having to know the rules about
+  validating deletions. This separation of concerns will avoid burdening
+  the already many-LoC `Timeline` type with even more responsibility.
+- Decouple the deletion queue from Tenant attachment lifetime: we may
+  "hibernate" an inactive tenant by tearing down its `Tenant`/`Timeline`
+  objects in the pageserver, without having to wait for deletions to be done.
+- Amortize the cost of I/O for the persistent queue, instead of having many
+  tiny queues.
+- Coalesce deletions into a smaller number of larger DeleteObjects calls
+
+Because of the cost of doing I/O for persistence, and the desire to coalesce
+generation validation requests across tenants, and coalesce deletions into
+larger DeleteObjects requests, there will be one deletion queue per pageserver
+rather than one per tenant. This has the added benefit that when deactivating
+a tenant, we do not have to drain their deletion queue: deletions can proceed
+for a tenant whose main `Tenant` object has been torn down.
+
+#### Flow of deletion
+
+The flow of a deletion is becomes:
+
+1. Need for deletion of an object (=> layer file) is identified.
+2. Unlink the object from all the places that reference it (=> `index_part.json`).
+3. Enqueue the deletion to a persistent queue.
+   Each entry is `tenant_id, attachment_generation, S3 key`.
+4. Validate & execute in batches:
+  4.1 For a batch of entries, call into control plane.
+  4.2 For the subset of entries that passed validation, execute a `DeleteObjects` S3 DELETE request for their S3 keys.
+
+As outlined in the Part 1 on correctness, it is critical that deletions are only
+executed once the key is not referenced anywhere in S3.
+This property is obviously upheld by the scheme above.
+
+#### We Accept Object Leakage In Acceptable Circumcstances
+
+If we crash in the flow above between (2) and (3), we lose track of unreferenced object.
+Further, enqueuing a single to the persistent queue may not be durable immediately to amortize cost of flush to disk.
+This is acceptable for now, it can be caught by [the scrubber](#cleaning-up-orphan-objects-scrubbing).
+
+There are various measures we can take to improve this in the future.
+1. Cap amount of time until enqueued entry becomes durable (timeout for flush-to-tisk)
+2. Proactively flush:
+    - On graceful shutdown, as we anticipate that some or
+      all of our attachments may be re-assigned while we are offline.
+    - On tenant detach.
+3. For each entry, keep track of whether it has passed (2).
+   Only admit entries to (4) one they have passed (2).
+   This requires re-writing / two queue entries (intent, commit) per deletion.
+
+The important take-away with any of the above is that it's not
+disastrous to leak objects in exceptional circumstances.
+
+#### Operations that may skip the queue
+
+Deletions of an entire timeline are [exempt](#Timeline-Deletion) from generation number validation. Once the
+control plane sends the deletion request, there is no requirement to retain the readability
+of any data within the timeline, and all objects within the timeline path may be deleted
+at any time from the control plane's deletion request onwards.
+
+Since deletions of smaller timelines won't have enough objects to compose a full sized
+DeleteObjects request, it is still useful to send these through the last part of the
+deletion pipeline to coalesce with other executing deletions: to enable this, the
+deletion queue should expose two input channels: one for deletions that must be
+processed in a generation-aware way, and a fast path for timeline deletions, where
+that fast path may skip validation and the persistent queue.
+
+### Cleaning up orphan objects (scrubbing)
+
+An orphan object is any object which is no longer referenced by a running node or by metadata.
+
+Examples of how orphan objects arise:
+
+- A node PUTs a layer object, then crashes before it writes the
+  index_part.json that references that layer.
+- A stale node carries on running for some time, and writes out an unbounded number of
+  objects while it believes itself to be the rightful writer for a tenant.
+- A pageserver crashes between un-linking an object from the index, and persisting
+  the object to its deletion queue.
+
+Orphan objects are functionally harmless, but have a small cost due to S3 capacity consumed. We
+may clean them up at some time in the future, but doing a ListObjectsv2 operation and cross
+referencing with the latest metadata to identify objects which are not referenced.
+
+Scrubbing will be done only by an attached pageserver (not some third party process), and deletions requested during scrub will go through the same
+validation as all other deletions: the attachment generation must be
+fresh. This avoids the possibility of a stale pageserver incorrectly
+thinking than an object written by a newer generation is stale, and deleting
+it.
+
+It is not strictly necessary that scrubbing be done by an attached
+pageserver: it could also be done externally. However, an external
+scrubber would still require the same validation procedure that
+a pageserver's deletion queue performs, before actually erasing
+objects.
+
+## Operational impact
+
+### Availability
+
+Coordination of generation numbers via the control plane introduce a dependency for certain
+operations:
+
+1. Starting new pageservers (or activating pageservers after a restart)
+2. Executing enqueued deletions
+3. Advertising updated `remote_consistent_lsn` to enable WAL trimming
+
+Item 1. would mean that some in-place restarts that previously would have resumed service even if the control plane were
+unavailable, will now not resume service to users until the control plane is available. We could
+avoid this by having a timeout on communication with the control plane, and after some timeout,
+resume service with the previous generation numbers (assuming this was persisted to disk). However,
+this is unlikely to be needed as the control plane is already an essential & highly available component. Also, having a node re-use an old generation number would complicate
+reasoning about the system, as it would break the invariant that a generation number uniquely identifies
+a tenant's attachment to a given pageserver _process_: it would merely identify the tenant's attachment
+to the pageserver _machine_ or its _on-disk-state_.
+
+Item 2. is a non-issue operationally: it's harmless to delay deletions, the only impact of objects pending deletion is
+the S3 capacity cost.
+
+Item 3. could be an issue if safekeepers are low on disk space and the control plane is unavailable for a long time. If this became an issue,
+we could adjust the safekeeper to delete segments from local disk sooner, as soon as they're uploaded to S3, rather than waiting for
+remote_consistent_lsn to advance.
+
+For a managed service, the general approach should be to make sure we are monitoring & respond fast enough
+that control plane outages are bounded in time.
+
+There is also the fact that control plane runs in a single region.
+The latency for distant regions is not a big concern for us because all request types added by this RFC are either infrequent or not in the way of the data path.
+However, we lose region isolation for the operations listed above.
+The ongoing work to split console and control will give us per-region control plane, and all operations in this RFC can be handled by these per-region control planes.
+With that in mind, we accept the trade-offs outlined in this paragraph.
+
+We will also implement an "escape hatch" config generation numbers, where in a major disaster outage,
+we may manually run pageservers with a hand-selected generation number, so that we can bring them online
+independently of a control plane.
+
+### Rollout
+
+Although there is coupling between components, we may deploy most of the new data plane components
+independently of the control plane: initially they can just use a static generation number.
+
+#### Phase 1
+
+The pageserver is deployed with some special config to:
+
+- Always act like everything is generation 1 and do not wait for a control plane issued generation on attach
+- Skip the places in deletion and remote_consistent_lsn updates where we would call into control plane
+
+#### Phase 2
+
+The control plane changes are deployed: control plane will now track and increment generation numbers.
+
+#### Phase 3
+
+The pageserver is deployed with its control-plane-dependent changes enabled: it will now require
+the control plane to service re-attach requests on startup, and handle generation
+validation requests.
+
+### On-disk backward compatibility
+
+Backward compatibility with existing data is straightforward:
+
+- When reading the index, we may assume that any layer whose metadata doesn't include
+  generations will have a path without generation suffix.
+- When locating the index file on attachment, we may use the "fallback" listing path
+  and if there is only an index without generation suffix, that is the one we load.
+
+It is not necessary to re-write existing layers: even new index files will be able
+to represent generation-less layers.
+
+### On-disk forward compatibility
+
+We will do a two phase rollout, probably over multiple releases because we will naturally
+have some of the read-side code ready before the overall functionality is ready:
+
+1. Deploy pageservers which understand the new index format and generation suffixes
+   in keys, but do not write objects with generation numbers in the keys.
+2. Deploy pageservers that write objects with generation numbers in the keys.
+
+Old pageservers will be oblivious to generation numbers. That means that they can't
+read objects with generation numbers in the name. This is why we must
+first step must deploy the ability to read, before the second step
+starts writing them.
+
+# Frequently Asked Questions
+
+## Why a generation _suffix_ rather than _prefix_?
+
+The choice is motivated by object listing, since one can list by prefix but not
+suffix.
+
+In [finding remote indices](#finding-the-remote-indices-for-timelines), we rely
+on being able to do a prefix listing for `<tenant>/<timeline>/index_part.json*`.
+That relies on the prefix listing.
+
+The converse case of using a generation prefix and listing by generation is
+not needed: one could imagine listing by generation while scrubbing (so that
+a particular generation's layers could be scrubbed), but this is not part
+of normal operations, and the [scrubber](#cleaning-up-orphan-objects-scrubbing) probably won't work that way anyway.
+
+## Wouldn't it be simpler to have a separate deletion queue per timeline?
+
+Functionally speaking, we could. That's how RemoteTimelineClient currently works,
+but this approach does not map well to a long-lived persistent queue with
+generation validation.
+
+Anything we do per-timeline generates tiny random I/O, on a pageserver with
+tens of thousands of timelines operating: to be ready for high scale, we should:
+
+- A) Amortize costs where we can (e.g. a shared deletion queue)
+- B) Expect to put tenants into a quiescent state while they're not
+  busy: i.e. we shouldn't keep a tenant alive to service its deletion queue.
+
+This was discussed in the [scope](#scope) part of the deletion queue section.
+
+# Appendix A: Examples of use in high availability/failover
+
+The generation numbers proposed in this RFC are adaptable to a variety of different
+failover scenarios and models. The sections below sketch how they would work in practice.
+
+### In-place restart of a pageserver
+
+"In-place" here means that the restart is done before any other element in the system
+has taken action in response to the node being down.
+
+- After restart, the node issues a re-attach request to the control plane, and
+  receives new generation numbers for all its attached tenants.
+- Tenants may be activated with the generation number in the re-attach response.
+- If any of its attachments were in fact stale (i.e. had be reassigned to another
+  node while this node was offline), then
+  - the re-attach response will inform the tenant about this by not including
+    the tenant of this by _not_ incrementing the generation for that attachment.
+  - This will implicitly block deletions in the tenant, but as an optimization
+    the pageserver should also proactively stop doing S3 uploads when it notices this stale-generation state.
+  - The control plane is expected to eventually detach this tenant from the
+    pageserver.
+
+If the control plane does not include a tenant in the re-attach response,
+but there is still local state for the tenant in the filesystem, the pageserver
+deletes the local state in response and does not load/active the tenant.
+See the [earlier section on pageserver startup](#pageserver-attachstartup-changes) for details.
+Control plane can use this mechanism to clean up a pageserver that has been
+down for so long that all its tenants were migrated away before it came back
+up again and asked for re-attach.
+
+### Failure of a pageserver
+
+In this context, read "failure" as the most ambiguous possible case, where
+a pageserver is unavailable to clients and control plane, but may still be executing and talking
+to S3.
+
+#### Case A: re-attachment to other nodes
+
+1. Let's say node 0 becomes unresponsive in a cluster of three nodes 0, 1, 2.
+2. Some external mechanism notices that the node is unavailable and initiates
+   movement of all tenants attached to that node to a different node according
+   to some distribution rule.
+   In this example, it would mean incrementing the generation
+   of all tenants that were attached to node 0, as each tenant's assigned pageserver changes.
+3. A tenant which is now attached to node 1 will _also_ still be attached to node
+   0, from the perspective of node 0. Node 0 will still be using its old generation,
+   node 1 will be using a newer generation.
+4. S3 writes will continue from nodes 0 and 1: there will be an index_part.json-00000001
+   \_and\* an index_part.json-00000002. Objects written under the old suffix
+   after the new attachment was created do not matter from the rest of the system's
+   perspective: the endpoints are reading from the new attachment location. Objects
+   written by node 0 are just garbage that can be cleaned up at leisure. Node 0 will
+   not do any deletions because it can't synchronize with control plane, or if it could,
+   its deletion queue processing would get errors for the validation requests.
+
+#### Case B: direct node replacement with same node_id and drive
+
+This is the scenario we would experience if running pageservers in some dynamic
+VM/container environment that would auto-replace a given node_id when it became
+unresponsive, with the node's storage supplied by some network block device
+that is attached to the replacement VM/container.
+
+1. Let's say node 0 fails, and there may be some other peers but they aren't relevant.
+2. Some external mechanism notices that the node is unavailable, and creates
+   a "new node 0" (Node 0b) which is a physically separate server. The original node 0
+   (Node 0a) may still be running, because we do not assume the environment fences nodes.
+3. On startup, node 0b re-attaches and gets higher generation numbers for
+   all tenants.
+4. S3 writes continue from nodes 0a and 0b, but the writes do not collide due to different
+   generation in the suffix, and the writes from node 0a are not visible to the rest
+   of the system because endpoints are reading only from node 0b.
+
+# Appendix B: interoperability with other features
+
+## Sharded Keyspace
+
+The design in this RFC maps neatly to a sharded keyspace design where subsets of the key space
+for a tenant are assigned to different pageservers:
+
+- the "unit of work" for attachments becomes something like a TenantShard rather than a Tenant
+- TenantShards get generation numbers just as Tenants do.
+- Write workload (ingest, compaction) for a tenant is spread out across pageservers via
+  TenantShards, but each TenantShard still has exactly one valid writer at a time.
+
+## Read replicas
+
+_This section is about a passive reader of S3 pageserver state, not a postgres
+read replica_
+
+For historical reads to LSNs below the remote persistent LSN, any node may act as a reader at any
+time: remote data is logically immutable data, and the use of deferred deletion in this RFC helps
+mitigate the fact that remote data is not _physically_ immutable (i.e. the actual data for a given
+page moves around as compaction happens).
+
+A read replica needs to be aware of generations in remote data in order to read the latest
+metadata (find the index_part.json with the latest suffix). It may either query this
+from the control plane, or find it with ListObjectsv2 request
+
+## Seamless migration
+
+To make tenant migration totally seamless, we will probably want to intentionally double-attach
+a tenant briefly, serving reads from the old node while waiting for the new node to be ready.
+
+This RFC enables that double-attachment: two nodes may be attached at the same time, with the migration destination
+having a higher generation number. The old node will be able to ingest and serve reads, but not
+do any deletes. The new node's attachment must also avoid deleting layers that the old node may
+still use. A new piece of state
+will be needed for this in the control plane's definition of an attachment.
+
+## Warm secondary locations
+
+To enable faster tenant movement after a pageserver is lost, we will probably want to spend some
+disk capacity on keeping standby locations populated with local disk data.
+
+There's no conflict between this RFC and that: implementing warm secondary locations on a per-tenant basis
+would be a separate change to the control plane to store standby location(s) for a tenant. Because
+the standbys do not write to S3, they do not need to be assigned generation numbers. When a tenant is
+re-attached to a standby location, that would increment the tenant attachment generation and this
+would work the same as any other attachment change, but with a warm cache.
+
+## Ephemeral node IDs
+
+This RFC intentionally avoids changing anything fundamental about how pageservers are identified
+and registered with the control plane, to avoid coupling the implementation of pageserver split
+brain protection with more fundamental changes in the management of the pageservers.
+
+Moving to ephemeral node IDs would provide an extra layer of
+resilience in the system, as it would prevent the control plane
+accidentally attaching to two physical nodes with the same
+generation, if somehow there were two physical nodes with
+the same node IDs (currently we rely on EC2 guarantees to
+eliminate this scenario). With ephemeral node IDs, there would be
+no possibility of that happening, no matter the behavior of
+underlying infrastructure.
+
+Nothing fundamental in the pageserver's handling of generations needs to change to handle ephemeral node IDs, since we hardly use the
+`node_id` anywhere. The `/re-attach` API would be extended
+to enable the pageserver to obtain its ephemeral ID, and provide
+some correlation identifier (e.g. EC instance ID), to help the
+control plane re-attach tenants to the same physical server that
+previously had them attached.

libs/pageserver_api/Cargo.toml

@@ -12,6 +12,7 @@ const_format.workspace = true
 anyhow.workspace = true
 bytes.workspace = true
 byteorder.workspace = true
+hex.workspace = true
 utils.workspace = true
 postgres_ffi.workspace = true
 enum-map.workspace = true

libs/pageserver_api/src/control_api.rs

@@ -0,0 +1,89 @@
+/// Types in this file are for pageserver's upward-facing API calls to the control plane
+use hex::FromHex;
+use serde::{Deserialize, Serialize};
+use utils::id::{NodeId, TenantId};
+
+/// TenantId's serialization is an array of u8, which is rather unfriendly
+/// for outside callers who aren't working with the native Rust TenantId.
+/// This class wraps it in serialization that is just the hex strict
+/// representation.
+#[derive(Eq, PartialEq, Clone, Hash)]
+pub struct HexTenantId(TenantId);
+
+impl HexTenantId {
+    pub fn new(t: TenantId) -> Self {
+        Self(t)
+    }
+
+    pub fn take(self) -> TenantId {
+        self.0
+    }
+}
+
+impl AsRef<TenantId> for HexTenantId {
+    fn as_ref(&self) -> &TenantId {
+        &self.0
+    }
+}
+
+impl Serialize for HexTenantId {
+    fn serialize<S>(&self, serializer: S) -> Result<S::Ok, S::Error>
+    where
+        S: serde::Serializer,
+    {
+        let hex = self.0.hex_encode();
+        serializer.collect_str(&hex)
+    }
+}
+
+impl<'de> Deserialize<'de> for HexTenantId {
+    fn deserialize<D>(deserializer: D) -> Result<Self, D::Error>
+    where
+        D: serde::Deserializer<'de>,
+    {
+        let string = String::deserialize(deserializer)?;
+        TenantId::from_hex(string)
+            .map(|t| HexTenantId::new(t))
+            .map_err(|e| serde::de::Error::custom(format!("{e}")))
+    }
+}
+
+// Top level s
+
+#[derive(Serialize, Deserialize)]
+pub struct ReAttachRequest {
+    pub node_id: NodeId,
+}
+
+#[derive(Serialize, Deserialize)]
+pub struct ReAttachResponseTenant {
+    pub id: HexTenantId,
+    pub generation: u32,
+}
+
+#[derive(Serialize, Deserialize)]
+pub struct ReAttachResponse {
+    pub tenants: Vec<ReAttachResponseTenant>,
+}
+
+#[derive(Serialize, Deserialize)]
+pub struct ValidateRequestTenant {
+    pub id: HexTenantId,
+    pub gen: u32,
+}
+
+#[derive(Serialize, Deserialize)]
+pub struct ValidateRequest {
+    pub tenants: Vec<ValidateRequestTenant>,
+}
+
+#[derive(Serialize, Deserialize)]
+pub struct ValidateResponse {
+    pub tenants: Vec<ValidateResponseTenant>,
+}
+
+#[derive(Serialize, Deserialize)]
+pub struct ValidateResponseTenant {
+    pub id: HexTenantId,
+    pub valid: bool,
+}

libs/pageserver_api/src/lib.rs

@@ -1,6 +1,7 @@
 use const_format::formatcp;
 
 /// Public API types
+pub mod control_api;
 pub mod models;
 pub mod reltag;
 

libs/pageserver_api/src/models.rs

@@ -194,6 +194,9 @@ pub struct TimelineCreateRequest {
 pub struct TenantCreateRequest {
     #[serde_as(as = "DisplayFromStr")]
     pub new_tenant_id: TenantId,
+    #[serde(default)]
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub generation: Option<u32>,
     #[serde(flatten)]
     pub config: TenantConfig, // as we have a flattened field, we should reject all unknown fields in it
 }
@@ -241,15 +244,6 @@ pub struct StatusResponse {
     pub id: NodeId,
 }
 
-impl TenantCreateRequest {
-    pub fn new(new_tenant_id: TenantId) -> TenantCreateRequest {
-        TenantCreateRequest {
-            new_tenant_id,
-            config: TenantConfig::default(),
-        }
-    }
-}
-
 #[serde_as]
 #[derive(Serialize, Deserialize, Debug)]
 #[serde(deny_unknown_fields)]
@@ -293,9 +287,11 @@ impl TenantConfigRequest {
     }
 }
 
-#[derive(Debug, Serialize, Deserialize)]
+#[derive(Debug, Deserialize)]
 pub struct TenantAttachRequest {
     pub config: TenantAttachConfig,
+    #[serde(default)]
+    pub generation: Option<u32>,
 }
 
 /// Newtype to enforce deny_unknown_fields on TenantConfig for

libs/remote_storage/src/lib.rs

@@ -13,13 +13,14 @@ use std::{
     collections::HashMap,
     fmt::Debug,
     num::{NonZeroU32, NonZeroUsize},
-    path::{Path, PathBuf},
+    path::{Path, PathBuf, StripPrefixError},
     pin::Pin,
     sync::Arc,
 };
 
 use anyhow::{bail, Context};
 
+use serde::{Deserialize, Serialize};
 use tokio::io;
 use toml_edit::Item;
 use tracing::info;
@@ -44,12 +45,34 @@ 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())
@@ -88,6 +111,15 @@ impl RemotePath {
     pub fn extension(&self) -> Option<&str> {
         self.0.extension()?.to_str()
     }
+
+    /// Unwrap the PathBuf that RemotePath wraps
+    pub fn take(self) -> PathBuf {
+        self.0
+    }
+
+    pub fn strip_prefix(&self, p: &RemotePath) -> Result<&Path, StripPrefixError> {
+        self.0.strip_prefix(&p.0)
+    }
 }
 
 /// Storage (potentially remote) API to manage its state.
@@ -166,6 +198,8 @@ 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),
 }
@@ -177,6 +211,7 @@ 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:?}"),
         }
     }
@@ -241,6 +276,18 @@ 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/local_fs.rs

@@ -148,21 +148,53 @@ impl RemoteStorage for LocalFs {
             Some(folder) => folder.with_base(&self.storage_root),
             None => self.storage_root.clone(),
         };
-        let mut files = vec![];
-        let mut directory_queue = vec![full_path.clone()];
 
+        // If we were given a directory, we may use it as our starting point.
+        // Otherwise, we must go up to the parent directory.  This is because
+        // S3 object list prefixes can be arbitrary strings, but when reading
+        // the local filesystem we need a directory to start calling read_dir on.
+        let mut initial_dir = full_path.clone();
+        match fs::metadata(full_path.clone()).await {
+            Err(e) => {
+                // It's not a file that exists: strip the prefix back to the parent directory
+                if matches!(e.kind(), ErrorKind::NotFound) {
+                    initial_dir.pop();
+                }
+            }
+            Ok(meta) => {
+                if !meta.is_dir() {
+                    // It's not a directory: strip back to the parent
+                    initial_dir.pop();
+                }
+            }
+        }
+
+        // Note that PathBuf starts_with only considers full path segments, but
+        // object prefixes are arbitrary strings, so we need the strings for doing
+        // starts_with later.
+        let prefix = full_path.to_string_lossy();
+
+        let mut files = vec![];
+        let mut directory_queue = vec![initial_dir.clone()];
         while let Some(cur_folder) = directory_queue.pop() {
             let mut entries = fs::read_dir(cur_folder.clone()).await?;
             while let Some(entry) = entries.next_entry().await? {
                 let file_name: PathBuf = entry.file_name().into();
                 let full_file_name = cur_folder.clone().join(&file_name);
-                let file_remote_path = self.local_file_to_relative_path(full_file_name.clone());
-                files.push(file_remote_path.clone());
-                if full_file_name.is_dir() {
-                    directory_queue.push(full_file_name);
+                if full_file_name
+                    .to_str()
+                    .map(|s| s.starts_with(prefix.as_ref()))
+                    .unwrap_or(false)
+                {
+                    let file_remote_path = self.local_file_to_relative_path(full_file_name.clone());
+                    files.push(file_remote_path.clone());
+                    if full_file_name.is_dir() {
+                        directory_queue.push(full_file_name);
+                    }
                 }
             }
         }
+
         Ok(files)
     }
 

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;
+use hyper::{Body, StatusCode};
 use scopeguard::ScopeGuard;
 use tokio::{
     io::{self, AsyncRead},
@@ -529,7 +529,16 @@ impl RemoteStorage for S3Bucket {
                     }
                 }
                 Err(e) => {
-                    return Err(e.into());
+                    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}"));
+                        }
+                    }
                 }
             }
         }

libs/safekeeper_api/src/models.rs

@@ -31,6 +31,8 @@ fn lsn_invalid() -> Lsn {
 #[serde_as]
 #[derive(Debug, Clone, Deserialize, Serialize)]
 pub struct SkTimelineInfo {
+    /// Term.
+    pub term: Option<u64>,
     /// Term of the last entry.
     pub last_log_term: Option<u64>,
     /// LSN of the last record.
@@ -58,4 +60,6 @@ pub struct SkTimelineInfo {
     /// A connection string to use for WAL receiving.
     #[serde(default)]
     pub safekeeper_connstr: Option<String>,
+    #[serde(default)]
+    pub http_connstr: Option<String>,
 }

libs/utils/Cargo.toml

@@ -38,6 +38,7 @@ url.workspace = true
 uuid.workspace = true
 
 pq_proto.workspace = true
+postgres_connection.workspace = true
 metrics.workspace = true
 workspace_hack.workspace = true
 

libs/utils/src/generation.rs

@@ -0,0 +1,121 @@
+use std::fmt::Display;
+
+use serde::{Deserialize, Serialize};
+
+#[derive(Copy, Clone, Debug, 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");
+            }
+        }
+    }
+
+    pub fn previous(&self) -> Self {
+        if let Self::Valid(v) = self {
+            Self::new(v - 1)
+        } else {
+            Self::none()
+        }
+    }
+
+    pub fn into(self) -> Option<u32> {
+        if let Self::Valid(v) = self {
+            Some(v)
+        } else {
+            None
+        }
+    }
+}
+
+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",
+            ))
+        }
+    }
+}
+
+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)?))
+    }
+}
+
+impl Display 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

@@ -24,6 +24,9 @@ pub enum ApiError {
     #[error("Precondition failed: {0}")]
     PreconditionFailed(Box<str>),
 
+    #[error("Shutting down")]
+    ShuttingDown,
+
     #[error(transparent)]
     InternalServerError(anyhow::Error),
 }
@@ -52,6 +55,10 @@ 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

@@ -50,7 +50,7 @@ impl Id {
         Id::from(tli_buf)
     }
 
-    fn hex_encode(&self) -> String {
+    pub fn hex_encode(&self) -> String {
         static HEX: &[u8] = b"0123456789abcdef";
 
         let mut buf = vec![0u8; self.0.len() * 2];
@@ -133,6 +133,10 @@ macro_rules! id_newtype {
             pub const fn from_array(b: [u8; 16]) -> Self {
                 $t(Id(b))
             }
+
+            pub fn hex_encode(&self) -> String {
+                self.0.hex_encode()
+            }
         }
 
         impl FromStr for $t {
@@ -244,13 +248,13 @@ 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, Serialize, Deserialize, PartialOrd, Ord)]
+#[derive(Clone, Copy, PartialEq, Eq, Hash, PartialOrd, Ord)]
 pub struct ConnectionId(Id);
 
 id_newtype!(ConnectionId);
 
 // A pair uniquely identifying Neon instance.
-#[derive(Debug, Clone, Copy, PartialOrd, Ord, PartialEq, Eq, Hash, Serialize, Deserialize)]
+#[derive(Debug, Clone, Copy, PartialOrd, Ord, PartialEq, Eq, Hash)]
 pub struct TenantTimelineId {
     pub tenant_id: TenantId,
     pub timeline_id: TimelineId,
@@ -273,6 +277,36 @@ 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;
 
@@ -58,6 +61,8 @@ pub mod serde_regex;
 
 pub mod pageserver_feedback;
 
+pub mod postgres_client;
+
 pub mod tracing_span_assert;
 
 pub mod rate_limit;

libs/utils/src/postgres_client.rs

@@ -0,0 +1,37 @@
+//! Postgres client connection code common to other crates (safekeeper and
+//! pageserver) which depends on tenant/timeline ids and thus not fitting into
+//! postgres_connection crate.
+
+use anyhow::Context;
+use postgres_connection::{parse_host_port, PgConnectionConfig};
+
+use crate::id::TenantTimelineId;
+
+/// Create client config for fetching WAL from safekeeper on particular timeline.
+/// listen_pg_addr_str is in form host:\[port\].
+pub fn wal_stream_connection_config(
+    TenantTimelineId {
+        tenant_id,
+        timeline_id,
+    }: TenantTimelineId,
+    listen_pg_addr_str: &str,
+    auth_token: Option<&str>,
+    availability_zone: Option<&str>,
+) -> anyhow::Result<PgConnectionConfig> {
+    let (host, port) =
+        parse_host_port(listen_pg_addr_str).context("Unable to parse listen_pg_addr_str")?;
+    let port = port.unwrap_or(5432);
+    let mut connstr = PgConnectionConfig::new_host_port(host, port)
+        .extend_options([
+            "-c".to_owned(),
+            format!("timeline_id={}", timeline_id),
+            format!("tenant_id={}", tenant_id),
+        ])
+        .set_password(auth_token.map(|s| s.to_owned()));
+
+    if let Some(availability_zone) = availability_zone {
+        connstr = connstr.extend_options([format!("availability_zone={}", availability_zone)]);
+    }
+
+    Ok(connstr)
+}

libs/vm_monitor/README.md

@@ -16,3 +16,19 @@ in the `neon-postgres` cgroup and set its `memory.{max,high}`.
 * See also: [`neondatabase/vm-monitor`](https://github.com/neondatabase/vm-monitor/),
 where initial development of the monitor happened. The repository is no longer
 maintained but the commit history may be useful for debugging.
+
+## Structure
+
+The `vm-monitor` is loosely comprised of a few systems. These are:
+* the server: this is just a simple `axum` server that accepts requests and
+upgrades them to websocket connections. The server only allows one connection at
+a time. This means that upon receiving a new connection, the server will terminate
+and old one if it exists.
+* the filecache: a struct that allows communication with the Postgres file cache.
+On startup, we connect to the filecache and hold on to the connection for the
+entire monitor lifetime.
+* the cgroup watcher: the `CgroupWatcher` manages the `neon-postgres` cgroup by
+listening for `memory.high` events and setting its `memory.{high,max}` values.
+* the runner: the runner marries the filecache and cgroup watcher together,
+communicating with the agent throught the `Dispatcher`, and then calling filecache
+and cgroup watcher functions as needed to upscale and downscale

libs/vm_monitor/src/cgroup.rs

@@ -634,7 +634,7 @@ impl CgroupWatcher {
             .context("failed to get memory subsystem")?
             .set_mem(cgroups_rs::memory::SetMemory {
                 low: None,
-                high: Some(MaxValue::Value(bytes.min(i64::MAX as u64) as i64)),
+                high: Some(MaxValue::Value(u64::min(bytes, i64::MAX as u64) as i64)),
                 min: None,
                 max: None,
             })
@@ -654,8 +654,10 @@ impl CgroupWatcher {
             .set_mem(cgroups_rs::memory::SetMemory {
                 min: None,
                 low: None,
-                high: Some(MaxValue::Value(limits.high.min(i64::MAX as u64) as i64)),
-                max: Some(MaxValue::Value(limits.max.min(i64::MAX as u64) as i64)),
+                high: Some(MaxValue::Value(
+                    u64::min(limits.high, i64::MAX as u64) as i64
+                )),
+                max: Some(MaxValue::Value(u64::min(limits.max, i64::MAX as u64) as i64)),
             })
             .context("failed to set memory limits")
     }

libs/vm_monitor/src/dispatcher.rs

@@ -1,7 +1,7 @@
 //! Managing the websocket connection and other signals in the monitor.
 //!
 //! Contains types that manage the interaction (not data interchange, see `protocol`)
-//! between informant and monitor, allowing us to to process and send messages in a
+//! between agent and monitor, allowing us to to process and send messages in a
 //! straightforward way. The dispatcher also manages that signals that come from
 //! the cgroup (requesting upscale), and the signals that go to the cgroup
 //! (notifying it of upscale).
@@ -24,16 +24,16 @@ use crate::protocol::{
 /// The central handler for all communications in the monitor.
 ///
 /// The dispatcher has two purposes:
-/// 1. Manage the connection to the informant, sending and receiving messages.
+/// 1. Manage the connection to the agent, sending and receiving messages.
 /// 2. Communicate with the cgroup manager, notifying it when upscale is received,
-///    and sending a message to the informant when the cgroup manager requests
+///    and sending a message to the agent when the cgroup manager requests
 ///    upscale.
 #[derive(Debug)]
 pub struct Dispatcher {
-    /// We read informant messages of of `source`
+    /// We read agent messages of of `source`
     pub(crate) source: SplitStream<WebSocket>,
 
-    /// We send messages to the informant through `sink`
+    /// We send messages to the agent through `sink`
     sink: SplitSink<WebSocket, Message>,
 
     /// Used to notify the cgroup when we are upscaled.
@@ -43,7 +43,7 @@ pub struct Dispatcher {
     /// we send an `UpscaleRequst` to the agent.
     pub(crate) request_upscale_events: mpsc::Receiver<()>,
 
-    /// The protocol version we have agreed to use with the informant. This is negotiated
+    /// The protocol version we have agreed to use with the agent. This is negotiated
     /// during the creation of the dispatcher, and should be the highest shared protocol
     /// version.
     ///
@@ -56,9 +56,9 @@ pub struct Dispatcher {
 impl Dispatcher {
     /// Creates a new dispatcher using the passed-in connection.
     ///
-    /// Performs a negotiation with the informant to determine the highest protocol
+    /// Performs a negotiation with the agent to determine the highest protocol
     /// version that both support. This consists of two steps:
-    /// 1. Wait for the informant to sent the range of protocols it supports.
+    /// 1. Wait for the agent to sent the range of protocols it supports.
     /// 2. Send a protocol version that works for us as well, or an error if there
     ///    is no compatible version.
     pub async fn new(
@@ -69,7 +69,7 @@ impl Dispatcher {
         let (mut sink, mut source) = stream.split();
 
         // Figure out the highest protocol version we both support
-        info!("waiting for informant to send protocol version range");
+        info!("waiting for agent to send protocol version range");
         let Some(message) = source.next().await else {
             bail!("websocket connection closed while performing protocol handshake")
         };
@@ -79,7 +79,7 @@ impl Dispatcher {
         let Message::Text(message_text) = message else {
             // All messages should be in text form, since we don't do any
             // pinging/ponging. See nhooyr/websocket's implementation and the
-            // informant/agent for more info
+            // agent for more info
             bail!("received non-text message during proocol handshake: {message:?}")
         };
 
@@ -88,32 +88,30 @@ impl Dispatcher {
             max: PROTOCOL_MAX_VERSION,
         };
 
-        let informant_range: ProtocolRange = serde_json::from_str(&message_text)
+        let agent_range: ProtocolRange = serde_json::from_str(&message_text)
             .context("failed to deserialize protocol version range")?;
 
-        info!(range = ?informant_range, "received protocol version range");
+        info!(range = ?agent_range, "received protocol version range");
 
-        let highest_shared_version = match monitor_range.highest_shared_version(&informant_range) {
+        let highest_shared_version = match monitor_range.highest_shared_version(&agent_range) {
             Ok(version) => {
                 sink.send(Message::Text(
                     serde_json::to_string(&ProtocolResponse::Version(version)).unwrap(),
                 ))
                 .await
-                .context("failed to notify informant of negotiated protocol version")?;
+                .context("failed to notify agent of negotiated protocol version")?;
                 version
             }
             Err(e) => {
                 sink.send(Message::Text(
                     serde_json::to_string(&ProtocolResponse::Error(format!(
                         "Received protocol version range {} which does not overlap with {}",
-                        informant_range, monitor_range
+                        agent_range, monitor_range
                     )))
                     .unwrap(),
                 ))
                 .await
-                .context(
-                    "failed to notify informant of no overlap between protocol version ranges",
-                )?;
+                .context("failed to notify agent of no overlap between protocol version ranges")?;
                 Err(e).context("error determining suitable protocol version range")?
             }
         };
@@ -137,7 +135,7 @@ impl Dispatcher {
             .context("failed to send resources and oneshot sender across channel")
     }
 
-    /// Send a message to the informant.
+    /// Send a message to the agent.
     ///
     /// Although this function is small, it has one major benefit: it is the only
     /// way to send data accross the connection, and you can only pass in a proper

libs/vm_monitor/src/filecache.rs

@@ -59,8 +59,8 @@ pub struct FileCacheConfig {
     spread_factor: f64,
 }
 
-impl Default for FileCacheConfig {
-    fn default() -> Self {
+impl FileCacheConfig {
+    pub fn default_in_memory() -> Self {
         Self {
             in_memory: true,
             // 75 %
@@ -71,9 +71,19 @@ impl Default for FileCacheConfig {
             spread_factor: 0.1,
         }
     }
-}
 
-impl FileCacheConfig {
+    pub fn default_on_disk() -> Self {
+        Self {
+            in_memory: false,
+            resource_multiplier: 0.75,
+            // 256 MiB - lower than when in memory because overcommitting is safe; if we don't have
+            // memory, the kernel will just evict from its page cache, rather than e.g. killing
+            // everything.
+            min_remaining_after_cache: NonZeroU64::new(256 * MiB).unwrap(),
+            spread_factor: 0.1,
+        }
+    }
+
     /// Make sure fields of the config are consistent.
     pub fn validate(&self) -> anyhow::Result<()> {
         // Single field validity
@@ -132,11 +142,11 @@ impl FileCacheConfig {
 
         // Conversions to ensure we don't overflow from floating-point ops
         let size_from_spread =
-            0_i64.max((available as f64 / (1.0 + self.spread_factor)) as i64) as u64;
+            i64::max(0, (available as f64 / (1.0 + self.spread_factor)) as i64) as u64;
 
         let size_from_normal = (total as f64 * self.resource_multiplier) as u64;
 
-        let byte_size = size_from_spread.min(size_from_normal);
+        let byte_size = u64::min(size_from_spread, size_from_normal);
 
         // The file cache operates in units of mebibytes, so the sizes we produce should
         // be rounded to a mebibyte. We round down to be conservative.
@@ -268,7 +278,7 @@ impl FileCacheState {
             .context("failed to extract max file cache size from query result")?;
 
         let max_mb = max_bytes / MiB;
-        let num_mb = (num_bytes / MiB).max(max_mb);
+        let num_mb = u64::min(num_bytes, max_bytes) / MiB;
 
         let capped = if num_bytes > max_bytes {
             " (capped by maximum size)"

libs/vm_monitor/src/lib.rs

@@ -39,6 +39,16 @@ pub struct Args {
     #[arg(short, long)]
     pub pgconnstr: Option<String>,
 
+    /// Flag to signal that the Postgres file cache is on disk (i.e. not in memory aside from the
+    /// kernel's page cache), and therefore should not count against available memory.
+    //
+    // NB: Ideally this flag would directly refer to whether the file cache is in memory (rather
+    // than a roundabout way, via whether it's on disk), but in order to be backwards compatible
+    // during the switch away from an in-memory file cache, we had to default to the previous
+    // behavior.
+    #[arg(long)]
+    pub file_cache_on_disk: bool,
+
     /// The address we should listen on for connection requests. For the
     /// agent, this is 0.0.0.0:10301. For the informant, this is 127.0.0.1:10369.
     #[arg(short, long)]
@@ -146,7 +156,7 @@ pub async fn start(args: &'static Args, token: CancellationToken) -> anyhow::Res
 
 /// Handles incoming websocket connections.
 ///
-/// If we are already to connected to an informant, we kill that old connection
+/// If we are already to connected to an agent, we kill that old connection
 /// and accept the new one.
 #[tracing::instrument(name = "/monitor", skip_all, fields(?args))]
 pub async fn ws_handler(
@@ -196,10 +206,10 @@ async fn start_monitor(
             return;
         }
     };
-    info!("connected to informant");
+    info!("connected to agent");
 
     match monitor.run().await {
         Ok(()) => info!("monitor was killed due to new connection"),
-        Err(e) => error!(error = ?e, "monitor terminated by itself"),
+        Err(e) => error!(error = ?e, "monitor terminated unexpectedly"),
     }
 }

libs/vm_monitor/src/protocol.rs

@@ -1,13 +1,13 @@
-//! Types representing protocols and actual informant-monitor messages.
+//! Types representing protocols and actual agent-monitor messages.
 //!
 //! The pervasive use of serde modifiers throughout this module is to ease
 //! serialization on the go side. Because go does not have enums (which model
 //! messages well), it is harder to model messages, and we accomodate that with
 //! serde.
 //!
-//! *Note*: the informant sends and receives messages in different ways.
+//! *Note*: the agent sends and receives messages in different ways.
 //!
-//! The informant serializes messages in the form and then sends them. The use
+//! The agent serializes messages in the form and then sends them. The use
 //! of `#[serde(tag = "type", content = "content")]` allows us to use `Type`
 //! to determine how to deserialize `Content`.
 //! ```ignore
@@ -25,9 +25,9 @@
 //!     Id   uint64
 //! }
 //! ```
-//! After reading the type field, the informant will decode the entire message
+//! After reading the type field, the agent will decode the entire message
 //! again, this time into the correct type using the embedded fields.
-//! Because the informant cannot just extract the json contained in a certain field
+//! Because the agent cannot just extract the json contained in a certain field
 //! (it initially deserializes to `map[string]interface{}`), we keep the fields
 //! at the top level, so the entire piece of json can be deserialized into a struct,
 //! such as a `DownscaleResult`, with the `Type` and `Id` fields ignored.
@@ -37,7 +37,7 @@ use std::cmp;
 
 use serde::{de::Error, Deserialize, Serialize};
 
-/// A Message we send to the informant.
+/// A Message we send to the agent.
 #[derive(Serialize, Deserialize, Debug, Clone)]
 pub struct OutboundMsg {
     #[serde(flatten)]
@@ -51,31 +51,31 @@ impl OutboundMsg {
     }
 }
 
-/// The different underlying message types we can send to the informant.
+/// The different underlying message types we can send to the agent.
 #[derive(Serialize, Deserialize, Debug, Clone)]
 #[serde(tag = "type")]
 pub enum OutboundMsgKind {
-    /// Indicates that the informant sent an invalid message, i.e, we couldn't
+    /// Indicates that the agent sent an invalid message, i.e, we couldn't
     /// properly deserialize it.
     InvalidMessage { error: String },
     /// Indicates that we experienced an internal error while processing a message.
     /// For example, if a cgroup operation fails while trying to handle an upscale,
     /// we return `InternalError`.
     InternalError { error: String },
-    /// Returned to the informant once we have finished handling an upscale. If the
+    /// Returned to the agent once we have finished handling an upscale. If the
     /// handling was unsuccessful, an `InternalError` will get returned instead.
     /// *Note*: this is a struct variant because of the way go serializes struct{}
     UpscaleConfirmation {},
     /// Indicates to the monitor that we are urgently requesting resources.
     /// *Note*: this is a struct variant because of the way go serializes struct{}
     UpscaleRequest {},
-    /// Returned to the informant once we have finished attempting to downscale. If
+    /// Returned to the agent once we have finished attempting to downscale. If
     /// an error occured trying to do so, an `InternalError` will get returned instead.
     /// However, if we are simply unsuccessful (for example, do to needing the resources),
     /// that gets included in the `DownscaleResult`.
     DownscaleResult {
         // FIXME for the future (once the informant is deprecated)
-        // As of the time of writing, the informant/agent version of this struct is
+        // As of the time of writing, the agent/informant version of this struct is
         // called api.DownscaleResult. This struct has uppercase fields which are
         // serialized as such. Thus, we serialize using uppercase names so we don't
         // have to make a breaking change to the agent<->informant protocol. Once
@@ -88,12 +88,12 @@ pub enum OutboundMsgKind {
         status: String,
     },
     /// Part of the bidirectional heartbeat. The heartbeat is initiated by the
-    /// informant.
+    /// agent.
     /// *Note*: this is a struct variant because of the way go serializes struct{}
     HealthCheck {},
 }
 
-/// A message received form the informant.
+/// A message received form the agent.
 #[derive(Serialize, Deserialize, Debug, Clone)]
 pub struct InboundMsg {
     #[serde(flatten)]
@@ -101,7 +101,7 @@ pub struct InboundMsg {
     pub(crate) id: usize,
 }
 
-/// The different underlying message types we can receive from the informant.
+/// The different underlying message types we can receive from the agent.
 #[derive(Serialize, Deserialize, Debug, Clone)]
 #[serde(tag = "type", content = "content")]
 pub enum InboundMsgKind {
@@ -120,14 +120,14 @@ pub enum InboundMsgKind {
     /// when done.
     DownscaleRequest { target: Resources },
     /// Part of the bidirectional heartbeat. The heartbeat is initiated by the
-    /// informant.
+    /// agent.
     /// *Note*: this is a struct variant because of the way go serializes struct{}
     HealthCheck {},
 }
 
 /// Represents the resources granted to a VM.
 #[derive(Serialize, Deserialize, Debug, Clone, Copy)]
-// Renamed because the agent/informant has multiple resources types:
+// Renamed because the agent has multiple resources types:
 // `Resources` (milliCPU/memory slots)
 // `Allocation` (vCPU/bytes) <- what we correspond to
 #[serde(rename(serialize = "Allocation", deserialize = "Allocation"))]
@@ -151,7 +151,7 @@ pub const PROTOCOL_MAX_VERSION: ProtocolVersion = ProtocolVersion::V1_0;
 pub struct ProtocolVersion(u8);
 
 impl ProtocolVersion {
-    /// Represents v1.0 of the informant<-> monitor protocol - the initial version
+    /// Represents v1.0 of the agent<-> monitor protocol - the initial version
     ///
     /// Currently the latest version.
     const V1_0: ProtocolVersion = ProtocolVersion(1);

libs/vm_monitor/src/runner.rs

@@ -1,4 +1,4 @@
-//! Exposes the `Runner`, which handles messages received from informant and
+//! Exposes the `Runner`, which handles messages received from agent and
 //! sends upscale requests.
 //!
 //! This is the "Monitor" part of the monitor binary and is the main entrypoint for
@@ -21,8 +21,8 @@ use crate::filecache::{FileCacheConfig, FileCacheState};
 use crate::protocol::{InboundMsg, InboundMsgKind, OutboundMsg, OutboundMsgKind, Resources};
 use crate::{bytes_to_mebibytes, get_total_system_memory, spawn_with_cancel, Args, MiB};
 
-/// Central struct that interacts with informant, dispatcher, and cgroup to handle
-/// signals from the informant.
+/// Central struct that interacts with agent, dispatcher, and cgroup to handle
+/// signals from the agent.
 #[derive(Debug)]
 pub struct Runner {
     config: Config,
@@ -110,10 +110,10 @@ impl Runner {
         // memory limits.
         if let Some(connstr) = &args.pgconnstr {
             info!("initializing file cache");
-            let config: FileCacheConfig = Default::default();
-            if !config.in_memory {
-                panic!("file cache not in-memory implemented")
-            }
+            let config = match args.file_cache_on_disk {
+                true => FileCacheConfig::default_on_disk(),
+                false => FileCacheConfig::default_in_memory(),
+            };
 
             let mut file_cache = FileCacheState::new(connstr, config, token.clone())
                 .await
@@ -140,7 +140,10 @@ impl Runner {
             if actual_size != new_size {
                 info!("file cache size actually got set to {actual_size}")
             }
-            file_cache_reserved_bytes = actual_size;
+            // Mark the resources given to the file cache as reserved, but only if it's in memory.
+            if !args.file_cache_on_disk {
+                file_cache_reserved_bytes = actual_size;
+            }
 
             state.filecache = Some(file_cache);
         }
@@ -227,18 +230,17 @@ impl Runner {
         let mut status = vec![];
         let mut file_cache_mem_usage = 0;
         if let Some(file_cache) = &mut self.filecache {
-            if !file_cache.config.in_memory {
-                panic!("file cache not in-memory unimplemented")
-            }
-
             let actual_usage = file_cache
                 .set_file_cache_size(expected_file_cache_mem_usage)
                 .await
                 .context("failed to set file cache size")?;
-            file_cache_mem_usage = actual_usage;
+            if file_cache.config.in_memory {
+                file_cache_mem_usage = actual_usage;
+            }
             let message = format!(
-                "set file cache size to {} MiB",
-                bytes_to_mebibytes(actual_usage)
+                "set file cache size to {} MiB (in memory = {})",
+                bytes_to_mebibytes(actual_usage),
+                file_cache.config.in_memory,
             );
             info!("downscale: {message}");
             status.push(message);
@@ -289,10 +291,6 @@ impl Runner {
         // Get the file cache's expected contribution to the memory usage
         let mut file_cache_mem_usage = 0;
         if let Some(file_cache) = &mut self.filecache {
-            if !file_cache.config.in_memory {
-                panic!("file cache not in-memory unimplemented");
-            }
-
             let expected_usage = file_cache.config.calculate_cache_size(usable_system_memory);
             info!(
                 target = bytes_to_mebibytes(expected_usage),
@@ -304,6 +302,9 @@ impl Runner {
                 .set_file_cache_size(expected_usage)
                 .await
                 .context("failed to set file cache size")?;
+            if file_cache.config.in_memory {
+                file_cache_mem_usage = actual_usage;
+            }
 
             if actual_usage != expected_usage {
                 warn!(
@@ -312,7 +313,6 @@ impl Runner {
                     bytes_to_mebibytes(actual_usage)
                 )
             }
-            file_cache_mem_usage = actual_usage;
         }
 
         if let Some(cgroup) = &self.cgroup {
@@ -371,7 +371,7 @@ impl Runner {
                 Ok(None)
             }
             InboundMsgKind::InternalError { error } => {
-                warn!(error, id, "informant experienced an internal error");
+                warn!(error, id, "agent experienced an internal error");
                 Ok(None)
             }
             InboundMsgKind::HealthCheck {} => {
@@ -405,10 +405,12 @@ impl Runner {
                         .await
                         .context("failed to send message")?;
                 }
-                // there is a message from the informant
+                // there is a message from the agent
                 msg = self.dispatcher.source.next() => {
                     if let Some(msg) = msg {
-                        info!(message = ?msg, "received message");
+                        // Don't use 'message' as a key as the string also uses
+                        // that for its key
+                        info!(?msg, "received message");
                         match msg {
                             Ok(msg) => {
                                 let message: InboundMsg = match msg {
@@ -417,8 +419,10 @@ impl Runner {
                                     }
                                     other => {
                                         warn!(
-                                            message = ?other,
-                                            "informant should only send text messages but received different type"
+                                            // Don't use 'message' as a key as the
+                                            // string also uses that for its key
+                                            msg = ?other,
+                                            "agent should only send text messages but received different type"
                                         );
                                         continue
                                     },
@@ -429,7 +433,7 @@ impl Runner {
                                     Ok(None) => continue,
                                     Err(e) => {
                                         let error = e.to_string();
-                                        warn!(%error, "error handling message");
+                                        warn!(?error, "error handling message");
                                         OutboundMsg::new(
                                             OutboundMsgKind::InternalError {
                                                 error

pageserver/ctl/src/layer_map_analyzer.rs

@@ -97,7 +97,7 @@ pub(crate) fn parse_filename(name: &str) -> Option<LayerFile> {
 // Finds the max_holes largest holes, ignoring any that are smaller than MIN_HOLE_LENGTH"
 async fn get_holes(path: &Path, max_holes: usize) -> Result<Vec<Hole>> {
     let file = FileBlockReader::new(VirtualFile::open(path)?);
-    let summary_blk = file.read_blk(0)?;
+    let summary_blk = file.read_blk(0).await?;
     let actual_summary = Summary::des_prefix(summary_blk.as_ref())?;
     let tree_reader = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(
         actual_summary.index_start_blk,

pageserver/ctl/src/layers.rs

@@ -48,7 +48,7 @@ async fn read_delta_file(path: impl AsRef<Path>) -> Result<()> {
     virtual_file::init(10);
     page_cache::init(100);
     let file = FileBlockReader::new(VirtualFile::open(path)?);
-    let summary_blk = file.read_blk(0)?;
+    let summary_blk = file.read_blk(0).await?;
     let actual_summary = Summary::des_prefix(summary_blk.as_ref())?;
     let tree_reader = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(
         actual_summary.index_start_blk,

pageserver/src/bin/pageserver.rs

@@ -2,12 +2,14 @@
 
 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;
@@ -349,6 +351,35 @@ 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");
@@ -386,6 +417,7 @@ fn start_pageserver(
         TenantSharedResources {
             broker_client: broker_client.clone(),
             remote_storage: remote_storage.clone(),
+            deletion_queue_client: deletion_queue.new_client(),
         },
         order,
     ))?;
@@ -482,6 +514,7 @@ fn start_pageserver(
             http_auth,
             broker_client.clone(),
             remote_storage,
+            deletion_queue.clone(),
             disk_usage_eviction_state,
         )?
         .build()
@@ -604,6 +637,36 @@ 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

@@ -204,6 +204,8 @@ pub struct PageServerConf {
     /// has it's initial logical size calculated. Not running background tasks for some seconds is
     /// not terrible.
     pub background_task_maximum_delay: Duration,
+
+    pub control_plane_api: Option<Url>,
 }
 
 /// We do not want to store this in a PageServerConf because the latter may be logged
@@ -278,6 +280,8 @@ struct PageServerConfigBuilder {
     ondemand_download_behavior_treat_error_as_warn: BuilderValue<bool>,
 
     background_task_maximum_delay: BuilderValue<Duration>,
+
+    control_plane_api: BuilderValue<Option<Url>>,
 }
 
 impl Default for PageServerConfigBuilder {
@@ -340,6 +344,8 @@ impl Default for PageServerConfigBuilder {
                 DEFAULT_BACKGROUND_TASK_MAXIMUM_DELAY,
             )
             .unwrap()),
+
+            control_plane_api: Set(None),
         }
     }
 }
@@ -468,6 +474,10 @@ impl PageServerConfigBuilder {
         self.background_task_maximum_delay = BuilderValue::Set(delay);
     }
 
+    pub fn control_plane_api(&mut self, api: Url) {
+        self.control_plane_api = BuilderValue::Set(Some(api))
+    }
+
     pub fn build(self) -> anyhow::Result<PageServerConf> {
         let concurrent_tenant_size_logical_size_queries = self
             .concurrent_tenant_size_logical_size_queries
@@ -553,6 +563,9 @@ impl PageServerConfigBuilder {
             background_task_maximum_delay: self
                 .background_task_maximum_delay
                 .ok_or(anyhow!("missing background_task_maximum_delay"))?,
+            control_plane_api: self
+                .control_plane_api
+                .ok_or(anyhow!("missing control_plane_api"))?,
         })
     }
 }
@@ -566,6 +579,27 @@ 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())
     }
@@ -643,23 +677,6 @@ 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)
@@ -758,6 +775,7 @@ impl PageServerConf {
                 },
                 "ondemand_download_behavior_treat_error_as_warn" => builder.ondemand_download_behavior_treat_error_as_warn(parse_toml_bool(key, item)?),
                 "background_task_maximum_delay" => builder.background_task_maximum_delay(parse_toml_duration(key, item)?),
+                "control_plane_api" => builder.control_plane_api(parse_toml_string(key, item)?.parse().context("failed to parse control plane URL")?),
                 _ => bail!("unrecognized pageserver option '{key}'"),
             }
         }
@@ -926,6 +944,7 @@ impl PageServerConf {
             test_remote_failures: 0,
             ondemand_download_behavior_treat_error_as_warn: false,
             background_task_maximum_delay: Duration::ZERO,
+            control_plane_api: None,
         }
     }
 }
@@ -1149,6 +1168,7 @@ background_task_maximum_delay = '334 s'
                 background_task_maximum_delay: humantime::parse_duration(
                     defaults::DEFAULT_BACKGROUND_TASK_MAXIMUM_DELAY
                 )?,
+                control_plane_api: None
             },
             "Correct defaults should be used when no config values are provided"
         );
@@ -1204,6 +1224,7 @@ background_task_maximum_delay = '334 s'
                 test_remote_failures: 0,
                 ondemand_download_behavior_treat_error_as_warn: false,
                 background_task_maximum_delay: Duration::from_secs(334),
+                control_plane_api: None
             },
             "Should be able to parse all basic config values correctly"
         );

pageserver/src/deletion_queue.rs

@@ -0,0 +1,850 @@
+mod backend;
+mod executor;
+mod frontend;
+
+use std::collections::HashMap;
+use std::path::PathBuf;
+
+use crate::metrics::DELETION_QUEUE_SUBMITTED;
+use crate::tenant::remote_timeline_client::remote_timeline_path;
+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::generation::Generation;
+use utils::id::{TenantId, 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 TenantDeletionList {
+    /// For each Timeline, a list of key fragments to append to the timeline remote path
+    /// when reconstructing a full key
+    timelines: HashMap<TimelineId, Vec<String>>,
+
+    /// The generation in which this deletion was emitted: note that this may not be the
+    /// same as the generation of any layers being deleted.  The generation of the layer
+    /// has already been absorbed into the keys in `objects`
+    generation: Generation,
+}
+
+#[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.  Each Tenant only appears once
+    /// with one unique generation ID: if someone tries to push a second generation
+    /// ID for the same tenant, we will start a new DeletionList.
+    tenants: HashMap<TenantId, TenantDeletionList>,
+
+    /// Avoid having to walk `tenants` to calculate size
+    size: usize,
+}
+
+#[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,
+            tenants: HashMap::new(),
+            size: 0,
+        }
+    }
+
+    fn drain(&mut self) -> Self {
+        let mut tenants = HashMap::new();
+        std::mem::swap(&mut self.tenants, &mut tenants);
+        let other = Self {
+            version: Self::VERSION_LATEST,
+            sequence: self.sequence,
+            tenants,
+            size: self.size,
+        };
+        self.size = 0;
+        other
+    }
+
+    fn is_empty(&self) -> bool {
+        self.tenants.is_empty()
+    }
+
+    fn len(&self) -> usize {
+        self.size
+    }
+
+    /// Returns true if the push was accepted, false if the caller must start a new
+    /// deletion list.
+    fn push(
+        &mut self,
+        tenant: &TenantId,
+        timeline: &TimelineId,
+        generation: Generation,
+        objects: &mut Vec<RemotePath>,
+    ) -> bool {
+        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 true;
+        }
+
+        let tenant_entry = self
+            .tenants
+            .entry(*tenant)
+            .or_insert_with(|| TenantDeletionList {
+                timelines: HashMap::new(),
+                generation: generation,
+            });
+
+        if tenant_entry.generation != generation {
+            // Only one generation per tenant per list: signal to
+            // caller to start a new list.
+            return false;
+        }
+
+        let timeline_entry = tenant_entry
+            .timelines
+            .entry(*timeline)
+            .or_insert_with(|| Vec::new());
+
+        let timeline_remote_path = remote_timeline_path(tenant, timeline);
+
+        self.size += objects.len();
+        timeline_entry.extend(objects.drain(..).map(|p| {
+            p.strip_prefix(&timeline_remote_path)
+                .expect("Timeline paths always start with the timeline prefix")
+                .to_string_lossy()
+                .to_string()
+        }));
+        true
+    }
+
+    fn take_paths(self) -> Vec<RemotePath> {
+        let mut result = Vec::new();
+        for (tenant, tenant_deletions) in self.tenants.into_iter() {
+            for (timeline, timeline_layers) in tenant_deletions.timelines.into_iter() {
+                let timeline_remote_path = remote_timeline_path(&tenant, &timeline);
+                result.extend(
+                    timeline_layers
+                        .into_iter()
+                        .map(|l| timeline_remote_path.join(&PathBuf::from(l))),
+                );
+            }
+        }
+
+        result
+    }
+}
+
+#[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,
+        generation: Generation,
+        layers: Vec<(LayerFileName, Generation)>,
+    ) -> Result<(), DeletionQueueError> {
+        DELETION_QUEUE_SUBMITTED.inc_by(layers.len() as u64);
+        self.do_push(FrontendQueueMessage::Delete(DeletionOp {
+            tenant_id,
+            timeline_id,
+            layers,
+            generation,
+            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,
+                cancel.clone(),
+            )),
+            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, remote_timeline_client::remote_timeline_path};
+
+    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 = remote_timeline_path(&tenant_id, &TIMELINE_ID);
+        let remote_timeline_path = ctx.remote_fs_dir.join(relative_remote_path.get_path());
+        let deletion_prefix = ctx.harness.conf.deletion_prefix();
+
+        // Exercise the distinction between the generation of the layers
+        // we delete, and the generation of the running Tenant.
+        let layer_generation = Generation::new(0xdeadbeef);
+        let now_generation = Generation::new(0xfeedbeef);
+
+        let remote_layer_file_name_1 =
+            format!("{}{}", layer_file_name_1, layer_generation.get_suffix());
+
+        // Inject a victim file to remote storage
+        info!("Writing");
+        std::fs::create_dir_all(&remote_timeline_path)?;
+        std::fs::write(
+            remote_timeline_path.join(remote_layer_file_name_1.clone()),
+            content,
+        )?;
+        assert_remote_files(&[&remote_layer_file_name_1], &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,
+            now_generation,
+            [(layer_file_name_1.clone(), layer_generation)].to_vec(),
+        ))?;
+        assert_remote_files(&[&remote_layer_file_name_1], &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(&[&remote_layer_file_name_1], &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 = remote_timeline_path(&tenant_id, &TIMELINE_ID);
+        let remote_timeline_path = ctx.remote_fs_dir.join(relative_remote_path.get_path());
+        let deletion_prefix = ctx.harness.conf.deletion_prefix();
+        let layer_generation = Generation::new(0xdeadbeef);
+        let now_generation = Generation::new(0xfeedbeef);
+        let remote_layer_file_name_1 =
+            format!("{}{}", layer_file_name_1, layer_generation.get_suffix());
+
+        // 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(remote_layer_file_name_1.clone()),
+            content,
+        )?;
+        ctx.runtime.block_on(client.push_layers(
+            tenant_id,
+            TIMELINE_ID,
+            now_generation,
+            [(layer_file_name_1.clone(), layer_generation)].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 crate::tenant::remote_timeline_client::remote_layer_path;
+
+    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>) -> 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 mut objects = op.objects;
+                                for (layer, generation) in op.layers {
+                                    objects.push(remote_layer_path(
+                                        &op.tenant_id,
+                                        &op.timeline_id,
+                                        &layer,
+                                        generation,
+                                    ));
+                                }
+
+                                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

@@ -0,0 +1,300 @@
+use std::collections::HashMap;
+use std::time::Duration;
+
+use futures::future::TryFutureExt;
+use pageserver_api::control_api::HexTenantId;
+use pageserver_api::control_api::{ValidateRequest, ValidateRequestTenant, ValidateResponse};
+use serde::de::DeserializeOwned;
+use tokio_util::sync::CancellationToken;
+use tracing::debug;
+use tracing::info;
+use tracing::warn;
+use utils::backoff;
+
+use crate::config::PageServerConf;
+use crate::metrics::DELETION_QUEUE_ERRORS;
+
+use super::executor::ExecutorMessage;
+use super::DeletionHeader;
+use super::DeletionList;
+use super::DeletionQueueError;
+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>,
+
+    cancel: CancellationToken,
+}
+
+#[derive(thiserror::Error, Debug)]
+enum ValidateCallError {
+    #[error("shutdown")]
+    Shutdown,
+    #[error("remote: {0}")]
+    Remote(reqwest::Error),
+}
+
+async fn retry_http_forever<T>(
+    url: &url::Url,
+    request: ValidateRequest,
+    cancel: CancellationToken,
+) -> Result<T, DeletionQueueError>
+where
+    T: DeserializeOwned,
+{
+    let client = reqwest::ClientBuilder::new()
+        .build()
+        .expect("Failed to construct http client");
+
+    let response = match backoff::retry(
+        || {
+            client
+                .post(url.clone())
+                .json(&request)
+                .send()
+                .map_err(|e| ValidateCallError::Remote(e))
+        },
+        |_| false,
+        3,
+        u32::MAX,
+        "calling control plane generation validation API",
+        backoff::Cancel::new(cancel.clone(), || ValidateCallError::Shutdown),
+    )
+    .await
+    {
+        Err(ValidateCallError::Shutdown) => {
+            return Err(DeletionQueueError::ShuttingDown);
+        }
+        Err(ValidateCallError::Remote(_)) => {
+            panic!("We retry forever");
+        }
+        Ok(r) => r,
+    };
+
+    // TODO: handle non-200 response
+    // TODO: handle decode error
+    Ok(response.json::<T>().await.unwrap())
+}
+
+impl BackendQueueWorker {
+    pub(super) fn new(
+        conf: &'static PageServerConf,
+        rx: tokio::sync::mpsc::Receiver<BackendQueueMessage>,
+        tx: tokio::sync::mpsc::Sender<ExecutorMessage>,
+        cancel: CancellationToken,
+    ) -> Self {
+        Self {
+            conf,
+            rx,
+            tx,
+            pending_lists: Vec::new(),
+            pending_key_count: 0,
+            executed_lists: Vec::new(),
+            cancel,
+        }
+    }
+
+    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 validate_lists(&mut self) -> Result<(), DeletionQueueError> {
+        let control_plane_api = match &self.conf.control_plane_api {
+            None => {
+                // Generations are not switched on yet.
+                return Ok(());
+            }
+            Some(api) => api,
+        };
+
+        let validate_path = control_plane_api
+            .join("validate")
+            .expect("Failed to build validate path");
+
+        for list in &mut self.pending_lists {
+            let request = ValidateRequest {
+                tenants: list
+                    .tenants
+                    .iter()
+                    .map(|(tid, tdl)| ValidateRequestTenant {
+                        id: HexTenantId::new(*tid),
+                        gen: tdl.generation.into().expect(
+                            "Generation should always be valid for a Tenant doing deletions",
+                        ),
+                    })
+                    .collect(),
+            };
+
+            // Retry forever, we cannot make progress until we get a response
+            let response: ValidateResponse =
+                retry_http_forever(&validate_path, request, self.cancel.clone()).await?;
+
+            let tenants_valid: HashMap<_, _> = response
+                .tenants
+                .into_iter()
+                .map(|t| (t.id.take(), t.valid))
+                .collect();
+
+            // Filter the list based on whether the server responded valid: true.
+            // If a tenant is omitted in the response, it has been deleted, and we should
+            // proceed with deletion.
+            list.tenants.retain(|tenant_id, _tenant| {
+                let r = tenants_valid.get(tenant_id).map(|v| *v).unwrap_or(true);
+                if !r {
+                    warn!("Dropping stale deletions for tenant {tenant_id}, objects may be leaked");
+                }
+                r
+            });
+        }
+
+        Ok(())
+    }
+
+    pub async fn flush(&mut self) {
+        // Issue any required generation validation calls to the control plane
+        if let Err(DeletionQueueError::ShuttingDown) = self.validate_lists().await {
+            warn!("Shutting down");
+            return;
+        }
+
+        // Submit all keys from pending DeletionLists into the executor
+        for list in self.pending_lists.drain(..) {
+            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.pending_key_count = 0;
+        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.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

@@ -0,0 +1,143 @@
+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

@@ -0,0 +1,376 @@
+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::generation::Generation;
+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::remote_timeline_client::remote_layer_path;
+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, Generation)>,
+    pub(super) objects: Vec<RemotePath>,
+
+    /// The _current_ generation of the Tenant attachment in which we are enqueuing
+    /// this deletion.
+    pub(super) generation: Generation,
+}
+
+#[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 onward_list = self.pending.drain();
+
+                // 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 mut layer_paths = Vec::new();
+                    for (layer, generation) in op.layers {
+                        layer_paths.push(remote_layer_path(
+                            &op.tenant_id,
+                            &op.timeline_id,
+                            &layer,
+                            generation,
+                        ));
+                    }
+                    layer_paths.extend(op.objects);
+
+                    if self.pending.push(
+                        &op.tenant_id,
+                        &op.timeline_id,
+                        op.generation,
+                        &mut layer_paths,
+                    ) == false
+                    {
+                        self.flush().await;
+                        let retry = self.pending.push(
+                            &op.tenant_id,
+                            &op.timeline_id,
+                            op.generation,
+                            &mut layer_paths,
+                        );
+                        if retry != true {
+                            // Unexpeted: after we flush, we should have
+                            // drained self.pending, so a conflict on
+                            // generation numbers should be impossible.
+                            tracing::error!(
+                                "Failed to enqueue deletions, leaking objects.  This is a bug."
+                            );
+                        }
+                    }
+                }
+                FrontendQueueMessage::Flush(op) => {
+                    if self.pending.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.len() > DELETION_LIST_TARGET_SIZE || !self.pending_flushes.is_empty() {
+                self.flush().await;
+            }
+        }
+        info!("Deletion queue shut down.");
+    }
+}

pageserver/src/http/openapi_spec.yml

@@ -52,6 +52,29 @@ 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
@@ -383,7 +406,6 @@ paths:
         schema:
           type: string
           format: hex
-
     post:
       description: |
         Schedules attach operation to happen in the background for the given tenant.
@@ -1020,6 +1042,9 @@ components:
       properties:
         config:
           $ref: '#/components/schemas/TenantConfig'
+        generation:
+          type: integer
+          description: Attachment generation number.
     TenantConfigRequest:
       allOf:
         - $ref: '#/components/schemas/TenantConfig'

pageserver/src/http/routes.rs

@@ -23,6 +23,7 @@ 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;
@@ -32,11 +33,13 @@ use crate::tenant::mgr::{
 };
 use crate::tenant::size::ModelInputs;
 use crate::tenant::storage_layer::LayerAccessStatsReset;
-use crate::tenant::{LogicalSizeCalculationCause, PageReconstructError, Timeline};
+use crate::tenant::timeline::Timeline;
+use crate::tenant::{LogicalSizeCalculationCause, PageReconstructError};
 use crate::{config::PageServerConf, tenant::mgr};
 use crate::{disk_usage_eviction_task, tenant};
 use utils::{
     auth::JwtAuth,
+    generation::Generation,
     http::{
         endpoint::{self, attach_openapi_ui, auth_middleware, check_permission_with},
         error::{ApiError, HttpErrorBody},
@@ -56,6 +59,7 @@ 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>,
 }
@@ -65,6 +69,7 @@ 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> {
@@ -78,6 +83,7 @@ impl State {
             allowlist_routes,
             remote_storage,
             broker_client,
+            deletion_queue,
             disk_usage_eviction_state,
         })
     }
@@ -472,7 +478,7 @@ async fn tenant_attach_handler(
     check_permission(&request, Some(tenant_id))?;
 
     let maybe_body: Option<TenantAttachRequest> = json_request_or_empty_body(&mut request).await?;
-    let tenant_conf = match maybe_body {
+    let tenant_conf = match &maybe_body {
         Some(request) => TenantConfOpt::try_from(&*request.config).map_err(ApiError::BadRequest)?,
         None => TenantConfOpt::default(),
     };
@@ -483,13 +489,30 @@ async fn tenant_attach_handler(
 
     let state = get_state(&request);
 
+    let generation = if state.conf.control_plane_api.is_some() {
+        // If we have been configured with a control plane URI, then generations are
+        // mandatory, as we will attempt to re-attach on startup.
+        maybe_body
+            .as_ref()
+            .map(|tar| tar.generation)
+            .flatten()
+            .map(|g| Generation::new(g))
+            .ok_or(ApiError::BadRequest(anyhow!(
+                "generation attribute missing"
+            )))?
+    } else {
+        Generation::none()
+    };
+
     if let Some(remote_storage) = &state.remote_storage {
         mgr::attach_tenant(
             state.conf,
             tenant_id,
+            generation,
             tenant_conf,
             state.broker_client.clone(),
             remote_storage.clone(),
+            &state.deletion_queue,
             &ctx,
         )
         .instrument(info_span!("tenant_attach", %tenant_id))
@@ -552,6 +575,7 @@ 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))
@@ -867,6 +891,12 @@ async fn tenant_create_handler(
     let tenant_conf =
         TenantConfOpt::try_from(&request_data.config).map_err(ApiError::BadRequest)?;
 
+    // TODO: make generation mandatory here once control plane supports it.
+    let generation = request_data
+        .generation
+        .map(|g| Generation::new(g))
+        .unwrap_or(Generation::none());
+
     let ctx = RequestContext::new(TaskKind::MgmtRequest, DownloadBehavior::Warn);
 
     let state = get_state(&request);
@@ -875,8 +905,10 @@ async fn tenant_create_handler(
         state.conf,
         tenant_conf,
         target_tenant_id,
+        generation,
         state.broker_client.clone(),
         state.remote_storage.clone(),
+        &state.deletion_queue,
         &ctx,
     )
     .instrument(info_span!("tenant_create", tenant_id = %target_tenant_id))
@@ -1117,6 +1149,48 @@ 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,
@@ -1326,6 +1400,7 @@ 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");
@@ -1355,6 +1430,7 @@ pub fn make_router(
                 conf,
                 auth,
                 remote_storage,
+                deletion_queue,
                 broker_client,
                 disk_usage_eviction_state,
             )
@@ -1439,6 +1515,9 @@ 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

@@ -3,6 +3,7 @@ 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

@@ -795,6 +795,31 @@ 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_cache.rs

@@ -75,10 +75,7 @@
 use std::{
     collections::{hash_map::Entry, HashMap},
     convert::TryInto,
-    sync::{
-        atomic::{AtomicU64, AtomicU8, AtomicUsize, Ordering},
-        RwLock, RwLockReadGuard, RwLockWriteGuard, TryLockError,
-    },
+    sync::atomic::{AtomicU64, AtomicU8, AtomicUsize, Ordering},
 };
 
 use anyhow::Context;
@@ -162,7 +159,7 @@ struct Version {
 }
 
 struct Slot {
-    inner: RwLock<SlotInner>,
+    inner: tokio::sync::RwLock<SlotInner>,
     usage_count: AtomicU8,
 }
 
@@ -203,6 +200,11 @@ impl Slot {
             Err(usage_count) => usage_count,
         }
     }
+
+    /// Sets the usage count to a specific value.
+    fn set_usage_count(&self, count: u8) {
+        self.usage_count.store(count, Ordering::Relaxed);
+    }
 }
 
 pub struct PageCache {
@@ -215,9 +217,9 @@ pub struct PageCache {
     ///
     /// If you add support for caching different kinds of objects, each object kind
     /// can have a separate mapping map, next to this field.
-    materialized_page_map: RwLock<HashMap<MaterializedPageHashKey, Vec<Version>>>,
+    materialized_page_map: std::sync::RwLock<HashMap<MaterializedPageHashKey, Vec<Version>>>,
 
-    immutable_page_map: RwLock<HashMap<(FileId, u32), usize>>,
+    immutable_page_map: std::sync::RwLock<HashMap<(FileId, u32), usize>>,
 
     /// The actual buffers with their metadata.
     slots: Box<[Slot]>,
@@ -233,7 +235,7 @@ pub struct PageCache {
 /// PageReadGuard is a "lease" on a buffer, for reading. The page is kept locked
 /// until the guard is dropped.
 ///
-pub struct PageReadGuard<'i>(RwLockReadGuard<'i, SlotInner>);
+pub struct PageReadGuard<'i>(tokio::sync::RwLockReadGuard<'i, SlotInner>);
 
 impl std::ops::Deref for PageReadGuard<'_> {
     type Target = [u8; PAGE_SZ];
@@ -260,9 +262,10 @@ impl AsRef<[u8; PAGE_SZ]> for PageReadGuard<'_> {
 /// to initialize.
 ///
 pub struct PageWriteGuard<'i> {
-    inner: RwLockWriteGuard<'i, SlotInner>,
+    inner: tokio::sync::RwLockWriteGuard<'i, SlotInner>,
 
     // Are the page contents currently valid?
+    // Used to mark pages as invalid that are assigned but not yet filled with data.
     valid: bool,
 }
 
@@ -337,7 +340,7 @@ impl PageCache {
     /// The 'lsn' is an upper bound, this will return the latest version of
     /// the given block, but not newer than 'lsn'. Returns the actual LSN of the
     /// returned page.
-    pub fn lookup_materialized_page(
+    pub async fn lookup_materialized_page(
         &self,
         tenant_id: TenantId,
         timeline_id: TimelineId,
@@ -357,7 +360,7 @@ impl PageCache {
             lsn,
         };
 
-        if let Some(guard) = self.try_lock_for_read(&mut cache_key) {
+        if let Some(guard) = self.try_lock_for_read(&mut cache_key).await {
             if let CacheKey::MaterializedPage {
                 hash_key: _,
                 lsn: available_lsn,
@@ -384,7 +387,7 @@ impl PageCache {
     ///
     /// Store an image of the given page in the cache.
     ///
-    pub fn memorize_materialized_page(
+    pub async fn memorize_materialized_page(
         &self,
         tenant_id: TenantId,
         timeline_id: TimelineId,
@@ -401,7 +404,7 @@ impl PageCache {
             lsn,
         };
 
-        match self.lock_for_write(&cache_key)? {
+        match self.lock_for_write(&cache_key).await? {
             WriteBufResult::Found(write_guard) => {
                 // We already had it in cache. Another thread must've put it there
                 // concurrently. Check that it had the same contents that we
@@ -419,31 +422,14 @@ impl PageCache {
 
     // Section 1.2: Public interface functions for working with immutable file pages.
 
-    pub fn read_immutable_buf(&self, file_id: FileId, blkno: u32) -> anyhow::Result<ReadBufResult> {
+    pub async fn read_immutable_buf(
+        &self,
+        file_id: FileId,
+        blkno: u32,
+    ) -> anyhow::Result<ReadBufResult> {
         let mut cache_key = CacheKey::ImmutableFilePage { file_id, blkno };
 
-        self.lock_for_read(&mut cache_key)
-    }
-
-    /// Immediately drop all buffers belonging to given file
-    pub fn drop_buffers_for_immutable(&self, drop_file_id: FileId) {
-        for slot_idx in 0..self.slots.len() {
-            let slot = &self.slots[slot_idx];
-
-            let mut inner = slot.inner.write().unwrap();
-            if let Some(key) = &inner.key {
-                match key {
-                    CacheKey::ImmutableFilePage { file_id, blkno: _ }
-                        if *file_id == drop_file_id =>
-                    {
-                        // remove mapping for old buffer
-                        self.remove_mapping(key);
-                        inner.key = None;
-                    }
-                    _ => {}
-                }
-            }
-        }
+        self.lock_for_read(&mut cache_key).await
     }
 
     //
@@ -463,14 +449,14 @@ impl PageCache {
     ///
     /// If no page is found, returns None and *cache_key is left unmodified.
     ///
-    fn try_lock_for_read(&self, cache_key: &mut CacheKey) -> Option<PageReadGuard> {
+    async fn try_lock_for_read(&self, cache_key: &mut CacheKey) -> Option<PageReadGuard> {
         let cache_key_orig = cache_key.clone();
         if let Some(slot_idx) = self.search_mapping(cache_key) {
             // The page was found in the mapping. Lock the slot, and re-check
             // that it's still what we expected (because we released the mapping
             // lock already, another thread could have evicted the page)
             let slot = &self.slots[slot_idx];
-            let inner = slot.inner.read().unwrap();
+            let inner = slot.inner.read().await;
             if inner.key.as_ref() == Some(cache_key) {
                 slot.inc_usage_count();
                 return Some(PageReadGuard(inner));
@@ -511,7 +497,7 @@ impl PageCache {
     /// }
     /// ```
     ///
-    fn lock_for_read(&self, cache_key: &mut CacheKey) -> anyhow::Result<ReadBufResult> {
+    async fn lock_for_read(&self, cache_key: &mut CacheKey) -> anyhow::Result<ReadBufResult> {
         let (read_access, hit) = match cache_key {
             CacheKey::MaterializedPage { .. } => {
                 unreachable!("Materialized pages use lookup_materialized_page")
@@ -526,7 +512,7 @@ impl PageCache {
         let mut is_first_iteration = true;
         loop {
             // First check if the key already exists in the cache.
-            if let Some(read_guard) = self.try_lock_for_read(cache_key) {
+            if let Some(read_guard) = self.try_lock_for_read(cache_key).await {
                 if is_first_iteration {
                     hit.inc();
                 }
@@ -556,7 +542,7 @@ impl PageCache {
             // Make the slot ready
             let slot = &self.slots[slot_idx];
             inner.key = Some(cache_key.clone());
-            slot.usage_count.store(1, Ordering::Relaxed);
+            slot.set_usage_count(1);
 
             return Ok(ReadBufResult::NotFound(PageWriteGuard {
                 inner,
@@ -569,13 +555,13 @@ impl PageCache {
     /// found, returns None.
     ///
     /// When locking a page for writing, the search criteria is always "exact".
-    fn try_lock_for_write(&self, cache_key: &CacheKey) -> Option<PageWriteGuard> {
+    async fn try_lock_for_write(&self, cache_key: &CacheKey) -> Option<PageWriteGuard> {
         if let Some(slot_idx) = self.search_mapping_for_write(cache_key) {
             // The page was found in the mapping. Lock the slot, and re-check
             // that it's still what we expected (because we don't released the mapping
             // lock already, another thread could have evicted the page)
             let slot = &self.slots[slot_idx];
-            let inner = slot.inner.write().unwrap();
+            let inner = slot.inner.write().await;
             if inner.key.as_ref() == Some(cache_key) {
                 slot.inc_usage_count();
                 return Some(PageWriteGuard { inner, valid: true });
@@ -588,10 +574,10 @@ impl PageCache {
     ///
     /// Similar to lock_for_read(), but the returned buffer is write-locked and
     /// may be modified by the caller even if it's already found in the cache.
-    fn lock_for_write(&self, cache_key: &CacheKey) -> anyhow::Result<WriteBufResult> {
+    async fn lock_for_write(&self, cache_key: &CacheKey) -> anyhow::Result<WriteBufResult> {
         loop {
             // First check if the key already exists in the cache.
-            if let Some(write_guard) = self.try_lock_for_write(cache_key) {
+            if let Some(write_guard) = self.try_lock_for_write(cache_key).await {
                 return Ok(WriteBufResult::Found(write_guard));
             }
 
@@ -617,7 +603,7 @@ impl PageCache {
             // Make the slot ready
             let slot = &self.slots[slot_idx];
             inner.key = Some(cache_key.clone());
-            slot.usage_count.store(1, Ordering::Relaxed);
+            slot.set_usage_count(1);
 
             return Ok(WriteBufResult::NotFound(PageWriteGuard {
                 inner,
@@ -772,7 +758,7 @@ impl PageCache {
     /// Find a slot to evict.
     ///
     /// On return, the slot is empty and write-locked.
-    fn find_victim(&self) -> anyhow::Result<(usize, RwLockWriteGuard<SlotInner>)> {
+    fn find_victim(&self) -> anyhow::Result<(usize, tokio::sync::RwLockWriteGuard<SlotInner>)> {
         let iter_limit = self.slots.len() * 10;
         let mut iters = 0;
         loop {
@@ -784,10 +770,7 @@ impl PageCache {
             if slot.dec_usage_count() == 0 {
                 let mut inner = match slot.inner.try_write() {
                     Ok(inner) => inner,
-                    Err(TryLockError::Poisoned(err)) => {
-                        anyhow::bail!("buffer lock was poisoned: {err:?}")
-                    }
-                    Err(TryLockError::WouldBlock) => {
+                    Err(_err) => {
                         // If we have looped through the whole buffer pool 10 times
                         // and still haven't found a victim buffer, something's wrong.
                         // Maybe all the buffers were in locked. That could happen in
@@ -816,6 +799,8 @@ impl PageCache {
     fn new(num_pages: usize) -> Self {
         assert!(num_pages > 0, "page cache size must be > 0");
 
+        // We use Box::leak here and into_boxed_slice to avoid leaking uninitialized
+        // memory that Vec's might contain.
         let page_buffer = Box::leak(vec![0u8; num_pages * PAGE_SZ].into_boxed_slice());
 
         let size_metrics = &crate::metrics::PAGE_CACHE_SIZE;
@@ -829,7 +814,7 @@ impl PageCache {
                 let buf: &mut [u8; PAGE_SZ] = chunk.try_into().unwrap();
 
                 Slot {
-                    inner: RwLock::new(SlotInner { key: None, buf }),
+                    inner: tokio::sync::RwLock::new(SlotInner { key: None, buf }),
                     usage_count: AtomicU8::new(0),
                 }
             })

pageserver/src/tenant.rs

@@ -59,6 +59,7 @@ 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;
@@ -85,6 +86,7 @@ pub use pageserver_api::models::TenantState;
 use toml_edit;
 use utils::{
     crashsafe,
+    generation::Generation,
     id::{TenantId, TimelineId},
     lsn::{Lsn, RecordLsn},
 };
@@ -119,7 +121,7 @@ mod span;
 
 pub mod metadata;
 mod par_fsync;
-mod remote_timeline_client;
+pub mod remote_timeline_client;
 pub mod storage_layer;
 
 pub mod config;
@@ -156,6 +158,7 @@ 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,
 }
 
 ///
@@ -178,6 +181,10 @@ pub struct Tenant {
     tenant_conf: Arc<RwLock<TenantConfOpt>>,
 
     tenant_id: TenantId,
+
+    // The remote storage generation, used to protect S3 objects from split-brain
+    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
@@ -191,6 +198,9 @@ 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>,
@@ -522,9 +532,11 @@ 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
@@ -538,7 +550,9 @@ 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
@@ -648,12 +662,8 @@ 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());
 
@@ -665,6 +675,7 @@ impl Tenant {
                 self.conf,
                 self.tenant_id,
                 timeline_id,
+                self.generation,
             );
             part_downloads.spawn(
                 async move {
@@ -698,10 +709,7 @@ impl Tenant {
             debug!("successfully downloaded index part for timeline {timeline_id}");
             match index_part {
                 MaybeDeletedIndexPart::IndexPart(index_part) => {
-                    timeline_ancestors.insert(
-                        timeline_id,
-                        index_part.parse_metadata().context("parse_metadata")?,
-                    );
+                    timeline_ancestors.insert(timeline_id, index_part.metadata.clone());
                     remote_index_and_client.insert(timeline_id, (index_part, client));
                 }
                 MaybeDeletedIndexPart::Deleted(index_part) => {
@@ -730,6 +738,7 @@ impl Tenant {
                 remote_metadata,
                 TimelineResources {
                     remote_client: Some(remote_client),
+                    deletion_queue_client: self.deletion_queue_client.clone(),
                 },
                 ctx,
             )
@@ -752,8 +761,9 @@ impl Tenant {
             DeleteTimelineFlow::resume_deletion(
                 Arc::clone(self),
                 timeline_id,
-                &index_part.parse_metadata().context("parse_metadata")?,
+                &index_part.metadata,
                 Some(remote_timeline_client),
+                self.deletion_queue_client.clone(),
                 None,
             )
             .await
@@ -854,6 +864,8 @@ impl Tenant {
             TenantConfOpt::default(),
             wal_redo_manager,
             tenant_id,
+            Generation::broken(),
+            None,
             None,
         ))
     }
@@ -871,6 +883,7 @@ 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>,
@@ -888,6 +901,7 @@ 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(
@@ -896,7 +910,9 @@ impl Tenant {
             tenant_conf,
             wal_redo_manager,
             tenant_id,
+            generation,
             remote_storage.clone(),
+            Some(deletion_queue_client),
         );
         let tenant = Arc::new(tenant);
 
@@ -1304,6 +1320,7 @@ impl Tenant {
                                 timeline_id,
                                 &local_metadata,
                                 Some(remote_client),
+                                self.deletion_queue_client.clone(),
                                 init_order,
                             )
                             .await
@@ -1314,10 +1331,7 @@ impl Tenant {
                         }
                     };
 
-                    let remote_metadata = index_part
-                        .parse_metadata()
-                        .context("parse_metadata")
-                        .map_err(LoadLocalTimelineError::Load)?;
+                    let remote_metadata = index_part.metadata.clone();
                     (
                         Some(RemoteStartupData {
                             index_part,
@@ -1356,6 +1370,7 @@ impl Tenant {
                         timeline_id,
                         &local_metadata,
                         None,
+                        None,
                         init_order,
                     )
                     .await
@@ -2280,6 +2295,7 @@ impl Tenant {
             ancestor,
             new_timeline_id,
             self.tenant_id,
+            self.generation,
             Arc::clone(&self.walredo_mgr),
             resources,
             pg_version,
@@ -2297,8 +2313,18 @@ 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 {
@@ -2355,6 +2381,7 @@ impl Tenant {
 
         Tenant {
             tenant_id,
+            generation,
             conf,
             // using now here is good enough approximation to catch tenants with really long
             // activation times.
@@ -2364,6 +2391,7 @@ 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)),
@@ -2937,13 +2965,17 @@ impl Tenant {
                 self.conf,
                 self.tenant_id,
                 timeline_id,
+                self.generation,
             );
             Some(remote_client)
         } else {
             None
         };
 
-        TimelineResources { remote_client }
+        TimelineResources {
+            remote_client,
+            deletion_queue_client: self.deletion_queue_client.clone(),
+        }
     }
 
     /// Creates intermediate timeline structure and its files.
@@ -3460,6 +3492,7 @@ pub mod harness {
         pub conf: &'static PageServerConf,
         pub tenant_conf: TenantConf,
         pub tenant_id: TenantId,
+        pub generation: Generation,
     }
 
     static LOG_HANDLE: OnceCell<()> = OnceCell::new();
@@ -3501,13 +3534,14 @@ 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)
+                self.try_load(&ctx, None, None)
                     .await
                     .expect("failed to load test tenant"),
                 ctx,
@@ -3518,6 +3552,7 @@ 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);
 
@@ -3527,7 +3562,9 @@ pub mod harness {
                 TenantConfOpt::from(self.tenant_conf),
                 walredo_mgr,
                 self.tenant_id,
+                self.generation,
                 remote_storage,
+                deletion_queue_client,
             ));
             tenant
                 .load(None, ctx)
@@ -4092,7 +4129,7 @@ mod tests {
         std::fs::write(metadata_path, metadata_bytes)?;
 
         let err = harness
-            .try_load(&ctx, None)
+            .try_load(&ctx, None, None)
             .await
             .err()
             .expect("should fail");
@@ -4107,7 +4144,7 @@ mod tests {
         let mut found_error_message = false;
         let mut err_source = err.source();
         while let Some(source) = err_source {
-            if source.to_string() == "metadata checksum mismatch" {
+            if source.to_string().contains("metadata checksum mismatch") {
                 found_error_message = true;
                 break;
             }

pageserver/src/tenant/blob_io.rs

@@ -33,7 +33,7 @@ impl<'a> BlockCursor<'a> {
         let mut blknum = (offset / PAGE_SZ as u64) as u32;
         let mut off = (offset % PAGE_SZ as u64) as usize;
 
-        let mut buf = self.read_blk(blknum)?;
+        let mut buf = self.read_blk(blknum).await?;
 
         // peek at the first byte, to determine if it's a 1- or 4-byte length
         let first_len_byte = buf[off];
@@ -49,7 +49,7 @@ impl<'a> BlockCursor<'a> {
                 // it is split across two pages
                 len_buf[..thislen].copy_from_slice(&buf[off..PAGE_SZ]);
                 blknum += 1;
-                buf = self.read_blk(blknum)?;
+                buf = self.read_blk(blknum).await?;
                 len_buf[thislen..].copy_from_slice(&buf[0..4 - thislen]);
                 off = 4 - thislen;
             } else {
@@ -70,7 +70,7 @@ impl<'a> BlockCursor<'a> {
             if page_remain == 0 {
                 // continue on next page
                 blknum += 1;
-                buf = self.read_blk(blknum)?;
+                buf = self.read_blk(blknum).await?;
                 off = 0;
                 page_remain = PAGE_SZ;
             }

pageserver/src/tenant/block_io.rs

@@ -39,7 +39,7 @@ pub enum BlockLease<'a> {
     PageReadGuard(PageReadGuard<'static>),
     EphemeralFileMutableTail(&'a [u8; PAGE_SZ]),
     #[cfg(test)]
-    Rc(std::rc::Rc<[u8; PAGE_SZ]>),
+    Arc(std::sync::Arc<[u8; PAGE_SZ]>),
 }
 
 impl From<PageReadGuard<'static>> for BlockLease<'static> {
@@ -49,9 +49,9 @@ impl From<PageReadGuard<'static>> for BlockLease<'static> {
 }
 
 #[cfg(test)]
-impl<'a> From<std::rc::Rc<[u8; PAGE_SZ]>> for BlockLease<'a> {
-    fn from(value: std::rc::Rc<[u8; PAGE_SZ]>) -> Self {
-        BlockLease::Rc(value)
+impl<'a> From<std::sync::Arc<[u8; PAGE_SZ]>> for BlockLease<'a> {
+    fn from(value: std::sync::Arc<[u8; PAGE_SZ]>) -> Self {
+        BlockLease::Arc(value)
     }
 }
 
@@ -63,7 +63,7 @@ impl<'a> Deref for BlockLease<'a> {
             BlockLease::PageReadGuard(v) => v.deref(),
             BlockLease::EphemeralFileMutableTail(v) => v,
             #[cfg(test)]
-            BlockLease::Rc(v) => v.deref(),
+            BlockLease::Arc(v) => v.deref(),
         }
     }
 }
@@ -83,13 +83,13 @@ pub(crate) enum BlockReaderRef<'a> {
 
 impl<'a> BlockReaderRef<'a> {
     #[inline(always)]
-    fn read_blk(&self, blknum: u32) -> Result<BlockLease, std::io::Error> {
+    async fn read_blk(&self, blknum: u32) -> Result<BlockLease, std::io::Error> {
         use BlockReaderRef::*;
         match self {
-            FileBlockReaderVirtual(r) => r.read_blk(blknum),
-            FileBlockReaderFile(r) => r.read_blk(blknum),
-            EphemeralFile(r) => r.read_blk(blknum),
-            Adapter(r) => r.read_blk(blknum),
+            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)]
             TestDisk(r) => r.read_blk(blknum),
         }
@@ -134,8 +134,8 @@ impl<'a> BlockCursor<'a> {
     /// access to the contents of the page. (For the page cache, the
     /// lease object represents a lock on the buffer.)
     #[inline(always)]
-    pub fn read_blk(&self, blknum: u32) -> Result<BlockLease, std::io::Error> {
-        self.reader.read_blk(blknum)
+    pub async fn read_blk(&self, blknum: u32) -> Result<BlockLease, std::io::Error> {
+        self.reader.read_blk(blknum).await
     }
 }
 
@@ -170,11 +170,12 @@ where
     /// Returns a "lease" object that can be used to
     /// access to the contents of the page. (For the page cache, the
     /// lease object represents a lock on the buffer.)
-    pub fn read_blk(&self, blknum: u32) -> Result<BlockLease, std::io::Error> {
+    pub async fn read_blk(&self, blknum: u32) -> Result<BlockLease, std::io::Error> {
         let cache = page_cache::get();
         loop {
             match cache
                 .read_immutable_buf(self.file_id, blknum)
+                .await
                 .map_err(|e| {
                     std::io::Error::new(
                         std::io::ErrorKind::Other,

pageserver/src/tenant/disk_btree.rs

@@ -262,7 +262,7 @@ where
         let block_cursor = self.reader.block_cursor();
         while let Some((node_blknum, opt_iter)) = stack.pop() {
             // Locate the node.
-            let node_buf = block_cursor.read_blk(self.start_blk + node_blknum)?;
+            let node_buf = block_cursor.read_blk(self.start_blk + node_blknum).await?;
 
             let node = OnDiskNode::deparse(node_buf.as_ref())?;
             let prefix_len = node.prefix_len as usize;
@@ -357,7 +357,7 @@ where
         let block_cursor = self.reader.block_cursor();
 
         while let Some((blknum, path, depth, child_idx, key_off)) = stack.pop() {
-            let blk = block_cursor.read_blk(self.start_blk + blknum)?;
+            let blk = block_cursor.read_blk(self.start_blk + blknum).await?;
             let buf: &[u8] = blk.as_ref();
             let node = OnDiskNode::<L>::deparse(buf)?;
 
@@ -704,7 +704,7 @@ pub(crate) mod tests {
         pub(crate) fn read_blk(&self, blknum: u32) -> io::Result<BlockLease> {
             let mut buf = [0u8; PAGE_SZ];
             buf.copy_from_slice(&self.blocks[blknum as usize]);
-            Ok(std::rc::Rc::new(buf).into())
+            Ok(std::sync::Arc::new(buf).into())
         }
     }
     impl BlockReader for TestDisk {

pageserver/src/tenant/ephemeral_file.rs

@@ -61,13 +61,14 @@ impl EphemeralFile {
         self.len
     }
 
-    pub(crate) fn read_blk(&self, blknum: u32) -> Result<BlockLease, io::Error> {
+    pub(crate) async fn read_blk(&self, blknum: u32) -> Result<BlockLease, io::Error> {
         let flushed_blknums = 0..self.len / PAGE_SZ as u64;
         if flushed_blknums.contains(&(blknum as u64)) {
             let cache = page_cache::get();
             loop {
                 match cache
                     .read_immutable_buf(self.page_cache_file_id, blknum)
+                    .await
                     .map_err(|e| {
                         std::io::Error::new(
                             std::io::ErrorKind::Other,
@@ -135,10 +136,13 @@ impl EphemeralFile {
                                 // 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.
                                 let cache = page_cache::get();
-                                match cache.read_immutable_buf(
-                                    self.ephemeral_file.page_cache_file_id,
-                                    self.blknum,
-                                ) {
+                                match cache
+                                    .read_immutable_buf(
+                                        self.ephemeral_file.page_cache_file_id,
+                                        self.blknum,
+                                    )
+                                    .await
+                                {
                                     Ok(page_cache::ReadBufResult::Found(_guard)) => {
                                         // This function takes &mut self, so, it shouldn't be possible to reach this point.
                                         unreachable!("we just wrote blknum {} and this function takes &mut self, so, no concurrent read_blk is possible", self.blknum);
@@ -221,9 +225,8 @@ pub fn is_ephemeral_file(filename: &str) -> bool {
 
 impl Drop for EphemeralFile {
     fn drop(&mut self) {
-        // drop all pages from page cache
-        let cache = page_cache::get();
-        cache.drop_buffers_for_immutable(self.page_cache_file_id);
+        // There might still be pages in the [`crate::page_cache`] for this file.
+        // We leave them there, [`crate::page_cache::PageCache::find_victim`] will evict them when needed.
 
         // unlink the file
         let res = std::fs::remove_file(&self.file.path);

pageserver/src/tenant/metadata.rs

@@ -12,7 +12,7 @@ use std::fs::{File, OpenOptions};
 use std::io::{self, Write};
 
 use anyhow::{bail, ensure, Context};
-use serde::{Deserialize, Serialize};
+use serde::{de::Error, Deserialize, Serialize, Serializer};
 use thiserror::Error;
 use tracing::info_span;
 use utils::bin_ser::SerializeError;
@@ -232,6 +232,28 @@ impl TimelineMetadata {
     }
 }
 
+impl<'de> Deserialize<'de> for TimelineMetadata {
+    fn deserialize<D>(deserializer: D) -> Result<Self, D::Error>
+    where
+        D: serde::Deserializer<'de>,
+    {
+        let bytes = Vec::<u8>::deserialize(deserializer)?;
+        Self::from_bytes(bytes.as_slice()).map_err(|e| D::Error::custom(format!("{e}")))
+    }
+}
+
+impl Serialize for TimelineMetadata {
+    fn serialize<S>(&self, serializer: S) -> Result<S::Ok, S::Error>
+    where
+        S: Serializer,
+    {
+        let bytes = self
+            .to_bytes()
+            .map_err(|e| serde::ser::Error::custom(format!("{e}")))?;
+        bytes.serialize(serializer)
+    }
+}
+
 /// Save timeline metadata to file
 pub fn save_metadata(
     conf: &'static PageServerConf,

pageserver/src/tenant/mgr.rs

@@ -1,10 +1,13 @@
 //! This module acts as a switchboard to access different repositories managed by this
 //! page server.
 
+use hyper::StatusCode;
+use pageserver_api::control_api::{HexTenantId, ReAttachRequest, ReAttachResponse};
 use std::collections::{hash_map, HashMap};
 use std::ffi::OsStr;
 use std::path::Path;
 use std::sync::Arc;
+use std::time::Duration;
 use tokio::fs;
 
 use anyhow::Context;
@@ -18,6 +21,7 @@ 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;
@@ -25,6 +29,7 @@ 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;
@@ -75,6 +80,78 @@ pub async fn init_tenant_mgr(
 
     let mut tenants = HashMap::new();
 
+    // If we are configured to use the control plane API, then it is the source of truth for what to attach
+    let tenant_generations = conf
+        .control_plane_api
+        .as_ref()
+        .map(|control_plane_api| async {
+            let client = reqwest::ClientBuilder::new()
+                .build()
+                .expect("Failed to construct http client");
+
+            // FIXME: it's awkward that join() requires the base to have a trailing slash, makes
+            // it easy to get a config wrong
+            assert!(
+                control_plane_api.as_str().ends_with("/"),
+                "control plane API needs trailing slash"
+            );
+
+            let re_attach_path = control_plane_api
+                .join("re-attach")
+                .expect("Failed to build re-attach path");
+            let request = ReAttachRequest { node_id: conf.id };
+
+            // TODO: we should have been passed a cancellation token, and use it to end
+            // this loop gracefully
+            loop {
+                let response = match client
+                    .post(re_attach_path.clone())
+                    .json(&request)
+                    .send()
+                    .await
+                {
+                    Err(e) => Err(anyhow::Error::from(e)),
+                    Ok(r) => {
+                        if r.status() == StatusCode::OK {
+                            r.json::<ReAttachResponse>()
+                                .await
+                                .map_err(|e| anyhow::Error::from(e))
+                        } else {
+                            Err(anyhow::anyhow!("Unexpected status {}", r.status()))
+                        }
+                    }
+                };
+
+                match response {
+                    Ok(res) => {
+                        tracing::info!(
+                            "Received re-attach response with {0} tenants",
+                            res.tenants.len()
+                        );
+
+                        // TODO: do something with it
+                        break res
+                            .tenants
+                            .into_iter()
+                            .map(|t| (t.id, t.generation))
+                            .collect::<HashMap<_, _>>();
+                    }
+                    Err(e) => {
+                        tracing::error!("Error re-attaching tenants, retrying: {e:#}");
+                        tokio::time::sleep(Duration::from_secs(1)).await;
+                    }
+                }
+            }
+        });
+
+    let tenant_generations = match tenant_generations {
+        Some(g) => Some(g.await),
+        None => {
+            info!("Control plane API not configured, tenant generations are disabled");
+            None
+        }
+    };
+
     let mut dir_entries = fs::read_dir(&tenants_dir)
         .await
         .with_context(|| format!("Failed to list tenants dir {tenants_dir:?}"))?;
@@ -122,9 +199,53 @@ pub async fn init_tenant_mgr(
                         continue;
                     }
 
+                    let tenant_id = match tenant_dir_path
+                        .file_name()
+                        .and_then(OsStr::to_str)
+                        .unwrap_or_default()
+                        .parse::<TenantId>()
+                    {
+                        Ok(id) => id,
+                        Err(_) => {
+                            warn!(
+                                "Invalid tenant path (garbage in our repo directory?): {0}",
+                                tenant_dir_path.display()
+                            );
+                            continue;
+                        }
+                    };
+
+                    let generation = if let Some(generations) = &tenant_generations {
+                        // We have a generation map: treat it as the authority for whether
+                        // this tenant is really attached.
+                        if let Some(gen) = generations.get(&HexTenantId::new(tenant_id)) {
+                            Generation::new(*gen)
+                        } else {
+                            info!("Detaching tenant {0}, control plane omitted it in re-attach response", tenant_id);
+                            if let Err(e) = fs::remove_dir_all(&tenant_dir_path).await {
+                                error!(
+                                    "Failed to remove detached tenant directory '{}': {:?}",
+                                    tenant_dir_path.display(),
+                                    e
+                                );
+                            }
+                            continue;
+                        }
+                    } else {
+                        // Legacy mode: no generation information, any tenant present
+                        // on local disk may activate
+                        info!(
+                            "Starting tenant {0} in legacy mode, no generation",
+                            tenant_dir_path.display()
+                        );
+                        Generation::none()
+                    };
+
                     match schedule_local_tenant_processing(
                         conf,
+                        tenant_id,
                         &tenant_dir_path,
+                        generation,
                         resources.clone(),
                         Some(init_order.clone()),
                         &TENANTS,
@@ -160,7 +281,9 @@ pub async fn init_tenant_mgr(
 
 pub(crate) fn schedule_local_tenant_processing(
     conf: &'static PageServerConf,
+    tenant_id: TenantId,
     tenant_path: &Path,
+    generation: Generation,
     resources: TenantSharedResources,
     init_order: Option<InitializationOrder>,
     tenants: &'static tokio::sync::RwLock<TenantsMap>,
@@ -181,15 +304,6 @@ pub(crate) fn schedule_local_tenant_processing(
         "Cannot load tenant from empty directory {tenant_path:?}"
     );
 
-    let tenant_id = tenant_path
-        .file_name()
-        .and_then(OsStr::to_str)
-        .unwrap_or_default()
-        .parse::<TenantId>()
-        .with_context(|| {
-            format!("Could not parse tenant id out of the tenant dir name in path {tenant_path:?}")
-        })?;
-
     let tenant_ignore_mark = conf.tenant_ignore_mark_file_path(&tenant_id);
     anyhow::ensure!(
         !conf.tenant_ignore_mark_file_path(&tenant_id).exists(),
@@ -202,9 +316,11 @@ pub(crate) fn schedule_local_tenant_processing(
             match Tenant::spawn_attach(
                 conf,
                 tenant_id,
+                generation,
                 resources.broker_client,
                 tenants,
                 remote_storage,
+                resources.deletion_queue_client,
                 ctx,
             ) {
                 Ok(tenant) => tenant,
@@ -224,7 +340,9 @@ 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, resources, init_order, tenants, ctx,
+        )
     };
     Ok(tenant)
 }
@@ -347,8 +465,10 @@ pub async fn create_tenant(
     conf: &'static PageServerConf,
     tenant_conf: TenantConfOpt,
     tenant_id: TenantId,
+    generation: Generation,
     broker_client: storage_broker::BrokerClientChannel,
     remote_storage: Option<GenericRemoteStorage>,
+    deletion_queue: &DeletionQueue,
     ctx: &RequestContext,
 ) -> Result<Arc<Tenant>, TenantMapInsertError> {
     tenant_map_insert(tenant_id, || {
@@ -362,9 +482,11 @@ 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)?;
+            schedule_local_tenant_processing(conf, tenant_id, &tenant_directory,
+                generation, tenant_resources, None, &TENANTS, ctx)?;
         // TODO: tenant object & its background loops remain, untracked in tenant map, if we fail here.
         //      See https://github.com/neondatabase/neon/issues/4233
 
@@ -513,6 +635,7 @@ 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, || {
@@ -526,8 +649,11 @@ 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)
+        // TODO: remove the `/load` API once generation support is complete:
+        // it becomes equivalent to attaching.
+        let new_tenant = schedule_local_tenant_processing(conf, tenant_id, &tenant_path, Generation::none(), resources, None,  &TENANTS, ctx)
             .with_context(|| {
                 format!("Failed to schedule tenant processing in path {tenant_path:?}")
             })?;
@@ -591,9 +717,11 @@ pub async fn list_tenants() -> Result<Vec<(TenantId, TenantState)>, TenantMapLis
 pub async fn attach_tenant(
     conf: &'static PageServerConf,
     tenant_id: TenantId,
+    generation: Generation,
     tenant_conf: TenantConfOpt,
     broker_client: storage_broker::BrokerClientChannel,
     remote_storage: GenericRemoteStorage,
+    deletion_queue: &DeletionQueue,
     ctx: &RequestContext,
 ) -> Result<(), TenantMapInsertError> {
     tenant_map_insert(tenant_id, || {
@@ -611,8 +739,9 @@ 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)?;
+        let attached_tenant = schedule_local_tenant_processing(conf, tenant_id, &tenant_dir, generation, resources, None, &TENANTS, ctx)?;
         // TODO: tenant object & its background loops remain, untracked in tenant map, if we fail here.
         //      See https://github.com/neondatabase/neon/issues/4233
 

pageserver/src/tenant/remote_timeline_client.rs

@@ -56,9 +56,11 @@
 //! # 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.
+//! 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.
 //!
 //! 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
@@ -70,10 +72,12 @@
 //! 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.)
+//! 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.
 //!
 //! All of this relies on the following invariants:
 //!
@@ -200,12 +204,11 @@
 //! [`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::Context;
+use anyhow::{bail, Context};
 use chrono::{NaiveDateTime, Utc};
 // re-export these
 pub use download::{is_temp_download_file, list_remote_timelines};
@@ -216,7 +219,7 @@ 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};
 
@@ -226,6 +229,7 @@ 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,
@@ -234,7 +238,6 @@ 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::{
     config::PageServerConf,
     task_mgr,
@@ -244,6 +247,7 @@ use crate::{
     tenant::upload_queue::{
         UploadOp, UploadQueue, UploadQueueInitialized, UploadQueueStopped, UploadTask,
     },
+    tenant::TIMELINES_SEGMENT_NAME,
 };
 
 use utils::id::{TenantId, TimelineId};
@@ -252,6 +256,7 @@ 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.
@@ -315,6 +320,7 @@ pub struct RemoteTimelineClient {
 
     tenant_id: TenantId,
     timeline_id: TimelineId,
+    generation: Generation,
 
     upload_queue: Mutex<UploadQueue>,
 
@@ -335,12 +341,14 @@ 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)),
@@ -453,6 +461,7 @@ impl RemoteTimelineClient {
             &self.storage_impl,
             &self.tenant_id,
             &self.timeline_id,
+            self.generation,
         )
         .measure_remote_op(
             self.tenant_id,
@@ -541,8 +550,7 @@ impl RemoteTimelineClient {
         // ahead of what's _actually_ on the remote during index upload.
         upload_queue.latest_metadata = metadata.clone();
 
-        let metadata_bytes = upload_queue.latest_metadata.to_bytes()?;
-        self.schedule_index_upload(upload_queue, metadata_bytes);
+        self.schedule_index_upload(upload_queue, upload_queue.latest_metadata.clone());
 
         Ok(())
     }
@@ -562,8 +570,7 @@ impl RemoteTimelineClient {
         let upload_queue = guard.initialized_mut()?;
 
         if upload_queue.latest_files_changes_since_metadata_upload_scheduled > 0 {
-            let metadata_bytes = upload_queue.latest_metadata.to_bytes()?;
-            self.schedule_index_upload(upload_queue, metadata_bytes);
+            self.schedule_index_upload(upload_queue, upload_queue.latest_metadata.clone());
         }
 
         Ok(())
@@ -573,7 +580,7 @@ impl RemoteTimelineClient {
     fn schedule_index_upload(
         self: &Arc<Self>,
         upload_queue: &mut UploadQueueInitialized,
-        metadata_bytes: Vec<u8>,
+        metadata: TimelineMetadata,
     ) {
         info!(
             "scheduling metadata upload with {} files ({} changed)",
@@ -586,7 +593,7 @@ impl RemoteTimelineClient {
         let index_part = IndexPart::new(
             upload_queue.latest_files.clone(),
             disk_consistent_lsn,
-            metadata_bytes,
+            metadata,
         );
         let op = UploadOp::UploadMetadata(index_part, disk_consistent_lsn);
         self.calls_unfinished_metric_begin(&op);
@@ -633,50 +640,66 @@ impl RemoteTimelineClient {
     /// deletion won't actually be performed, until any previously scheduled
     /// upload operations, and the index file upload, have completed
     /// successfully.
-    pub fn schedule_layer_file_deletion(
+    pub async fn schedule_layer_file_deletion(
         self: &Arc<Self>,
         names: &[LayerFileName],
+        deletion_queue_client: &DeletionQueueClient,
     ) -> anyhow::Result<()> {
-        let mut guard = self.upload_queue.lock().unwrap();
-        let upload_queue = guard.initialized_mut()?;
+        // Synchronous update of upload queues under mutex
+        let with_generations = {
+            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_bytes = upload_queue.latest_metadata.to_bytes()?;
+            // 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.
-        //
-        // 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 = || {
-            for name in names {
-                upload_queue.latest_files.remove(name);
-                upload_queue.latest_files_changes_since_metadata_upload_scheduled += 1;
-            }
+            // Decorate our list of names with each name's generation, dropping
+            // makes that are unexpectedly missing from our metadata.
+            let with_generations: Vec<_> = names
+                .into_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.clone(), meta.generation))
+                    } else {
+                        // This is unexpected: latest_files is meant to be kept up to
+                        // date.  We can't delete the layer if we have forgotten what
+                        // generation it was in.
+                        warn!("Deleting layer {name} not found in latest_files list");
+                        None
+                    }
+                })
+                .collect();
 
             if upload_queue.latest_files_changes_since_metadata_upload_scheduled > 0 {
-                self.schedule_index_upload(upload_queue, metadata_bytes);
+                self.schedule_index_upload(upload_queue, metadata);
             }
 
-            // schedule the actual deletions
-            for name in names {
-                let op = UploadOp::Delete(Delete {
-                    file_kind: RemoteOpFileKind::Layer,
-                    layer_file_name: name.clone(),
-                    scheduled_from_timeline_delete: false,
-                });
-                self.calls_unfinished_metric_begin(&op);
-                upload_queue.queued_operations.push_back(op);
-                info!("scheduled layer file deletion {name}");
-            }
-
-            // Launch the tasks immediately, if possible
-            self.launch_queued_tasks(upload_queue);
+            with_generations
         };
-        no_bail_here();
+
+        // 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})");
+        }
+
+        // Enqueue deletions
+        deletion_queue_client
+            .push_layers(
+                self.tenant_id,
+                self.timeline_id,
+                self.generation,
+                with_generations,
+            )
+            .await?;
         Ok(())
     }
 
@@ -762,10 +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,
                 )
             },
@@ -802,12 +825,13 @@ 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>) -> anyhow::Result<()> {
+    pub(crate) async fn delete_all(
+        self: &Arc<Self>,
+        deletion_queue: &DeletionQueueClient,
+    ) -> anyhow::Result<()> {
         debug_assert_current_span_has_tenant_and_timeline_id();
 
-        let (mut receiver, deletions_queued) = {
-            let mut deletions_queued = 0;
-
+        let layers: Vec<_> = {
             let mut locked = self.upload_queue.lock().unwrap();
             let stopped = locked.stopped_mut()?;
 
@@ -819,40 +843,29 @@ impl RemoteTimelineClient {
 
             stopped
                 .upload_queue_for_deletion
-                .queued_operations
-                .reserve(stopped.upload_queue_for_deletion.latest_files.len());
-
-            // schedule the actual deletions
-            for name in stopped.upload_queue_for_deletion.latest_files.keys() {
-                let op = UploadOp::Delete(Delete {
-                    file_kind: RemoteOpFileKind::Layer,
-                    layer_file_name: name.clone(),
-                    scheduled_from_timeline_delete: true,
-                });
-                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,
-            )
+                .latest_files
+                .drain()
+                .map(|kv| (kv.0, kv.1.generation))
+                .collect()
         };
 
-        receiver.changed().await.context("upload queue shut down")?;
+        let layer_deletion_count = layers.len();
+
+        let layer_paths = layers
+            .into_iter()
+            .map(|(layer, generation)| {
+                remote_layer_path(&self.tenant_id, &self.timeline_id, &layer, generation)
+            })
+            .collect();
+        deletion_queue.push_immediate(layer_paths).await?;
 
         // 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)?;
+        let timeline_storage_path = remote_timeline_path(&self.tenant_id, &self.timeline_id);
+
+        // 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 remaining = backoff::retry(
             || async {
@@ -881,17 +894,9 @@ impl RemoteTimelineClient {
             })
             .collect();
 
+        let not_referenced_count = remaining.len();
         if !remaining.is_empty() {
-            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")?;
+            deletion_queue.push_immediate(remaining).await?;
         }
 
         fail::fail_point!("timeline-delete-before-index-delete", |_| {
@@ -902,18 +907,14 @@ impl RemoteTimelineClient {
 
         let index_file_path = timeline_storage_path.join(Path::new(IndexPart::FILE_NAME));
 
-        debug!("deleting index part");
+        debug!("enqueuing index part deletion");
+        deletion_queue
+            .push_immediate([index_file_path].to_vec())
+            .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")?;
+        // 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?;
 
         fail::fail_point!("timeline-delete-after-index-delete", |_| {
             Err(anyhow::anyhow!(
@@ -921,7 +922,7 @@ impl RemoteTimelineClient {
             ))?
         });
 
-        info!(prefix=%timeline_storage_path, referenced=deletions_queued, not_referenced=%remaining.len(), "done deleting in timeline prefix, including index_part.json");
+        info!(prefix=%timeline_storage_path, referenced=layer_deletion_count, not_referenced=%not_referenced_count, "done deleting in timeline prefix, including index_part.json");
 
         Ok(())
     }
@@ -944,10 +945,6 @@ 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(),
             };
@@ -975,9 +972,6 @@ 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;
@@ -1056,15 +1050,17 @@ 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,
@@ -1086,10 +1082,10 @@ 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(
@@ -1109,21 +1105,6 @@ 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)
-                        .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
@@ -1183,15 +1164,7 @@ 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(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::Stopped(_) => { None }
                 UploadQueue::Initialized(qi) => { Some(qi) }
             };
 
@@ -1213,9 +1186,6 @@ 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!(),
             };
 
@@ -1247,13 +1217,6 @@ 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;
@@ -1313,7 +1276,6 @@ 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(),
                     };
@@ -1334,9 +1296,7 @@ impl RemoteTimelineClient {
 
                 // consistency check
                 assert_eq!(
-                    qi.num_inprogress_layer_uploads
-                        + qi.num_inprogress_metadata_uploads
-                        + qi.num_inprogress_deletions,
+                    qi.num_inprogress_layer_uploads + qi.num_inprogress_metadata_uploads,
                     qi.inprogress_tasks.len()
                 );
 
@@ -1361,14 +1321,84 @@ 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,
+    generation: Generation,
+) -> RemotePath {
+    // Generation-aware key format
+    let path = format!(
+        "tenants/{tenant_id}/{TIMELINES_SEGMENT_NAME}/{timeline_id}/{0}{1}",
+        layer_file_name.file_name(),
+        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: Option<Generation>,
+) -> anyhow::Result<RemotePath> {
+    let stripped = local_path
+        .strip_prefix(&conf.workdir)
+        .context("Failed to strip workdir prefix")?;
+
+    let suffixed = if let Some(generation) = generation {
+        format!(
+            "{0}{1}",
+            stripped.to_string_lossy(),
+            generation.get_suffix()
+        )
+    } else {
+        stripped.to_string_lossy().to_string()
+    };
+
+    RemotePath::new(&PathBuf::from(suffixed)).with_context(|| {
+        format!(
+            "Failed 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,
     };
@@ -1410,8 +1440,11 @@ 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();
@@ -1432,6 +1465,7 @@ mod tests {
         tenant_ctx: RequestContext,
         remote_fs_dir: PathBuf,
         client: Arc<RemoteTimelineClient>,
+        deletion_queue: MockDeletionQueue,
     }
 
     impl TestSetup {
@@ -1462,6 +1496,8 @@ 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 {
@@ -1469,7 +1505,8 @@ mod tests {
                 runtime: tokio::runtime::Handle::current(),
                 tenant_id: harness.tenant_id,
                 timeline_id: TIMELINE_ID,
-                storage_impl: storage,
+                generation,
+                storage_impl: storage.clone(),
                 upload_queue: Mutex::new(UploadQueue::Uninitialized),
                 metrics: Arc::new(RemoteTimelineClientMetrics::new(
                     &harness.tenant_id,
@@ -1477,6 +1514,8 @@ mod tests {
                 )),
             });
 
+            let deletion_queue = MockDeletionQueue::new(Some(storage));
+
             Ok(Self {
                 harness,
                 tenant,
@@ -1484,6 +1523,7 @@ mod tests {
                 tenant_ctx: ctx,
                 remote_fs_dir,
                 client,
+                deletion_queue,
             })
         }
     }
@@ -1512,6 +1552,7 @@ 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);
@@ -1527,6 +1568,8 @@ 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();
@@ -1546,13 +1589,13 @@ 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();
 
@@ -1610,30 +1653,23 @@ mod tests {
                 &layer_file_name_2.file_name(),
             ],
         );
-        let downloaded_metadata = index_part.parse_metadata().unwrap();
-        assert_eq!(downloaded_metadata, metadata);
+        assert_eq!(index_part.metadata, metadata);
 
         // Schedule upload and then a deletion. Check that the deletion is queued
         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();
-        client
-            .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, 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_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(),
@@ -1641,10 +1677,50 @@ mod tests {
                 "index_part.json",
             ],
             &remote_timeline_dir,
+            generation,
         );
 
-        // Finish them
+        client
+            .schedule_layer_file_deletion(
+                &[layer_file_name_1.clone()],
+                &deletion_queue.new_client(),
+            )
+            .await
+            .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
+            );
+        }
+        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
         client.wait_completion().await.unwrap();
+        deletion_queue.pump().await;
+
+        // 1 layer was deleted
+        assert_eq!(deletion_queue.get_executed(), 1);
 
         assert_remote_files(
             &[
@@ -1653,6 +1729,7 @@ mod tests {
                 "index_part.json",
             ],
             &remote_timeline_dir,
+            generation,
         );
     }
 
@@ -1705,12 +1782,14 @@ 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();
 
@@ -1745,4 +1824,23 @@ mod tests {
             }
         );
     }
+
+    // #[tokio::test]
+    // async fn index_part_download() {
+    //     let TestSetup {
+    //         harness,
+    //         tenant: _tenant,
+    //         timeline: _timeline,
+    //         client,
+    //         ..
+    //     } = TestSetup::new("index_part_download").await.unwrap();
+
+    //     let example_index_part = IndexPart {
+    //         version: 3,
+    //         timeline_layers: HashSet::new(),
+    //         layer_metadata:
+
+    //     }
+
+    // }
 }

pageserver/src/tenant/remote_timeline_client/delete.rs

@@ -1,29 +0,0 @@
-//! 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;
-
-pub(super) async fn delete_layer<'a>(
-    conf: &'static PageServerConf,
-    storage: &'a GenericRemoteStorage,
-    local_layer_path: &'a Path,
-) -> 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 = conf.remote_path(local_layer_path)?;
-
-    // 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!("Failed to 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 remote_storage::{DownloadError, GenericRemoteStorage};
+use crate::tenant::Generation;
+use remote_storage::{DownloadError, GenericRemoteStorage, RemotePath};
 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 +43,16 @@ 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.generation,
+    );
 
     // Perform a rename inspired by durable_rename from file_utils.c.
     // The sequence:
@@ -173,21 +178,19 @@ 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?;
 
@@ -221,46 +224,140 @@ pub async fn list_remote_timelines<'a>(
     Ok(timeline_ids)
 }
 
+async fn do_download_index_part(
+    local_path: &Path,
+    storage: &GenericRemoteStorage,
+    tenant_id: &TenantId,
+    timeline_id: &TimelineId,
+    index_generation: Generation,
+) -> Result<IndexPart, DownloadError> {
+    let remote_path = remote_index_path(tenant_id, timeline_id, index_generation);
+
+    let index_part_bytes = download_retry(
+        || storage.download_all(&remote_path),
+        &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 {local_path:?}"))
+        .map_err(DownloadError::Other)?;
+
+    Ok(index_part)
+}
+
 pub(super) async fn download_index_part(
     conf: &'static PageServerConf,
     storage: &GenericRemoteStorage,
     tenant_id: &TenantId,
     timeline_id: &TimelineId,
+    my_generation: Generation,
 ) -> Result<IndexPart, DownloadError> {
-    let index_part_path = conf
+    let local_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 index_part_bytes = download_retry(
-        || async {
-            let mut index_part_download = storage.download(&part_storage_path).await?;
+    if my_generation.is_none() {
+        // Operating without generations: just fetch the generation-less path
+        return do_download_index_part(&local_path, storage, tenant_id, timeline_id, my_generation)
+            .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!("Failed to download an index part into file {index_part_path:?}")
-            })
-            .map_err(DownloadError::Other)?;
-            Ok(index_part_bytes)
-        },
-        &format!("download {part_storage_path:?}"),
+    let previous_gen = my_generation.previous();
+    let r_previous =
+        do_download_index_part(&local_path, storage, tenant_id, timeline_id, previous_gen).await;
+
+    match r_previous {
+        Ok(index_part) => {
+            tracing::debug!("Found index_part from previous generation {previous_gen}");
+            return Ok(index_part);
+        }
+        Err(e) => {
+            if matches!(e, DownloadError::NotFound) {
+                tracing::debug!("No index_part found from previous generation {previous_gen}, falling back to listing");
+            } else {
+                return Err(e);
+            }
+        }
+    };
+
+    /// Given the key of an index, parse out the generation part of the name
+    fn parse_generation(path: RemotePath) -> Option<Generation> {
+        let path = path.take();
+        let file_name = match path.file_name() {
+            Some(f) => f,
+            None => {
+                // Unexpected: we should be seeing index_part.json paths only
+                tracing::warn!("Malformed index key {0}", path.display());
+                return None;
+            }
+        };
+
+        let file_name_str = match file_name.to_str() {
+            Some(s) => s,
+            None => {
+                tracing::warn!("Malformed index key {0}", path.display());
+                return None;
+            }
+        };
+
+        match file_name_str.split_once("-") {
+            Some((_, gen_suffix)) => u32::from_str_radix(gen_suffix, 16)
+                .map(|g| Generation::new(g))
+                .ok(),
+            None => None,
+        }
+    }
+
+    // Fallback: we did not find an index_part.json from the previous generation, so
+    // we will list all the index_part objects and pick the most recent.
+    let index_prefix = remote_index_path(tenant_id, timeline_id, Generation::none());
+    let indices = backoff::retry(
+        || async { storage.list_files(Some(&index_prefix)).await },
+        |_| false,
+        FAILED_DOWNLOAD_WARN_THRESHOLD,
+        FAILED_REMOTE_OP_RETRIES,
+        "listing index_part files",
+        // TODO: use a cancellation token (https://github.com/neondatabase/neon/issues/5066)
+        backoff::Cancel::new(CancellationToken::new(), || -> anyhow::Error {
+            unreachable!()
+        }),
     )
-    .await?;
+    .await
+    .map_err(|e| DownloadError::Other(e))?;
 
-    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:?}")
-        })
-        .map_err(DownloadError::Other)?;
+    let mut generations: Vec<_> = indices
+        .into_iter()
+        .filter_map(|k| parse_generation(k))
+        .filter(|g| g <= &my_generation)
+        .collect();
 
-    Ok(index_part)
+    generations.sort();
+    match generations.last() {
+        Some(g) => {
+            tracing::debug!("Found index_part in generation {g} (my generation {my_generation})");
+            do_download_index_part(&local_path, storage, tenant_id, timeline_id, *g).await
+        }
+        None => {
+            // This is not an error: the timeline may be newly created, or we may be
+            // upgrading and have no historical index_part with a generation suffix.
+            // Fall back to trying to load the un-suffixed index_part.json.
+            tracing::info!(
+                "No index_part.json-* found when loading {}/{} in generation {}",
+                tenant_id,
+                timeline_id,
+                my_generation
+            );
+            return do_download_index_part(
+                &local_path,
+                storage,
+                tenant_id,
+                timeline_id,
+                Generation::none(),
+            )
+            .await;
+        }
+    }
 }
 
 /// Helper function to handle retries for a download operation.

pageserver/src/tenant/remote_timeline_client/index.rs

@@ -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 {
@@ -77,7 +84,9 @@ pub struct IndexPart {
     // private because internally we would read from metadata instead.
     #[serde_as(as = "DisplayFromStr")]
     disk_consistent_lsn: Lsn,
-    metadata_bytes: Vec<u8>,
+
+    #[serde(rename = "metadata_bytes")]
+    pub metadata: TimelineMetadata,
 }
 
 impl IndexPart {
@@ -95,7 +104,7 @@ impl IndexPart {
     pub fn new(
         layers_and_metadata: HashMap<LayerFileName, LayerFileMetadata>,
         disk_consistent_lsn: Lsn,
-        metadata_bytes: Vec<u8>,
+        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());
@@ -111,14 +120,10 @@ impl IndexPart {
             timeline_layers,
             layer_metadata,
             disk_consistent_lsn,
-            metadata_bytes,
+            metadata,
             deleted_at: None,
         }
     }
-
-    pub fn parse_metadata(&self) -> anyhow::Result<TimelineMetadata> {
-        TimelineMetadata::from_bytes(&self.metadata_bytes)
-    }
 }
 
 impl TryFrom<&UploadQueueInitialized> for IndexPart {
@@ -126,26 +131,31 @@ impl TryFrom<&UploadQueueInitialized> for IndexPart {
 
     fn try_from(upload_queue: &UploadQueueInitialized) -> Result<Self, Self::Error> {
         let disk_consistent_lsn = upload_queue.latest_metadata.disk_consistent_lsn();
-        let metadata_bytes = upload_queue.latest_metadata.to_bytes()?;
+        let metadata = upload_queue.latest_metadata.clone();
 
         Ok(Self::new(
             upload_queue.latest_files.clone(),
             disk_consistent_lsn,
-            metadata_bytes,
+            metadata,
         ))
     }
 }
 
 /// 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,
+
+    #[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 {
         IndexLayerMetadata {
             file_size: other.file_size,
+            generation: other.generation,
         }
     }
 }
@@ -174,15 +184,17 @@ mod tests {
             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_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].to_vec(),
+            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: None,
         };
 
@@ -201,7 +213,7 @@ mod tests {
                 "000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__00000000016B59D8-00000000016B5A51": { "file_size": 9007199254741001 }
             },
             "disk_consistent_lsn":"0/16960E8",
-            "metadata_bytes":[112,11,159,210,0,54,0,4,0,0,0,0,1,105,96,232,1,0,0,0,0,1,105,96,112,0,0,0,0,0,0,0,0,0,0,0,0,0,1,105,96,112,0,0,0,0,1,105,96,112,0,0,0,14,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0]
+            "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]
         }"#;
 
         let expected = IndexPart {
@@ -211,15 +223,17 @@ mod tests {
             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_bytes: [112,11,159,210,0,54,0,4,0,0,0,0,1,105,96,232,1,0,0,0,0,1,105,96,112,0,0,0,0,0,0,0,0,0,0,0,0,0,1,105,96,112,0,0,0,0,1,105,96,112,0,0,0,14,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0].to_vec(),
+            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: None,
         };
 
@@ -238,7 +252,7 @@ mod tests {
                 "000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__00000000016B59D8-00000000016B5A51": { "file_size": 9007199254741001 }
             },
             "disk_consistent_lsn":"0/16960E8",
-            "metadata_bytes":[112,11,159,210,0,54,0,4,0,0,0,0,1,105,96,232,1,0,0,0,0,1,105,96,112,0,0,0,0,0,0,0,0,0,0,0,0,0,1,105,96,112,0,0,0,0,1,105,96,112,0,0,0,14,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],
+            "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"
         }"#;
 
@@ -249,15 +263,17 @@ mod tests {
             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_bytes: [112,11,159,210,0,54,0,4,0,0,0,0,1,105,96,232,1,0,0,0,0,1,105,96,112,0,0,0,0,0,0,0,0,0,0,0,0,0,1,105,96,112,0,0,0,0,1,105,96,112,0,0,0,14,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0].to_vec(),
+            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())
         };
@@ -281,7 +297,7 @@ mod tests {
             timeline_layers: HashSet::new(),
             layer_metadata: HashMap::new(),
             disk_consistent_lsn: "0/2532648".parse::<Lsn>().unwrap(),
-            metadata_bytes: [
+            metadata: TimelineMetadata::from_bytes(&[
                 136, 151, 49, 208, 0, 70, 0, 4, 0, 0, 0, 0, 2, 83, 38, 72, 1, 0, 0, 0, 0, 2, 83,
                 38, 32, 1, 87, 198, 240, 135, 97, 119, 45, 125, 38, 29, 155, 161, 140, 141, 255,
                 210, 0, 0, 0, 0, 2, 83, 38, 72, 0, 0, 0, 0, 1, 73, 240, 192, 0, 0, 0, 0, 1, 73,
@@ -302,8 +318,8 @@ mod tests {
                 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
                 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
                 0, 0,
-            ]
-            .to_vec(),
+            ])
+            .unwrap(),
             deleted_at: None,
         };
 

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");
@@ -32,13 +36,9 @@ pub(super) async fn upload_index_part<'a>(
     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}'"))
 }
@@ -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, Some(generation))?;
     let source_file_res = fs::File::open(&source_path).await;
     let source_file = match source_file_res {
         Ok(source_file) => source_file,

pageserver/src/tenant/storage_layer/delta_layer.rs

@@ -467,7 +467,7 @@ impl DeltaLayer {
             PathOrConf::Path(_) => None,
         };
 
-        let loaded = DeltaLayerInner::load(&path, summary)?;
+        let loaded = DeltaLayerInner::load(&path, summary).await?;
 
         if let PathOrConf::Path(ref path) = self.path_or_conf {
             // not production code
@@ -841,12 +841,15 @@ impl Drop for DeltaLayerWriter {
 }
 
 impl DeltaLayerInner {
-    pub(super) fn load(path: &std::path::Path, summary: Option<Summary>) -> anyhow::Result<Self> {
+    pub(super) async fn load(
+        path: &std::path::Path,
+        summary: Option<Summary>,
+    ) -> anyhow::Result<Self> {
         let file = VirtualFile::open(path)
             .with_context(|| format!("Failed to open file '{}'", path.display()))?;
         let file = FileBlockReader::new(file);
 
-        let summary_blk = file.read_blk(0)?;
+        let summary_blk = file.read_blk(0).await?;
         let actual_summary = Summary::des_prefix(summary_blk.as_ref())?;
 
         if let Some(mut expected_summary) = summary {
@@ -1028,7 +1031,7 @@ impl<'a> ValueRef<'a> {
 pub(crate) struct Adapter<T>(T);
 
 impl<T: AsRef<DeltaLayerInner>> Adapter<T> {
-    pub(crate) fn read_blk(&self, blknum: u32) -> Result<BlockLease, std::io::Error> {
-        self.0.as_ref().file.read_blk(blknum)
+    pub(crate) async fn read_blk(&self, blknum: u32) -> Result<BlockLease, std::io::Error> {
+        self.0.as_ref().file.read_blk(blknum).await
     }
 }

pageserver/src/tenant/storage_layer/image_layer.rs

@@ -349,7 +349,8 @@ impl ImageLayer {
             PathOrConf::Path(_) => None,
         };
 
-        let loaded = ImageLayerInner::load(&path, self.desc.image_layer_lsn(), expected_summary)?;
+        let loaded =
+            ImageLayerInner::load(&path, self.desc.image_layer_lsn(), expected_summary).await?;
 
         if let PathOrConf::Path(ref path) = self.path_or_conf {
             // not production code
@@ -432,7 +433,7 @@ impl ImageLayer {
 }
 
 impl ImageLayerInner {
-    pub(super) fn load(
+    pub(super) async fn load(
         path: &std::path::Path,
         lsn: Lsn,
         summary: Option<Summary>,
@@ -440,7 +441,7 @@ impl ImageLayerInner {
         let file = VirtualFile::open(path)
             .with_context(|| format!("Failed to open file '{}'", path.display()))?;
         let file = FileBlockReader::new(file);
-        let summary_blk = file.read_blk(0)?;
+        let summary_blk = file.read_blk(0).await?;
         let actual_summary = Summary::des_prefix(summary_blk.as_ref())?;
 
         if let Some(mut expected_summary) = summary {

pageserver/src/tenant/timeline.rs

@@ -38,6 +38,7 @@ 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::{
@@ -67,6 +68,7 @@ use postgres_connection::PgConnectionConfig;
 use postgres_ffi::to_pg_timestamp;
 use utils::{
     completion,
+    generation::Generation,
     id::{TenantId, TimelineId},
     lsn::{AtomicLsn, Lsn, RecordLsn},
     seqwait::SeqWait,
@@ -141,6 +143,7 @@ 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 {
@@ -152,6 +155,9 @@ 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
+    generation: Generation,
+
     pub pg_version: u32,
 
     /// The tuple has two elements.
@@ -195,6 +201,9 @@ 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
@@ -465,7 +474,7 @@ impl Timeline {
         // The cached image can be returned directly if there is no WAL between the cached image
         // and requested LSN. The cached image can also be used to reduce the amount of WAL needed
         // for redo.
-        let cached_page_img = match self.lookup_cached_page(&key, lsn) {
+        let cached_page_img = match self.lookup_cached_page(&key, lsn).await {
             Some((cached_lsn, cached_img)) => {
                 match cached_lsn.cmp(&lsn) {
                     Ordering::Less => {} // there might be WAL between cached_lsn and lsn, we need to check
@@ -494,6 +503,7 @@ impl Timeline {
 
         RECONSTRUCT_TIME
             .observe_closure_duration(|| self.reconstruct_value(key, lsn, reconstruct_state))
+            .await
     }
 
     /// Get last or prev record separately. Same as get_last_record_rlsn().last/prev.
@@ -1198,7 +1208,7 @@ 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(
@@ -1261,6 +1271,18 @@ 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)]
@@ -1376,6 +1398,7 @@ 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,
@@ -1405,6 +1428,7 @@ 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),
@@ -1413,6 +1437,7 @@ 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 {
@@ -1614,7 +1639,10 @@ impl Timeline {
         let (conf, tenant_id, timeline_id) = (self.conf, self.tenant_id, self.timeline_id);
         let span = tracing::Span::current();
 
-        let (loaded_layers, needs_upload, total_physical_size) = tokio::task::spawn_blocking({
+        // 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();
                 let discovered = init::scan_timeline_dir(&timeline_path)?;
@@ -1655,11 +1683,16 @@ 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();
+                let mut needs_cleanup = Vec::new();
                 let mut total_physical_size = 0;
 
                 for (name, decision) in decided {
@@ -1675,14 +1708,10 @@ impl Timeline {
                         Err(FutureLayer { local }) => {
                             if local.is_some() {
                                 path.push(name.file_name());
-                                init::cleanup_future_layer(&path, name, disk_consistent_lsn)?;
+                                init::cleanup_future_layer(&path, &name, disk_consistent_lsn)?;
                                 path.pop();
-                            } else {
-                                // we cannot do anything for remote layers, but not continuing to
-                                // process it will leave it out index_part.json as well.
                             }
-                            //
-                            // we do not currently schedule deletions for these.
+                            needs_cleanup.push(name);
                             continue;
                         }
                     };
@@ -1736,7 +1765,11 @@ impl Timeline {
 
                     loaded_layers.push(layer);
                 }
-                Ok((loaded_layers, needs_upload, total_physical_size))
+                Ok((
+                    loaded_layers,
+                    (needs_upload, needs_cleanup),
+                    total_physical_size,
+                ))
             }
         })
         .await
@@ -1748,9 +1781,15 @@ 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_index_upload_for_file_changes()?;
             // Tenant::create_timeline will wait for these uploads to happen before returning, or
             // on retry.
@@ -2261,7 +2300,15 @@ impl Timeline {
                         )));
                     }
                 }
-                ancestor.wait_lsn(timeline.ancestor_lsn, ctx).await?;
+                ancestor
+                    .wait_lsn(timeline.ancestor_lsn, ctx)
+                    .await
+                    .with_context(|| {
+                        format!(
+                            "wait for lsn {} on ancestor timeline_id={}",
+                            timeline.ancestor_lsn, ancestor.timeline_id
+                        )
+                    })?;
 
                 timeline_owned = ancestor;
                 timeline = &*timeline_owned;
@@ -2440,13 +2487,14 @@ impl Timeline {
         }
     }
 
-    fn lookup_cached_page(&self, key: &Key, lsn: Lsn) -> Option<(Lsn, Bytes)> {
+    async fn lookup_cached_page(&self, key: &Key, lsn: Lsn) -> Option<(Lsn, Bytes)> {
         let cache = page_cache::get();
 
         // FIXME: It's pointless to check the cache for things that are not 8kB pages.
         // We should look at the key to determine if it's a cacheable object
-        let (lsn, read_guard) =
-            cache.lookup_materialized_page(self.tenant_id, self.timeline_id, key, lsn)?;
+        let (lsn, read_guard) = cache
+            .lookup_materialized_page(self.tenant_id, self.timeline_id, key, lsn)
+            .await?;
         let img = Bytes::from(read_guard.to_vec());
         Some((lsn, img))
     }
@@ -2656,7 +2704,7 @@ impl Timeline {
                 (
                     HashMap::from([(
                         layer.filename(),
-                        LayerFileMetadata::new(layer.layer_desc().file_size),
+                        LayerFileMetadata::new(layer.layer_desc().file_size, self.generation),
                     )]),
                     Some(layer),
                 )
@@ -3052,7 +3100,10 @@ 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
@@ -3727,7 +3778,7 @@ 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),
                 )?;
             }
 
@@ -3736,7 +3787,10 @@ 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,
@@ -3776,7 +3830,13 @@ impl Timeline {
 
         // Also schedule the deletions in remote storage
         if let Some(remote_client) = &self.remote_client {
-            remote_client.schedule_layer_file_deletion(&layer_names_to_delete)?;
+            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?;
         }
 
         Ok(())
@@ -4110,7 +4170,15 @@ impl Timeline {
             }
 
             if let Some(remote_client) = &self.remote_client {
-                remote_client.schedule_layer_file_deletion(&layer_names_to_delete)?;
+                // 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?;
             }
 
             apply.flush();
@@ -4128,7 +4196,7 @@ impl Timeline {
     ///
     /// Reconstruct a value, using the given base image and WAL records in 'data'.
     ///
-    fn reconstruct_value(
+    async fn reconstruct_value(
         &self,
         key: Key,
         request_lsn: Lsn,
@@ -4197,6 +4265,7 @@ impl Timeline {
                             last_rec_lsn,
                             &img,
                         )
+                        .await
                         .context("Materialized page memoization failed")
                     {
                         return Err(PageReconstructError::from(e));
@@ -4699,6 +4768,7 @@ 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};
@@ -4721,9 +4791,17 @@ 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)).await.unwrap();
+        let tenant = harness
+            .try_load(
+                &ctx,
+                Some(remote_storage),
+                Some(deletion_queue.new_client()),
+            )
+            .await
+            .unwrap();
         let timeline = tenant
             .create_test_timeline(TimelineId::generate(), Lsn(0x10), 14, &ctx)
             .await
@@ -4786,9 +4864,17 @@ 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)).await.unwrap();
+        let tenant = harness
+            .try_load(
+                &ctx,
+                Some(remote_storage),
+                Some(deletion_queue.new_client()),
+            )
+            .await
+            .unwrap();
         let timeline = tenant
             .create_test_timeline(TimelineId::generate(), Lsn(0x10), 14, &ctx)
             .await

pageserver/src/tenant/timeline/delete.rs

@@ -14,6 +14,7 @@ use utils::{
 
 use crate::{
     config::PageServerConf,
+    deletion_queue::DeletionQueueClient,
     task_mgr::{self, TaskKind},
     tenant::{
         metadata::TimelineMetadata,
@@ -238,15 +239,6 @@ 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.
@@ -407,6 +399,7 @@ 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.
@@ -416,7 +409,10 @@ impl DeleteTimelineFlow {
                 timeline_id,
                 local_metadata,
                 None, // Ancestor is not needed for deletion.
-                TimelineResources { remote_client },
+                TimelineResources {
+                    remote_client,
+                    deletion_queue_client,
+                },
                 init_order,
                 // Important. We dont pass ancestor above because it can be missing.
                 // Thus we need to skip the validation here.
@@ -559,7 +555,7 @@ impl DeleteTimelineFlow {
     ) -> Result<(), DeleteTimelineError> {
         delete_local_layer_files(conf, tenant.tenant_id, timeline).await?;
 
-        delete_remote_layers_and_index(timeline).await?;
+        timeline.delete_all_remote().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),
@@ -183,7 +197,7 @@ pub(super) fn cleanup_local_file_for_remote(
 
 pub(super) fn cleanup_future_layer(
     path: &Path,
-    name: LayerFileName,
+    name: &LayerFileName,
     disk_consistent_lsn: Lsn,
 ) -> anyhow::Result<()> {
     use LayerFileName::*;

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

@@ -31,10 +31,11 @@ use storage_broker::Streaming;
 use tokio::select;
 use tracing::*;
 
-use postgres_connection::{parse_host_port, PgConnectionConfig};
+use postgres_connection::PgConnectionConfig;
 use utils::backoff::{
     exponential_backoff, DEFAULT_BASE_BACKOFF_SECONDS, DEFAULT_MAX_BACKOFF_SECONDS,
 };
+use utils::postgres_client::wal_stream_connection_config;
 use utils::{
     id::{NodeId, TenantTimelineId},
     lsn::Lsn,
@@ -879,33 +880,6 @@ impl ReconnectReason {
     }
 }
 
-fn wal_stream_connection_config(
-    TenantTimelineId {
-        tenant_id,
-        timeline_id,
-    }: TenantTimelineId,
-    listen_pg_addr_str: &str,
-    auth_token: Option<&str>,
-    availability_zone: Option<&str>,
-) -> anyhow::Result<PgConnectionConfig> {
-    let (host, port) =
-        parse_host_port(listen_pg_addr_str).context("Unable to parse listen_pg_addr_str")?;
-    let port = port.unwrap_or(5432);
-    let mut connstr = PgConnectionConfig::new_host_port(host, port)
-        .extend_options([
-            "-c".to_owned(),
-            format!("timeline_id={}", timeline_id),
-            format!("tenant_id={}", tenant_id),
-        ])
-        .set_password(auth_token.map(|s| s.to_owned()));
-
-    if let Some(availability_zone) = availability_zone {
-        connstr = connstr.extend_options([format!("availability_zone={}", availability_zone)]);
-    }
-
-    Ok(connstr)
-}
-
 #[cfg(test)]
 mod tests {
     use super::*;
@@ -921,6 +895,7 @@ mod tests {
             timeline: SafekeeperTimelineInfo {
                 safekeeper_id: 0,
                 tenant_timeline_id: None,
+                term: 0,
                 last_log_term: 0,
                 flush_lsn: 0,
                 commit_lsn,
@@ -929,6 +904,7 @@ mod tests {
                 peer_horizon_lsn: 0,
                 local_start_lsn: 0,
                 safekeeper_connstr: safekeeper_connstr.to_owned(),
+                http_connstr: safekeeper_connstr.to_owned(),
                 availability_zone: None,
             },
             latest_update,

pageserver/src/tenant/upload_queue.rs

@@ -1,5 +1,3 @@
-use crate::metrics::RemoteOpFileKind;
-
 use super::storage_layer::LayerFileName;
 use crate::tenant::metadata::TimelineMetadata;
 use crate::tenant::remote_timeline_client::index::IndexPart;
@@ -62,7 +60,6 @@ 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
@@ -120,7 +117,6 @@ 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(),
         };
@@ -148,22 +144,20 @@ impl UploadQueue {
             );
         }
 
-        let index_part_metadata = index_part.parse_metadata()?;
         info!(
             "initializing upload queue with remote index_part.disk_consistent_lsn: {}",
-            index_part_metadata.disk_consistent_lsn()
+            index_part.metadata.disk_consistent_lsn()
         );
 
         let state = UploadQueueInitialized {
             latest_files: files,
             latest_files_changes_since_metadata_upload_scheduled: 0,
-            latest_metadata: index_part_metadata.clone(),
-            last_uploaded_consistent_lsn: index_part_metadata.disk_consistent_lsn(),
+            latest_metadata: index_part.metadata.clone(),
+            last_uploaded_consistent_lsn: index_part.metadata.disk_consistent_lsn(),
             // what follows are boring default initializations
             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(),
         };
@@ -201,13 +195,6 @@ 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,
-}
-
 #[derive(Debug)]
 pub(crate) enum UploadOp {
     /// Upload a layer file
@@ -216,9 +203,6 @@ 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<()>),
 }
@@ -234,13 +218,9 @@ impl std::fmt::Display for UploadOp {
                     metadata.file_size()
                 )
             }
-            UploadOp::UploadMetadata(_, lsn) => write!(f, "UploadMetadata(lsn: {})", lsn),
-            UploadOp::Delete(delete) => write!(
-                f,
-                "Delete(path: {}, scheduled_from_timeline_delete: {})",
-                delete.layer_file_name.file_name(),
-                delete.scheduled_from_timeline_delete
-            ),
+            UploadOp::UploadMetadata(_, lsn) => {
+                write!(f, "UploadMetadata(lsn: {})", lsn)
+            }
             UploadOp::Barrier(_) => write!(f, "Barrier"),
         }
     }

pageserver/src/test.log

@@ -0,0 +1 @@
+-bash: scripts/pytest: No such file or directory

safekeeper/src/bin/safekeeper.rs

@@ -341,21 +341,35 @@ async fn start_safekeeper(conf: SafeKeeperConf) -> Result<()> {
 
     let (wal_backup_launcher_tx, wal_backup_launcher_rx) = mpsc::channel(100);
 
-    // Load all timelines from disk to memory.
-    GlobalTimelines::init(conf.clone(), wal_backup_launcher_tx)?;
-
     // Keep handles to main tasks to die if any of them disappears.
     let mut tasks_handles: FuturesUnordered<BoxFuture<(String, JoinTaskRes)>> =
         FuturesUnordered::new();
 
+    // Start wal backup launcher before loading timelines as we'll notify it
+    // through the channel about timelines which need offloading, not draining
+    // the channel would cause deadlock.
+    let current_thread_rt = conf
+        .current_thread_runtime
+        .then(|| Handle::try_current().expect("no runtime in main"));
+    let conf_ = conf.clone();
+    let wal_backup_handle = current_thread_rt
+        .as_ref()
+        .unwrap_or_else(|| WAL_BACKUP_RUNTIME.handle())
+        .spawn(wal_backup::wal_backup_launcher_task_main(
+            conf_,
+            wal_backup_launcher_rx,
+        ))
+        .map(|res| ("WAL backup launcher".to_owned(), res));
+    tasks_handles.push(Box::pin(wal_backup_handle));
+
+    // Load all timelines from disk to memory.
+    GlobalTimelines::init(conf.clone(), wal_backup_launcher_tx).await?;
+
     let conf_ = conf.clone();
     // Run everything in current thread rt, if asked.
     if conf.current_thread_runtime {
         info!("running in current thread runtime");
     }
-    let current_thread_rt = conf
-        .current_thread_runtime
-        .then(|| Handle::try_current().expect("no runtime in main"));
 
     let wal_service_handle = current_thread_rt
         .as_ref()
@@ -408,17 +422,6 @@ async fn start_safekeeper(conf: SafeKeeperConf) -> Result<()> {
         .map(|res| ("WAL remover".to_owned(), res));
     tasks_handles.push(Box::pin(wal_remover_handle));
 
-    let conf_ = conf.clone();
-    let wal_backup_handle = current_thread_rt
-        .as_ref()
-        .unwrap_or_else(|| WAL_BACKUP_RUNTIME.handle())
-        .spawn(wal_backup::wal_backup_launcher_task_main(
-            conf_,
-            wal_backup_launcher_rx,
-        ))
-        .map(|res| ("WAL backup launcher".to_owned(), res));
-    tasks_handles.push(Box::pin(wal_backup_handle));
-
     set_build_info_metric(GIT_VERSION);
 
     // TODO: update tokio-stream, convert to real async Stream with

safekeeper/src/control_file_upgrade.rs

@@ -1,7 +1,6 @@
 //! Code to deal with safekeeper control file upgrades
 use crate::safekeeper::{
-    AcceptorState, PersistedPeers, PgUuid, SafeKeeperState, ServerInfo, Term, TermHistory,
-    TermSwitchEntry,
+    AcceptorState, PersistedPeers, PgUuid, SafeKeeperState, ServerInfo, Term, TermHistory, TermLsn,
 };
 use anyhow::{bail, Result};
 use pq_proto::SystemId;
@@ -145,7 +144,7 @@ pub fn upgrade_control_file(buf: &[u8], version: u32) -> Result<SafeKeeperState>
         let oldstate = SafeKeeperStateV1::des(&buf[..buf.len()])?;
         let ac = AcceptorState {
             term: oldstate.acceptor_state.term,
-            term_history: TermHistory(vec![TermSwitchEntry {
+            term_history: TermHistory(vec![TermLsn {
                 term: oldstate.acceptor_state.epoch,
                 lsn: Lsn(0),
             }]),

safekeeper/src/http/routes.rs

@@ -19,6 +19,7 @@ use crate::receive_wal::WalReceiverState;
 use crate::safekeeper::ServerInfo;
 use crate::safekeeper::Term;
 use crate::send_wal::WalSenderState;
+use crate::timeline::PeerInfo;
 use crate::{debug_dump, pull_timeline};
 
 use crate::timelines_global_map::TimelineDeleteForceResult;
@@ -101,6 +102,7 @@ pub struct TimelineStatus {
     pub peer_horizon_lsn: Lsn,
     #[serde_as(as = "DisplayFromStr")]
     pub remote_consistent_lsn: Lsn,
+    pub peers: Vec<PeerInfo>,
     pub walsenders: Vec<WalSenderState>,
     pub walreceivers: Vec<WalReceiverState>,
 }
@@ -140,6 +142,7 @@ async fn timeline_status_handler(request: Request<Body>) -> Result<Response<Body
         term_history,
     };
 
+    let conf = get_conf(&request);
     // Note: we report in memory values which can be lost.
     let status = TimelineStatus {
         tenant_id: ttid.tenant_id,
@@ -153,6 +156,7 @@ async fn timeline_status_handler(request: Request<Body>) -> Result<Response<Body
         backup_lsn: inmem.backup_lsn,
         peer_horizon_lsn: inmem.peer_horizon_lsn,
         remote_consistent_lsn: tli.get_walsenders().get_remote_consistent_lsn(),
+        peers: tli.get_peers(conf).await,
         walsenders: tli.get_walsenders().get_all(),
         walreceivers: tli.get_walreceivers().get_all(),
     };
@@ -282,12 +286,14 @@ async fn record_safekeeper_info(mut request: Request<Body>) -> Result<Response<B
             tenant_id: ttid.tenant_id.as_ref().to_owned(),
             timeline_id: ttid.timeline_id.as_ref().to_owned(),
         }),
+        term: sk_info.term.unwrap_or(0),
         last_log_term: sk_info.last_log_term.unwrap_or(0),
         flush_lsn: sk_info.flush_lsn.0,
         commit_lsn: sk_info.commit_lsn.0,
         remote_consistent_lsn: sk_info.remote_consistent_lsn.0,
         peer_horizon_lsn: sk_info.peer_horizon_lsn.0,
         safekeeper_connstr: sk_info.safekeeper_connstr.unwrap_or_else(|| "".to_owned()),
+        http_connstr: sk_info.http_connstr.unwrap_or_else(|| "".to_owned()),
         backup_lsn: sk_info.backup_lsn.0,
         local_start_lsn: sk_info.local_start_lsn.0,
         availability_zone: None,

safekeeper/src/json_ctrl.rs

@@ -21,7 +21,7 @@ use crate::safekeeper::{AcceptorProposerMessage, AppendResponse, ServerInfo};
 use crate::safekeeper::{
     AppendRequest, AppendRequestHeader, ProposerAcceptorMessage, ProposerElected,
 };
-use crate::safekeeper::{SafeKeeperState, Term, TermHistory, TermSwitchEntry};
+use crate::safekeeper::{SafeKeeperState, Term, TermHistory, TermLsn};
 use crate::timeline::Timeline;
 use crate::GlobalTimelines;
 use postgres_backend::PostgresBackend;
@@ -119,7 +119,7 @@ async fn send_proposer_elected(tli: &Arc<Timeline>, term: Term, lsn: Lsn) -> any
     let history = tli.get_state().await.1.acceptor_state.term_history;
     let history = history.up_to(lsn.checked_sub(1u64).unwrap());
     let mut history_entries = history.0;
-    history_entries.push(TermSwitchEntry { term, lsn });
+    history_entries.push(TermLsn { term, lsn });
     let history = TermHistory(history_entries);
 
     let proposer_elected_request = ProposerAcceptorMessage::Elected(ProposerElected {

safekeeper/src/lib.rs

@@ -19,6 +19,7 @@ pub mod json_ctrl;
 pub mod metrics;
 pub mod pull_timeline;
 pub mod receive_wal;
+pub mod recovery;
 pub mod remove_wal;
 pub mod safekeeper;
 pub mod send_wal;

safekeeper/src/pull_timeline.rs

@@ -227,7 +227,9 @@ async fn pull_timeline(status: TimelineStatus, host: String) -> Result<Response>
     tokio::fs::create_dir_all(conf.tenant_dir(&ttid.tenant_id)).await?;
     tokio::fs::rename(tli_dir_path, &timeline_path).await?;
 
-    let tli = GlobalTimelines::load_timeline(ttid).context("Failed to load timeline after copy")?;
+    let tli = GlobalTimelines::load_timeline(ttid)
+        .await
+        .context("Failed to load timeline after copy")?;
 
     info!(
         "Loaded timeline {}, flush_lsn={}",

safekeeper/src/recovery.rs

@@ -0,0 +1,40 @@
+//! This module implements pulling WAL from peer safekeepers if compute can't
+//! provide it, i.e. safekeeper lags too much.
+
+use std::sync::Arc;
+
+use tokio::{select, time::sleep, time::Duration};
+use tracing::{info, instrument};
+
+use crate::{timeline::Timeline, SafeKeeperConf};
+
+/// Entrypoint for per timeline task which always runs, checking whether
+/// recovery for this safekeeper is needed and starting it if so.
+#[instrument(name = "recovery task", skip_all, fields(ttid = %tli.ttid))]
+pub async fn recovery_main(tli: Arc<Timeline>, _conf: SafeKeeperConf) {
+    info!("started");
+    let mut cancellation_rx = match tli.get_cancellation_rx() {
+        Ok(rx) => rx,
+        Err(_) => {
+            info!("timeline canceled during task start");
+            return;
+        }
+    };
+
+    select! {
+        _ = recovery_main_loop(tli) => { unreachable!() }
+        _ = cancellation_rx.changed() => {
+            info!("stopped");
+        }
+    }
+}
+
+const CHECK_INTERVAL_MS: u64 = 2000;
+
+/// Check regularly whether we need to start recovery.
+async fn recovery_main_loop(_tli: Arc<Timeline>) {
+    let check_duration = Duration::from_millis(CHECK_INTERVAL_MS);
+    loop {
+        sleep(check_duration).await;
+    }
+}

safekeeper/src/safekeeper.rs

@@ -34,22 +34,33 @@ pub const UNKNOWN_SERVER_VERSION: u32 = 0;
 
 /// Consensus logical timestamp.
 pub type Term = u64;
-const INVALID_TERM: Term = 0;
+pub const INVALID_TERM: Term = 0;
 
-#[derive(Debug, Clone, Copy, Serialize, Deserialize)]
-pub struct TermSwitchEntry {
+#[derive(Debug, Clone, Copy, Serialize, Deserialize, PartialEq, Eq, PartialOrd, Ord)]
+pub struct TermLsn {
     pub term: Term,
     pub lsn: Lsn,
 }
+
+// Creation from tuple provides less typing (e.g. for unit tests).
+impl From<(Term, Lsn)> for TermLsn {
+    fn from(pair: (Term, Lsn)) -> TermLsn {
+        TermLsn {
+            term: pair.0,
+            lsn: pair.1,
+        }
+    }
+}
+
 #[derive(Clone, Serialize, Deserialize)]
-pub struct TermHistory(pub Vec<TermSwitchEntry>);
+pub struct TermHistory(pub Vec<TermLsn>);
 
 impl TermHistory {
     pub fn empty() -> TermHistory {
         TermHistory(Vec::new())
     }
 
-    // Parse TermHistory as n_entries followed by TermSwitchEntry pairs
+    // Parse TermHistory as n_entries followed by TermLsn pairs
     pub fn from_bytes(bytes: &mut Bytes) -> Result<TermHistory> {
         if bytes.remaining() < 4 {
             bail!("TermHistory misses len");
@@ -60,7 +71,7 @@ impl TermHistory {
             if bytes.remaining() < 16 {
                 bail!("TermHistory is incomplete");
             }
-            res.push(TermSwitchEntry {
+            res.push(TermLsn {
                 term: bytes.get_u64_le(),
                 lsn: bytes.get_u64_le().into(),
             })
@@ -557,12 +568,17 @@ where
             .up_to(self.flush_lsn())
     }
 
+    /// Get current term.
+    pub fn get_term(&self) -> Term {
+        self.state.acceptor_state.term
+    }
+
     pub fn get_epoch(&self) -> Term {
         self.state.acceptor_state.get_epoch(self.flush_lsn())
     }
 
     /// wal_store wrapper avoiding commit_lsn <= flush_lsn violation when we don't have WAL yet.
-    fn flush_lsn(&self) -> Lsn {
+    pub fn flush_lsn(&self) -> Lsn {
         max(self.wal_store.flush_lsn(), self.state.timeline_start_lsn)
     }
 
@@ -1138,7 +1154,7 @@ mod tests {
         let pem = ProposerElected {
             term: 1,
             start_streaming_at: Lsn(1),
-            term_history: TermHistory(vec![TermSwitchEntry {
+            term_history: TermHistory(vec![TermLsn {
                 term: 1,
                 lsn: Lsn(3),
             }]),

safekeeper/src/send_wal.rs

@@ -2,12 +2,12 @@
 //! with the "START_REPLICATION" message, and registry of walsenders.
 
 use crate::handler::SafekeeperPostgresHandler;
-use crate::safekeeper::Term;
+use crate::safekeeper::{Term, TermLsn};
 use crate::timeline::Timeline;
 use crate::wal_service::ConnectionId;
 use crate::wal_storage::WalReader;
 use crate::GlobalTimelines;
-use anyhow::Context as AnyhowContext;
+use anyhow::{bail, Context as AnyhowContext};
 use bytes::Bytes;
 use parking_lot::Mutex;
 use postgres_backend::PostgresBackend;
@@ -390,26 +390,25 @@ impl SafekeeperPostgresHandler {
             self.appname.clone(),
         ));
 
-        let commit_lsn_watch_rx = tli.get_commit_lsn_watch_rx();
-
-        // Walproposer gets special handling: safekeeper must give proposer all
-        // local WAL till the end, whether committed or not (walproposer will
-        // hang otherwise). That's because walproposer runs the consensus and
-        // synchronizes safekeepers on the most advanced one.
+        // Walsender can operate in one of two modes which we select by
+        // application_name: give only committed WAL (used by pageserver) or all
+        // existing WAL (up to flush_lsn, used by walproposer or peer recovery).
+        // The second case is always driven by a consensus leader which term
+        // must generally be also supplied. However we're sloppy to do this in
+        // walproposer recovery which will be removed soon. So TODO is to make
+        // it not Option'al then.
         //
-        // There is a small risk of this WAL getting concurrently garbaged if
-        // another compute rises which collects majority and starts fixing log
-        // on this safekeeper itself. That's ok as (old) proposer will never be
-        // able to commit such WAL.
-        let stop_pos: Option<Lsn> = if self.is_walproposer_recovery() {
-            let wal_end = tli.get_flush_lsn().await;
-            Some(wal_end)
+        // Fetching WAL without term in recovery creates a small risk of this
+        // WAL getting concurrently garbaged if another compute rises which
+        // collects majority and starts fixing log on this safekeeper itself.
+        // That's ok as (old) proposer will never be able to commit such WAL.
+        let end_watch = if self.is_walproposer_recovery() {
+            EndWatch::Flush(tli.get_term_flush_lsn_watch_rx())
         } else {
-            None
+            EndWatch::Commit(tli.get_commit_lsn_watch_rx())
         };
-
-        // take the latest commit_lsn if don't have stop_pos
-        let end_pos = stop_pos.unwrap_or(*commit_lsn_watch_rx.borrow());
+        // we don't check term here; it will be checked on first waiting/WAL reading anyway.
+        let end_pos = end_watch.get();
 
         if end_pos < start_pos {
             warn!(
@@ -419,8 +418,10 @@ impl SafekeeperPostgresHandler {
         }
 
         info!(
-            "starting streaming from {:?} till {:?}, available WAL ends at {}",
-            start_pos, stop_pos, end_pos
+            "starting streaming from {:?}, available WAL ends at {}, recovery={}",
+            start_pos,
+            end_pos,
+            matches!(end_watch, EndWatch::Flush(_))
         );
 
         // switch to copy
@@ -445,9 +446,8 @@ impl SafekeeperPostgresHandler {
             appname,
             start_pos,
             end_pos,
-            stop_pos,
             term,
-            commit_lsn_watch_rx,
+            end_watch,
             ws_guard: ws_guard.clone(),
             wal_reader,
             send_buf: [0; MAX_SEND_SIZE],
@@ -466,6 +466,32 @@ impl SafekeeperPostgresHandler {
     }
 }
 
+/// Walsender streams either up to commit_lsn (normally) or flush_lsn in the
+/// given term (recovery by walproposer or peer safekeeper).
+enum EndWatch {
+    Commit(Receiver<Lsn>),
+    Flush(Receiver<TermLsn>),
+}
+
+impl EndWatch {
+    /// Get current end of WAL.
+    fn get(&self) -> Lsn {
+        match self {
+            EndWatch::Commit(r) => *r.borrow(),
+            EndWatch::Flush(r) => r.borrow().lsn,
+        }
+    }
+
+    /// Wait for the update.
+    async fn changed(&mut self) -> anyhow::Result<()> {
+        match self {
+            EndWatch::Commit(r) => r.changed().await?,
+            EndWatch::Flush(r) => r.changed().await?,
+        }
+        Ok(())
+    }
+}
+
 /// A half driving sending WAL.
 struct WalSender<'a, IO> {
     pgb: &'a mut PostgresBackend<IO>,
@@ -480,14 +506,12 @@ struct WalSender<'a, IO> {
     // We send this LSN to the receiver as wal_end, so that it knows how much
     // WAL this safekeeper has. This LSN should be as fresh as possible.
     end_pos: Lsn,
-    // If present, terminate after reaching this position; used by walproposer
-    // in recovery.
-    stop_pos: Option<Lsn>,
     /// When streaming uncommitted part, the term the client acts as the leader
     /// in. Streaming is stopped if local term changes to a different (higher)
     /// value.
     term: Option<Term>,
-    commit_lsn_watch_rx: Receiver<Lsn>,
+    /// Watch channel receiver to learn end of available WAL (and wait for its advancement).
+    end_watch: EndWatch,
     ws_guard: Arc<WalSenderGuard>,
     wal_reader: WalReader,
     // buffer for readling WAL into to send it
@@ -497,29 +521,20 @@ struct WalSender<'a, IO> {
 impl<IO: AsyncRead + AsyncWrite + Unpin> WalSender<'_, IO> {
     /// Send WAL until
     /// - an error occurs
-    /// - if we are streaming to walproposer, we've streamed until stop_pos
-    ///   (recovery finished)
-    /// - receiver is caughtup and there is no computes
+    /// - receiver is caughtup and there is no computes (if streaming up to commit_lsn)
     ///
     /// Err(CopyStreamHandlerEnd) is always returned; Result is used only for ?
     /// convenience.
     async fn run(&mut self) -> Result<(), CopyStreamHandlerEnd> {
         loop {
-            // If we are streaming to walproposer, check it is time to stop.
-            if let Some(stop_pos) = self.stop_pos {
-                if self.start_pos >= stop_pos {
-                    // recovery finished
-                    return Err(CopyStreamHandlerEnd::ServerInitiated(format!(
-                        "ending streaming to walproposer at {}, recovery finished",
-                        self.start_pos
-                    )));
-                }
-            } else {
-                // Wait for the next portion if it is not there yet, or just
-                // update our end of WAL available for sending value, we
-                // communicate it to the receiver.
-                self.wait_wal().await?;
-            }
+            // Wait for the next portion if it is not there yet, or just
+            // update our end of WAL available for sending value, we
+            // communicate it to the receiver.
+            self.wait_wal().await?;
+            assert!(
+                self.end_pos > self.start_pos,
+                "nothing to send after waiting for WAL"
+            );
 
             // try to send as much as available, capped by MAX_SEND_SIZE
             let mut send_size = self
@@ -567,7 +582,7 @@ impl<IO: AsyncRead + AsyncWrite + Unpin> WalSender<'_, IO> {
     /// exit in the meanwhile
     async fn wait_wal(&mut self) -> Result<(), CopyStreamHandlerEnd> {
         loop {
-            self.end_pos = *self.commit_lsn_watch_rx.borrow();
+            self.end_pos = self.end_watch.get();
             if self.end_pos > self.start_pos {
                 // We have something to send.
                 trace!("got end_pos {:?}, streaming", self.end_pos);
@@ -575,27 +590,31 @@ impl<IO: AsyncRead + AsyncWrite + Unpin> WalSender<'_, IO> {
             }
 
             // Wait for WAL to appear, now self.end_pos == self.start_pos.
-            if let Some(lsn) = wait_for_lsn(&mut self.commit_lsn_watch_rx, self.start_pos).await? {
+            if let Some(lsn) = wait_for_lsn(&mut self.end_watch, self.term, self.start_pos).await? {
                 self.end_pos = lsn;
                 trace!("got end_pos {:?}, streaming", self.end_pos);
                 return Ok(());
             }
 
-            // Timed out waiting for WAL, check for termination and send KA
-            if let Some(remote_consistent_lsn) = self
-                .ws_guard
-                .walsenders
-                .get_ws_remote_consistent_lsn(self.ws_guard.id)
-            {
-                if self.tli.should_walsender_stop(remote_consistent_lsn).await {
-                    // Terminate if there is nothing more to send.
-                    // Note that "ending streaming" part of the string is used by
-                    // pageserver to identify WalReceiverError::SuccessfulCompletion,
-                    // do not change this string without updating pageserver.
-                    return Err(CopyStreamHandlerEnd::ServerInitiated(format!(
+            // Timed out waiting for WAL, check for termination and send KA.
+            // Check for termination only if we are streaming up to commit_lsn
+            // (to pageserver).
+            if let EndWatch::Commit(_) = self.end_watch {
+                if let Some(remote_consistent_lsn) = self
+                    .ws_guard
+                    .walsenders
+                    .get_ws_remote_consistent_lsn(self.ws_guard.id)
+                {
+                    if self.tli.should_walsender_stop(remote_consistent_lsn).await {
+                        // Terminate if there is nothing more to send.
+                        // Note that "ending streaming" part of the string is used by
+                        // pageserver to identify WalReceiverError::SuccessfulCompletion,
+                        // do not change this string without updating pageserver.
+                        return Err(CopyStreamHandlerEnd::ServerInitiated(format!(
                         "ending streaming to {:?} at {}, receiver is caughtup and there is no computes",
                         self.appname, self.start_pos,
                     )));
+                    }
                 }
             }
 
@@ -663,22 +682,32 @@ impl<IO: AsyncRead + AsyncWrite + Unpin> ReplyReader<IO> {
 
 const POLL_STATE_TIMEOUT: Duration = Duration::from_secs(1);
 
-/// Wait until we have commit_lsn > lsn or timeout expires. Returns
-/// - Ok(Some(commit_lsn)) if needed lsn is successfully observed;
+/// Wait until we have available WAL > start_pos or timeout expires. Returns
+/// - Ok(Some(end_pos)) if needed lsn is successfully observed;
 /// - Ok(None) if timeout expired;
-/// - Err in case of error (if watch channel is in trouble, shouldn't happen).
-async fn wait_for_lsn(rx: &mut Receiver<Lsn>, lsn: Lsn) -> anyhow::Result<Option<Lsn>> {
+/// - Err in case of error -- only if 1) term changed while fetching in recovery
+///   mode 2) watch channel closed, which must never happen.
+async fn wait_for_lsn(
+    rx: &mut EndWatch,
+    client_term: Option<Term>,
+    start_pos: Lsn,
+) -> anyhow::Result<Option<Lsn>> {
     let res = timeout(POLL_STATE_TIMEOUT, async move {
-        let mut commit_lsn;
         loop {
-            rx.changed().await?;
-            commit_lsn = *rx.borrow();
-            if commit_lsn > lsn {
-                break;
+            let end_pos = rx.get();
+            if end_pos > start_pos {
+                return Ok(end_pos);
             }
+            if let EndWatch::Flush(rx) = rx {
+                let curr_term = rx.borrow().term;
+                if let Some(client_term) = client_term {
+                    if curr_term != client_term {
+                        bail!("term changed: requested {}, now {}", client_term, curr_term);
+                    }
+                }
+            }
+            rx.changed().await?;
         }
-
-        Ok(commit_lsn)
     })
     .await;
 

safekeeper/src/timeline.rs

@@ -3,8 +3,11 @@
 
 use anyhow::{anyhow, bail, Result};
 use postgres_ffi::XLogSegNo;
+use serde::{Deserialize, Serialize};
+use serde_with::serde_as;
 use tokio::fs;
 
+use serde_with::DisplayFromStr;
 use std::cmp::max;
 use std::path::PathBuf;
 use std::sync::Arc;
@@ -24,9 +27,10 @@ use storage_broker::proto::SafekeeperTimelineInfo;
 use storage_broker::proto::TenantTimelineId as ProtoTenantTimelineId;
 
 use crate::receive_wal::WalReceivers;
+use crate::recovery::recovery_main;
 use crate::safekeeper::{
     AcceptorProposerMessage, ProposerAcceptorMessage, SafeKeeper, SafeKeeperState,
-    SafekeeperMemState, ServerInfo, Term,
+    SafekeeperMemState, ServerInfo, Term, TermLsn, INVALID_TERM,
 };
 use crate::send_wal::WalSenders;
 use crate::{control_file, safekeeper::UNKNOWN_SERVER_VERSION};
@@ -37,18 +41,25 @@ use crate::SafeKeeperConf;
 use crate::{debug_dump, wal_storage};
 
 /// Things safekeeper should know about timeline state on peers.
-#[derive(Debug, Clone)]
+#[serde_as]
+#[derive(Debug, Clone, Serialize, Deserialize)]
 pub struct PeerInfo {
     pub sk_id: NodeId,
     /// Term of the last entry.
     _last_log_term: Term,
     /// LSN of the last record.
+    #[serde_as(as = "DisplayFromStr")]
     _flush_lsn: Lsn,
+    #[serde_as(as = "DisplayFromStr")]
     pub commit_lsn: Lsn,
     /// Since which LSN safekeeper has WAL. TODO: remove this once we fill new
     /// sk since backup_lsn.
+    #[serde_as(as = "DisplayFromStr")]
     pub local_start_lsn: Lsn,
-    /// When info was received.
+    /// When info was received. Serde annotations are not very useful but make
+    /// the code compile -- we don't rely on this field externally.
+    #[serde(skip)]
+    #[serde(default = "Instant::now")]
     ts: Instant,
 }
 
@@ -237,8 +248,9 @@ impl SharedState {
                 tenant_id: ttid.tenant_id.as_ref().to_owned(),
                 timeline_id: ttid.timeline_id.as_ref().to_owned(),
             }),
+            term: self.sk.state.acceptor_state.term,
             last_log_term: self.sk.get_epoch(),
-            flush_lsn: self.sk.wal_store.flush_lsn().0,
+            flush_lsn: self.sk.flush_lsn().0,
             // note: this value is not flushed to control file yet and can be lost
             commit_lsn: self.sk.inmem.commit_lsn.0,
             remote_consistent_lsn: remote_consistent_lsn.0,
@@ -247,6 +259,7 @@ impl SharedState {
                 .advertise_pg_addr
                 .to_owned()
                 .unwrap_or(conf.listen_pg_addr.clone()),
+            http_connstr: conf.listen_http_addr.to_owned(),
             backup_lsn: self.sk.inmem.backup_lsn.0,
             local_start_lsn: self.sk.state.local_start_lsn.0,
             availability_zone: conf.availability_zone.clone(),
@@ -296,6 +309,13 @@ pub struct Timeline {
     commit_lsn_watch_tx: watch::Sender<Lsn>,
     commit_lsn_watch_rx: watch::Receiver<Lsn>,
 
+    /// Broadcasts (current term, flush_lsn) updates, walsender is interested in
+    /// them when sending in recovery mode (to walproposer or peers). Note: this
+    /// is just a notification, WAL reading should always done with lock held as
+    /// term can change otherwise.
+    term_flush_lsn_watch_tx: watch::Sender<TermLsn>,
+    term_flush_lsn_watch_rx: watch::Receiver<TermLsn>,
+
     /// Safekeeper and other state, that should remain consistent and
     /// synchronized with the disk. This is tokio mutex as we write WAL to disk
     /// while holding it, ensuring that consensus checks are in order.
@@ -317,16 +337,20 @@ pub struct Timeline {
 impl Timeline {
     /// Load existing timeline from disk.
     pub fn load_timeline(
-        conf: SafeKeeperConf,
+        conf: &SafeKeeperConf,
         ttid: TenantTimelineId,
         wal_backup_launcher_tx: Sender<TenantTimelineId>,
     ) -> Result<Timeline> {
         let _enter = info_span!("load_timeline", timeline = %ttid.timeline_id).entered();
 
-        let shared_state = SharedState::restore(&conf, &ttid)?;
+        let shared_state = SharedState::restore(conf, &ttid)?;
         let rcl = shared_state.sk.state.remote_consistent_lsn;
         let (commit_lsn_watch_tx, commit_lsn_watch_rx) =
             watch::channel(shared_state.sk.state.commit_lsn);
+        let (term_flush_lsn_watch_tx, term_flush_lsn_watch_rx) = watch::channel(TermLsn::from((
+            shared_state.sk.get_term(),
+            shared_state.sk.flush_lsn(),
+        )));
         let (cancellation_tx, cancellation_rx) = watch::channel(false);
 
         Ok(Timeline {
@@ -334,6 +358,8 @@ impl Timeline {
             wal_backup_launcher_tx,
             commit_lsn_watch_tx,
             commit_lsn_watch_rx,
+            term_flush_lsn_watch_tx,
+            term_flush_lsn_watch_rx,
             mutex: Mutex::new(shared_state),
             walsenders: WalSenders::new(rcl),
             walreceivers: WalReceivers::new(),
@@ -345,7 +371,7 @@ impl Timeline {
 
     /// Create a new timeline, which is not yet persisted to disk.
     pub fn create_empty(
-        conf: SafeKeeperConf,
+        conf: &SafeKeeperConf,
         ttid: TenantTimelineId,
         wal_backup_launcher_tx: Sender<TenantTimelineId>,
         server_info: ServerInfo,
@@ -353,6 +379,8 @@ impl Timeline {
         local_start_lsn: Lsn,
     ) -> Result<Timeline> {
         let (commit_lsn_watch_tx, commit_lsn_watch_rx) = watch::channel(Lsn::INVALID);
+        let (term_flush_lsn_watch_tx, term_flush_lsn_watch_rx) =
+            watch::channel(TermLsn::from((INVALID_TERM, Lsn::INVALID)));
         let (cancellation_tx, cancellation_rx) = watch::channel(false);
         let state = SafeKeeperState::new(&ttid, server_info, vec![], commit_lsn, local_start_lsn);
 
@@ -361,7 +389,9 @@ impl Timeline {
             wal_backup_launcher_tx,
             commit_lsn_watch_tx,
             commit_lsn_watch_rx,
-            mutex: Mutex::new(SharedState::create_new(&conf, &ttid, state)?),
+            term_flush_lsn_watch_tx,
+            term_flush_lsn_watch_rx,
+            mutex: Mutex::new(SharedState::create_new(conf, &ttid, state)?),
             walsenders: WalSenders::new(Lsn(0)),
             walreceivers: WalReceivers::new(),
             cancellation_rx,
@@ -370,12 +400,16 @@ impl Timeline {
         })
     }
 
-    /// Initialize fresh timeline on disk and start background tasks. If bootstrap
+    /// Initialize fresh timeline on disk and start background tasks. If init
     /// fails, timeline is cancelled and cannot be used anymore.
     ///
-    /// Bootstrap is transactional, so if it fails, created files will be deleted,
+    /// Init is transactional, so if it fails, created files will be deleted,
     /// and state on disk should remain unchanged.
-    pub async fn bootstrap(&self, shared_state: &mut MutexGuard<'_, SharedState>) -> Result<()> {
+    pub async fn init_new(
+        self: &Arc<Timeline>,
+        shared_state: &mut MutexGuard<'_, SharedState>,
+        conf: &SafeKeeperConf,
+    ) -> Result<()> {
         match fs::metadata(&self.timeline_dir).await {
             Ok(_) => {
                 // Timeline directory exists on disk, we should leave state unchanged
@@ -391,7 +425,7 @@ impl Timeline {
         // Create timeline directory.
         fs::create_dir_all(&self.timeline_dir).await?;
 
-        // Write timeline to disk and TODO: start background tasks.
+        // Write timeline to disk and start background tasks.
         if let Err(e) = shared_state.sk.persist().await {
             // Bootstrap failed, cancel timeline and remove timeline directory.
             self.cancel(shared_state);
@@ -405,12 +439,16 @@ impl Timeline {
 
             return Err(e);
         }
-
-        // TODO: add more initialization steps here
-        self.update_status(shared_state);
+        self.bootstrap(conf);
         Ok(())
     }
 
+    /// Bootstrap new or existing timeline starting background stasks.
+    pub fn bootstrap(self: &Arc<Timeline>, conf: &SafeKeeperConf) {
+        // Start recovery task which always runs on the timeline.
+        tokio::spawn(recovery_main(self.clone(), conf.clone()));
+    }
+
     /// Delete timeline from disk completely, by removing timeline directory. Background
     /// timeline activities will stop eventually.
     pub async fn delete_from_disk(
@@ -444,6 +482,16 @@ impl Timeline {
         *self.cancellation_rx.borrow()
     }
 
+    /// Returns watch channel which gets value when timeline is cancelled. It is
+    /// guaranteed to have not cancelled value observed (errors otherwise).
+    pub fn get_cancellation_rx(&self) -> Result<watch::Receiver<bool>> {
+        let rx = self.cancellation_rx.clone();
+        if *rx.borrow() {
+            bail!(TimelineError::Cancelled(self.ttid));
+        }
+        Ok(rx)
+    }
+
     /// Take a writing mutual exclusive lock on timeline shared_state.
     pub async fn write_shared_state(&self) -> MutexGuard<SharedState> {
         self.mutex.lock().await
@@ -520,6 +568,11 @@ impl Timeline {
         self.commit_lsn_watch_rx.clone()
     }
 
+    /// Returns term_flush_lsn watch channel.
+    pub fn get_term_flush_lsn_watch_rx(&self) -> watch::Receiver<TermLsn> {
+        self.term_flush_lsn_watch_rx.clone()
+    }
+
     /// Pass arrived message to the safekeeper.
     pub async fn process_msg(
         &self,
@@ -531,6 +584,7 @@ impl Timeline {
 
         let mut rmsg: Option<AcceptorProposerMessage>;
         let commit_lsn: Lsn;
+        let term_flush_lsn: TermLsn;
         {
             let mut shared_state = self.write_shared_state().await;
             rmsg = shared_state.sk.process_msg(msg).await?;
@@ -544,8 +598,11 @@ impl Timeline {
             }
 
             commit_lsn = shared_state.sk.inmem.commit_lsn;
+            term_flush_lsn =
+                TermLsn::from((shared_state.sk.get_term(), shared_state.sk.flush_lsn()));
         }
         self.commit_lsn_watch_tx.send(commit_lsn)?;
+        self.term_flush_lsn_watch_tx.send(term_flush_lsn)?;
         Ok(rmsg)
     }
 

safekeeper/src/timelines_global_map.rs

@@ -11,7 +11,7 @@ use serde::Serialize;
 use std::collections::HashMap;
 use std::path::PathBuf;
 use std::str::FromStr;
-use std::sync::{Arc, Mutex, MutexGuard};
+use std::sync::{Arc, Mutex};
 use tokio::sync::mpsc::Sender;
 use tracing::*;
 use utils::id::{TenantId, TenantTimelineId, TimelineId};
@@ -71,19 +71,23 @@ pub struct GlobalTimelines;
 
 impl GlobalTimelines {
     /// Inject dependencies needed for the timeline constructors and load all timelines to memory.
-    pub fn init(
+    pub async fn init(
         conf: SafeKeeperConf,
         wal_backup_launcher_tx: Sender<TenantTimelineId>,
     ) -> Result<()> {
-        let mut state = TIMELINES_STATE.lock().unwrap();
-        assert!(state.wal_backup_launcher_tx.is_none());
-        state.wal_backup_launcher_tx = Some(wal_backup_launcher_tx);
-        state.conf = Some(conf);
+        // clippy isn't smart enough to understand that drop(state) releases the
+        // lock, so use explicit block
+        let tenants_dir = {
+            let mut state = TIMELINES_STATE.lock().unwrap();
+            assert!(state.wal_backup_launcher_tx.is_none());
+            state.wal_backup_launcher_tx = Some(wal_backup_launcher_tx);
+            state.conf = Some(conf);
 
-        // Iterate through all directories and load tenants for all directories
-        // named as a valid tenant_id.
+            // Iterate through all directories and load tenants for all directories
+            // named as a valid tenant_id.
+            state.get_conf().workdir.clone()
+        };
         let mut tenant_count = 0;
-        let tenants_dir = state.get_conf().workdir.clone();
         for tenants_dir_entry in std::fs::read_dir(&tenants_dir)
             .with_context(|| format!("failed to list tenants dir {}", tenants_dir.display()))?
         {
@@ -93,7 +97,7 @@ impl GlobalTimelines {
                         TenantId::from_str(tenants_dir_entry.file_name().to_str().unwrap_or(""))
                     {
                         tenant_count += 1;
-                        GlobalTimelines::load_tenant_timelines(&mut state, tenant_id)?;
+                        GlobalTimelines::load_tenant_timelines(tenant_id).await?;
                     }
                 }
                 Err(e) => error!(
@@ -108,7 +112,7 @@ impl GlobalTimelines {
         info!(
             "found {} tenants directories, successfully loaded {} timelines",
             tenant_count,
-            state.timelines.len()
+            TIMELINES_STATE.lock().unwrap().timelines.len()
         );
         Ok(())
     }
@@ -116,17 +120,21 @@ impl GlobalTimelines {
     /// Loads all timelines for the given tenant to memory. Returns fs::read_dir
     /// errors if any.
     ///
-    /// Note: This function (and all reading/loading below) is sync because
-    /// timelines are loaded while holding GlobalTimelinesState lock. Which is
-    /// fine as this is called only from single threaded main runtime on boot,
-    /// but clippy complains anyway, and suppressing that isn't trivial as async
-    /// is the keyword, ha. That only other user is pull_timeline.rs for which
-    /// being blocked is not that bad, and we can do spawn_blocking.
-    fn load_tenant_timelines(
-        state: &mut MutexGuard<'_, GlobalTimelinesState>,
-        tenant_id: TenantId,
-    ) -> Result<()> {
-        let timelines_dir = state.get_conf().tenant_dir(&tenant_id);
+    /// It is async for update_status_notify sake. Since TIMELINES_STATE lock is
+    /// sync and there is no important reason to make it async (it is always
+    /// held for a short while) we just lock and unlock it for each timeline --
+    /// this function is called during init when nothing else is running, so
+    /// this is fine.
+    async fn load_tenant_timelines(tenant_id: TenantId) -> Result<()> {
+        let (conf, wal_backup_launcher_tx) = {
+            let state = TIMELINES_STATE.lock().unwrap();
+            (
+                state.get_conf().clone(),
+                state.wal_backup_launcher_tx.as_ref().unwrap().clone(),
+            )
+        };
+
+        let timelines_dir = conf.tenant_dir(&tenant_id);
         for timelines_dir_entry in std::fs::read_dir(&timelines_dir)
             .with_context(|| format!("failed to list timelines dir {}", timelines_dir.display()))?
         {
@@ -136,13 +144,16 @@ impl GlobalTimelines {
                         TimelineId::from_str(timeline_dir_entry.file_name().to_str().unwrap_or(""))
                     {
                         let ttid = TenantTimelineId::new(tenant_id, timeline_id);
-                        match Timeline::load_timeline(
-                            state.get_conf().clone(),
-                            ttid,
-                            state.wal_backup_launcher_tx.as_ref().unwrap().clone(),
-                        ) {
+                        match Timeline::load_timeline(&conf, ttid, wal_backup_launcher_tx.clone()) {
                             Ok(timeline) => {
-                                state.timelines.insert(ttid, Arc::new(timeline));
+                                let tli = Arc::new(timeline);
+                                TIMELINES_STATE
+                                    .lock()
+                                    .unwrap()
+                                    .timelines
+                                    .insert(ttid, tli.clone());
+                                tli.bootstrap(&conf);
+                                tli.update_status_notify().await.unwrap();
                             }
                             // If we can't load a timeline, it's most likely because of a corrupted
                             // directory. We will log an error and won't allow to delete/recreate
@@ -168,18 +179,22 @@ impl GlobalTimelines {
     }
 
     /// Load timeline from disk to the memory.
-    pub fn load_timeline(ttid: TenantTimelineId) -> Result<Arc<Timeline>> {
+    pub async fn load_timeline(ttid: TenantTimelineId) -> Result<Arc<Timeline>> {
         let (conf, wal_backup_launcher_tx) = TIMELINES_STATE.lock().unwrap().get_dependencies();
 
-        match Timeline::load_timeline(conf, ttid, wal_backup_launcher_tx) {
+        match Timeline::load_timeline(&conf, ttid, wal_backup_launcher_tx) {
             Ok(timeline) => {
                 let tli = Arc::new(timeline);
+
                 // TODO: prevent concurrent timeline creation/loading
                 TIMELINES_STATE
                     .lock()
                     .unwrap()
                     .timelines
                     .insert(ttid, tli.clone());
+
+                tli.bootstrap(&conf);
+
                 Ok(tli)
             }
             // If we can't load a timeline, it's bad. Caller will figure it out.
@@ -217,7 +232,7 @@ impl GlobalTimelines {
         info!("creating new timeline {}", ttid);
 
         let timeline = Arc::new(Timeline::create_empty(
-            conf,
+            &conf,
             ttid,
             wal_backup_launcher_tx,
             server_info,
@@ -240,23 +255,24 @@ impl GlobalTimelines {
             // Write the new timeline to the disk and start background workers.
             // Bootstrap is transactional, so if it fails, the timeline will be deleted,
             // and the state on disk should remain unchanged.
-            if let Err(e) = timeline.bootstrap(&mut shared_state).await {
-                // Note: the most likely reason for bootstrap failure is that the timeline
+            if let Err(e) = timeline.init_new(&mut shared_state, &conf).await {
+                // Note: the most likely reason for init failure is that the timeline
                 // directory already exists on disk. This happens when timeline is corrupted
                 // and wasn't loaded from disk on startup because of that. We want to preserve
                 // the timeline directory in this case, for further inspection.
 
                 // TODO: this is an unusual error, perhaps we should send it to sentry
                 // TODO: compute will try to create timeline every second, we should add backoff
-                error!("failed to bootstrap timeline {}: {}", ttid, e);
+                error!("failed to init new timeline {}: {}", ttid, e);
 
-                // Timeline failed to bootstrap, it cannot be used. Remove it from the map.
+                // Timeline failed to init, it cannot be used. Remove it from the map.
                 TIMELINES_STATE.lock().unwrap().timelines.remove(&ttid);
                 return Err(e);
             }
             // We are done with bootstrap, release the lock, return the timeline.
             // {} block forces release before .await
         }
+        timeline.update_status_notify().await?;
         timeline.wal_backup_launcher_tx.send(timeline.ttid).await?;
         Ok(timeline)
     }

scripts/flaky_tests.py

@@ -12,25 +12,26 @@ import psycopg2.extras
 # We call the test "flaky" if it failed at least once on the main branch in the last N=10 days.
 FLAKY_TESTS_QUERY = """
     SELECT
-        DISTINCT parent_suite, suite, test
+        DISTINCT parent_suite, suite, REGEXP_REPLACE(test, '(release|debug)-pg(\\d+)-?', '') as deparametrized_test
     FROM
         (
             SELECT
-                revision,
-                jsonb_array_elements(data -> 'children') -> 'name' as parent_suite,
-                jsonb_array_elements(jsonb_array_elements(data -> 'children') -> 'children') -> 'name' as suite,
-                jsonb_array_elements(jsonb_array_elements(jsonb_array_elements(data -> 'children') -> 'children') -> 'children') -> 'name' as test,
-                jsonb_array_elements(jsonb_array_elements(jsonb_array_elements(data -> 'children') -> 'children') -> 'children') -> 'status' as status,
-                jsonb_array_elements(jsonb_array_elements(jsonb_array_elements(data -> 'children') -> 'children') -> 'children') -> 'retriesStatusChange' as retries_status_change,
-                to_timestamp((jsonb_array_elements(jsonb_array_elements(jsonb_array_elements(data -> 'children') -> 'children') -> 'children') -> 'time' -> 'start')::bigint / 1000)::date as timestamp
+                reference,
+                jsonb_array_elements(data -> 'children') ->> 'name' as parent_suite,
+                jsonb_array_elements(jsonb_array_elements(data -> 'children') -> 'children') ->> 'name' as suite,
+                jsonb_array_elements(jsonb_array_elements(jsonb_array_elements(data -> 'children') -> 'children') -> 'children') ->> 'name' as test,
+                jsonb_array_elements(jsonb_array_elements(jsonb_array_elements(data -> 'children') -> 'children') -> 'children') ->> 'status' as status,
+                jsonb_array_elements(jsonb_array_elements(jsonb_array_elements(data -> 'children') -> 'children') -> 'children') ->> 'retriesStatusChange' as retries_status_change,
+                to_timestamp((jsonb_array_elements(jsonb_array_elements(jsonb_array_elements(data -> 'children') -> 'children') -> 'children') -> 'time' ->> 'start')::bigint / 1000)::date as timestamp
             FROM
                 regress_test_results
-            WHERE
-                reference = 'refs/heads/main'
         ) data
     WHERE
         timestamp > CURRENT_DATE - INTERVAL '%s' day
-        AND (status::text IN ('"failed"', '"broken"') OR retries_status_change::boolean)
+        AND (
+            (status IN ('failed', 'broken') AND reference = 'refs/heads/main')
+            OR retries_status_change::boolean
+        )
     ;
 """
 
@@ -40,6 +41,9 @@ def main(args: argparse.Namespace):
     interval_days = args.days
     output = args.output
 
+    build_type = args.build_type
+    pg_version = args.pg_version
+
     res: DefaultDict[str, DefaultDict[str, Dict[str, bool]]]
     res = defaultdict(lambda: defaultdict(dict))
 
@@ -55,8 +59,21 @@ def main(args: argparse.Namespace):
         rows = []
 
     for row in rows:
-        logging.info(f"\t{row['parent_suite'].replace('.', '/')}/{row['suite']}.py::{row['test']}")
-        res[row["parent_suite"]][row["suite"]][row["test"]] = True
+        # We don't want to automatically rerun tests in a performance suite
+        if row["parent_suite"] != "test_runner.regress":
+            continue
+
+        deparametrized_test = row["deparametrized_test"]
+        dash_if_needed = "" if deparametrized_test.endswith("[]") else "-"
+        parametrized_test = deparametrized_test.replace(
+            "[",
+            f"[{build_type}-pg{pg_version}{dash_if_needed}",
+        )
+        res[row["parent_suite"]][row["suite"]][parametrized_test] = True
+
+        logging.info(
+            f"\t{row['parent_suite'].replace('.', '/')}/{row['suite']}.py::{parametrized_test}"
+        )
 
     logging.info(f"saving results to {output.name}")
     json.dump(res, output, indent=2)
@@ -77,6 +94,18 @@ if __name__ == "__main__":
         type=int,
         help="how many days to look back for flaky tests (default: 10)",
     )
+    parser.add_argument(
+        "--build-type",
+        required=True,
+        type=str,
+        help="for which build type to create list of flaky tests (debug or release)",
+    )
+    parser.add_argument(
+        "--pg-version",
+        required=True,
+        type=int,
+        help="for which Postgres version to create list of flaky tests (14, 15, etc.)",
+    )
     parser.add_argument(
         "connstr",
         help="connection string to the test results database",

storage_broker/benches/rps.rs

@@ -125,6 +125,7 @@ async fn publish(client: Option<BrokerClientChannel>, n_keys: u64) {
                     tenant_id: vec![0xFF; 16],
                     timeline_id: tli_from_u64(counter % n_keys),
                 }),
+                term: 0,
                 last_log_term: 0,
                 flush_lsn: counter,
                 commit_lsn: 2,
@@ -132,6 +133,7 @@ async fn publish(client: Option<BrokerClientChannel>, n_keys: u64) {
                 remote_consistent_lsn: 4,
                 peer_horizon_lsn: 5,
                 safekeeper_connstr: "zenith-1-sk-1.local:7676".to_owned(),
+                http_connstr: "zenith-1-sk-1.local:7677".to_owned(),
                 local_start_lsn: 0,
                 availability_zone: None,
             };

storage_broker/proto/broker.proto

@@ -22,6 +22,8 @@ message SubscribeSafekeeperInfoRequest {
 message SafekeeperTimelineInfo {
     uint64 safekeeper_id = 1;
     TenantTimelineId tenant_timeline_id = 2;
+    // Safekeeper term
+    uint64 term = 12;
     // Term of the last entry.
     uint64 last_log_term = 3;
     // LSN of the last record.
@@ -36,6 +38,8 @@ message SafekeeperTimelineInfo {
     uint64 local_start_lsn = 9;
     // A connection string to use for WAL receiving.
     string safekeeper_connstr = 10;
+    // HTTP endpoint connection string
+    string http_connstr = 13;
     // Availability zone of a safekeeper.
     optional string availability_zone = 11;
 }

storage_broker/src/bin/storage_broker.rs

@@ -519,6 +519,7 @@ mod tests {
                 tenant_id: vec![0x00; 16],
                 timeline_id,
             }),
+            term: 0,
             last_log_term: 0,
             flush_lsn: 1,
             commit_lsn: 2,
@@ -526,6 +527,7 @@ mod tests {
             remote_consistent_lsn: 4,
             peer_horizon_lsn: 5,
             safekeeper_connstr: "neon-1-sk-1.local:7676".to_owned(),
+            http_connstr: "neon-1-sk-1.local:7677".to_owned(),
             local_start_lsn: 0,
             availability_zone: None,
         }

test_runner/fixtures/neon_fixtures.py

@@ -428,6 +428,7 @@ class NeonEnvBuilder:
         preserve_database_files: bool = False,
         initial_tenant: Optional[TenantId] = None,
         initial_timeline: Optional[TimelineId] = None,
+        enable_generations: bool = False,
     ):
         self.repo_dir = repo_dir
         self.rust_log_override = rust_log_override
@@ -454,6 +455,7 @@ class NeonEnvBuilder:
         self.preserve_database_files = preserve_database_files
         self.initial_tenant = initial_tenant or TenantId.generate()
         self.initial_timeline = initial_timeline or TimelineId.generate()
+        self.enable_generations = False
 
     def init_configs(self) -> NeonEnv:
         # Cannot create more than one environment from one builder
@@ -713,6 +715,9 @@ class NeonEnvBuilder:
                 sk.stop(immediate=True)
             self.env.pageserver.stop(immediate=True)
 
+            if self.env.attachment_service is not None:
+                self.env.attachment_service.stop(immediate=True)
+
             cleanup_error = None
             try:
                 self.cleanup_remote_storage()
@@ -766,6 +771,8 @@ class NeonEnv:
         the tenant id
     """
 
+    PAGESERVER_ID = 1
+
     def __init__(self, config: NeonEnvBuilder):
         self.repo_dir = config.repo_dir
         self.rust_log_override = config.rust_log_override
@@ -789,6 +796,14 @@ class NeonEnv:
         self.initial_tenant = config.initial_tenant
         self.initial_timeline = config.initial_timeline
 
+        if config.enable_generations:
+            attachment_service_port = self.port_distributor.get_port()
+            self.control_plane_api: Optional[str] = f"http://127.0.0.1:{attachment_service_port}"
+            self.attachment_service: Optional[NeonAttachmentService] = NeonAttachmentService(self)
+        else:
+            self.control_plane_api = None
+            self.attachment_service = None
+
         # Create a config file corresponding to the options
         toml = textwrap.dedent(
             f"""
@@ -814,7 +829,7 @@ class NeonEnv:
         toml += textwrap.dedent(
             f"""
             [pageserver]
-            id=1
+            id={self.PAGESERVER_ID}
             listen_pg_addr = 'localhost:{pageserver_port.pg}'
             listen_http_addr = 'localhost:{pageserver_port.http}'
             pg_auth_type = '{pg_auth_type}'
@@ -822,6 +837,13 @@ class NeonEnv:
         """
         )
 
+        if self.control_plane_api is not None:
+            toml += textwrap.dedent(
+                f"""
+                control_plane_api = '{self.control_plane_api}'
+            """
+            )
+
         # Create a corresponding NeonPageserver object
         self.pageserver = NeonPageserver(
             self, port=pageserver_port, config_override=config.pageserver_config_override
@@ -868,6 +890,9 @@ class NeonEnv:
     def start(self):
         # Start up broker, pageserver and all safekeepers
         self.broker.try_start()
+
+        if self.attachment_service is not None:
+            self.attachment_service.start()
         self.pageserver.start()
 
         for safekeeper in self.safekeepers:
@@ -1289,6 +1314,16 @@ class NeonCli(AbstractNeonCli):
             res.check_returncode()
             return res
 
+    def attachment_service_start(self):
+        cmd = ["attachment_service", "start"]
+        return self.raw_cli(cmd)
+
+    def attachment_service_stop(self, immediate: bool):
+        cmd = ["attachment_service", "stop"]
+        if immediate:
+            cmd.extend(["-m", "immediate"])
+        return self.raw_cli(cmd)
+
     def pageserver_start(
         self,
         overrides: Tuple[str, ...] = (),
@@ -1470,6 +1505,33 @@ class ComputeCtl(AbstractNeonCli):
     COMMAND = "compute_ctl"
 
 
+class NeonAttachmentService:
+    def __init__(self, env: NeonEnv):
+        self.env = env
+
+    def start(self):
+        self.env.neon_cli.attachment_service_start()
+        self.running = True
+        return self
+
+    def stop(self, immediate: bool = False) -> "NeonAttachmentService":
+        if self.running:
+            self.env.neon_cli.attachment_service_stop(immediate)
+            self.running = False
+        return self
+
+    def __enter__(self) -> "NeonAttachmentService":
+        return self
+
+    def __exit__(
+        self,
+        exc_type: Optional[Type[BaseException]],
+        exc: Optional[BaseException],
+        tb: Optional[TracebackType],
+    ):
+        self.stop(immediate=True)
+
+
 class NeonPageserver(PgProtocol):
     """
     An object representing a running pageserver.
@@ -1633,6 +1695,26 @@ class NeonPageserver(PgProtocol):
 
         return None
 
+    def tenant_attach(
+        self, tenant_id: TenantId, config: None | Dict[str, Any] = None, config_null: bool = False
+    ):
+        """
+        Tenant attachment passes through here to acquire a generation number before proceeding
+        to call into the pageserver HTTP client.
+        """
+        if self.env.attachment_service is not None:
+            response = requests.post(
+                f"{self.env.control_plane_api}/attach_hook",
+                json={"tenant_id": str(tenant_id), "pageserver_id": self.env.PAGESERVER_ID},
+            )
+            response.raise_for_status()
+            generation = response.json()["gen"]
+        else:
+            generation = None
+
+        client = self.env.pageserver.http_client()
+        return client.tenant_attach(tenant_id, config, config_null, generation=generation)
+
 
 def append_pageserver_param_overrides(
     params_to_update: List[str],

test_runner/fixtures/pageserver/http.py

@@ -186,18 +186,25 @@ class PageserverHttpClient(requests.Session):
         return TenantId(new_tenant_id)
 
     def tenant_attach(
-        self, tenant_id: TenantId, config: None | Dict[str, Any] = None, config_null: bool = False
+        self,
+        tenant_id: TenantId,
+        config: None | Dict[str, Any] = None,
+        config_null: bool = False,
+        generation: Optional[int] = None,
     ):
         if config_null:
             assert config is None
-            body = "null"
+            body: Any = None
         else:
             # null-config is prohibited by the API
             config = config or {}
-            body = json.dumps({"config": config})
+            body = {"config": config}
+            if generation is not None:
+                body.update({"generation": generation})
+
         res = self.post(
             f"http://localhost:{self.port}/v1/tenant/{tenant_id}/attach",
-            data=body,
+            data=json.dumps(body),
             headers={"Content-Type": "application/json"},
         )
         self.verbose_error(res)
@@ -613,3 +620,8 @@ 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/fixtures/pageserver/utils.py

@@ -233,10 +233,19 @@ if TYPE_CHECKING:
 
 def assert_prefix_empty(neon_env_builder: "NeonEnvBuilder", prefix: Optional[str] = None):
     response = list_prefix(neon_env_builder, prefix)
-    objects = response.get("Contents")
-    assert (
-        response["KeyCount"] == 0
-    ), f"remote dir with prefix {prefix} is not empty after deletion: {objects}"
+    keys = response["KeyCount"]
+    objects = response.get("Contents", [])
+
+    if keys != 0 and len(objects) == 0:
+        # this has been seen in one case with mock_s3:
+        # https://neon-github-public-dev.s3.amazonaws.com/reports/pr-4938/6000769714/index.html#suites/3556ed71f2d69272a7014df6dcb02317/ca01e4f4d8d9a11f
+        # looking at moto impl, it might be there's a race with common prefix (sub directory) not going away with deletes
+        common_prefixes = response.get("CommonPrefixes", [])
+        log.warn(
+            f"contradicting ListObjectsV2 response with KeyCount={keys} and Contents={objects}, CommonPrefixes={common_prefixes}"
+        )
+
+    assert keys == 0, f"remote dir with prefix {prefix} is not empty after deletion: {objects}"
 
 
 def assert_prefix_not_empty(neon_env_builder: "NeonEnvBuilder", prefix: Optional[str] = None):

test_runner/regress/test_pageserver_restart.py

@@ -7,7 +7,10 @@ from fixtures.neon_fixtures import NeonEnvBuilder
 
 # Test restarting page server, while safekeeper and compute node keep
 # running.
-def test_pageserver_restart(neon_env_builder: NeonEnvBuilder):
+@pytest.mark.parametrize("generations", [True, False])
+def test_pageserver_restart(neon_env_builder: NeonEnvBuilder, generations: bool):
+    neon_env_builder.enable_generations = generations
+
     env = neon_env_builder.init_start()
 
     env.neon_cli.create_branch("test_pageserver_restart")

test_runner/regress/test_remote_storage.py

@@ -12,7 +12,10 @@ 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
@@ -52,9 +55,9 @@ from requests import ReadTimeout
 #
 # The tests are done for all types of remote storage pageserver supports.
 @pytest.mark.parametrize("remote_storage_kind", available_remote_storages())
+@pytest.mark.parametrize("generations", [True, False])
 def test_remote_storage_backup_and_restore(
-    neon_env_builder: NeonEnvBuilder,
-    remote_storage_kind: RemoteStorageKind,
+    neon_env_builder: NeonEnvBuilder, remote_storage_kind: RemoteStorageKind, generations: bool
 ):
     # Use this test to check more realistic SK ids: some etcd key parsing bugs were related,
     # and this test needs SK to write data to pageserver, so it will be visible
@@ -65,6 +68,8 @@ def test_remote_storage_backup_and_restore(
         test_name="test_remote_storage_backup_and_restore",
     )
 
+    neon_env_builder.enable_generations = generations
+
     # Exercise retry code path by making all uploads and downloads fail for the
     # first time. The retries print INFO-messages to the log; we will check
     # that they are present after the test.
@@ -155,7 +160,8 @@ def test_remote_storage_backup_and_restore(
     # background task to load the tenant. In that background task,
     # listing the remote timelines will fail because of the failpoint,
     # and the tenant will be marked as Broken.
-    client.tenant_attach(tenant_id)
+    # client.tenant_attach(tenant_id)
+    env.pageserver.tenant_attach(tenant_id)
 
     tenant_info = wait_until_tenant_state(pageserver_http, tenant_id, "Broken", 15)
     assert tenant_info["attachment_status"] == {
@@ -165,7 +171,7 @@ def test_remote_storage_backup_and_restore(
 
     # Ensure that even though the tenant is broken, we can't attach it again.
     with pytest.raises(Exception, match=f"tenant {tenant_id} already exists, state: Broken"):
-        client.tenant_attach(tenant_id)
+        env.pageserver.tenant_attach(tenant_id)
 
     # Restart again, this implicitly clears the failpoint.
     # test_remote_failures=1 remains active, though, as it's in the pageserver config.
@@ -183,7 +189,7 @@ def test_remote_storage_backup_and_restore(
     # Ensure that the pageserver remembers that the tenant was attaching, by
     # trying to attach it again. It should fail.
     with pytest.raises(Exception, match=f"tenant {tenant_id} already exists, state:"):
-        client.tenant_attach(tenant_id)
+        env.pageserver.tenant_attach(tenant_id)
     log.info("waiting for tenant to become active. this should be quick with on-demand download")
 
     wait_until_tenant_active(
@@ -250,35 +256,20 @@ def test_remote_storage_upload_queue_retries(
 
     client = env.pageserver.http_client()
 
-    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):
+    def configure_storage_write_failpoints(action):
         client.configure_failpoints(
             [
                 ("before-upload-layer", action),
                 ("before-upload-index", action),
-                ("before-delete-layer", action),
             ]
         )
 
-    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(
+    def configure_storage_delete_failpoints(action):
+        client.configure_failpoints(
             [
-                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",
+                ("deletion-queue-before-execute", action),
             ]
         )
-        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(
@@ -291,23 +282,52 @@ def test_remote_storage_upload_queue_retries(
         assert val is not None, "expecting metric to be present"
         return int(val)
 
-    # 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_deletions_executed() -> int:
+        executed = client.get_metric_value("pageserver_deletion_queue_executed_total")
+        if executed is None:
+            return 0
+        else:
+            return int(executed)
 
-    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)
+    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
 
     # let all future operations queue up
-    configure_storage_sync_failpoints("return")
+    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"
+    )
 
     # Create more churn to generate all upload ops.
     # The checkpoint / compact / gc ops will block because they call remote_client.wait_completion().
@@ -315,38 +335,77 @@ def test_remote_storage_upload_queue_retries(
     churn_thread_result = [False]
 
     def churn_while_failpoints_active(result):
-        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
+        generate_uploads_and_deletions(
+            env, init=False, tenant_id=tenant_id, timeline_id=timeline_id, data="d"
+        )
         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()
 
-    # 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)
+    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))
 
-    # unblock churn operations
-    configure_storage_sync_failpoints("off")
+        # 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))
 
-    # ... 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)
+        # 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")
 
-    # 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]
 
@@ -364,7 +423,7 @@ def test_remote_storage_upload_queue_retries(
     env.pageserver.start()
     client = env.pageserver.http_client()
 
-    client.tenant_attach(tenant_id)
+    env.pageserver.tenant_attach(tenant_id)
 
     wait_until_tenant_active(client, tenant_id)
 
@@ -432,7 +491,6 @@ 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():
@@ -502,7 +560,7 @@ def test_remote_timeline_client_calls_started_metric(
     env.pageserver.start()
     client = env.pageserver.http_client()
 
-    client.tenant_attach(tenant_id)
+    env.pageserver.tenant_attach(tenant_id)
 
     wait_until_tenant_active(client, tenant_id)
 
@@ -930,4 +988,154 @@ 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

@@ -47,6 +47,15 @@ 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(
@@ -91,7 +100,9 @@ def test_tenant_delete_smoke(
 
     iterations = poll_for_remote_storage_iterations(remote_storage_kind)
 
-    tenant_delete_wait_completed(ps_http, tenant_id, iterations)
+    # 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_path = env.tenant_dir(tenant_id=tenant_id)
     assert not tenant_path.exists()
@@ -201,6 +212,17 @@ 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

@@ -488,7 +488,14 @@ 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)
 
-    wait_timeline_detail_404(ps_http, env.initial_tenant, leaf_timeline_id, iterations=4)
+    # 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,
+    )
 
     assert (
         not leaf_timeline_path.exists()
@@ -534,7 +541,7 @@ def test_timeline_delete_fail_before_local_delete(neon_env_builder: NeonEnvBuild
     wait_until(
         2,
         0.5,
-        lambda: assert_prefix_empty(neon_env_builder),
+        lambda: assert_prefix_empty(neon_env_builder, prefix="/tenants"),
     )
 
 
@@ -688,7 +695,7 @@ 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=2)
+    wait_timeline_detail_404(ps_http, env.initial_tenant, child_timeline_id, iterations=4)
 
 
 @pytest.mark.parametrize(
@@ -772,7 +779,11 @@ 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))
+    wait_until(
+        2,
+        0.5,
+        lambda: assert_prefix_empty(neon_env_builder, "/tenants"),
+    )
 
 
 def test_delete_orphaned_objects(
@@ -827,6 +838,8 @@ 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(

vm-cgconfig.conf

@@ -1,12 +0,0 @@
-# Configuration for cgroups in VM compute nodes
-group neon-postgres {
-    perm {
-        admin {
-            uid = vm-informant;
-        }
-        task {
-            gid = users;
-        }
-    }
-    memory {}
-}