proxy: parse proxy protocol header for health checks

## Problem
We attempted validation for cancelled errors under the assumption that
if vectored get fails, sequential get will too.
That's not right 100% of times though because sequential get may have
the values cached and slip them through
even when shutting down.

## Summary of changes
Don't validate if either search impl failed due to tenant shutdown.

.github/ISSUE_TEMPLATE/epic-template.md

@@ -16,9 +16,9 @@ assignees: ''
 
 ## Implementation ideas
 
-
+## Tasks
 ```[tasklist]
-### Tasks
+- [ ] Example Task
 ```
 
 

.github/workflows/build_and_test.yml

@@ -1132,11 +1132,9 @@ jobs:
               -f branch=main \
               -f dockerTag=${{needs.tag.outputs.build-tag}}
           elif [[ "$GITHUB_REF_NAME" == "release-proxy" ]]; then
-            gh workflow --repo neondatabase/aws run deploy-prod.yml --ref main \
+            gh workflow --repo neondatabase/aws run deploy-proxy-prod.yml --ref main \
               -f deployPgSniRouter=true \
               -f deployProxy=true \
-              -f deployStorage=false \
-              -f deployStorageBroker=false \
               -f branch=main \
               -f dockerTag=${{needs.tag.outputs.build-tag}}
           else

.github/workflows/release.yml

@@ -97,7 +97,7 @@ jobs:
           **Please merge this Pull Request using 'Create a merge commit' button**
         EOF
 
-        gh pr create --title "Proxy release ${RELEASE_DATE}}" \
+        gh pr create --title "Proxy release ${RELEASE_DATE}" \
                      --body-file "body.md" \
                      --head "${RELEASE_BRANCH}" \
                      --base "release-proxy"

Cargo.lock

@@ -25,9 +25,9 @@ checksum = "f26201604c87b1e01bd3d98f8d5d9a8fcbb815e8cedb41ffccbeb4bf593a35fe"
 
 [[package]]
 name = "ahash"
-version = "0.8.5"
+version = "0.8.9"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "cd7d5a2cecb58716e47d67d5703a249964b14c7be1ec3cad3affc295b2d1c35d"
+checksum = "d713b3834d76b85304d4d525563c1276e2e30dc97cc67bfb4585a4a29fc2c89f"
 dependencies = [
  "cfg-if",
  "const-random",
@@ -1389,9 +1389,9 @@ dependencies = [
 
 [[package]]
 name = "crc32c"
-version = "0.6.3"
+version = "0.6.5"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "3dfea2db42e9927a3845fb268a10a72faed6d416065f77873f05e411457c363e"
+checksum = "89254598aa9b9fa608de44b3ae54c810f0f06d755e24c50177f1f8f31ff50ce2"
 dependencies = [
  "rustc_version",
 ]
@@ -2959,9 +2959,9 @@ dependencies = [
 
 [[package]]
 name = "mio"
-version = "0.8.10"
+version = "0.8.11"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "8f3d0b296e374a4e6f3c7b0a1f5a51d748a0d34c85e7dc48fc3fa9a87657fe09"
+checksum = "a4a650543ca06a924e8b371db273b2756685faae30f8487da1b56505a8f78b0c"
 dependencies = [
  "libc",
  "log",
@@ -4216,7 +4216,6 @@ dependencies = [
  "thiserror",
  "tikv-jemalloc-ctl",
  "tikv-jemallocator",
- "tls-listener",
  "tokio",
  "tokio-postgres",
  "tokio-postgres-rustls",
@@ -5794,25 +5793,11 @@ version = "0.1.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "1f3ccbac311fea05f86f61904b462b55fb3df8837a366dfc601a0161d0532f20"
 
-[[package]]
-name = "tls-listener"
-version = "0.7.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "81294c017957a1a69794f506723519255879e15a870507faf45dfed288b763dd"
-dependencies = [
- "futures-util",
- "hyper",
- "pin-project-lite",
- "thiserror",
- "tokio",
- "tokio-rustls",
-]
-
 [[package]]
 name = "tokio"
-version = "1.34.0"
+version = "1.36.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d0c014766411e834f7af5b8f4cf46257aab4036ca95e9d2c144a10f59ad6f5b9"
+checksum = "61285f6515fa018fb2d1e46eb21223fff441ee8db5d0f1435e8ab4f5cdb80931"
 dependencies = [
  "backtrace",
  "bytes",

Cargo.toml

@@ -156,7 +156,6 @@ test-context = "0.1"
 thiserror = "1.0"
 tikv-jemallocator = "0.5"
 tikv-jemalloc-ctl = "0.5"
-tls-listener = { version = "0.7", features = ["rustls", "hyper-h1"] }
 tokio = { version = "1.17", features = ["macros"] }
 tokio-epoll-uring = { git = "https://github.com/neondatabase/tokio-epoll-uring.git" , branch = "main" }
 tokio-io-timeout = "1.2.0"

README.md

@@ -230,6 +230,8 @@ postgres=# select * from t;
 > cargo neon stop
 ```
 
+More advanced usages can be found at [Control Plane and Neon Local](./control_plane/README.md).
+
 #### Handling build failures
 
 If you encounter errors during setting up the initial tenant, it's best to stop everything (`cargo neon stop`) and remove the `.neon` directory. Then fix the problems, and start the setup again.

clippy.toml

@@ -3,3 +3,10 @@ disallowed-methods = [
     # Allow this for now, to deny it later once we stop using Handle::block_on completely
     # "tokio::runtime::Handle::block_on",
 ]
+
+disallowed-macros = [
+    # use std::pin::pin
+    "futures::pin_mut",
+    # cannot disallow this, because clippy finds used from tokio macros
+    #"tokio::pin",
+]

compute_tools/src/compute.rs

@@ -17,9 +17,8 @@ use chrono::{DateTime, Utc};
 use futures::future::join_all;
 use futures::stream::FuturesUnordered;
 use futures::StreamExt;
+use postgres::error::SqlState;
 use postgres::{Client, NoTls};
-use tokio;
-use tokio_postgres;
 use tracing::{debug, error, info, instrument, warn};
 use utils::id::{TenantId, TimelineId};
 use utils::lsn::Lsn;
@@ -765,6 +764,26 @@ impl ComputeNode {
         Ok((pg, logs_handle))
     }
 
+    /// Do post configuration of the already started Postgres. This function spawns a background thread to
+    /// configure the database after applying the compute spec. Currently, it upgrades the neon extension
+    /// version. In the future, it may upgrade all 3rd-party extensions.
+    #[instrument(skip_all)]
+    pub fn post_apply_config(&self) -> Result<()> {
+        let connstr = self.connstr.clone();
+        thread::spawn(move || {
+            let func = || {
+                let mut client = Client::connect(connstr.as_str(), NoTls)?;
+                handle_neon_extension_upgrade(&mut client)
+                    .context("handle_neon_extension_upgrade")?;
+                Ok::<_, anyhow::Error>(())
+            };
+            if let Err(err) = func() {
+                error!("error while post_apply_config: {err:#}");
+            }
+        });
+        Ok(())
+    }
+
     /// Do initial configuration of the already started Postgres.
     #[instrument(skip_all)]
     pub fn apply_config(&self, compute_state: &ComputeState) -> Result<()> {
@@ -776,27 +795,34 @@ impl ComputeNode {
         // but we can create a new one and grant it all privileges.
         let connstr = self.connstr.clone();
         let mut client = match Client::connect(connstr.as_str(), NoTls) {
-            Err(e) => {
-                info!(
-                    "cannot connect to postgres: {}, retrying with `zenith_admin` username",
-                    e
-                );
-                let mut zenith_admin_connstr = connstr.clone();
+            Err(e) => match e.code() {
+                Some(&SqlState::INVALID_PASSWORD)
+                | Some(&SqlState::INVALID_AUTHORIZATION_SPECIFICATION) => {
+                    // connect with zenith_admin if cloud_admin could not authenticate
+                    info!(
+                        "cannot connect to postgres: {}, retrying with `zenith_admin` username",
+                        e
+                    );
+                    let mut zenith_admin_connstr = connstr.clone();
 
-                zenith_admin_connstr
-                    .set_username("zenith_admin")
-                    .map_err(|_| anyhow::anyhow!("invalid connstr"))?;
+                    zenith_admin_connstr
+                        .set_username("zenith_admin")
+                        .map_err(|_| anyhow::anyhow!("invalid connstr"))?;
 
-                let mut client = Client::connect(zenith_admin_connstr.as_str(), NoTls)?;
-                // Disable forwarding so that users don't get a cloud_admin role
-                client.simple_query("SET neon.forward_ddl = false")?;
-                client.simple_query("CREATE USER cloud_admin WITH SUPERUSER")?;
-                client.simple_query("GRANT zenith_admin TO cloud_admin")?;
-                drop(client);
+                    let mut client =
+                        Client::connect(zenith_admin_connstr.as_str(), NoTls)
+                            .context("broken cloud_admin credential: tried connecting with cloud_admin but could not authenticate, and zenith_admin does not work either")?;
+                    // Disable forwarding so that users don't get a cloud_admin role
+                    client.simple_query("SET neon.forward_ddl = false")?;
+                    client.simple_query("CREATE USER cloud_admin WITH SUPERUSER")?;
+                    client.simple_query("GRANT zenith_admin TO cloud_admin")?;
+                    drop(client);
 
-                // reconnect with connstring with expected name
-                Client::connect(connstr.as_str(), NoTls)?
-            }
+                    // reconnect with connstring with expected name
+                    Client::connect(connstr.as_str(), NoTls)?
+                }
+                _ => return Err(e.into()),
+            },
             Ok(client) => client,
         };
 
@@ -992,18 +1018,21 @@ impl ComputeNode {
         let pg_process = self.start_postgres(pspec.storage_auth_token.clone())?;
 
         let config_time = Utc::now();
-        if pspec.spec.mode == ComputeMode::Primary && !pspec.spec.skip_pg_catalog_updates {
-            let pgdata_path = Path::new(&self.pgdata);
-            // temporarily reset max_cluster_size in config
-            // to avoid the possibility of hitting the limit, while we are applying config:
-            // creating new extensions, roles, etc...
-            config::compute_ctl_temp_override_create(pgdata_path, "neon.max_cluster_size=-1")?;
-            self.pg_reload_conf()?;
+        if pspec.spec.mode == ComputeMode::Primary {
+            if !pspec.spec.skip_pg_catalog_updates {
+                let pgdata_path = Path::new(&self.pgdata);
+                // temporarily reset max_cluster_size in config
+                // to avoid the possibility of hitting the limit, while we are applying config:
+                // creating new extensions, roles, etc...
+                config::compute_ctl_temp_override_create(pgdata_path, "neon.max_cluster_size=-1")?;
+                self.pg_reload_conf()?;
 
-            self.apply_config(&compute_state)?;
+                self.apply_config(&compute_state)?;
 
-            config::compute_ctl_temp_override_remove(pgdata_path)?;
-            self.pg_reload_conf()?;
+                config::compute_ctl_temp_override_remove(pgdata_path)?;
+                self.pg_reload_conf()?;
+            }
+            self.post_apply_config()?;
         }
 
         let startup_end_time = Utc::now();

compute_tools/src/extension_server.rs

@@ -71,7 +71,7 @@ More specifically, here is an example ext_index.json
     }
 }
 */
-use anyhow::{self, Result};
+use anyhow::Result;
 use anyhow::{bail, Context};
 use bytes::Bytes;
 use compute_api::spec::RemoteExtSpec;

compute_tools/src/http/api.rs

@@ -13,8 +13,6 @@ use compute_api::responses::{ComputeStatus, ComputeStatusResponse, GenericAPIErr
 use anyhow::Result;
 use hyper::service::{make_service_fn, service_fn};
 use hyper::{Body, Method, Request, Response, Server, StatusCode};
-use num_cpus;
-use serde_json;
 use tokio::task;
 use tracing::{error, info, warn};
 use tracing_utils::http::OtelName;

compute_tools/src/spec.rs

@@ -744,7 +744,17 @@ pub fn handle_extension_neon(client: &mut Client) -> Result<()> {
     // - extension was just installed
     // - extension was already installed and is up to date
     let query = "ALTER EXTENSION neon UPDATE";
-    info!("update neon extension schema with query: {}", query);
+    info!("update neon extension version with query: {}", query);
+    client.simple_query(query)?;
+
+    Ok(())
+}
+
+#[instrument(skip_all)]
+pub fn handle_neon_extension_upgrade(client: &mut Client) -> Result<()> {
+    info!("handle neon extension upgrade");
+    let query = "ALTER EXTENSION neon UPDATE";
+    info!("update neon extension version with query: {}", query);
     client.simple_query(query)?;
 
     Ok(())

control_plane/README.md

@@ -0,0 +1,26 @@
+# Control Plane and Neon Local
+
+This crate contains tools to start a Neon development environment locally. This utility can be used with the `cargo neon` command.
+
+## Example: Start with Postgres 16
+
+To create and start a local development environment with Postgres 16, you will need to provide `--pg-version` flag to 3 of the start-up commands.
+
+```shell
+cargo neon init --pg-version 16
+cargo neon start
+cargo neon tenant create --set-default --pg-version 16
+cargo neon endpoint create main --pg-version 16
+cargo neon endpoint start main
+```
+
+## Example: Create Test User and Database
+
+By default, `cargo neon` starts an endpoint with `cloud_admin` and `postgres` database. If you want to have a role and a database similar to what we have on the cloud service, you can do it with the following commands when starting an endpoint.
+
+```shell
+cargo neon endpoint create main --pg-version 16 --update-catalog true
+cargo neon endpoint start main --create-test-user true
+```
+
+The first command creates `neon_superuser` and necessary roles. The second command creates `test` user and `neondb` database. You will see a connection string that connects you to the test user after running the second command.

control_plane/attachment_service/migrations/2024-02-29-094122_generations_null/down.sql

@@ -0,0 +1,2 @@
+ALTER TABLE tenant_shards ALTER generation SET NOT NULL;
+ALTER TABLE tenant_shards ALTER generation_pageserver SET NOT NULL;

control_plane/attachment_service/migrations/2024-02-29-094122_generations_null/up.sql

@@ -0,0 +1,4 @@
+
+
+ALTER TABLE tenant_shards ALTER generation DROP NOT NULL;
+ALTER TABLE tenant_shards ALTER generation_pageserver DROP NOT NULL;

control_plane/attachment_service/src/compute_hook.rs

@@ -3,7 +3,7 @@ use std::{collections::HashMap, time::Duration};
 use control_plane::endpoint::{ComputeControlPlane, EndpointStatus};
 use control_plane::local_env::LocalEnv;
 use hyper::{Method, StatusCode};
-use pageserver_api::shard::{ShardIndex, ShardNumber, TenantShardId};
+use pageserver_api::shard::{ShardCount, ShardNumber, ShardStripeSize, TenantShardId};
 use postgres_connection::parse_host_port;
 use serde::{Deserialize, Serialize};
 use tokio_util::sync::CancellationToken;
@@ -19,8 +19,66 @@ const SLOWDOWN_DELAY: Duration = Duration::from_secs(5);
 
 pub(crate) const API_CONCURRENCY: usize = 32;
 
-pub(super) struct ComputeHookTenant {
-    shards: Vec<(ShardIndex, NodeId)>,
+struct ShardedComputeHookTenant {
+    stripe_size: ShardStripeSize,
+    shard_count: ShardCount,
+    shards: Vec<(ShardNumber, NodeId)>,
+}
+
+enum ComputeHookTenant {
+    Unsharded(NodeId),
+    Sharded(ShardedComputeHookTenant),
+}
+
+impl ComputeHookTenant {
+    /// Construct with at least one shard's information
+    fn new(tenant_shard_id: TenantShardId, stripe_size: ShardStripeSize, node_id: NodeId) -> Self {
+        if tenant_shard_id.shard_count.count() > 1 {
+            Self::Sharded(ShardedComputeHookTenant {
+                shards: vec![(tenant_shard_id.shard_number, node_id)],
+                stripe_size,
+                shard_count: tenant_shard_id.shard_count,
+            })
+        } else {
+            Self::Unsharded(node_id)
+        }
+    }
+
+    /// Set one shard's location.  If stripe size or shard count have changed, Self is reset
+    /// and drops existing content.
+    fn update(
+        &mut self,
+        tenant_shard_id: TenantShardId,
+        stripe_size: ShardStripeSize,
+        node_id: NodeId,
+    ) {
+        match self {
+            Self::Unsharded(existing_node_id) if tenant_shard_id.shard_count.count() == 1 => {
+                *existing_node_id = node_id
+            }
+            Self::Sharded(sharded_tenant)
+                if sharded_tenant.stripe_size == stripe_size
+                    && sharded_tenant.shard_count == tenant_shard_id.shard_count =>
+            {
+                if let Some(existing) = sharded_tenant
+                    .shards
+                    .iter()
+                    .position(|s| s.0 == tenant_shard_id.shard_number)
+                {
+                    sharded_tenant.shards.get_mut(existing).unwrap().1 = node_id;
+                } else {
+                    sharded_tenant
+                        .shards
+                        .push((tenant_shard_id.shard_number, node_id));
+                    sharded_tenant.shards.sort_by_key(|s| s.0)
+                }
+            }
+            _ => {
+                // Shard count changed: reset struct.
+                *self = Self::new(tenant_shard_id, stripe_size, node_id);
+            }
+        }
+    }
 }
 
 #[derive(Serialize, Deserialize, Debug)]
@@ -33,6 +91,7 @@ struct ComputeHookNotifyRequestShard {
 #[derive(Serialize, Deserialize, Debug)]
 struct ComputeHookNotifyRequest {
     tenant_id: TenantId,
+    stripe_size: Option<ShardStripeSize>,
     shards: Vec<ComputeHookNotifyRequestShard>,
 }
 
@@ -63,42 +122,43 @@ pub(crate) enum NotifyError {
 }
 
 impl ComputeHookTenant {
-    async fn maybe_reconfigure(&mut self, tenant_id: TenantId) -> Option<ComputeHookNotifyRequest> {
-        // Find the highest shard count and drop any shards that aren't
-        // for that shard count.
-        let shard_count = self.shards.iter().map(|(k, _v)| k.shard_count).max();
-        let Some(shard_count) = shard_count else {
-            // No shards, nothing to do.
-            tracing::info!("ComputeHookTenant::maybe_reconfigure: no shards");
-            return None;
-        };
-
-        self.shards.retain(|(k, _v)| k.shard_count == shard_count);
-        self.shards
-            .sort_by_key(|(shard, _node_id)| shard.shard_number);
-
-        if self.shards.len() == shard_count.count() as usize || shard_count.is_unsharded() {
-            // We have pageservers for all the shards: emit a configuration update
-            return Some(ComputeHookNotifyRequest {
+    fn maybe_reconfigure(&self, tenant_id: TenantId) -> Option<ComputeHookNotifyRequest> {
+        match self {
+            Self::Unsharded(node_id) => Some(ComputeHookNotifyRequest {
                 tenant_id,
-                shards: self
-                    .shards
-                    .iter()
-                    .map(|(shard, node_id)| ComputeHookNotifyRequestShard {
-                        shard_number: shard.shard_number,
-                        node_id: *node_id,
-                    })
-                    .collect(),
-            });
-        } else {
-            tracing::info!(
-                "ComputeHookTenant::maybe_reconfigure: not enough shards ({}/{})",
-                self.shards.len(),
-                shard_count.count()
-            );
-        }
+                shards: vec![ComputeHookNotifyRequestShard {
+                    shard_number: ShardNumber(0),
+                    node_id: *node_id,
+                }],
+                stripe_size: None,
+            }),
+            Self::Sharded(sharded_tenant)
+                if sharded_tenant.shards.len() == sharded_tenant.shard_count.count() as usize =>
+            {
+                Some(ComputeHookNotifyRequest {
+                    tenant_id,
+                    shards: sharded_tenant
+                        .shards
+                        .iter()
+                        .map(|(shard_number, node_id)| ComputeHookNotifyRequestShard {
+                            shard_number: *shard_number,
+                            node_id: *node_id,
+                        })
+                        .collect(),
+                    stripe_size: Some(sharded_tenant.stripe_size),
+                })
+            }
+            Self::Sharded(sharded_tenant) => {
+                // Sharded tenant doesn't yet have information for all its shards
 
-        None
+                tracing::info!(
+                    "ComputeHookTenant::maybe_reconfigure: not enough shards ({}/{})",
+                    sharded_tenant.shards.len(),
+                    sharded_tenant.shard_count.count()
+                );
+                None
+            }
+        }
     }
 }
 
@@ -139,7 +199,11 @@ impl ComputeHook {
         };
         let cplane =
             ComputeControlPlane::load(env.clone()).expect("Error loading compute control plane");
-        let ComputeHookNotifyRequest { tenant_id, shards } = reconfigure_request;
+        let ComputeHookNotifyRequest {
+            tenant_id,
+            shards,
+            stripe_size,
+        } = reconfigure_request;
 
         let compute_pageservers = shards
             .into_iter()
@@ -156,7 +220,9 @@ impl ComputeHook {
         for (endpoint_name, endpoint) in &cplane.endpoints {
             if endpoint.tenant_id == tenant_id && endpoint.status() == EndpointStatus::Running {
                 tracing::info!("Reconfiguring endpoint {}", endpoint_name,);
-                endpoint.reconfigure(compute_pageservers.clone()).await?;
+                endpoint
+                    .reconfigure(compute_pageservers.clone(), stripe_size)
+                    .await?;
             }
         }
 
@@ -271,30 +337,26 @@ impl ComputeHook {
         &self,
         tenant_shard_id: TenantShardId,
         node_id: NodeId,
+        stripe_size: ShardStripeSize,
         cancel: &CancellationToken,
     ) -> Result<(), NotifyError> {
         let mut locked = self.state.lock().await;
-        let entry = locked
-            .entry(tenant_shard_id.tenant_id)
-            .or_insert_with(|| ComputeHookTenant { shards: Vec::new() });
 
-        let shard_index = ShardIndex {
-            shard_count: tenant_shard_id.shard_count,
-            shard_number: tenant_shard_id.shard_number,
+        use std::collections::hash_map::Entry;
+        let tenant = match locked.entry(tenant_shard_id.tenant_id) {
+            Entry::Vacant(e) => e.insert(ComputeHookTenant::new(
+                tenant_shard_id,
+                stripe_size,
+                node_id,
+            )),
+            Entry::Occupied(e) => {
+                let tenant = e.into_mut();
+                tenant.update(tenant_shard_id, stripe_size, node_id);
+                tenant
+            }
         };
 
-        let mut set = false;
-        for (existing_shard, existing_node) in &mut entry.shards {
-            if *existing_shard == shard_index {
-                *existing_node = node_id;
-                set = true;
-            }
-        }
-        if !set {
-            entry.shards.push((shard_index, node_id));
-        }
-
-        let reconfigure_request = entry.maybe_reconfigure(tenant_shard_id.tenant_id).await;
+        let reconfigure_request = tenant.maybe_reconfigure(tenant_shard_id.tenant_id);
         let Some(reconfigure_request) = reconfigure_request else {
             // The tenant doesn't yet have pageservers for all its shards: we won't notify anything
             // until it does.
@@ -316,3 +378,85 @@ impl ComputeHook {
         }
     }
 }
+
+#[cfg(test)]
+pub(crate) mod tests {
+    use pageserver_api::shard::{ShardCount, ShardNumber};
+    use utils::id::TenantId;
+
+    use super::*;
+
+    #[test]
+    fn tenant_updates() -> anyhow::Result<()> {
+        let tenant_id = TenantId::generate();
+        let mut tenant_state = ComputeHookTenant::new(
+            TenantShardId {
+                tenant_id,
+                shard_count: ShardCount::new(0),
+                shard_number: ShardNumber(0),
+            },
+            ShardStripeSize(12345),
+            NodeId(1),
+        );
+
+        // An unsharded tenant is always ready to emit a notification
+        assert!(tenant_state.maybe_reconfigure(tenant_id).is_some());
+        assert_eq!(
+            tenant_state
+                .maybe_reconfigure(tenant_id)
+                .unwrap()
+                .shards
+                .len(),
+            1
+        );
+        assert!(tenant_state
+            .maybe_reconfigure(tenant_id)
+            .unwrap()
+            .stripe_size
+            .is_none());
+
+        // Writing the first shard of a multi-sharded situation (i.e. in a split)
+        // resets the tenant state and puts it in an non-notifying state (need to
+        // see all shards)
+        tenant_state.update(
+            TenantShardId {
+                tenant_id,
+                shard_count: ShardCount::new(2),
+                shard_number: ShardNumber(1),
+            },
+            ShardStripeSize(32768),
+            NodeId(1),
+        );
+        assert!(tenant_state.maybe_reconfigure(tenant_id).is_none());
+
+        // Writing the second shard makes it ready to notify
+        tenant_state.update(
+            TenantShardId {
+                tenant_id,
+                shard_count: ShardCount::new(2),
+                shard_number: ShardNumber(0),
+            },
+            ShardStripeSize(32768),
+            NodeId(1),
+        );
+
+        assert!(tenant_state.maybe_reconfigure(tenant_id).is_some());
+        assert_eq!(
+            tenant_state
+                .maybe_reconfigure(tenant_id)
+                .unwrap()
+                .shards
+                .len(),
+            2
+        );
+        assert_eq!(
+            tenant_state
+                .maybe_reconfigure(tenant_id)
+                .unwrap()
+                .stripe_size,
+            Some(ShardStripeSize(32768))
+        );
+
+        Ok(())
+    }
+}

control_plane/attachment_service/src/http.rs

@@ -1,9 +1,10 @@
 use crate::reconciler::ReconcileError;
 use crate::service::{Service, STARTUP_RECONCILE_TIMEOUT};
+use crate::PlacementPolicy;
 use hyper::{Body, Request, Response};
 use hyper::{StatusCode, Uri};
 use pageserver_api::models::{
-    TenantCreateRequest, TenantLocationConfigRequest, TenantShardSplitRequest,
+    TenantConfigRequest, TenantCreateRequest, TenantLocationConfigRequest, TenantShardSplitRequest,
     TenantTimeTravelRequest, TimelineCreateRequest,
 };
 use pageserver_api::shard::TenantShardId;
@@ -117,9 +118,14 @@ async fn handle_tenant_create(
     check_permissions(&req, Scope::PageServerApi)?;
 
     let create_req = json_request::<TenantCreateRequest>(&mut req).await?;
+
+    // TODO: enable specifying this.  Using Single as a default helps legacy tests to work (they
+    // have no expectation of HA).
+    let placement_policy = PlacementPolicy::Single;
+
     json_response(
         StatusCode::CREATED,
-        service.tenant_create(create_req).await?,
+        service.tenant_create(create_req, placement_policy).await?,
     )
 }
 
@@ -185,6 +191,27 @@ async fn handle_tenant_location_config(
     )
 }
 
+async fn handle_tenant_config_set(
+    service: Arc<Service>,
+    mut req: Request<Body>,
+) -> Result<Response<Body>, ApiError> {
+    check_permissions(&req, Scope::PageServerApi)?;
+
+    let config_req = json_request::<TenantConfigRequest>(&mut req).await?;
+
+    json_response(StatusCode::OK, service.tenant_config_set(config_req).await?)
+}
+
+async fn handle_tenant_config_get(
+    service: Arc<Service>,
+    req: Request<Body>,
+) -> Result<Response<Body>, ApiError> {
+    let tenant_id: TenantId = parse_request_param(&req, "tenant_id")?;
+    check_permissions(&req, Scope::PageServerApi)?;
+
+    json_response(StatusCode::OK, service.tenant_config_get(tenant_id)?)
+}
+
 async fn handle_tenant_time_travel_remote_storage(
     service: Arc<Service>,
     mut req: Request<Body>,
@@ -216,7 +243,15 @@ async fn handle_tenant_time_travel_remote_storage(
             done_if_after_raw,
         )
         .await?;
+    json_response(StatusCode::OK, ())
+}
 
+async fn handle_tenant_secondary_download(
+    service: Arc<Service>,
+    req: Request<Body>,
+) -> Result<Response<Body>, ApiError> {
+    let tenant_id: TenantId = parse_request_param(&req, "tenant_id")?;
+    service.tenant_secondary_download(tenant_id).await?;
     json_response(StatusCode::OK, ())
 }
 
@@ -551,12 +586,21 @@ pub fn make_router(
         .delete("/v1/tenant/:tenant_id", |r| {
             tenant_service_handler(r, handle_tenant_delete)
         })
+        .put("/v1/tenant/config", |r| {
+            tenant_service_handler(r, handle_tenant_config_set)
+        })
+        .get("/v1/tenant/:tenant_id/config", |r| {
+            tenant_service_handler(r, handle_tenant_config_get)
+        })
         .put("/v1/tenant/:tenant_id/location_config", |r| {
             tenant_service_handler(r, handle_tenant_location_config)
         })
         .put("/v1/tenant/:tenant_id/time_travel_remote_storage", |r| {
             tenant_service_handler(r, handle_tenant_time_travel_remote_storage)
         })
+        .post("/v1/tenant/:tenant_id/secondary/download", |r| {
+            tenant_service_handler(r, handle_tenant_secondary_download)
+        })
         // Timeline operations
         .delete("/v1/tenant/:tenant_id/timeline/:timeline_id", |r| {
             tenant_service_handler(r, handle_tenant_timeline_delete)

control_plane/attachment_service/src/lib.rs

@@ -13,14 +13,20 @@ mod schema;
 pub mod service;
 mod tenant_state;
 
-#[derive(Clone, Serialize, Deserialize, Debug)]
+#[derive(Clone, Serialize, Deserialize, Debug, PartialEq, Eq)]
 enum PlacementPolicy {
     /// Cheapest way to attach a tenant: just one pageserver, no secondary
     Single,
     /// Production-ready way to attach a tenant: one attached pageserver and
     /// some number of secondaries.
     Double(usize),
-    /// Do not attach to any pageservers
+    /// Create one secondary mode locations. This is useful when onboarding
+    /// a tenant, or for an idle tenant that we might want to bring online quickly.
+    Secondary,
+
+    /// Do not attach to any pageservers.  This is appropriate for tenants that
+    /// have been idle for a long time, where we do not mind some delay in making
+    /// them available in future.
     Detached,
 }
 

control_plane/attachment_service/src/main.rs

@@ -9,7 +9,7 @@ use attachment_service::http::make_router;
 use attachment_service::metrics::preinitialize_metrics;
 use attachment_service::persistence::Persistence;
 use attachment_service::service::{Config, Service};
-use aws_config::{self, BehaviorVersion, Region};
+use aws_config::{BehaviorVersion, Region};
 use camino::Utf8PathBuf;
 use clap::Parser;
 use diesel::Connection;

control_plane/attachment_service/src/persistence.rs

@@ -7,8 +7,10 @@ use self::split_state::SplitState;
 use camino::Utf8Path;
 use camino::Utf8PathBuf;
 use diesel::pg::PgConnection;
-use diesel::prelude::*;
-use diesel::Connection;
+use diesel::{
+    Connection, ExpressionMethods, Insertable, QueryDsl, QueryResult, Queryable, RunQueryDsl,
+    Selectable, SelectableHelper,
+};
 use pageserver_api::controller_api::NodeSchedulingPolicy;
 use pageserver_api::models::TenantConfig;
 use pageserver_api::shard::{ShardCount, ShardNumber, TenantShardId};
@@ -331,7 +333,15 @@ impl Persistence {
                 shard_number: ShardNumber(tsp.shard_number as u8),
                 shard_count: ShardCount::new(tsp.shard_count as u8),
             };
-            result.insert(tenant_shard_id, Generation::new(tsp.generation as u32));
+
+            let Some(g) = tsp.generation else {
+                // If the generation_pageserver column was non-NULL, then the generation column should also be non-NULL:
+                // we only set generation_pageserver when setting generation.
+                return Err(DatabaseError::Logical(
+                    "Generation should always be set after incrementing".to_string(),
+                ));
+            };
+            result.insert(tenant_shard_id, Generation::new(g as u32));
         }
 
         Ok(result)
@@ -364,7 +374,85 @@ impl Persistence {
             })
             .await?;
 
-        Ok(Generation::new(updated.generation as u32))
+        // Generation is always non-null in the rseult: if the generation column had been NULL, then we
+        // should have experienced an SQL Confilict error while executing a query that tries to increment it.
+        debug_assert!(updated.generation.is_some());
+        let Some(g) = updated.generation else {
+            return Err(DatabaseError::Logical(
+                "Generation should always be set after incrementing".to_string(),
+            )
+            .into());
+        };
+
+        Ok(Generation::new(g as u32))
+    }
+
+    /// For use when updating a persistent property of a tenant, such as its config or placement_policy.
+    ///
+    /// Do not use this for settting generation, unless in the special onboarding code path (/location_config)
+    /// API: use [`Self::increment_generation`] instead.  Setting the generation via this route is a one-time thing
+    /// that we only do the first time a tenant is set to an attached policy via /location_config.
+    pub(crate) async fn update_tenant_shard(
+        &self,
+        tenant_shard_id: TenantShardId,
+        input_placement_policy: PlacementPolicy,
+        input_config: TenantConfig,
+        input_generation: Option<Generation>,
+    ) -> DatabaseResult<()> {
+        use crate::schema::tenant_shards::dsl::*;
+
+        self.with_conn(move |conn| {
+            let query = diesel::update(tenant_shards)
+                .filter(tenant_id.eq(tenant_shard_id.tenant_id.to_string()))
+                .filter(shard_number.eq(tenant_shard_id.shard_number.0 as i32))
+                .filter(shard_count.eq(tenant_shard_id.shard_count.literal() as i32));
+
+            if let Some(input_generation) = input_generation {
+                // Update includes generation column
+                query
+                    .set((
+                        generation.eq(Some(input_generation.into().unwrap() as i32)),
+                        placement_policy
+                            .eq(serde_json::to_string(&input_placement_policy).unwrap()),
+                        config.eq(serde_json::to_string(&input_config).unwrap()),
+                    ))
+                    .execute(conn)?;
+            } else {
+                // Update does not include generation column
+                query
+                    .set((
+                        placement_policy
+                            .eq(serde_json::to_string(&input_placement_policy).unwrap()),
+                        config.eq(serde_json::to_string(&input_config).unwrap()),
+                    ))
+                    .execute(conn)?;
+            }
+
+            Ok(())
+        })
+        .await?;
+
+        Ok(())
+    }
+
+    pub(crate) async fn update_tenant_config(
+        &self,
+        input_tenant_id: TenantId,
+        input_config: TenantConfig,
+    ) -> DatabaseResult<()> {
+        use crate::schema::tenant_shards::dsl::*;
+
+        self.with_conn(move |conn| {
+            diesel::update(tenant_shards)
+                .filter(tenant_id.eq(input_tenant_id.to_string()))
+                .set((config.eq(serde_json::to_string(&input_config).unwrap()),))
+                .execute(conn)?;
+
+            Ok(())
+        })
+        .await?;
+
+        Ok(())
     }
 
     pub(crate) async fn detach(&self, tenant_shard_id: TenantShardId) -> anyhow::Result<()> {
@@ -375,7 +463,7 @@ impl Persistence {
                 .filter(shard_number.eq(tenant_shard_id.shard_number.0 as i32))
                 .filter(shard_count.eq(tenant_shard_id.shard_count.literal() as i32))
                 .set((
-                    generation_pageserver.eq(i64::MAX),
+                    generation_pageserver.eq(Option::<i64>::None),
                     placement_policy.eq(serde_json::to_string(&PlacementPolicy::Detached).unwrap()),
                 ))
                 .execute(conn)?;
@@ -501,12 +589,15 @@ pub(crate) struct TenantShardPersistence {
     pub(crate) shard_stripe_size: i32,
 
     // Latest generation number: next time we attach, increment this
-    // and use the incremented number when attaching
-    pub(crate) generation: i32,
+    // and use the incremented number when attaching.
+    //
+    // Generation is only None when first onboarding a tenant, where it may
+    // be in PlacementPolicy::Secondary and therefore have no valid generation state.
+    pub(crate) generation: Option<i32>,
 
     // Currently attached pageserver
     #[serde(rename = "pageserver")]
-    pub(crate) generation_pageserver: i64,
+    pub(crate) generation_pageserver: Option<i64>,
 
     #[serde(default)]
     pub(crate) placement_policy: String,

control_plane/attachment_service/src/reconciler.rs

@@ -26,7 +26,7 @@ pub(super) struct Reconciler {
     /// of a tenant's state from when we spawned a reconcile task.
     pub(super) tenant_shard_id: TenantShardId,
     pub(crate) shard: ShardIdentity,
-    pub(crate) generation: Generation,
+    pub(crate) generation: Option<Generation>,
     pub(crate) intent: TargetState,
     pub(crate) config: TenantConfig,
     pub(crate) observed: ObservedState,
@@ -104,6 +104,7 @@ impl Reconciler {
         node_id: NodeId,
         config: LocationConfig,
         flush_ms: Option<Duration>,
+        lazy: bool,
     ) -> anyhow::Result<()> {
         let node = self
             .pageservers
@@ -118,7 +119,7 @@ impl Reconciler {
         let client =
             mgmt_api::Client::new(node.base_url(), self.service_config.jwt_token.as_deref());
         client
-            .location_config(self.tenant_shard_id, config.clone(), flush_ms)
+            .location_config(self.tenant_shard_id, config.clone(), flush_ms, lazy)
             .await?;
         tracing::info!("location_config({}) complete: {:?}", node_id, config);
 
@@ -312,11 +313,16 @@ impl Reconciler {
             &self.shard,
             &self.config,
             LocationConfigMode::AttachedStale,
-            Some(self.generation),
+            self.generation,
             None,
         );
-        self.location_config(origin_ps_id, stale_conf, Some(Duration::from_secs(10)))
-            .await?;
+        self.location_config(
+            origin_ps_id,
+            stale_conf,
+            Some(Duration::from_secs(10)),
+            false,
+        )
+        .await?;
 
         let baseline_lsns = Some(self.get_lsns(self.tenant_shard_id, &origin_ps_id).await?);
 
@@ -335,21 +341,23 @@ impl Reconciler {
         }
 
         // Increment generation before attaching to new pageserver
-        self.generation = self
-            .persistence
-            .increment_generation(self.tenant_shard_id, dest_ps_id)
-            .await?;
+        self.generation = Some(
+            self.persistence
+                .increment_generation(self.tenant_shard_id, dest_ps_id)
+                .await?,
+        );
 
         let dest_conf = build_location_config(
             &self.shard,
             &self.config,
             LocationConfigMode::AttachedMulti,
-            Some(self.generation),
+            self.generation,
             None,
         );
 
         tracing::info!("🔁 Attaching to pageserver {}", dest_ps_id);
-        self.location_config(dest_ps_id, dest_conf, None).await?;
+        self.location_config(dest_ps_id, dest_conf, None, false)
+            .await?;
 
         if let Some(baseline) = baseline_lsns {
             tracing::info!("🕑 Waiting for LSN to catch up...");
@@ -381,7 +389,7 @@ impl Reconciler {
             None,
             Some(LocationConfigSecondary { warm: true }),
         );
-        self.location_config(origin_ps_id, origin_secondary_conf.clone(), None)
+        self.location_config(origin_ps_id, origin_secondary_conf.clone(), None, false)
             .await?;
         // TODO: we should also be setting the ObservedState on earlier API calls, in case we fail
         // partway through.  In fact, all location conf API calls should be in a wrapper that sets
@@ -401,10 +409,10 @@ impl Reconciler {
             &self.shard,
             &self.config,
             LocationConfigMode::AttachedSingle,
-            Some(self.generation),
+            self.generation,
             None,
         );
-        self.location_config(dest_ps_id, dest_final_conf.clone(), None)
+        self.location_config(dest_ps_id, dest_final_conf.clone(), None, false)
             .await?;
         self.observed.locations.insert(
             dest_ps_id,
@@ -433,24 +441,67 @@ impl Reconciler {
 
         // If the attached pageserver is not attached, do so now.
         if let Some(node_id) = self.intent.attached {
-            let mut wanted_conf =
-                attached_location_conf(self.generation, &self.shard, &self.config);
+            // If we are in an attached policy, then generation must have been set (null generations
+            // are only present when a tenant is initially loaded with a secondary policy)
+            debug_assert!(self.generation.is_some());
+            let Some(generation) = self.generation else {
+                return Err(ReconcileError::Other(anyhow::anyhow!(
+                    "Attempted to attach with NULL generation"
+                )));
+            };
+
+            let mut wanted_conf = attached_location_conf(generation, &self.shard, &self.config);
             match self.observed.locations.get(&node_id) {
                 Some(conf) if conf.conf.as_ref() == Some(&wanted_conf) => {
                     // Nothing to do
                     tracing::info!(%node_id, "Observed configuration already correct.")
                 }
-                _ => {
+                observed => {
                     // In all cases other than a matching observed configuration, we will
                     // reconcile this location.  This includes locations with different configurations, as well
                     // as locations with unknown (None) observed state.
-                    self.generation = self
-                        .persistence
-                        .increment_generation(self.tenant_shard_id, node_id)
-                        .await?;
-                    wanted_conf.generation = self.generation.into();
+
+                    // The general case is to increment the generation.  However, there are cases
+                    // where this is not necessary:
+                    // - if we are only updating the TenantConf part of the location
+                    // - if we are only changing the attachment mode (e.g. going to attachedmulti or attachedstale)
+                    //   and the location was already in the correct generation
+                    let increment_generation = match observed {
+                        None => true,
+                        Some(ObservedStateLocation { conf: None }) => true,
+                        Some(ObservedStateLocation {
+                            conf: Some(observed),
+                        }) => {
+                            let generations_match = observed.generation == wanted_conf.generation;
+
+                            use LocationConfigMode::*;
+                            let mode_transition_requires_gen_inc =
+                                match (observed.mode, wanted_conf.mode) {
+                                    // Usually the short-lived attachment modes (multi and stale) are only used
+                                    // in the case of [`Self::live_migrate`], but it is simple to handle them correctly
+                                    // here too.  Locations are allowed to go Single->Stale and Multi->Single within the same generation.
+                                    (AttachedSingle, AttachedStale) => false,
+                                    (AttachedMulti, AttachedSingle) => false,
+                                    (lhs, rhs) => lhs != rhs,
+                                };
+
+                            !generations_match || mode_transition_requires_gen_inc
+                        }
+                    };
+
+                    if increment_generation {
+                        let generation = self
+                            .persistence
+                            .increment_generation(self.tenant_shard_id, node_id)
+                            .await?;
+                        self.generation = Some(generation);
+                        wanted_conf.generation = generation.into();
+                    }
                     tracing::info!(%node_id, "Observed configuration requires update.");
-                    self.location_config(node_id, wanted_conf, None).await?;
+                    // Use lazy=true, because we may run many of Self concurrently, and do not want to
+                    // overload the pageserver with logical size calculations.
+                    self.location_config(node_id, wanted_conf, None, true)
+                        .await?;
                     self.compute_notify().await?;
                 }
             }
@@ -502,7 +553,7 @@ impl Reconciler {
             if self.cancel.is_cancelled() {
                 return Err(ReconcileError::Cancel);
             }
-            self.location_config(node_id, conf, None).await?;
+            self.location_config(node_id, conf, None, false).await?;
         }
 
         Ok(())
@@ -514,7 +565,12 @@ impl Reconciler {
         if let Some(node_id) = self.intent.attached {
             let result = self
                 .compute_hook
-                .notify(self.tenant_shard_id, node_id, &self.cancel)
+                .notify(
+                    self.tenant_shard_id,
+                    node_id,
+                    self.shard.stripe_size,
+                    &self.cancel,
+                )
                 .await;
             if let Err(e) = &result {
                 // It is up to the caller whether they want to drop out on this error, but they don't have to:

control_plane/attachment_service/src/scheduler.rs

@@ -284,7 +284,6 @@ pub(crate) mod test_utils {
 #[cfg(test)]
 mod tests {
     use super::*;
-    use utils::id::NodeId;
 
     use crate::tenant_state::IntentState;
     #[test]

control_plane/attachment_service/src/schema.rs

@@ -17,8 +17,8 @@ diesel::table! {
         shard_number -> Int4,
         shard_count -> Int4,
         shard_stripe_size -> Int4,
-        generation -> Int4,
-        generation_pageserver -> Int8,
+        generation -> Nullable<Int4>,
+        generation_pageserver -> Nullable<Int8>,
         placement_policy -> Varchar,
         splitting -> Int2,
         config -> Text,

control_plane/attachment_service/src/service.rs

@@ -14,10 +14,13 @@ use control_plane::attachment_service::{
 use diesel::result::DatabaseErrorKind;
 use futures::{stream::FuturesUnordered, StreamExt};
 use hyper::StatusCode;
-use pageserver_api::controller_api::{
-    NodeAvailability, NodeConfigureRequest, NodeRegisterRequest, NodeSchedulingPolicy,
-    TenantCreateResponse, TenantCreateResponseShard, TenantLocateResponse,
-    TenantLocateResponseShard, TenantShardMigrateRequest, TenantShardMigrateResponse,
+use pageserver_api::{
+    controller_api::{
+        NodeAvailability, NodeConfigureRequest, NodeRegisterRequest, NodeSchedulingPolicy,
+        TenantCreateResponse, TenantCreateResponseShard, TenantLocateResponse,
+        TenantLocateResponseShard, TenantShardMigrateRequest, TenantShardMigrateResponse,
+    },
+    models::TenantConfigRequest,
 };
 use pageserver_api::{
     models::{
@@ -65,6 +68,11 @@ const SHORT_RECONCILE_TIMEOUT: Duration = Duration::from_secs(5);
 // some data in it.
 const RECONCILE_TIMEOUT: Duration = Duration::from_secs(30);
 
+// If we receive a call using Secondary mode initially, it will omit generation.  We will initialize
+// tenant shards into this generation, and as long as it remains in this generation, we will accept
+// input generation from future requests as authoritative.
+const INITIAL_GENERATION: Generation = Generation::new(0);
+
 /// How long [`Service::startup_reconcile`] is allowed to take before it should give
 /// up on unresponsive pageservers and proceed.
 pub(crate) const STARTUP_RECONCILE_TIMEOUT: Duration = Duration::from_secs(30);
@@ -167,6 +175,21 @@ impl From<ReconcileWaitError> for ApiError {
     }
 }
 
+#[allow(clippy::large_enum_variant)]
+enum TenantCreateOrUpdate {
+    Create((TenantCreateRequest, PlacementPolicy)),
+    Update(Vec<ShardUpdate>),
+}
+
+struct ShardUpdate {
+    tenant_shard_id: TenantShardId,
+    placement_policy: PlacementPolicy,
+    tenant_config: TenantConfig,
+
+    /// If this is None, generation is not updated.
+    generation: Option<Generation>,
+}
+
 impl Service {
     pub fn get_config(&self) -> &Config {
         &self.config
@@ -260,7 +283,11 @@ impl Service {
                     // emit a compute notification for this. In the case where our observed state does not
                     // yet match our intent, we will eventually reconcile, and that will emit a compute notification.
                     if let Some(attached_at) = tenant_state.stably_attached() {
-                        compute_notifications.push((*tenant_shard_id, attached_at));
+                        compute_notifications.push((
+                            *tenant_shard_id,
+                            attached_at,
+                            tenant_state.shard.stripe_size,
+                        ));
                     }
                 }
             }
@@ -445,6 +472,7 @@ impl Service {
                         tenant_conf: models::TenantConfig::default(),
                     },
                     None,
+                    false,
                 )
                 .await
             {
@@ -469,7 +497,7 @@ impl Service {
     /// Returns a set of any shards for which notifications where not acked within the deadline.
     async fn compute_notify_many(
         &self,
-        notifications: Vec<(TenantShardId, NodeId)>,
+        notifications: Vec<(TenantShardId, NodeId, ShardStripeSize)>,
         deadline: Instant,
     ) -> HashSet<TenantShardId> {
         let compute_hook = self.inner.read().unwrap().compute_hook.clone();
@@ -480,11 +508,14 @@ impl Service {
         // Construct an async stream of futures to invoke the compute notify function: we do this
         // in order to subsequently use .buffered() on the stream to execute with bounded parallelism.
         let mut stream = futures::stream::iter(notifications.into_iter())
-            .map(|(tenant_shard_id, node_id)| {
+            .map(|(tenant_shard_id, node_id, stripe_size)| {
                 let compute_hook = compute_hook.clone();
                 let cancel = self.cancel.clone();
                 async move {
-                    if let Err(e) = compute_hook.notify(tenant_shard_id, node_id, &cancel).await {
+                    if let Err(e) = compute_hook
+                        .notify(tenant_shard_id, node_id, stripe_size, &cancel)
+                        .await
+                    {
                         tracing::error!(
                             %tenant_shard_id,
                             %node_id,
@@ -571,6 +602,9 @@ impl Service {
         // the shard so that a future [`TenantState::maybe_reconcile`] will try again.
         tenant.pending_compute_notification = result.pending_compute_notification;
 
+        // Let the TenantState know it is idle.
+        tenant.reconcile_complete(result.sequence);
+
         match result.result {
             Ok(()) => {
                 for (node_id, loc) in &result.observed.locations {
@@ -661,8 +695,8 @@ impl Service {
             // after when pageservers start up and register.
             let mut node_ids = HashSet::new();
             for tsp in &tenant_shard_persistence {
-                if tsp.generation_pageserver != i64::MAX {
-                    node_ids.insert(tsp.generation_pageserver);
+                if let Some(node_id) = tsp.generation_pageserver {
+                    node_ids.insert(node_id);
                 }
             }
             for node_id in node_ids {
@@ -699,18 +733,15 @@ impl Service {
             // We will populate intent properly later in [`Self::startup_reconcile`], initially populate
             // it with what we can infer: the node for which a generation was most recently issued.
             let mut intent = IntentState::new();
-            if tsp.generation_pageserver != i64::MAX {
-                intent.set_attached(
-                    &mut scheduler,
-                    Some(NodeId(tsp.generation_pageserver as u64)),
-                );
+            if let Some(generation_pageserver) = tsp.generation_pageserver {
+                intent.set_attached(&mut scheduler, Some(NodeId(generation_pageserver as u64)));
             }
 
             let new_tenant = TenantState {
                 tenant_shard_id,
                 shard: shard_identity,
                 sequence: Sequence::initial(),
-                generation: Generation::new(tsp.generation as u32),
+                generation: tsp.generation.map(|g| Generation::new(g as u32)),
                 policy: serde_json::from_str(&tsp.placement_policy).unwrap(),
                 intent,
                 observed: ObservedState::new(),
@@ -790,8 +821,8 @@ impl Service {
                 shard_number: attach_req.tenant_shard_id.shard_number.0 as i32,
                 shard_count: attach_req.tenant_shard_id.shard_count.literal() as i32,
                 shard_stripe_size: 0,
-                generation: 0,
-                generation_pageserver: i64::MAX,
+                generation: Some(0),
+                generation_pageserver: None,
                 placement_policy: serde_json::to_string(&PlacementPolicy::default()).unwrap(),
                 config: serde_json::to_string(&TenantConfig::default()).unwrap(),
                 splitting: SplitState::default(),
@@ -846,7 +877,7 @@ impl Service {
             .expect("Checked for existence above");
 
         if let Some(new_generation) = new_generation {
-            tenant_state.generation = new_generation;
+            tenant_state.generation = Some(new_generation);
         } else {
             // This is a detach notification.  We must update placement policy to avoid re-attaching
             // during background scheduling/reconciliation, or during attachment service restart.
@@ -896,7 +927,7 @@ impl Service {
                     node_id,
                     ObservedStateLocation {
                         conf: Some(attached_location_conf(
-                            tenant_state.generation,
+                            tenant_state.generation.unwrap(),
                             &tenant_state.shard,
                             &tenant_state.config,
                         )),
@@ -910,7 +941,7 @@ impl Service {
         Ok(AttachHookResponse {
             gen: attach_req
                 .node_id
-                .map(|_| tenant_state.generation.into().unwrap()),
+                .map(|_| tenant_state.generation.expect("Test hook, not used on tenants that are mid-onboarding with a NULL generation").into().unwrap()),
         })
     }
 
@@ -923,7 +954,7 @@ impl Service {
             attachment: tenant_state.and_then(|s| {
                 s.intent
                     .get_attached()
-                    .map(|ps| (s.generation.into().unwrap(), ps))
+                    .map(|ps| (s.generation.expect("Test hook, not used on tenants that are mid-onboarding with a NULL generation").into().unwrap(), ps))
             }),
         }
     }
@@ -973,7 +1004,17 @@ impl Service {
                 continue;
             };
 
-            shard_state.generation = std::cmp::max(shard_state.generation, new_gen);
+            // If [`Persistence::re_attach`] selected this shard, it must have alread
+            // had a generation set.
+            debug_assert!(shard_state.generation.is_some());
+            let Some(old_gen) = shard_state.generation else {
+                // Should never happen:  would only return incremented generation
+                // for a tenant that already had a non-null generation.
+                return Err(ApiError::InternalServerError(anyhow::anyhow!(
+                    "Generation must be set while re-attaching"
+                )));
+            };
+            shard_state.generation = Some(std::cmp::max(old_gen, new_gen));
             if let Some(observed) = shard_state
                 .observed
                 .locations
@@ -1003,7 +1044,7 @@ impl Service {
 
         for req_tenant in validate_req.tenants {
             if let Some(tenant_state) = locked.tenants.get(&req_tenant.id) {
-                let valid = tenant_state.generation == Generation::new(req_tenant.gen);
+                let valid = tenant_state.generation == Some(Generation::new(req_tenant.gen));
                 tracing::info!(
                     "handle_validate: {}(gen {}): valid={valid} (latest {:?})",
                     req_tenant.id,
@@ -1030,8 +1071,9 @@ impl Service {
     pub(crate) async fn tenant_create(
         &self,
         create_req: TenantCreateRequest,
+        placement_policy: PlacementPolicy,
     ) -> Result<TenantCreateResponse, ApiError> {
-        let (response, waiters) = self.do_tenant_create(create_req).await?;
+        let (response, waiters) = self.do_tenant_create(create_req, placement_policy).await?;
 
         self.await_waiters(waiters, SHORT_RECONCILE_TIMEOUT).await?;
         Ok(response)
@@ -1040,6 +1082,7 @@ impl Service {
     pub(crate) async fn do_tenant_create(
         &self,
         create_req: TenantCreateRequest,
+        placement_policy: PlacementPolicy,
     ) -> Result<(TenantCreateResponse, Vec<ReconcilerWaiter>), ApiError> {
         // This service expects to handle sharding itself: it is an error to try and directly create
         // a particular shard here.
@@ -1065,9 +1108,27 @@ impl Service {
             })
             .collect::<Vec<_>>();
 
-        // TODO: enable specifying this.  Using Single as a default helps legacy tests to work (they
-        // have no expectation of HA).
-        let placement_policy: PlacementPolicy = PlacementPolicy::Single;
+        // If the caller specifies a None generation, it means "start from default".  This is different
+        // to [`Self::tenant_location_config`], where a None generation is used to represent
+        // an incompletely-onboarded tenant.
+        let initial_generation = if matches!(placement_policy, PlacementPolicy::Secondary) {
+            tracing::info!(
+                "tenant_create: secondary mode, generation is_some={}",
+                create_req.generation.is_some()
+            );
+            create_req.generation.map(Generation::new)
+        } else {
+            tracing::info!(
+                "tenant_create: not secondary mode, generation is_some={}",
+                create_req.generation.is_some()
+            );
+            Some(
+                create_req
+                    .generation
+                    .map(Generation::new)
+                    .unwrap_or(INITIAL_GENERATION),
+            )
+        };
 
         // Ordering: we persist tenant shards before creating them on the pageserver.  This enables a caller
         // to clean up after themselves by issuing a tenant deletion if something goes wrong and we restart
@@ -1079,8 +1140,10 @@ impl Service {
                 shard_number: tenant_shard_id.shard_number.0 as i32,
                 shard_count: tenant_shard_id.shard_count.literal() as i32,
                 shard_stripe_size: create_req.shard_parameters.stripe_size.0 as i32,
-                generation: create_req.generation.map(|g| g as i32).unwrap_or(0),
-                generation_pageserver: i64::MAX,
+                generation: initial_generation.map(|g| g.into().unwrap() as i32),
+                // The pageserver is not known until scheduling happens: we will set this column when
+                // incrementing the generation the first time we attach to a pageserver.
+                generation_pageserver: None,
                 placement_policy: serde_json::to_string(&placement_policy).unwrap(),
                 config: serde_json::to_string(&create_req.config).unwrap(),
                 splitting: SplitState::default(),
@@ -1096,9 +1159,12 @@ impl Service {
 
         let (waiters, response_shards) = {
             let mut locked = self.inner.write().unwrap();
-            let (_nodes, tenants, scheduler) = locked.parts_mut();
+            let result_tx = locked.result_tx.clone();
+            let compute_hook = locked.compute_hook.clone();
+            let (nodes, tenants, scheduler) = locked.parts_mut();
 
             let mut response_shards = Vec::new();
+            let mut schcedule_error = None;
 
             for tenant_shard_id in create_ids {
                 tracing::info!("Creating shard {tenant_shard_id}...");
@@ -1120,65 +1186,69 @@ impl Service {
                             ))
                         })?;
 
-                        response_shards.push(TenantCreateResponseShard {
-                            shard_id: tenant_shard_id,
-                            node_id: entry
+                        if let Some(node_id) = entry.get().intent.get_attached() {
+                            let generation = entry
                                 .get()
-                                .intent
-                                .get_attached()
-                                .expect("We just set pageserver if it was None"),
-                            generation: entry.get().generation.into().unwrap(),
-                        });
+                                .generation
+                                .expect("Generation is set when in attached mode");
+                            response_shards.push(TenantCreateResponseShard {
+                                shard_id: tenant_shard_id,
+                                node_id: *node_id,
+                                generation: generation.into().unwrap(),
+                            });
+                        }
 
                         continue;
                     }
                     Entry::Vacant(entry) => {
-                        let mut state = TenantState::new(
+                        let state = entry.insert(TenantState::new(
                             tenant_shard_id,
                             ShardIdentity::from_params(
                                 tenant_shard_id.shard_number,
                                 &create_req.shard_parameters,
                             ),
                             placement_policy.clone(),
-                        );
+                        ));
 
-                        if let Some(create_gen) = create_req.generation {
-                            state.generation = Generation::new(create_gen);
-                        }
+                        state.generation = initial_generation;
                         state.config = create_req.config.clone();
+                        if let Err(e) = state.schedule(scheduler) {
+                            schcedule_error = Some(e);
+                        }
 
-                        state.schedule(scheduler).map_err(|e| {
-                            ApiError::Conflict(format!(
-                                "Failed to schedule shard {tenant_shard_id}: {e}"
-                            ))
-                        })?;
-
-                        response_shards.push(TenantCreateResponseShard {
-                            shard_id: tenant_shard_id,
-                            node_id: state
-                                .intent
-                                .get_attached()
-                                .expect("We just set pageserver if it was None"),
-                            generation: state.generation.into().unwrap(),
-                        });
-                        entry.insert(state)
+                        // Only include shards in result if we are attaching: the purpose
+                        // of the response is to tell the caller where the shards are attached.
+                        if let Some(node_id) = state.intent.get_attached() {
+                            let generation = state
+                                .generation
+                                .expect("Generation is set when in attached mode");
+                            response_shards.push(TenantCreateResponseShard {
+                                shard_id: tenant_shard_id,
+                                node_id: *node_id,
+                                generation: generation.into().unwrap(),
+                            });
+                        }
                     }
                 };
             }
 
-            // Take a snapshot of pageservers
-            let pageservers = locked.nodes.clone();
+            // If we failed to schedule shards, then they are still created in the controller,
+            // but we return an error to the requester to avoid a silent failure when someone
+            // tries to e.g. create a tenant whose placement policy requires more nodes than
+            // are present in the system.  We do this here rather than in the above loop, to
+            // avoid situations where we only create a subset of shards in the tenant.
+            if let Some(e) = schcedule_error {
+                return Err(ApiError::Conflict(format!(
+                    "Failed to schedule shard(s): {e}"
+                )));
+            }
 
-            let result_tx = locked.result_tx.clone();
-            let compute_hook = locked.compute_hook.clone();
-
-            let waiters = locked
-                .tenants
+            let waiters = tenants
                 .range_mut(TenantShardId::tenant_range(tenant_id))
                 .filter_map(|(_shard_id, shard)| {
                     shard.maybe_reconcile(
                         result_tx.clone(),
-                        &pageservers,
+                        nodes,
                         &compute_hook,
                         &self.config,
                         &self.persistence,
@@ -1214,12 +1284,114 @@ impl Service {
         Ok(())
     }
 
-    /// This API is used by the cloud control plane to do coarse-grained control of tenants:
-    /// - Call with mode Attached* to upsert the tenant.
-    /// - Call with mode Detached to switch to PolicyMode::Detached
+    /// Part of [`Self::tenant_location_config`]: dissect an incoming location config request,
+    /// and transform it into either a tenant creation of a series of shard updates.
+    fn tenant_location_config_prepare(
+        &self,
+        tenant_id: TenantId,
+        req: TenantLocationConfigRequest,
+    ) -> TenantCreateOrUpdate {
+        let mut updates = Vec::new();
+        let mut locked = self.inner.write().unwrap();
+        let (nodes, tenants, _scheduler) = locked.parts_mut();
+
+        // Use location config mode as an indicator of policy.
+        let placement_policy = match req.config.mode {
+            LocationConfigMode::Detached => PlacementPolicy::Detached,
+            LocationConfigMode::Secondary => PlacementPolicy::Secondary,
+            LocationConfigMode::AttachedMulti
+            | LocationConfigMode::AttachedSingle
+            | LocationConfigMode::AttachedStale => {
+                if nodes.len() > 1 {
+                    PlacementPolicy::Double(1)
+                } else {
+                    // Convenience for dev/test: if we just have one pageserver, import
+                    // tenants into Single mode so that scheduling will succeed.
+                    PlacementPolicy::Single
+                }
+            }
+        };
+
+        let mut create = true;
+        for (shard_id, shard) in tenants.range_mut(TenantShardId::tenant_range(tenant_id)) {
+            // Saw an existing shard: this is not a creation
+            create = false;
+
+            // Shards may have initially been created by a Secondary request, where we
+            // would have left generation as None.
+            //
+            // We only update generation the first time we see an attached-mode request,
+            // and if there is no existing generation set. The caller is responsible for
+            // ensuring that no non-storage-controller pageserver ever uses a higher
+            // generation than they passed in here.
+            use LocationConfigMode::*;
+            let set_generation = match req.config.mode {
+                AttachedMulti | AttachedSingle | AttachedStale if shard.generation.is_none() => {
+                    req.config.generation.map(Generation::new)
+                }
+                _ => None,
+            };
+
+            if shard.policy != placement_policy
+                || shard.config != req.config.tenant_conf
+                || set_generation.is_some()
+            {
+                updates.push(ShardUpdate {
+                    tenant_shard_id: *shard_id,
+                    placement_policy: placement_policy.clone(),
+                    tenant_config: req.config.tenant_conf.clone(),
+                    generation: set_generation,
+                });
+            }
+        }
+
+        if create {
+            use LocationConfigMode::*;
+            let generation = match req.config.mode {
+                AttachedMulti | AttachedSingle | AttachedStale => req.config.generation,
+                // If a caller provided a generation in a non-attached request, ignore it
+                // and leave our generation as None: this enables a subsequent update to set
+                // the generation when setting an attached mode for the first time.
+                _ => None,
+            };
+
+            TenantCreateOrUpdate::Create(
+                // Synthesize a creation request
+                (
+                    TenantCreateRequest {
+                        new_tenant_id: TenantShardId::unsharded(tenant_id),
+                        generation,
+                        shard_parameters: ShardParameters {
+                            // Must preserve the incoming shard_count do distinguish unsharded (0)
+                            // from single-sharded (1): this distinction appears in the S3 keys of the tenant.
+                            count: req.tenant_id.shard_count,
+                            // We only import un-sharded or single-sharded tenants, so stripe
+                            // size can be made up arbitrarily here.
+                            stripe_size: ShardParameters::DEFAULT_STRIPE_SIZE,
+                        },
+                        config: req.config.tenant_conf,
+                    },
+                    placement_policy,
+                ),
+            )
+        } else {
+            TenantCreateOrUpdate::Update(updates)
+        }
+    }
+
+    /// This API is used by the cloud control plane to migrate unsharded tenants that it created
+    /// directly with pageservers into this service.
     ///
-    /// In future, calling with mode Secondary may switch to a detach-lite mode in which a tenant only has
-    /// secondary locations.
+    /// Cloud control plane MUST NOT continue issuing GENERATION NUMBERS for this tenant once it
+    /// has attempted to call this API. Failure to oblige to this rule may lead to S3 corruption.
+    /// Think of the first attempt to call this API as a transfer of absolute authority over the
+    /// tenant's source of generation numbers.
+    ///
+    /// The mode in this request coarse-grained control of tenants:
+    /// - Call with mode Attached* to upsert the tenant.
+    /// - Call with mode Secondary to either onboard a tenant without attaching it, or
+    ///   to set an existing tenant to PolicyMode::Secondary
+    /// - Call with mode Detached to switch to PolicyMode::Detached
     pub(crate) async fn tenant_location_config(
         &self,
         tenant_id: TenantId,
@@ -1231,131 +1403,104 @@ impl Service {
             )));
         }
 
-        let mut waiters = Vec::new();
-        let mut result = TenantLocationConfigResponse { shards: Vec::new() };
-        let maybe_create = {
-            let mut locked = self.inner.write().unwrap();
-            let result_tx = locked.result_tx.clone();
-            let compute_hook = locked.compute_hook.clone();
-            let (nodes, tenants, scheduler) = locked.parts_mut();
+        // First check if this is a creation or an update
+        let create_or_update = self.tenant_location_config_prepare(tenant_id, req);
 
-            // Maybe we have existing shards
-            let mut create = true;
-            for (shard_id, shard) in tenants.range_mut(TenantShardId::tenant_range(tenant_id)) {
-                // Saw an existing shard: this is not a creation
-                create = false;
+        let mut result = TenantLocationConfigResponse {
+            shards: Vec::new(),
+            stripe_size: None,
+        };
+        let waiters = match create_or_update {
+            TenantCreateOrUpdate::Create((create_req, placement_policy)) => {
+                let (create_resp, waiters) =
+                    self.do_tenant_create(create_req, placement_policy).await?;
+                result.shards = create_resp
+                    .shards
+                    .into_iter()
+                    .map(|s| TenantShardLocation {
+                        node_id: s.node_id,
+                        shard_id: s.shard_id,
+                    })
+                    .collect();
+                waiters
+            }
+            TenantCreateOrUpdate::Update(updates) => {
+                // Persist updates
+                // Ordering: write to the database before applying changes in-memory, so that
+                // we will not appear time-travel backwards on a restart.
+                for ShardUpdate {
+                    tenant_shard_id,
+                    placement_policy,
+                    tenant_config,
+                    generation,
+                } in &updates
+                {
+                    self.persistence
+                        .update_tenant_shard(
+                            *tenant_shard_id,
+                            placement_policy.clone(),
+                            tenant_config.clone(),
+                            *generation,
+                        )
+                        .await?;
+                }
 
-                // Note that for existing tenants we do _not_ respect the generation in the request: this is likely
-                // to be stale.  Once a tenant is created in this service, our view of generation is authoritative, and
-                // callers' generations may be ignored.  This represents a one-way migration of tenants from the outer
-                // cloud control plane into this service.
+                // Apply updates in-memory
+                let mut waiters = Vec::new();
+                {
+                    let mut locked = self.inner.write().unwrap();
+                    let result_tx = locked.result_tx.clone();
+                    let compute_hook = locked.compute_hook.clone();
+                    let (nodes, tenants, scheduler) = locked.parts_mut();
 
-                // Use location config mode as an indicator of policy: if they ask for
-                // attached we go to default HA attached mode.  If they ask for secondary
-                // we go to secondary-only mode.  If they ask for detached we detach.
-                match req.config.mode {
-                    LocationConfigMode::Detached => {
-                        shard.policy = PlacementPolicy::Detached;
-                    }
-                    LocationConfigMode::Secondary => {
-                        // TODO: implement secondary-only mode.
-                        todo!();
-                    }
-                    LocationConfigMode::AttachedMulti
-                    | LocationConfigMode::AttachedSingle
-                    | LocationConfigMode::AttachedStale => {
-                        // TODO: persistence for changes in policy
-                        if nodes.len() > 1 {
-                            shard.policy = PlacementPolicy::Double(1)
-                        } else {
-                            // Convenience for dev/test: if we just have one pageserver, import
-                            // tenants into Single mode so that scheduling will succeed.
-                            shard.policy = PlacementPolicy::Single
+                    for ShardUpdate {
+                        tenant_shard_id,
+                        placement_policy,
+                        tenant_config,
+                        generation: update_generation,
+                    } in updates
+                    {
+                        let Some(shard) = tenants.get_mut(&tenant_shard_id) else {
+                            tracing::warn!("Shard {tenant_shard_id} removed while updating");
+                            continue;
+                        };
+
+                        // Update stripe size
+                        if result.stripe_size.is_none() && shard.shard.count.count() > 1 {
+                            result.stripe_size = Some(shard.shard.stripe_size);
+                        }
+
+                        shard.policy = placement_policy;
+                        shard.config = tenant_config;
+                        if let Some(generation) = update_generation {
+                            shard.generation = Some(generation);
+                        }
+
+                        shard.schedule(scheduler)?;
+
+                        let maybe_waiter = shard.maybe_reconcile(
+                            result_tx.clone(),
+                            nodes,
+                            &compute_hook,
+                            &self.config,
+                            &self.persistence,
+                            &self.gate,
+                            &self.cancel,
+                        );
+                        if let Some(waiter) = maybe_waiter {
+                            waiters.push(waiter);
+                        }
+
+                        if let Some(node_id) = shard.intent.get_attached() {
+                            result.shards.push(TenantShardLocation {
+                                shard_id: tenant_shard_id,
+                                node_id: *node_id,
+                            })
                         }
                     }
                 }
-
-                shard.schedule(scheduler)?;
-
-                let maybe_waiter = shard.maybe_reconcile(
-                    result_tx.clone(),
-                    nodes,
-                    &compute_hook,
-                    &self.config,
-                    &self.persistence,
-                    &self.gate,
-                    &self.cancel,
-                );
-                if let Some(waiter) = maybe_waiter {
-                    waiters.push(waiter);
-                }
-
-                if let Some(node_id) = shard.intent.get_attached() {
-                    result.shards.push(TenantShardLocation {
-                        shard_id: *shard_id,
-                        node_id: *node_id,
-                    })
-                }
+                waiters
             }
-
-            if create {
-                // Validate request mode
-                match req.config.mode {
-                    LocationConfigMode::Detached | LocationConfigMode::Secondary => {
-                        // When using this API to onboard an existing tenant to this service, it must start in
-                        // an attached state, because we need the request to come with a generation
-                        return Err(ApiError::BadRequest(anyhow::anyhow!(
-                            "Imported tenant must be in attached mode"
-                        )));
-                    }
-
-                    LocationConfigMode::AttachedMulti
-                    | LocationConfigMode::AttachedSingle
-                    | LocationConfigMode::AttachedStale => {
-                        // Pass
-                    }
-                }
-
-                // Validate request generation
-                let Some(generation) = req.config.generation else {
-                    // We can only import attached tenants, because we need the request to come with a generation
-                    return Err(ApiError::BadRequest(anyhow::anyhow!(
-                        "Generation is mandatory when importing tenant"
-                    )));
-                };
-
-                // Synthesize a creation request
-                Some(TenantCreateRequest {
-                    new_tenant_id: TenantShardId::unsharded(tenant_id),
-                    generation: Some(generation),
-                    shard_parameters: ShardParameters {
-                        // Must preserve the incoming shard_count do distinguish unsharded (0)
-                        // from single-sharded (1): this distinction appears in the S3 keys of the tenant.
-                        count: req.tenant_id.shard_count,
-                        // We only import un-sharded or single-sharded tenants, so stripe
-                        // size can be made up arbitrarily here.
-                        stripe_size: ShardParameters::DEFAULT_STRIPE_SIZE,
-                    },
-                    config: req.config.tenant_conf,
-                })
-            } else {
-                None
-            }
-        };
-
-        let waiters = if let Some(create_req) = maybe_create {
-            let (create_resp, waiters) = self.do_tenant_create(create_req).await?;
-            result.shards = create_resp
-                .shards
-                .into_iter()
-                .map(|s| TenantShardLocation {
-                    node_id: s.node_id,
-                    shard_id: s.shard_id,
-                })
-                .collect();
-            waiters
-        } else {
-            waiters
         };
 
         if let Err(e) = self.await_waiters(waiters, SHORT_RECONCILE_TIMEOUT).await {
@@ -1375,6 +1520,91 @@ impl Service {
         Ok(result)
     }
 
+    pub(crate) async fn tenant_config_set(&self, req: TenantConfigRequest) -> Result<(), ApiError> {
+        let tenant_id = req.tenant_id;
+        let config = req.config;
+
+        self.persistence
+            .update_tenant_config(req.tenant_id, config.clone())
+            .await?;
+
+        let waiters = {
+            let mut waiters = Vec::new();
+            let mut locked = self.inner.write().unwrap();
+            let result_tx = locked.result_tx.clone();
+            let compute_hook = locked.compute_hook.clone();
+            let (nodes, tenants, _scheduler) = locked.parts_mut();
+            for (_shard_id, shard) in tenants.range_mut(TenantShardId::tenant_range(tenant_id)) {
+                shard.config = config.clone();
+                if let Some(waiter) = shard.maybe_reconcile(
+                    result_tx.clone(),
+                    nodes,
+                    &compute_hook,
+                    &self.config,
+                    &self.persistence,
+                    &self.gate,
+                    &self.cancel,
+                ) {
+                    waiters.push(waiter);
+                }
+            }
+            waiters
+        };
+
+        if let Err(e) = self.await_waiters(waiters, SHORT_RECONCILE_TIMEOUT).await {
+            // Treat this as success because we have stored the configuration.  If e.g.
+            // a node was unavailable at this time, it should not stop us accepting a
+            // configuration change.
+            tracing::warn!(%tenant_id, "Accepted configuration update but reconciliation failed: {e}");
+        }
+
+        Ok(())
+    }
+
+    pub(crate) fn tenant_config_get(
+        &self,
+        tenant_id: TenantId,
+    ) -> Result<HashMap<&str, serde_json::Value>, ApiError> {
+        let config = {
+            let locked = self.inner.read().unwrap();
+
+            match locked
+                .tenants
+                .range(TenantShardId::tenant_range(tenant_id))
+                .next()
+            {
+                Some((_tenant_shard_id, shard)) => shard.config.clone(),
+                None => {
+                    return Err(ApiError::NotFound(
+                        anyhow::anyhow!("Tenant not found").into(),
+                    ))
+                }
+            }
+        };
+
+        // Unlike the pageserver, we do not have a set of global defaults: the config is
+        // entirely per-tenant.  Therefore the distinction between `tenant_specific_overrides`
+        // and `effective_config` in the response is meaningless, but we retain that syntax
+        // in order to remain compatible with the pageserver API.
+
+        let response = HashMap::from([
+            (
+                "tenant_specific_overrides",
+                serde_json::to_value(&config)
+                    .context("serializing tenant specific overrides")
+                    .map_err(ApiError::InternalServerError)?,
+            ),
+            (
+                "effective_config",
+                serde_json::to_value(&config)
+                    .context("serializing effective config")
+                    .map_err(ApiError::InternalServerError)?,
+            ),
+        ]);
+
+        Ok(response)
+    }
+
     pub(crate) async fn tenant_time_travel_remote_storage(
         &self,
         time_travel_req: &TenantTimeTravelRequest,
@@ -1460,6 +1690,60 @@ impl Service {
                         })?;
             }
         }
+        Ok(())
+    }
+
+    pub(crate) async fn tenant_secondary_download(
+        &self,
+        tenant_id: TenantId,
+    ) -> Result<(), ApiError> {
+        // Acquire lock and yield the collection of shard-node tuples which we will send requests onward to
+        let targets = {
+            let locked = self.inner.read().unwrap();
+            let mut targets = Vec::new();
+
+            for (tenant_shard_id, shard) in
+                locked.tenants.range(TenantShardId::tenant_range(tenant_id))
+            {
+                for node_id in shard.intent.get_secondary() {
+                    let node = locked
+                        .nodes
+                        .get(node_id)
+                        .expect("Pageservers may not be deleted while referenced");
+
+                    targets.push((*tenant_shard_id, node.clone()));
+                }
+            }
+            targets
+        };
+
+        // TODO: this API, and the underlying pageserver API, should take a timeout argument so that for long running
+        // downloads, they can return a clean 202 response instead of the HTTP client timing out.
+
+        // Issue concurrent requests to all shards' locations
+        let mut futs = FuturesUnordered::new();
+        for (tenant_shard_id, node) in targets {
+            let client = mgmt_api::Client::new(node.base_url(), self.config.jwt_token.as_deref());
+            futs.push(async move {
+                let result = client.tenant_secondary_download(tenant_shard_id).await;
+                (result, node)
+            })
+        }
+
+        // Handle any errors returned by pageservers.  This includes cases like this request racing with
+        // a scheduling operation, such that the tenant shard we're calling doesn't exist on that pageserver any more, as
+        // well as more general cases like 503s, 500s, or timeouts.
+        while let Some((result, node)) = futs.next().await {
+            let Err(e) = result else { continue };
+
+            // Secondary downloads are always advisory: if something fails, we nevertheless report success, so that whoever
+            // is calling us will proceed with whatever migration they're doing, albeit with a slightly less warm cache
+            // than they had hoped for.
+            tracing::warn!(
+                "Ignoring tenant secondary download error from pageserver {}: {e}",
+                node.id,
+            );
+        }
 
         Ok(())
     }
@@ -2039,8 +2323,8 @@ impl Service {
                     // Note: this generation is a placeholder, [`Persistence::begin_shard_split`] will
                     // populate the correct generation as part of its transaction, to protect us
                     // against racing with changes in the state of the parent.
-                    generation: 0,
-                    generation_pageserver: target.node.id.0 as i64,
+                    generation: None,
+                    generation_pageserver: Some(target.node.id.0 as i64),
                     placement_policy: serde_json::to_string(&policy).unwrap(),
                     // TODO: get the config out of the map
                     config: serde_json::to_string(&TenantConfig::default()).unwrap(),
@@ -2161,7 +2445,8 @@ impl Service {
                         .expect("It was present, we just split it");
                     let old_attached = old_state.intent.get_attached().unwrap();
                     old_state.intent.clear(scheduler);
-                    (old_attached, old_state.generation, old_state.config.clone())
+                    let generation = old_state.generation.expect("Shard must have been attached");
+                    (old_attached, generation, old_state.config.clone())
                 };
 
                 for child in child_ids {
@@ -2182,14 +2467,14 @@ impl Service {
                     child_state.observed = ObservedState {
                         locations: child_observed,
                     };
-                    child_state.generation = generation;
+                    child_state.generation = Some(generation);
                     child_state.config = config.clone();
 
                     // The child's TenantState::splitting is intentionally left at the default value of Idle,
                     // as at this point in the split process we have succeeded and this part is infallible:
                     // we will never need to do any special recovery from this state.
 
-                    child_locations.push((child, pageserver));
+                    child_locations.push((child, pageserver, child_shard.stripe_size));
 
                     tenants.insert(child, child_state);
                     response.new_shards.push(child);
@@ -2199,8 +2484,11 @@ impl Service {
 
         // Send compute notifications for all the new shards
         let mut failed_notifications = Vec::new();
-        for (child_id, child_ps) in child_locations {
-            if let Err(e) = compute_hook.notify(child_id, child_ps, &self.cancel).await {
+        for (child_id, child_ps, stripe_size) in child_locations {
+            if let Err(e) = compute_hook
+                .notify(child_id, child_ps, stripe_size, &self.cancel)
+                .await
+            {
                 tracing::warn!("Failed to update compute of {}->{} during split, proceeding anyway to complete split ({e})",
                         child_id, child_ps);
                 failed_notifications.push(child_id);
@@ -2231,6 +2519,19 @@ impl Service {
             let compute_hook = locked.compute_hook.clone();
             let (nodes, tenants, scheduler) = locked.parts_mut();
 
+            let Some(node) = nodes.get(&migrate_req.node_id) else {
+                return Err(ApiError::BadRequest(anyhow::anyhow!(
+                    "Node {} not found",
+                    migrate_req.node_id
+                )));
+            };
+
+            if node.availability != NodeAvailability::Active {
+                // Warn but proceed: the caller may intend to manually adjust the placement of
+                // a shard even if the node is down, e.g. if intervening during an incident.
+                tracing::warn!("Migrating to an unavailable node ({})", node.id);
+            }
+
             let Some(shard) = tenants.get_mut(&tenant_shard_id) else {
                 return Err(ApiError::NotFound(
                     anyhow::anyhow!("Tenant shard not found").into(),
@@ -2247,6 +2548,7 @@ impl Service {
                 match shard.policy {
                     PlacementPolicy::Single => {
                         shard.intent.clear_secondary(scheduler);
+                        shard.intent.set_attached(scheduler, Some(migrate_req.node_id));
                     }
                     PlacementPolicy::Double(_n) => {
                         // If our new attached node was a secondary, it no longer should be.
@@ -2256,6 +2558,12 @@ impl Service {
                         if let Some(old_attached) = old_attached {
                             shard.intent.push_secondary(scheduler, old_attached);
                         }
+
+                        shard.intent.set_attached(scheduler, Some(migrate_req.node_id));
+                    }
+                    PlacementPolicy::Secondary => {
+                        shard.intent.clear(scheduler);
+                        shard.intent.push_secondary(scheduler, migrate_req.node_id);
                     }
                     PlacementPolicy::Detached => {
                         return Err(ApiError::BadRequest(anyhow::anyhow!(
@@ -2263,9 +2571,6 @@ impl Service {
                         )))
                     }
                 }
-                shard
-                    .intent
-                    .set_attached(scheduler, Some(migrate_req.node_id));
 
                 tracing::info!("Migrating: new intent {:?}", shard.intent);
                 shard.sequence = shard.sequence.next();
@@ -2356,6 +2661,18 @@ impl Service {
                 .map(|t| t.to_persistent())
                 .collect::<Vec<_>>();
 
+            // This method can only validate the state of an idle system: if a reconcile is in
+            // progress, fail out early to avoid giving false errors on state that won't match
+            // between database and memory under a ReconcileResult is processed.
+            for t in locked.tenants.values() {
+                if t.reconciler.is_some() {
+                    return Err(ApiError::InternalServerError(anyhow::anyhow!(
+                        "Shard {} reconciliation in progress",
+                        t.tenant_shard_id
+                    )));
+                }
+            }
+
             (expect_nodes, expect_shards)
         };
 
@@ -2593,7 +2910,7 @@ impl Service {
                     observed_loc.conf = None;
                 }
 
-                if tenant_state.intent.notify_offline(config_req.node_id) {
+                if tenant_state.intent.demote_attached(config_req.node_id) {
                     tenant_state.sequence = tenant_state.sequence.next();
                     match tenant_state.schedule(scheduler) {
                         Err(e) => {
@@ -2660,6 +2977,9 @@ impl Service {
     /// Helper for methods that will try and call pageserver APIs for
     /// a tenant, such as timeline CRUD: they cannot proceed unless the tenant
     /// is attached somewhere.
+    ///
+    /// TODO: this doesn't actually ensure attached unless the PlacementPolicy is
+    /// an attached policy.  We should error out if it isn't.
     fn ensure_attached_schedule(
         &self,
         mut locked: std::sync::RwLockWriteGuard<'_, ServiceState>,

control_plane/attachment_service/src/tenant_state.rs

@@ -53,8 +53,11 @@ pub(crate) struct TenantState {
     pub(crate) sequence: Sequence,
 
     // Latest generation number: next time we attach, increment this
-    // and use the incremented number when attaching
-    pub(crate) generation: Generation,
+    // and use the incremented number when attaching.
+    //
+    // None represents an incompletely onboarded tenant via the [`Service::location_config`]
+    // API, where this tenant may only run in PlacementPolicy::Secondary.
+    pub(crate) generation: Option<Generation>,
 
     // High level description of how the tenant should be set up.  Provided
     // externally.
@@ -181,6 +184,13 @@ impl IntentState {
         }
     }
 
+    /// Remove the last secondary node from the list of secondaries
+    pub(crate) fn pop_secondary(&mut self, scheduler: &mut Scheduler) {
+        if let Some(node_id) = self.secondary.pop() {
+            scheduler.node_dec_ref(node_id);
+        }
+    }
+
     pub(crate) fn clear(&mut self, scheduler: &mut Scheduler) {
         if let Some(old_attached) = self.attached.take() {
             scheduler.node_dec_ref(old_attached);
@@ -208,11 +218,13 @@ impl IntentState {
         &self.secondary
     }
 
-    /// When a node goes offline, we update intents to avoid using it
-    /// as their attached pageserver.
+    /// If the node is in use as the attached location, demote it into
+    /// the list of secondary locations.  This is used when a node goes offline,
+    /// and we want to use a different node for attachment, but not permanently
+    /// forget the location on the offline node.
     ///
     /// Returns true if a change was made
-    pub(crate) fn notify_offline(&mut self, node_id: NodeId) -> bool {
+    pub(crate) fn demote_attached(&mut self, node_id: NodeId) -> bool {
         if self.attached == Some(node_id) {
             // TODO: when scheduler starts tracking attached + secondary counts separately, we will
             // need to call into it here.
@@ -315,7 +327,7 @@ pub(crate) struct ReconcileResult {
     pub(crate) result: Result<(), ReconcileError>,
 
     pub(crate) tenant_shard_id: TenantShardId,
-    pub(crate) generation: Generation,
+    pub(crate) generation: Option<Generation>,
     pub(crate) observed: ObservedState,
 
     /// Set [`TenantState::pending_compute_notification`] from this flag
@@ -340,7 +352,7 @@ impl TenantState {
             tenant_shard_id,
             policy,
             intent: IntentState::default(),
-            generation: Generation::new(0),
+            generation: Some(Generation::new(0)),
             shard,
             observed: ObservedState::default(),
             config: TenantConfig::default(),
@@ -438,10 +450,16 @@ impl TenantState {
         // more work on the same pageservers we're already using.
         let mut modified = false;
 
+        // Add/remove nodes to fulfil policy
         use PlacementPolicy::*;
         match self.policy {
             Single => {
                 // Should have exactly one attached, and zero secondaries
+                if !self.intent.secondary.is_empty() {
+                    self.intent.clear_secondary(scheduler);
+                    modified = true;
+                }
+
                 let (modified_attached, _attached_node_id) = self.schedule_attached(scheduler)?;
                 modified |= modified_attached;
 
@@ -451,6 +469,23 @@ impl TenantState {
                 }
             }
             Double(secondary_count) => {
+                let retain_secondaries = if self.intent.attached.is_none()
+                    && scheduler.node_preferred(&self.intent.secondary).is_some()
+                {
+                    // If we have no attached, and one of the secondaries is elegible to be promoted, retain
+                    // one more secondary than we usually would, as one of them will become attached futher down this function.
+                    secondary_count + 1
+                } else {
+                    secondary_count
+                };
+
+                while self.intent.secondary.len() > retain_secondaries {
+                    // We have no particular preference for one secondary location over another: just
+                    // arbitrarily drop from the end
+                    self.intent.pop_secondary(scheduler);
+                    modified = true;
+                }
+
                 // Should have exactly one attached, and N secondaries
                 let (modified_attached, attached_node_id) = self.schedule_attached(scheduler)?;
                 modified |= modified_attached;
@@ -463,15 +498,28 @@ impl TenantState {
                     modified = true;
                 }
             }
-            Detached => {
-                // Should have no attached or secondary pageservers
-                if self.intent.attached.is_some() {
-                    self.intent.set_attached(scheduler, None);
+            Secondary => {
+                if let Some(node_id) = self.intent.get_attached() {
+                    // Populate secondary by demoting the attached node
+                    self.intent.demote_attached(*node_id);
+                    modified = true;
+                } else if self.intent.secondary.is_empty() {
+                    // Populate secondary by scheduling a fresh node
+                    let node_id = scheduler.schedule_shard(&[])?;
+                    self.intent.push_secondary(scheduler, node_id);
                     modified = true;
                 }
-
-                if !self.intent.secondary.is_empty() {
-                    self.intent.clear_secondary(scheduler);
+                while self.intent.secondary.len() > 1 {
+                    // We have no particular preference for one secondary location over another: just
+                    // arbitrarily drop from the end
+                    self.intent.pop_secondary(scheduler);
+                    modified = true;
+                }
+            }
+            Detached => {
+                // Never add locations in this mode
+                if self.intent.get_attached().is_some() || !self.intent.get_secondary().is_empty() {
+                    self.intent.clear(scheduler);
                     modified = true;
                 }
             }
@@ -518,7 +566,12 @@ impl TenantState {
 
     fn dirty(&self) -> bool {
         if let Some(node_id) = self.intent.attached {
-            let wanted_conf = attached_location_conf(self.generation, &self.shard, &self.config);
+            // Maybe panic: it is a severe bug if we try to attach while generation is null.
+            let generation = self
+                .generation
+                .expect("Attempted to enter attached state without a generation");
+
+            let wanted_conf = attached_location_conf(generation, &self.shard, &self.config);
             match self.observed.locations.get(&node_id) {
                 Some(conf) if conf.conf.as_ref() == Some(&wanted_conf) => {}
                 Some(_) | None => {
@@ -596,6 +649,10 @@ impl TenantState {
         // Reconcile already in flight for the current sequence?
         if let Some(handle) = &self.reconciler {
             if handle.sequence == self.sequence {
+                tracing::info!(
+                    "Reconciliation already in progress for sequence {:?}",
+                    self.sequence,
+                );
                 return Some(ReconcilerWaiter {
                     tenant_shard_id: self.tenant_shard_id,
                     seq_wait: self.waiter.clone(),
@@ -615,6 +672,10 @@ impl TenantState {
             return None;
         };
 
+        // Advance the sequence before spawning a reconciler, so that sequence waiters
+        // can distinguish between before+after the reconcile completes.
+        self.sequence = self.sequence.next();
+
         let reconciler_cancel = cancel.child_token();
         let mut reconciler = Reconciler {
             tenant_shard_id: self.tenant_shard_id,
@@ -716,6 +777,17 @@ impl TenantState {
         })
     }
 
+    /// Called when a ReconcileResult has been emitted and the service is updating
+    /// our state: if the result is from a sequence >= my ReconcileHandle, then drop
+    /// the handle to indicate there is no longer a reconciliation in progress.
+    pub(crate) fn reconcile_complete(&mut self, sequence: Sequence) {
+        if let Some(reconcile_handle) = &self.reconciler {
+            if reconcile_handle.sequence <= sequence {
+                self.reconciler = None;
+            }
+        }
+    }
+
     // If we had any state at all referring to this node ID, drop it.  Does not
     // attempt to reschedule.
     pub(crate) fn deref_node(&mut self, node_id: NodeId) {
@@ -736,13 +808,8 @@ impl TenantState {
             shard_number: self.tenant_shard_id.shard_number.0 as i32,
             shard_count: self.tenant_shard_id.shard_count.literal() as i32,
             shard_stripe_size: self.shard.stripe_size.0 as i32,
-            generation: self.generation.into().unwrap_or(0) as i32,
-            generation_pageserver: self
-                .intent
-                .get_attached()
-                .map(|n| n.0 as i64)
-                .unwrap_or(i64::MAX),
-
+            generation: self.generation.map(|g| g.into().unwrap_or(0) as i32),
+            generation_pageserver: self.intent.get_attached().map(|n| n.0 as i64),
             placement_policy: serde_json::to_string(&self.policy).unwrap(),
             config: serde_json::to_string(&self.config).unwrap(),
             splitting: SplitState::default(),
@@ -805,8 +872,10 @@ pub(crate) mod tests {
         assert_ne!(attached_node_id, secondary_node_id);
 
         // Notifying the attached node is offline should demote it to a secondary
-        let changed = tenant_state.intent.notify_offline(attached_node_id);
+        let changed = tenant_state.intent.demote_attached(attached_node_id);
         assert!(changed);
+        assert!(tenant_state.intent.attached.is_none());
+        assert_eq!(tenant_state.intent.secondary.len(), 2);
 
         // Update the scheduler state to indicate the node is offline
         nodes.get_mut(&attached_node_id).unwrap().availability = NodeAvailability::Offline;

control_plane/src/attachment_service.rs

@@ -200,7 +200,7 @@ impl AttachmentService {
                 "localhost",
                 "-p",
                 &format!("{}", self.postgres_port),
-                &DB_NAME,
+                DB_NAME,
             ])
             .output()
             .await

control_plane/src/bin/neon_local.rs

@@ -1024,7 +1024,7 @@ async fn handle_endpoint(ep_match: &ArgMatches, env: &local_env::LocalEnv) -> Re
                         })
                         .collect::<Vec<_>>()
                 };
-            endpoint.reconfigure(pageservers).await?;
+            endpoint.reconfigure(pageservers, None).await?;
         }
         "stop" => {
             let endpoint_id = sub_args

control_plane/src/endpoint.rs

@@ -52,6 +52,7 @@ use compute_api::spec::RemoteExtSpec;
 use compute_api::spec::Role;
 use nix::sys::signal::kill;
 use nix::sys::signal::Signal;
+use pageserver_api::shard::ShardStripeSize;
 use serde::{Deserialize, Serialize};
 use url::Host;
 use utils::id::{NodeId, TenantId, TimelineId};
@@ -605,7 +606,7 @@ impl Endpoint {
         let conn_str = self.connstr("cloud_admin", "postgres");
         println!("Starting postgres node at '{}'", conn_str);
         if create_test_user {
-            let conn_str = self.connstr("user", "neondb");
+            let conn_str = self.connstr("test", "neondb");
             println!("Also at '{}'", conn_str);
         }
         let mut cmd = Command::new(self.env.neon_distrib_dir.join("compute_ctl"));
@@ -735,7 +736,11 @@ impl Endpoint {
         }
     }
 
-    pub async fn reconfigure(&self, mut pageservers: Vec<(Host, u16)>) -> Result<()> {
+    pub async fn reconfigure(
+        &self,
+        mut pageservers: Vec<(Host, u16)>,
+        stripe_size: Option<ShardStripeSize>,
+    ) -> Result<()> {
         let mut spec: ComputeSpec = {
             let spec_path = self.endpoint_path().join("spec.json");
             let file = std::fs::File::open(spec_path)?;
@@ -765,6 +770,9 @@ impl Endpoint {
         let pageserver_connstr = Self::build_pageserver_connstr(&pageservers);
         assert!(!pageserver_connstr.is_empty());
         spec.pageserver_connstring = Some(pageserver_connstr);
+        if stripe_size.is_some() {
+            spec.shard_stripe_size = stripe_size.map(|s| s.0 as usize);
+        }
 
         let client = reqwest::Client::new();
         let response = client

control_plane/src/pageserver.rs

@@ -537,10 +537,11 @@ impl PageServerNode {
         tenant_shard_id: TenantShardId,
         config: LocationConfig,
         flush_ms: Option<Duration>,
+        lazy: bool,
     ) -> anyhow::Result<()> {
         Ok(self
             .http_client
-            .location_config(tenant_shard_id, config, flush_ms)
+            .location_config(tenant_shard_id, config, flush_ms, lazy)
             .await?)
     }
 
@@ -605,7 +606,7 @@ impl PageServerNode {
                 eprintln!("connection error: {}", e);
             }
         });
-        tokio::pin!(client);
+        let client = std::pin::pin!(client);
 
         // Init base reader
         let (start_lsn, base_tarfile_path) = base;

docs/rfcs/002-storage.md

@@ -1,4 +1,4 @@
-# Zenith storage node — alternative
+# Neon storage node — alternative
 
 ## **Design considerations**
 

docs/rfcs/003-laptop-cli.md

@@ -1,6 +1,6 @@
 # Command line interface (end-user)
 
-Zenith CLI as it is described here mostly resides on the same conceptual level as pg_ctl/initdb/pg_recvxlog/etc and replaces some of them in an opinionated way. I would also suggest bundling our patched postgres inside zenith distribution at least at the start.
+Neon CLI as it is described here mostly resides on the same conceptual level as pg_ctl/initdb/pg_recvxlog/etc and replaces some of them in an opinionated way. I would also suggest bundling our patched postgres inside neon distribution at least at the start.
 
 This proposal is focused on managing local installations. For cluster operations, different tooling would be needed. The point of integration between the two is storage URL: no matter how complex cluster setup is it may provide an endpoint where the user may push snapshots.
 
@@ -8,40 +8,40 @@ The most important concept here is a snapshot, which can be created/pushed/pulle
 
 # Possible usage scenarios
 
-## Install zenith, run a postgres
+## Install neon, run a postgres
 
 ```
-> brew install pg-zenith 
-> zenith pg create # creates pgdata with default pattern pgdata$i
-> zenith pg list
+> brew install pg-neon 
+> neon pg create # creates pgdata with default pattern pgdata$i
+> neon pg list
 ID            PGDATA        USED    STORAGE            ENDPOINT
-primary1      pgdata1       0G      zenith-local       localhost:5432
+primary1      pgdata1       0G      neon-local       localhost:5432
 ```
 
-## Import standalone postgres to zenith
+## Import standalone postgres to neon
 
 ```
-> zenith snapshot import --from=basebackup://replication@localhost:5432/ oldpg
+> neon snapshot import --from=basebackup://replication@localhost:5432/ oldpg
 [====================------------] 60% | 20MB/s
-> zenith snapshot list
+> neon snapshot list
 ID          SIZE        PARENT
 oldpg       5G          -
 
-> zenith pg create --snapshot oldpg
+> neon pg create --snapshot oldpg
 Started postgres on localhost:5432
 
-> zenith pg list
+> neon pg list
 ID            PGDATA        USED    STORAGE            ENDPOINT
-primary1      pgdata1       5G      zenith-local       localhost:5432
+primary1      pgdata1       5G      neon-local       localhost:5432
 
-> zenith snapshot destroy oldpg
+> neon snapshot destroy oldpg
 Ok
 ```
 
 Also, we may start snapshot import implicitly by looking at snapshot schema
 
 ```
-> zenith pg create --snapshot basebackup://replication@localhost:5432/
+> neon pg create --snapshot basebackup://replication@localhost:5432/
 Downloading snapshot... Done.
 Started postgres on localhost:5432
 Destroying snapshot... Done.
@@ -52,39 +52,39 @@ Destroying snapshot... Done.
 Since we may export the whole snapshot as one big file (tar of basebackup, maybe with some manifest) it may be shared over conventional means: http, ssh, [git+lfs](https://docs.github.com/en/github/managing-large-files/about-git-large-file-storage).
 
 ```
-> zenith pg create --snapshot http://learn-postgres.com/movies_db.zenith movies
+> neon pg create --snapshot http://learn-postgres.com/movies_db.neon movies
 ```
 
 ## Create snapshot and push it to the cloud
 
 ```
-> zenith snapshot create pgdata1@snap1
-> zenith snapshot push --to ssh://stas@zenith.tech pgdata1@snap1
+> neon snapshot create pgdata1@snap1
+> neon snapshot push --to ssh://stas@neon.tech pgdata1@snap1
 ```
 
 ## Rollback database to the snapshot
 
-One way to rollback the database is just to init a new database from the snapshot and destroy the old one. But creating a new database from a snapshot would require a copy of that snapshot which is time consuming operation. Another option that would be cool to support is the ability to create the copy-on-write database from the snapshot without copying data, and store updated pages in a separate location, however that way would have performance implications. So to properly rollback the database to the older state we have `zenith pg checkout`.
+One way to rollback the database is just to init a new database from the snapshot and destroy the old one. But creating a new database from a snapshot would require a copy of that snapshot which is time consuming operation. Another option that would be cool to support is the ability to create the copy-on-write database from the snapshot without copying data, and store updated pages in a separate location, however that way would have performance implications. So to properly rollback the database to the older state we have `neon pg checkout`.
 
 ```
-> zenith pg list
+> neon pg list
 ID            PGDATA        USED    STORAGE            ENDPOINT
-primary1      pgdata1       5G      zenith-local       localhost:5432
+primary1      pgdata1       5G      neon-local       localhost:5432
 
-> zenith snapshot create pgdata1@snap1
+> neon snapshot create pgdata1@snap1
 
-> zenith snapshot list
+> neon snapshot list
 ID                    SIZE        PARENT
 oldpg                 5G          -
 pgdata1@snap1         6G          -
 pgdata1@CURRENT       6G          -
 
-> zenith pg checkout pgdata1@snap1
+> neon pg checkout pgdata1@snap1
 Stopping postgres on pgdata1.
 Rolling back pgdata1@CURRENT to pgdata1@snap1.
 Starting postgres on pgdata1.
 
-> zenith snapshot list
+> neon snapshot list
 ID                    SIZE        PARENT
 oldpg                 5G          -
 pgdata1@snap1         6G          -
@@ -99,7 +99,7 @@ Some notes: pgdata1@CURRENT -- implicit snapshot representing the current state
 PITR area acts like a continuous snapshot where you can reset the database to any point in time within this area (by area I mean some TTL period or some size limit, both possibly infinite).
 
 ```
-> zenith pitr create --storage s3tank --ttl 30d --name pitr_last_month
+> neon pitr create --storage s3tank --ttl 30d --name pitr_last_month
 ```
 
 Resetting the database to some state in past would require creating a snapshot on some lsn / time in this pirt area.
@@ -108,29 +108,29 @@ Resetting the database to some state in past would require creating a snapshot o
 
 ## storage
 
-Storage is either zenith pagestore or s3. Users may create a database in a pagestore and create/move *snapshots* and *pitr regions* in both pagestore and s3. Storage is a concept similar to `git remote`. After installation, I imagine one local storage is available by default.
+Storage is either neon pagestore or s3. Users may create a database in a pagestore and create/move *snapshots* and *pitr regions* in both pagestore and s3. Storage is a concept similar to `git remote`. After installation, I imagine one local storage is available by default.
 
-**zenith storage attach** -t [native|s3] -c key=value -n name
+**neon storage attach** -t [native|s3] -c key=value -n name
 
-Attaches/initializes storage. For --type=s3, user credentials and path should be provided. For --type=native we may support --path=/local/path and --url=zenith.tech/stas/mystore. Other possible term for native is 'zstore'.
+Attaches/initializes storage. For --type=s3, user credentials and path should be provided. For --type=native we may support --path=/local/path and --url=neon.tech/stas/mystore. Other possible term for native is 'zstore'.
 
 
-**zenith storage list**
+**neon storage list**
 
 Show currently attached storages. For example:
 
 ```
-> zenith storage list
+> neon storage list
 NAME            USED    TYPE                OPTIONS          PATH
-local           5.1G    zenith-local                         /opt/zenith/store/local
-local.compr     20.4G   zenith-local        compression=on    /opt/zenith/store/local.compr
-zcloud          60G     zenith-remote                        zenith.tech/stas/mystore
+local           5.1G    neon-local                         /opt/neon/store/local
+local.compr     20.4G   neon-local        compression=on    /opt/neon/store/local.compr
+zcloud          60G     neon-remote                        neon.tech/stas/mystore
 s3tank          80G     S3
 ```
 
-**zenith storage detach**
+**neon storage detach**
 
-**zenith storage show**
+**neon storage show**
 
 
 
@@ -140,29 +140,29 @@ Manages postgres data directories and can start postgres instances with proper c
 
 Pg is a term for a single postgres running on some data. I'm trying to avoid separation of datadir management and postgres instance management -- both that concepts bundled here together.
 
-**zenith pg create** [--no-start --snapshot --cow] -s storage-name -n pgdata
+**neon pg create** [--no-start --snapshot --cow] -s storage-name -n pgdata
 
 Creates (initializes) new data directory in given storage and starts postgres. I imagine that storage for this operation may be only local and data movement to remote location happens through snapshots/pitr.
 
 --no-start: just init datadir without creating 
 
---snapshot snap: init from the snapshot. Snap is a name or URL (zenith.tech/stas/mystore/snap1)
+--snapshot snap: init from the snapshot. Snap is a name or URL (neon.tech/stas/mystore/snap1)
 
 --cow: initialize Copy-on-Write data directory on top of some snapshot (makes sense if it is a snapshot of currently running a database)
 
-**zenith pg destroy**
+**neon pg destroy**
 
-**zenith pg start** [--replica] pgdata
+**neon pg start** [--replica] pgdata
 
 Start postgres with proper extensions preloaded/installed.
 
-**zenith pg checkout**
+**neon pg checkout**
 
 Rollback data directory to some previous snapshot. 
 
-**zenith pg stop** pg_id
+**neon pg stop** pg_id
 
-**zenith pg list**
+**neon pg list**
 
 ```
 ROLE                 PGDATA        USED    STORAGE            ENDPOINT
@@ -173,7 +173,7 @@ primary              my_pg2        3.2G    local.compr        localhost:5435
 -                    my_pg3        9.2G    local.compr        -
 ```
 
-**zenith pg show**
+**neon pg show**
 
 ```
 my_pg:
@@ -194,7 +194,7 @@ my_pg:
 
 ```
 
-**zenith pg start-rest/graphql** pgdata
+**neon pg start-rest/graphql** pgdata
 
 Starts REST/GraphQL proxy on top of postgres master. Not sure we should do that, just an idea.
 
@@ -203,35 +203,35 @@ Starts REST/GraphQL proxy on top of postgres master. Not sure we should do that,
 
 Snapshot creation is cheap -- no actual data is copied, we just start retaining old pages. Snapshot size means the amount of retained data, not all data. Snapshot name looks like pgdata_name@tag_name. tag_name is set by the user during snapshot creation. There are some reserved tag names: CURRENT represents the current state of the data directory; HEAD{i} represents the data directory state that resided in the database before i-th checkout.
 
-**zenith snapshot create** pgdata_name@snap_name
+**neon snapshot create** pgdata_name@snap_name
 
 Creates a new snapshot in the same storage where pgdata_name exists.
 
-**zenith snapshot push** --to url pgdata_name@snap_name
+**neon snapshot push** --to url pgdata_name@snap_name
 
-Produces binary stream of a given snapshot. Under the hood starts temp read-only postgres over this snapshot and sends basebackup stream. Receiving side should start `zenith snapshot recv` before push happens. If url has some special schema like zenith:// receiving side may require auth start `zenith snapshot recv` on the go.
+Produces binary stream of a given snapshot. Under the hood starts temp read-only postgres over this snapshot and sends basebackup stream. Receiving side should start `neon snapshot recv` before push happens. If url has some special schema like neon:// receiving side may require auth start `neon snapshot recv` on the go.
 
-**zenith snapshot recv**
+**neon snapshot recv**
 
 Starts a port listening for a basebackup stream, prints connection info to stdout (so that user may use that in push command), and expects data on that socket.
 
-**zenith snapshot pull** --from url or path
+**neon snapshot pull** --from url or path
 
-Connects to a remote zenith/s3/file and pulls snapshot. The remote site should be zenith service or files in our format.
+Connects to a remote neon/s3/file and pulls snapshot. The remote site should be neon service or files in our format.
 
-**zenith snapshot import** --from basebackup://<...>  or path
+**neon snapshot import** --from basebackup://<...>  or path
 
 Creates a new snapshot out of running postgres via basebackup protocol or basebackup files.
 
-**zenith snapshot export**
+**neon snapshot export**
 
-Starts read-only postgres over this snapshot and exports data in some format (pg_dump, or COPY TO on some/all tables). One of the options may be zenith own format which is handy for us (but I think just tar of basebackup would be okay).
+Starts read-only postgres over this snapshot and exports data in some format (pg_dump, or COPY TO on some/all tables). One of the options may be neon own format which is handy for us (but I think just tar of basebackup would be okay).
 
-**zenith snapshot diff** snap1 snap2
+**neon snapshot diff** snap1 snap2
 
 Shows size of data changed between two snapshots. We also may provide options to diff schema/data in tables. To do that start temp read-only postgreses.
 
-**zenith snapshot destroy**
+**neon snapshot destroy**
 
 ## pitr
 
@@ -239,7 +239,7 @@ Pitr represents wal stream and ttl policy for that stream
 
 XXX: any suggestions on a better name?
 
-**zenith pitr create** name
+**neon pitr create** name
 
 --ttl = inf | period
 
@@ -247,21 +247,21 @@ XXX: any suggestions on a better name?
 
 --storage = storage_name
 
-**zenith pitr extract-snapshot** pitr_name --lsn xxx
+**neon pitr extract-snapshot** pitr_name --lsn xxx
 
 Creates a snapshot out of some lsn in PITR area. The obtained snapshot may be managed with snapshot routines (move/send/export)
 
-**zenith pitr gc** pitr_name
+**neon pitr gc** pitr_name
 
 Force garbage collection on some PITR area.
 
-**zenith pitr list**
+**neon pitr list**
 
-**zenith pitr destroy**
+**neon pitr destroy**
 
 
 ## console
 
-**zenith console**
+**neon console**
 
 Opens browser targeted at web console with the more or less same functionality as described here.

docs/rfcs/004-durability.md

@@ -6,7 +6,7 @@ When do we consider the WAL record as durable, so that we can
 acknowledge the commit to the client and be reasonably certain that we
 will not lose the transaction?
 
-Zenith uses a group of WAL safekeeper nodes to hold the generated WAL.
+Neon uses a group of WAL safekeeper nodes to hold the generated WAL.
 A WAL record is considered durable, when it has been written to a
 majority of WAL safekeeper nodes. In this document, I use 5
 safekeepers, because I have five fingers. A WAL record is durable,

docs/rfcs/005-zenith_local.md

@@ -1,23 +1,23 @@
-# Zenith local
+# Neon local
 
-Here I list some objectives to keep in mind when discussing zenith-local design and a proposal that brings all components together.  Your comments on both parts are very welcome.
+Here I list some objectives to keep in mind when discussing neon-local design and a proposal that brings all components together.  Your comments on both parts are very welcome.
 
 #### Why do we need it?
 - For distribution - this easy to use binary will help us to build adoption among developers.
 - For internal use - to test all components together.
 
-In my understanding, we consider it to be just a mock-up version of zenith-cloud.
+In my understanding, we consider it to be just a mock-up version of neon-cloud.
 > Question: How much should we care about durability and security issues for a local setup?
 
 
 #### Why is it better than a simple local postgres?
 
-- Easy one-line setup. As simple as `cargo install zenith && zenith start`
+- Easy one-line setup. As simple as `cargo install neon && neon start`
 
 - Quick and cheap creation of compute nodes over the same storage.
 > Question: How can we describe a use-case for this feature?
 
-- Zenith-local can work with S3 directly. 
+- Neon-local can work with S3 directly. 
 
 - Push and pull images (snapshots) to remote S3 to exchange data with other users.
 
@@ -31,50 +31,50 @@ Ideally, just one binary that incorporates all elements we need.
 
 #### Components:
 
-- **zenith-CLI** - interface for end-users.  Turns commands to REST requests and handles responses to show them in a user-friendly way.  
-CLI proposal is here https://github.com/libzenith/rfcs/blob/003-laptop-cli.md/003-laptop-cli.md
-WIP code is here: https://github.com/libzenith/postgres/tree/main/pageserver/src/bin/cli
+- **neon-CLI** - interface for end-users.  Turns commands to REST requests and handles responses to show them in a user-friendly way.  
+CLI proposal is here https://github.com/neondatabase/rfcs/blob/003-laptop-cli.md/003-laptop-cli.md
+WIP code is here: https://github.com/neondatabase/postgres/tree/main/pageserver/src/bin/cli
 
-- **zenith-console** - WEB UI with same functionality as CLI.
+- **neon-console** - WEB UI with same functionality as CLI.
 >Note: not for the first release.
 
-- **zenith-local** - entrypoint. Service that starts all other components and handles REST API requests. See REST API proposal below.
-    > Idea: spawn all other components as child processes, so that we could shutdown everything by stopping zenith-local.
+- **neon-local** - entrypoint. Service that starts all other components and handles REST API requests. See REST API proposal below.
+    > Idea: spawn all other components as child processes, so that we could shutdown everything by stopping neon-local.
 
-- **zenith-pageserver** - consists of a storage and WAL-replaying service (modified PG in current implementation).
+- **neon-pageserver** - consists of a storage and WAL-replaying service (modified PG in current implementation).
 > Question: Probably, for local setup we should be able to bypass page-storage and interact directly with S3 to avoid double caching in shared buffers and page-server?
 
-WIP code is here: https://github.com/libzenith/postgres/tree/main/pageserver/src
+WIP code is here: https://github.com/neondatabase/postgres/tree/main/pageserver/src
 
-- **zenith-S3** - stores base images of the database and WAL in S3 object storage. Import and export images from/to zenith.
+- **neon-S3** - stores base images of the database and WAL in S3 object storage. Import and export images from/to neon.
 > Question: How should it operate in a local setup? Will we manage it ourselves or ask user to provide credentials for existing S3 object storage (i.e. minio)?
 > Question: Do we use it together with local page store or they are interchangeable?
 
 WIP code is ???
 
-- **zenith-safekeeper** - receives WAL from postgres, stores it durably, answers to Postgres that "sync" is succeed.
+- **neon-safekeeper** - receives WAL from postgres, stores it durably, answers to Postgres that "sync" is succeed.
 > Question: How should it operate in a local setup? In my understanding it should push WAL directly to S3 (if we use it) or store all data locally (if we use local page storage). The latter option seems meaningless (extra overhead and no gain), but it is still good to test the system.
 
-WIP code is here: https://github.com/libzenith/postgres/tree/main/src/bin/safekeeper
+WIP code is here: https://github.com/neondatabase/postgres/tree/main/src/bin/safekeeper
 
-- **zenith-computenode** - bottomless PostgreSQL, ideally upstream, but for a start - our modified version. User can quickly create and destroy them and work with it as a regular postgres database.
+- **neon-computenode** - bottomless PostgreSQL, ideally upstream, but for a start - our modified version. User can quickly create and destroy them and work with it as a regular postgres database.
  
- WIP code is in main branch and here: https://github.com/libzenith/postgres/commits/compute_node
+ WIP code is in main branch and here: https://github.com/neondatabase/postgres/commits/compute_node
 
 #### REST API:
 
 Service endpoint: `http://localhost:3000`
 
 Resources:
-- /storages - Where data lives: zenith-pageserver or zenith-s3
-- /pgs - Postgres - zenith-computenode
+- /storages - Where data lives: neon-pageserver or neon-s3
+- /pgs - Postgres - neon-computenode
 - /snapshots - snapshots **TODO**
 
->Question: Do we want to extend this API to manage zenith components? I.e. start page-server, manage safekeepers and so on? Or they will be hardcoded to just start once and for all?
+>Question: Do we want to extend this API to manage neon components? I.e. start page-server, manage safekeepers and so on? Or they will be hardcoded to just start once and for all?
 
 Methods and their mapping to CLI:
 
-- /storages - zenith-pageserver or zenith-s3
+- /storages - neon-pageserver or neon-s3
 
 CLI  | REST API
 ------------- | -------------
@@ -84,7 +84,7 @@ storage list | GET /storages
 storage show -n name | GET /storages/:storage_name 
 
 
-- /pgs - zenith-computenode
+- /pgs - neon-computenode
 
 CLI  | REST API
 ------------- | -------------

docs/rfcs/006-laptop-cli-v2-CLI.md

@@ -1,45 +1,45 @@
-Zenith CLI allows you to operate database clusters (catalog clusters) and their commit history locally and in the cloud. Since ANSI calls them catalog clusters and cluster is a loaded term in the modern infrastructure we will call it "catalog".
+Neon CLI allows you to operate database clusters (catalog clusters) and their commit history locally and in the cloud. Since ANSI calls them catalog clusters and cluster is a loaded term in the modern infrastructure we will call it "catalog".
 
 # CLI v2 (after chatting with Carl)
 
-Zenith introduces the notion of a repository.
+Neon introduces the notion of a repository.
 
 ```bash
-zenith init
-zenith clone zenith://zenith.tech/piedpiper/northwind -- clones a repo to the northwind directory
+neon init
+neon clone neon://neon.tech/piedpiper/northwind -- clones a repo to the northwind directory
 ```
 
 Once you have a cluster catalog you can explore it
 
 ```bash
-zenith log -- returns a list of commits
-zenith status -- returns if there are changes in the catalog that can be committed
-zenith commit -- commits the changes and generates a new commit hash
-zenith branch experimental <hash> -- creates a branch called testdb based on a given commit hash
+neon log -- returns a list of commits
+neon status -- returns if there are changes in the catalog that can be committed
+neon commit -- commits the changes and generates a new commit hash
+neon branch experimental <hash> -- creates a branch called testdb based on a given commit hash
 ```
 
 To make changes in the catalog you need to run compute nodes
 
 ```bash
 -- here is how you a compute node
-zenith start /home/pipedpiper/northwind:main -- starts a compute instance
-zenith start zenith://zenith.tech/northwind:main -- starts a compute instance in the cloud
+neon start /home/pipedpiper/northwind:main -- starts a compute instance
+neon start neon://neon.tech/northwind:main -- starts a compute instance in the cloud
 -- you can start a compute node against any hash or branch
-zenith start /home/pipedpiper/northwind:experimental --port 8008 -- start another compute instance (on different port)
+neon start /home/pipedpiper/northwind:experimental --port 8008 -- start another compute instance (on different port)
 -- you can start a compute node against any hash or branch
-zenith start /home/pipedpiper/northwind:<hash> --port 8009 -- start another compute instance (on different port)
+neon start /home/pipedpiper/northwind:<hash> --port 8009 -- start another compute instance (on different port)
 
 -- After running some DML you can run 
--- zenith status and see how there are two WAL streams one on top of 
+-- neon status and see how there are two WAL streams one on top of 
 -- the main branch
-zenith status 
+neon status 
 -- and another on top of the experimental branch
-zenith status -b experimental
+neon status -b experimental
 
 -- you can commit each branch separately
-zenith commit main
+neon commit main
 -- or
-zenith commit -c /home/pipedpiper/northwind:experimental
+neon commit -c /home/pipedpiper/northwind:experimental
 ```
 
 Starting compute instances against cloud environments
@@ -47,18 +47,18 @@ Starting compute instances against cloud environments
 ```bash
 -- you can start a compute instance against the cloud environment
 -- in this case all of the changes will be streamed into the cloud
-zenith start https://zenith:tech/pipedpiper/northwind:main
-zenith start https://zenith:tech/pipedpiper/northwind:main
-zenith status -c https://zenith:tech/pipedpiper/northwind:main
-zenith commit -c https://zenith:tech/pipedpiper/northwind:main
-zenith branch -c https://zenith:tech/pipedpiper/northwind:<hash> experimental
+neon start https://neon:tecj/pipedpiper/northwind:main
+neon start https://neon:tecj/pipedpiper/northwind:main
+neon status -c https://neon:tecj/pipedpiper/northwind:main
+neon commit -c https://neon:tecj/pipedpiper/northwind:main
+neon branch -c https://neon:tecj/pipedpiper/northwind:<hash> experimental
 ```
 
 Pushing data into the cloud
 
 ```bash
 -- pull all the commits from the cloud
-zenith pull
+neon pull
 -- push all the commits to the cloud
-zenith push
+neon push
 ```

docs/rfcs/006-laptop-cli-v2-repository-structure.md

@@ -1,14 +1,14 @@
 # Repository format
 
-A Zenith repository is similar to a traditional PostgreSQL backup
+A Neon repository is similar to a traditional PostgreSQL backup
 archive, like a WAL-G bucket or pgbarman backup catalogue. It holds
 multiple versions of a PostgreSQL database cluster.
 
-The distinguishing feature is that you can launch a Zenith Postgres
+The distinguishing feature is that you can launch a Neon Postgres
 server directly against a branch in the repository, without having to
-"restore" it first. Also, Zenith manages the storage automatically,
+"restore" it first. Also, Neon manages the storage automatically,
 there is no separation between full and incremental backups nor WAL
-archive. Zenith relies heavily on the WAL, and uses concepts similar
+archive. Neon relies heavily on the WAL, and uses concepts similar
 to incremental backups and WAL archiving internally, but it is hidden
 from the user.
 
@@ -19,15 +19,15 @@ efficient. Just something to get us started.
 
 The repository directory looks like this:
 
-    .zenith/timelines/4543be3daeab2ed4e58a285cbb8dd1fce6970f8c/wal/
-    .zenith/timelines/4543be3daeab2ed4e58a285cbb8dd1fce6970f8c/snapshots/<lsn>/
-    .zenith/timelines/4543be3daeab2ed4e58a285cbb8dd1fce6970f8c/history
+    .neon/timelines/4543be3daeab2ed4e58a285cbb8dd1fce6970f8c/wal/
+    .neon/timelines/4543be3daeab2ed4e58a285cbb8dd1fce6970f8c/snapshots/<lsn>/
+    .neon/timelines/4543be3daeab2ed4e58a285cbb8dd1fce6970f8c/history
     
-    .zenith/refs/branches/mybranch
-    .zenith/refs/tags/foo
-    .zenith/refs/tags/bar
+    .neon/refs/branches/mybranch
+    .neon/refs/tags/foo
+    .neon/refs/tags/bar
     
-    .zenith/datadirs/<timeline uuid>
+    .neon/datadirs/<timeline uuid>
 
 ### Timelines
 
@@ -39,7 +39,7 @@ All WAL is generated on a timeline. You can launch a read-only node
 against a tag or arbitrary LSN on a timeline, but in order to write,
 you need to create a timeline.
 
-Each timeline is stored in a directory under .zenith/timelines. It
+Each timeline is stored in a directory under .neon/timelines. It
 consists of a WAL archive, containing all the WAL in the standard
 PostgreSQL format, under the wal/ subdirectory.
 
@@ -66,18 +66,18 @@ contains the UUID of the timeline (and LSN, for tags).
 
 ### Datadirs
 
-.zenith/datadirs contains PostgreSQL data directories. You can launch
+.neon/datadirs contains PostgreSQL data directories. You can launch
 a Postgres instance on one of them with:
 
 ```
-  postgres -D .zenith/datadirs/4543be3daeab2ed4e58a285cbb8dd1fce6970f8c
+  postgres -D .neon/datadirs/4543be3daeab2ed4e58a285cbb8dd1fce6970f8c
 ```
 
 All the actual data is kept in the timeline directories, under
-.zenith/timelines. The data directories are only needed for active
+.neon/timelines. The data directories are only needed for active
 PostgreQSL instances. After an instance is stopped, the data directory
-can be safely removed. "zenith start" will recreate it quickly from
-the data in .zenith/timelines, if it's missing.
+can be safely removed. "neon start" will recreate it quickly from
+the data in .neon/timelines, if it's missing.
 
 ## Version 2
 
@@ -103,14 +103,14 @@ more advanced. The exact format is TODO. But it should support:
 
 ### Garbage collection
 
-When you run "zenith gc", old timelines that are no longer needed are
+When you run "neon gc", old timelines that are no longer needed are
 removed. That involves collecting the list of "unreachable" objects,
 starting from the named branches and tags.
 
 Also, if enough WAL has been generated on a timeline since last
 snapshot, a new snapshot or delta is created.
 
-### zenith push/pull
+### neon push/pull
 
 Compare the tags and branches on both servers, and copy missing ones.
 For each branch, compare the timeline it points to in both servers. If
@@ -123,7 +123,7 @@ every time you start up an instance? Then you would detect that the
 timelines have diverged. That would match with the "epoch" concept
 that we have in the WAL safekeeper
 
-### zenith checkout/commit
+### neon checkout/commit
 
 In this format, there is no concept of a "working tree", and hence no
 concept of checking out or committing. All modifications are done on
@@ -134,7 +134,7 @@ You can easily fork off a temporary timeline to emulate a "working tree".
 You can later remove it and have it garbage collected, or to "commit",
 re-point the branch to the new timeline.
 
-If we want to have a worktree and "zenith checkout/commit" concept, we can
+If we want to have a worktree and "neon checkout/commit" concept, we can
 emulate that with a temporary timeline. Create the temporary timeline at
-"zenith checkout", and have "zenith commit" modify the branch to point to
+"neon checkout", and have "neon commit" modify the branch to point to
 the new timeline.

docs/rfcs/007-serverless-on-laptop.md

@@ -4,27 +4,27 @@ How it works now
 1. Create repository, start page server on it
 
 ```
-$ zenith init
+$ neon init
 ...
 created main branch
-new zenith repository was created in .zenith
+new neon repository was created in .neon
 
-$ zenith pageserver start
-Starting pageserver at '127.0.0.1:64000' in .zenith
+$ neon pageserver start
+Starting pageserver at '127.0.0.1:64000' in .neon
 Page server started
 ```
 
 2. Create a branch, and start a Postgres instance on it
 
 ```
-$ zenith branch heikki main
+$ neon branch heikki main
 branching at end of WAL: 0/15ECF68
 
-$ zenith pg create heikki
+$ neon pg create heikki
 Initializing Postgres on timeline 76cf9279915be7797095241638e64644...
-Extracting base backup to create postgres instance: path=.zenith/pgdatadirs/pg1 port=55432
+Extracting base backup to create postgres instance: path=.neon/pgdatadirs/pg1 port=55432
 
-$ zenith pg start pg1
+$ neon pg start pg1
 Starting postgres node at 'host=127.0.0.1 port=55432 user=heikki'
 waiting for server to start.... done
 server started
@@ -52,20 +52,20 @@ serverless on your laptop, so that the workflow becomes just:
 1. Create repository, start page server on it (same as before)
 
 ```
-$ zenith init
+$ neon init
 ...
 created main branch
-new zenith repository was created in .zenith
+new neon repository was created in .neon
 
-$ zenith pageserver start
-Starting pageserver at '127.0.0.1:64000' in .zenith
+$ neon pageserver start
+Starting pageserver at '127.0.0.1:64000' in .neon
 Page server started
 ```
 
 2. Create branch
 
 ```
-$ zenith branch heikki main
+$ neon branch heikki main
 branching at end of WAL: 0/15ECF68
 ```
 

docs/rfcs/008-push-pull.md

@@ -7,22 +7,22 @@ Here is a proposal about implementing push/pull mechanics between pageservers. W
 The origin represents connection info for some remote pageserver. Let's use here same commands as git uses except using explicit list subcommand (git uses `origin -v` for that).
 
 ```
-zenith origin add <name> <connection_uri>
-zenith origin list
-zenith origin remove <name>
+neon origin add <name> <connection_uri>
+neon origin list
+neon origin remove <name>
 ```
 
 Connection URI a string of form `postgresql://user:pass@hostname:port` (https://www.postgresql.org/docs/13/libpq-connect.html#id-1.7.3.8.3.6). We can start with libpq password auth and later add support for client certs or require ssh as transport or invent some other kind of transport.
 
-Behind the scenes, this commands may update toml file inside .zenith directory.
+Behind the scenes, this commands may update toml file inside .neon directory.
 
 ## Push
 
 ### Pushing branch
 
 ```
-zenith push mybranch cloudserver # push to eponymous branch in cloudserver
-zenith push mybranch cloudserver:otherbranch # push to a different branch in cloudserver
+neon push mybranch cloudserver # push to eponymous branch in cloudserver
+neon push mybranch cloudserver:otherbranch # push to a different branch in cloudserver
 ```
 
 Exact mechanics would be slightly different in the following situations:

docs/rfcs/009-snapshot-first-storage-cli.md

@@ -2,7 +2,7 @@ While working on export/import commands, I understood that they fit really well
 
 We may think about backups as snapshots in a different format (i.e plain pgdata format, basebackup tar format, WAL-G format (if they want to support it) and so on). They use same storage API, the only difference is the code that packs/unpacks files.
 
-Even if zenith aims to maintains durability using it's own snapshots, backups will be useful for uploading data from postgres to zenith.
+Even if neon aims to maintains durability using it's own snapshots, backups will be useful for uploading data from postgres to neon.
 
 So here is an attempt to design consistent CLI for different usage scenarios:
 
@@ -16,8 +16,8 @@ Save`storage_dest` and other parameters in config.
 Push snapshots to `storage_dest` in background.
 
 ```
-zenith init --storage_dest=S3_PREFIX
-zenith start
+neon init --storage_dest=S3_PREFIX
+neon start
 ```
 
 #### 2. Restart pageserver (manually or crash-recovery).
@@ -25,7 +25,7 @@ Take `storage_dest` from pageserver config, start pageserver from latest snapsho
 Push snapshots to `storage_dest` in background.
 
 ```
-zenith start
+neon start
 ```
 
 #### 3. Import.
@@ -35,22 +35,22 @@ Do not save `snapshot_path` and `snapshot_format` in config, as it is a one-time
 Save`storage_dest` parameters in config.
 Push snapshots to `storage_dest` in background.
 ```
-//I.e. we want to start zenith on top of existing $PGDATA and use s3 as a persistent storage.
-zenith init --snapshot_path=FILE_PREFIX --snapshot_format=pgdata --storage_dest=S3_PREFIX
-zenith start
+//I.e. we want to start neon on top of existing $PGDATA and use s3 as a persistent storage.
+neon init --snapshot_path=FILE_PREFIX --snapshot_format=pgdata --storage_dest=S3_PREFIX
+neon start
 ```
 How to pass credentials needed for `snapshot_path`?
 
 #### 4. Export.
 Manually push snapshot to `snapshot_path` which differs from `storage_dest`
-Optionally set `snapshot_format`, which can be plain pgdata format or zenith format.
+Optionally set `snapshot_format`, which can be plain pgdata format or neon format.
 ```
-zenith export --snapshot_path=FILE_PREFIX --snapshot_format=pgdata
+neon export --snapshot_path=FILE_PREFIX --snapshot_format=pgdata
 ```
 
 #### Notes and questions
 - safekeeper s3_offload should use same (similar) syntax for storage. How to set it in UI?
-- Why do we need `zenith init` as a separate command? Can't we init everything at first start?
+- Why do we need `neon init` as a separate command? Can't we init everything at first start?
 - We can think of better names for all options.
 - Export to plain postgres format will be useless, if we are not 100% compatible on page level.
 I can recall at least one such difference - PD_WAL_LOGGED flag in pages.

docs/rfcs/013-term-history.md

@@ -9,7 +9,7 @@ receival and this might lag behind `term`; safekeeper switches to epoch `n` when
 it has received all committed log records from all `< n` terms. This roughly
 corresponds to proposed in
 
-https://github.com/zenithdb/rfcs/pull/3/files
+https://github.com/neondatabase/rfcs/pull/3/files
 
 
 This makes our biggest our difference from Raft. In Raft, every log record is

docs/rfcs/014-safekeepers-gossip.md

@@ -1,6 +1,6 @@
 # Safekeeper gossip
 
-Extracted from this [PR](https://github.com/zenithdb/rfcs/pull/13)
+Extracted from this [PR](https://github.com/neondatabase/rfcs/pull/13)
 
 ## Motivation
 

docs/rfcs/015-storage-messaging.md

@@ -2,7 +2,7 @@
 
 Created on 19.01.22
 
-Initially created [here](https://github.com/zenithdb/rfcs/pull/16) by @kelvich.
+Initially created [here](https://github.com/neondatabase/rfcs/pull/16) by @kelvich.
 
 That it is an alternative to (014-safekeeper-gossip)[]
 
@@ -292,4 +292,4 @@ But with an etcd we are in a bit different situation:
 1. We don't need persistency and strong consistency guarantees for the data we store in the etcd
 2. etcd uses Grpc as a protocol, and messages are pretty simple
 
-So it looks like implementing in-mem store with etcd interface is straightforward thing _if we will want that in future_. At the same time, we can avoid implementing it right now, and we will be able to run local zenith installation with etcd running somewhere in the background (as opposed to building and running console, which in turn requires Postgres).
+So it looks like implementing in-mem store with etcd interface is straightforward thing _if we will want that in future_. At the same time, we can avoid implementing it right now, and we will be able to run local neon installation with etcd running somewhere in the background (as opposed to building and running console, which in turn requires Postgres).

libs/pageserver_api/src/models.rs

@@ -14,7 +14,6 @@ use byteorder::{BigEndian, ReadBytesExt};
 use postgres_ffi::BLCKSZ;
 use serde::{Deserialize, Serialize};
 use serde_with::serde_as;
-use strum_macros;
 use utils::{
     completion,
     history_buffer::HistoryBufferWithDropCounter,
@@ -436,6 +435,8 @@ pub struct TenantShardLocation {
 #[serde(deny_unknown_fields)]
 pub struct TenantLocationConfigResponse {
     pub shards: Vec<TenantShardLocation>,
+    // If the shards' ShardCount count is >1, stripe_size will be set.
+    pub stripe_size: Option<ShardStripeSize>,
 }
 
 #[derive(Serialize, Deserialize, Debug)]
@@ -1077,7 +1078,6 @@ impl PagestreamBeMessage {
 
 #[cfg(test)]
 mod tests {
-    use bytes::Buf;
     use serde_json::json;
 
     use super::*;

libs/pageserver_api/src/shard.rs

@@ -6,7 +6,6 @@ use crate::{
 };
 use hex::FromHex;
 use serde::{Deserialize, Serialize};
-use thiserror;
 use utils::id::TenantId;
 
 #[derive(Ord, PartialOrd, Eq, PartialEq, Clone, Copy, Serialize, Deserialize, Debug, Hash)]
@@ -656,10 +655,7 @@ fn key_to_shard_number(count: ShardCount, stripe_size: ShardStripeSize, key: &Ke
 
 #[cfg(test)]
 mod tests {
-    use std::str::FromStr;
-
-    use bincode;
-    use utils::{id::TenantId, Hex};
+    use utils::Hex;
 
     use super::*;
 

libs/postgres_backend/src/lib.rs

@@ -6,7 +6,6 @@
 #![deny(clippy::undocumented_unsafe_blocks)]
 use anyhow::Context;
 use bytes::Bytes;
-use futures::pin_mut;
 use serde::{Deserialize, Serialize};
 use std::io::ErrorKind;
 use std::net::SocketAddr;
@@ -378,8 +377,7 @@ impl<IO: AsyncRead + AsyncWrite + Unpin> PostgresBackend<IO> {
         &mut self,
         cx: &mut std::task::Context<'_>,
     ) -> Poll<Result<(), std::io::Error>> {
-        let flush_fut = self.flush();
-        pin_mut!(flush_fut);
+        let flush_fut = std::pin::pin!(self.flush());
         flush_fut.poll(cx)
     }
 

libs/remote_storage/src/local_fs.rs

@@ -623,9 +623,7 @@ fn file_exists(file_path: &Utf8Path) -> anyhow::Result<bool> {
 mod fs_tests {
     use super::*;
 
-    use bytes::Bytes;
     use camino_tempfile::tempdir;
-    use futures_util::Stream;
     use std::{collections::HashMap, io::Write};
 
     async fn read_and_check_metadata(

libs/remote_storage/src/s3_bucket.rs

@@ -1040,7 +1040,7 @@ mod tests {
             Some("test/prefix/"),
             Some("/test/prefix/"),
         ];
-        let expected_outputs = vec![
+        let expected_outputs = [
             vec!["", "some/path", "some/path"],
             vec!["/", "/some/path", "/some/path"],
             vec![

libs/utils/src/auth.rs

@@ -1,7 +1,6 @@
 // For details about authentication see docs/authentication.md
 
 use arc_swap::ArcSwap;
-use serde;
 use std::{borrow::Cow, fmt::Display, fs, sync::Arc};
 
 use anyhow::Result;

libs/utils/src/completion.rs

@@ -4,7 +4,9 @@ use tokio_util::task::{task_tracker::TaskTrackerToken, TaskTracker};
 ///
 /// Can be cloned, moved and kept around in futures as "guard objects".
 #[derive(Clone)]
-pub struct Completion(TaskTrackerToken);
+pub struct Completion {
+    _token: TaskTrackerToken,
+}
 
 /// Barrier will wait until all clones of [`Completion`] have been dropped.
 #[derive(Clone)]
@@ -49,5 +51,5 @@ pub fn channel() -> (Completion, Barrier) {
     tracker.close();
 
     let token = tracker.token();
-    (Completion(token), Barrier(tracker))
+    (Completion { _token: token }, Barrier(tracker))
 }

libs/utils/src/generation.rs

@@ -45,7 +45,7 @@ impl Generation {
         Self::Broken
     }
 
-    pub fn new(v: u32) -> Self {
+    pub const fn new(v: u32) -> Self {
         Self::Valid(v)
     }
 

libs/utils/src/http/endpoint.rs

@@ -9,7 +9,7 @@ use metrics::{register_int_counter, Encoder, IntCounter, TextEncoder};
 use once_cell::sync::Lazy;
 use routerify::ext::RequestExt;
 use routerify::{Middleware, RequestInfo, Router, RouterBuilder};
-use tracing::{self, debug, info, info_span, warn, Instrument};
+use tracing::{debug, info, info_span, warn, Instrument};
 
 use std::future::Future;
 use std::str::FromStr;
@@ -156,6 +156,10 @@ pub struct ChannelWriter {
     buffer: BytesMut,
     pub tx: mpsc::Sender<std::io::Result<Bytes>>,
     written: usize,
+    /// Time spent waiting for the channel to make progress. It is not the same as time to upload a
+    /// buffer because we cannot know anything about that, but this should allow us to understand
+    /// the actual time taken without the time spent `std::thread::park`ed.
+    wait_time: std::time::Duration,
 }
 
 impl ChannelWriter {
@@ -168,6 +172,7 @@ impl ChannelWriter {
             buffer: BytesMut::with_capacity(buf_len).split_off(buf_len / 2),
             tx,
             written: 0,
+            wait_time: std::time::Duration::ZERO,
         }
     }
 
@@ -180,6 +185,8 @@ impl ChannelWriter {
         tracing::trace!(n, "flushing");
         let ready = self.buffer.split().freeze();
 
+        let wait_started_at = std::time::Instant::now();
+
         // not ideal to call from blocking code to block_on, but we are sure that this
         // operation does not spawn_blocking other tasks
         let res: Result<(), ()> = tokio::runtime::Handle::current().block_on(async {
@@ -192,6 +199,9 @@ impl ChannelWriter {
             // sending it to the client.
             Ok(())
         });
+
+        self.wait_time += wait_started_at.elapsed();
+
         if res.is_err() {
             return Err(std::io::ErrorKind::BrokenPipe.into());
         }
@@ -202,6 +212,10 @@ impl ChannelWriter {
     pub fn flushed_bytes(&self) -> usize {
         self.written
     }
+
+    pub fn wait_time(&self) -> std::time::Duration {
+        self.wait_time
+    }
 }
 
 impl std::io::Write for ChannelWriter {
@@ -252,22 +266,52 @@ async fn prometheus_metrics_handler(_req: Request<Body>) -> Result<Response<Body
 
     let span = info_span!("blocking");
     tokio::task::spawn_blocking(move || {
+        // there are situations where we lose scraped metrics under load, try to gather some clues
+        // since all nodes are queried this, keep the message count low.
+        let spawned_at = std::time::Instant::now();
+
         let _span = span.entered();
+
         let metrics = metrics::gather();
+
+        let gathered_at = std::time::Instant::now();
+
         let res = encoder
             .encode(&metrics, &mut writer)
             .and_then(|_| writer.flush().map_err(|e| e.into()));
 
+        // this instant is not when we finally got the full response sent, sending is done by hyper
+        // in another task.
+        let encoded_at = std::time::Instant::now();
+
+        let spawned_in = spawned_at - started_at;
+        let collected_in = gathered_at - spawned_at;
+        // remove the wait time here in case the tcp connection was clogged
+        let encoded_in = encoded_at - gathered_at - writer.wait_time();
+        let total = encoded_at - started_at;
+
         match res {
             Ok(()) => {
                 tracing::info!(
                     bytes = writer.flushed_bytes(),
-                    elapsed_ms = started_at.elapsed().as_millis(),
+                    total_ms = total.as_millis(),
+                    spawning_ms = spawned_in.as_millis(),
+                    collection_ms = collected_in.as_millis(),
+                    encoding_ms = encoded_in.as_millis(),
                     "responded /metrics"
                 );
             }
             Err(e) => {
-                tracing::warn!("failed to write out /metrics response: {e:#}");
+                // there is a chance that this error is not the BrokenPipe we generate in the writer
+                // for "closed connection", but it is highly unlikely.
+                tracing::warn!(
+                    after_bytes = writer.flushed_bytes(),
+                    total_ms = total.as_millis(),
+                    spawning_ms = spawned_in.as_millis(),
+                    collection_ms = collected_in.as_millis(),
+                    encoding_ms = encoded_in.as_millis(),
+                    "failed to write out /metrics response: {e:?}"
+                );
                 // semantics of this error are quite... unclear. we want to error the stream out to
                 // abort the response to somehow notify the client that we failed.
                 //

libs/utils/src/lsn.rs

@@ -415,7 +415,6 @@ mod tests {
 
     use super::*;
 
-    use serde::ser::Serialize;
     use serde_assert::{Deserializer, Serializer, Token, Tokens};
 
     #[test]

libs/utils/src/seqwait.rs

@@ -1,6 +1,6 @@
 #![warn(missing_docs)]
 
-use std::cmp::{Eq, Ordering, PartialOrd};
+use std::cmp::{Eq, Ordering};
 use std::collections::BinaryHeap;
 use std::fmt::Debug;
 use std::mem;
@@ -249,7 +249,6 @@ where
 mod tests {
     use super::*;
     use std::sync::Arc;
-    use std::time::Duration;
 
     impl MonotonicCounter<i32> for i32 {
         fn cnt_advance(&mut self, val: i32) {

libs/utils/src/simple_rcu.rs

@@ -221,7 +221,7 @@ impl RcuWaitList {
 #[cfg(test)]
 mod tests {
     use super::*;
-    use std::sync::{Arc, Mutex};
+    use std::sync::Mutex;
     use std::time::Duration;
 
     #[tokio::test]

libs/utils/src/sync/heavier_once_cell.rs

@@ -239,7 +239,6 @@ mod tests {
     use std::{
         convert::Infallible,
         pin::{pin, Pin},
-        sync::atomic::{AtomicUsize, Ordering},
         time::Duration,
     };
 

pageserver/client/src/mgmt_api.rs

@@ -251,21 +251,30 @@ impl Client {
         tenant_shard_id: TenantShardId,
         config: LocationConfig,
         flush_ms: Option<std::time::Duration>,
+        lazy: bool,
     ) -> Result<()> {
         let req_body = TenantLocationConfigRequest {
             tenant_id: tenant_shard_id,
             config,
         };
-        let path = format!(
+
+        let mut path = reqwest::Url::parse(&format!(
             "{}/v1/tenant/{}/location_config",
             self.mgmt_api_endpoint, tenant_shard_id
-        );
-        let path = if let Some(flush_ms) = flush_ms {
-            format!("{}?flush_ms={}", path, flush_ms.as_millis())
-        } else {
-            path
-        };
-        self.request(Method::PUT, &path, &req_body).await?;
+        ))
+        // Should always work: mgmt_api_endpoint is configuration, not user input.
+        .expect("Cannot build URL");
+
+        if lazy {
+            path.query_pairs_mut().append_pair("lazy", "true");
+        }
+
+        if let Some(flush_ms) = flush_ms {
+            path.query_pairs_mut()
+                .append_pair("flush_ms", &format!("{}", flush_ms.as_millis()));
+        }
+
+        self.request(Method::PUT, path, &req_body).await?;
         Ok(())
     }
 

pageserver/compaction/src/compact_tiered.rs

@@ -63,7 +63,7 @@ pub async fn compact_tiered<E: CompactionJobExecutor>(
         );
 
         // Identify the range of LSNs that belong to this level. We assume that
-        // each file in this level span an LSN range up to 1.75x target file
+        // each file in this level spans an LSN range up to 1.75x target file
         // size. That should give us enough slop that if we created a slightly
         // oversized L0 layer, e.g. because flushing the in-memory layer was
         // delayed for some reason, we don't consider the oversized layer to
@@ -248,7 +248,6 @@ enum CompactionStrategy {
     CreateImage,
 }
 
-#[allow(dead_code)] // Todo
 struct CompactionJob<E: CompactionJobExecutor> {
     key_range: Range<E::Key>,
     lsn_range: Range<Lsn>,
@@ -345,7 +344,7 @@ where
     ///
     /// TODO: Currently, this is called exactly once for the level, and we
     /// decide whether to create new image layers to cover the whole level, or
-    /// write a new set of delta. In the future, this should try to partition
+    /// write a new set of deltas. In the future, this should try to partition
     /// the key space, and make the decision separately for each partition.
     async fn divide_job(&mut self, job_id: JobId, ctx: &E::RequestContext) -> anyhow::Result<()> {
         let job = &self.jobs[job_id.0];
@@ -709,18 +708,6 @@ where
     }
 }
 
-// Sliding window through keyspace and values
-//
-// This is used to decide what layer to write next, from the beginning of the window.
-//
-// Candidates:
-//
-// 1. Create an image layer, snapping to previous images
-// 2. Create a delta layer, snapping to previous images
-// 3. Create an image layer, snapping to
-//
-//
-
 // Take previous partitioning, based on the image layers below.
 //
 // Candidate is at the front:
@@ -739,6 +726,10 @@ struct WindowElement<K> {
     last_key: K,  // inclusive
     accum_size: u64,
 }
+
+// Sliding window through keyspace and values
+//
+// This is used to decide what layer to write next, from the beginning of the window.
 struct Window<K> {
     elems: VecDeque<WindowElement<K>>,
 

pageserver/compaction/src/helpers.rs

@@ -6,7 +6,6 @@ use futures::future::BoxFuture;
 use futures::{Stream, StreamExt};
 use itertools::Itertools;
 use pin_project_lite::pin_project;
-use std::cmp::Ord;
 use std::collections::BinaryHeap;
 use std::collections::VecDeque;
 use std::future::Future;

pageserver/compaction/src/identify_levels.rs

@@ -1,5 +1,5 @@
-//! An LSM tree consists of multiple levels, each exponential larger than the
-//! previous level. And each level consists of be multiple "tiers". With tiered
+//! An LSM tree consists of multiple levels, each exponentially larger than the
+//! previous level. And each level consists of multiple "tiers". With tiered
 //! compaction, a level is compacted when it has accumulated more than N tiers,
 //! forming one tier on the next level.
 //!
@@ -170,13 +170,6 @@ where
     })
 }
 
-// helper struct used in depth()
-struct Event<K> {
-    key: K,
-    layer_idx: usize,
-    start: bool,
-}
-
 impl<L> Level<L> {
     /// Count the number of deltas stacked on each other.
     pub fn depth<K>(&self) -> u64
@@ -184,6 +177,11 @@ impl<L> Level<L> {
         K: CompactionKey,
         L: CompactionLayer<K>,
     {
+        struct Event<K> {
+            key: K,
+            layer_idx: usize,
+            start: bool,
+        }
         let mut events: Vec<Event<K>> = Vec::new();
         for (idx, l) in self.layers.iter().enumerate() {
             events.push(Event {
@@ -202,7 +200,7 @@ impl<L> Level<L> {
         // Sweep the key space left to right. Stop at each distinct key, and
         // count the number of deltas on top of the highest image at that key.
         //
-        // This is a little enefficient, as we walk through the active_set on
+        // This is a little inefficient, as we walk through the active_set on
         // every key. We could increment/decrement a counter on each step
         // instead, but that'd require a bit more complex bookkeeping.
         let mut active_set: BTreeSet<(Lsn, bool, usize)> = BTreeSet::new();
@@ -236,6 +234,7 @@ impl<L> Level<L> {
                 }
             }
         }
+        debug_assert_eq!(active_set, BTreeSet::new());
         max_depth
     }
 }

pageserver/compaction/src/interface.rs

@@ -4,12 +4,12 @@
 //! All the heavy lifting is done by the create_image and create_delta
 //! functions that the implementor provides.
 use async_trait::async_trait;
+use futures::Future;
 use pageserver_api::{key::Key, keyspace::key_range_size};
 use std::ops::Range;
 use utils::lsn::Lsn;
 
 /// Public interface. This is the main thing that the implementor needs to provide
-#[async_trait]
 pub trait CompactionJobExecutor {
     // Type system.
     //
@@ -17,8 +17,7 @@ pub trait CompactionJobExecutor {
     // compaction doesn't distinguish whether they are stored locally or
     // remotely.
     //
-    // The keyspace is defined by CompactionKey trait.
-    //
+    // The keyspace is defined by the CompactionKey trait.
     type Key: CompactionKey;
 
     type Layer: CompactionLayer<Self::Key> + Clone;
@@ -35,27 +34,27 @@ pub trait CompactionJobExecutor {
     // ----
 
     /// Return all layers that overlap the given bounding box.
-    async fn get_layers(
+    fn get_layers(
         &mut self,
         key_range: &Range<Self::Key>,
         lsn_range: &Range<Lsn>,
         ctx: &Self::RequestContext,
-    ) -> anyhow::Result<Vec<Self::Layer>>;
+    ) -> impl Future<Output = anyhow::Result<Vec<Self::Layer>>> + Send;
 
-    async fn get_keyspace(
+    fn get_keyspace(
         &mut self,
         key_range: &Range<Self::Key>,
         lsn: Lsn,
         ctx: &Self::RequestContext,
-    ) -> anyhow::Result<CompactionKeySpace<Self::Key>>;
+    ) -> impl Future<Output = anyhow::Result<CompactionKeySpace<Self::Key>>> + Send;
 
     /// NB: This is a pretty expensive operation. In the real pageserver
     /// implementation, it downloads the layer, and keeps it resident
     /// until the DeltaLayer is dropped.
-    async fn downcast_delta_layer(
+    fn downcast_delta_layer(
         &self,
         layer: &Self::Layer,
-    ) -> anyhow::Result<Option<Self::DeltaLayer>>;
+    ) -> impl Future<Output = anyhow::Result<Option<Self::DeltaLayer>>> + Send;
 
     // ----
     // Functions to execute the plan
@@ -63,33 +62,33 @@ pub trait CompactionJobExecutor {
 
     /// Create a new image layer, materializing all the values in the key range,
     /// at given 'lsn'.
-    async fn create_image(
+    fn create_image(
         &mut self,
         lsn: Lsn,
         key_range: &Range<Self::Key>,
         ctx: &Self::RequestContext,
-    ) -> anyhow::Result<()>;
+    ) -> impl Future<Output = anyhow::Result<()>> + Send;
 
     /// Create a new delta layer, containing all the values from 'input_layers'
     /// in the given key and LSN range.
-    async fn create_delta(
+    fn create_delta(
         &mut self,
         lsn_range: &Range<Lsn>,
         key_range: &Range<Self::Key>,
         input_layers: &[Self::DeltaLayer],
         ctx: &Self::RequestContext,
-    ) -> anyhow::Result<()>;
+    ) -> impl Future<Output = anyhow::Result<()>> + Send;
 
     /// Delete a layer. The compaction implementation will call this only after
     /// all the create_image() or create_delta() calls that deletion of this
     /// layer depends on have finished. But if the implementor has extra lazy
-    /// background tasks, like uploading the index json file to remote storage,
+    /// background tasks, like uploading the index json file to remote storage.
     /// it is the implementation's responsibility to track those.
-    async fn delete_layer(
+    fn delete_layer(
         &mut self,
         layer: &Self::Layer,
         ctx: &Self::RequestContext,
-    ) -> anyhow::Result<()>;
+    ) -> impl Future<Output = anyhow::Result<()>> + Send;
 }
 
 pub trait CompactionKey: std::cmp::Ord + Clone + Copy + std::fmt::Display {

pageserver/compaction/src/simulator.rs

@@ -429,7 +429,6 @@ impl From<&Arc<MockImageLayer>> for MockLayer {
     }
 }
 
-#[async_trait]
 impl interface::CompactionJobExecutor for MockTimeline {
     type Key = Key;
     type Layer = MockLayer;

pageserver/src/config.rs

@@ -20,7 +20,6 @@ use std::num::NonZeroUsize;
 use std::str::FromStr;
 use std::sync::Arc;
 use std::time::Duration;
-use toml_edit;
 use toml_edit::{Document, Item};
 
 use camino::{Utf8Path, Utf8PathBuf};
@@ -1203,10 +1202,7 @@ impl ConfigurableSemaphore {
 
 #[cfg(test)]
 mod tests {
-    use std::{
-        fs,
-        num::{NonZeroU32, NonZeroUsize},
-    };
+    use std::{fs, num::NonZeroU32};
 
     use camino_tempfile::{tempdir, Utf8TempDir};
     use pageserver_api::models::EvictionPolicy;

pageserver/src/consumption_metrics/metrics/tests.rs

@@ -1,7 +1,5 @@
 use super::*;
 use std::collections::HashMap;
-use std::time::SystemTime;
-use utils::lsn::Lsn;
 
 #[test]
 fn startup_collected_timeline_metrics_before_advancing() {

pageserver/src/context.rs

@@ -88,13 +88,16 @@
 
 use crate::task_mgr::TaskKind;
 
+pub(crate) mod optional_counter;
+
 // The main structure of this module, see module-level comment.
-#[derive(Clone, Debug)]
+#[derive(Debug)]
 pub struct RequestContext {
     task_kind: TaskKind,
     download_behavior: DownloadBehavior,
     access_stats_behavior: AccessStatsBehavior,
     page_content_kind: PageContentKind,
+    pub micros_spent_throttled: optional_counter::MicroSecondsCounterU32,
 }
 
 /// The kind of access to the page cache.
@@ -150,6 +153,7 @@ impl RequestContextBuilder {
                 download_behavior: DownloadBehavior::Download,
                 access_stats_behavior: AccessStatsBehavior::Update,
                 page_content_kind: PageContentKind::Unknown,
+                micros_spent_throttled: Default::default(),
             },
         }
     }
@@ -163,6 +167,7 @@ impl RequestContextBuilder {
                 download_behavior: original.download_behavior,
                 access_stats_behavior: original.access_stats_behavior,
                 page_content_kind: original.page_content_kind,
+                micros_spent_throttled: Default::default(),
             },
         }
     }

pageserver/src/context/optional_counter.rs

@@ -0,0 +1,101 @@
+use std::{
+    sync::atomic::{AtomicU32, Ordering},
+    time::Duration,
+};
+
+#[derive(Debug)]
+pub struct CounterU32 {
+    inner: AtomicU32,
+}
+impl Default for CounterU32 {
+    fn default() -> Self {
+        Self {
+            inner: AtomicU32::new(u32::MAX),
+        }
+    }
+}
+impl CounterU32 {
+    pub fn open(&self) -> Result<(), &'static str> {
+        match self
+            .inner
+            .compare_exchange(u32::MAX, 0, Ordering::Relaxed, Ordering::Relaxed)
+        {
+            Ok(_) => Ok(()),
+            Err(_) => Err("open() called on clsoed state"),
+        }
+    }
+    pub fn close(&self) -> Result<u32, &'static str> {
+        match self.inner.swap(u32::MAX, Ordering::Relaxed) {
+            u32::MAX => Err("close() called on closed state"),
+            x => Ok(x),
+        }
+    }
+
+    pub fn add(&self, count: u32) -> Result<(), &'static str> {
+        if count == 0 {
+            return Ok(());
+        }
+        let mut had_err = None;
+        self.inner
+            .fetch_update(Ordering::Relaxed, Ordering::Relaxed, |cur| match cur {
+                u32::MAX => {
+                    had_err = Some("add() called on closed state");
+                    None
+                }
+                x => {
+                    let (new, overflowed) = x.overflowing_add(count);
+                    if new == u32::MAX || overflowed {
+                        had_err = Some("add() overflowed the counter");
+                        None
+                    } else {
+                        Some(new)
+                    }
+                }
+            })
+            .map_err(|_| had_err.expect("we set it whenever the function returns None"))
+            .map(|_| ())
+    }
+}
+
+#[derive(Default, Debug)]
+pub struct MicroSecondsCounterU32 {
+    inner: CounterU32,
+}
+
+impl MicroSecondsCounterU32 {
+    pub fn open(&self) -> Result<(), &'static str> {
+        self.inner.open()
+    }
+    pub fn add(&self, duration: Duration) -> Result<(), &'static str> {
+        match duration.as_micros().try_into() {
+            Ok(x) => self.inner.add(x),
+            Err(_) => Err("add(): duration conversion error"),
+        }
+    }
+    pub fn close_and_checked_sub_from(&self, from: Duration) -> Result<Duration, &'static str> {
+        let val = self.inner.close()?;
+        let val = Duration::from_micros(val as u64);
+        let subbed = match from.checked_sub(val) {
+            Some(v) => v,
+            None => return Err("Duration::checked_sub"),
+        };
+        Ok(subbed)
+    }
+}
+
+#[cfg(test)]
+mod tests {
+
+    use super::*;
+
+    #[test]
+    fn test_basic() {
+        let counter = MicroSecondsCounterU32::default();
+        counter.open().unwrap();
+        counter.add(Duration::from_micros(23)).unwrap();
+        let res = counter
+            .close_and_checked_sub_from(Duration::from_micros(42))
+            .unwrap();
+        assert_eq!(res, Duration::from_micros(42 - 23));
+    }
+}

pageserver/src/deletion_queue.rs

@@ -20,10 +20,9 @@ use remote_storage::{GenericRemoteStorage, RemotePath};
 use serde::Deserialize;
 use serde::Serialize;
 use thiserror::Error;
-use tokio;
 use tokio_util::sync::CancellationToken;
 use tracing::Instrument;
-use tracing::{self, debug, error};
+use tracing::{debug, error};
 use utils::crashsafe::path_with_suffix_extension;
 use utils::generation::Generation;
 use utils::id::TimelineId;
@@ -726,7 +725,7 @@ mod test {
     use camino::Utf8Path;
     use hex_literal::hex;
     use pageserver_api::shard::ShardIndex;
-    use std::{io::ErrorKind, time::Duration};
+    use std::io::ErrorKind;
     use tracing::info;
 
     use remote_storage::{RemoteStorageConfig, RemoteStorageKind};
@@ -735,10 +734,7 @@ mod test {
     use crate::{
         control_plane_client::RetryForeverError,
         repository::Key,
-        tenant::{
-            harness::TenantHarness, remote_timeline_client::remote_timeline_path,
-            storage_layer::DeltaFileName,
-        },
+        tenant::{harness::TenantHarness, storage_layer::DeltaFileName},
     };
 
     use super::*;
@@ -1161,13 +1157,8 @@ mod test {
 pub(crate) mod mock {
     use tracing::info;
 
-    use crate::tenant::remote_timeline_client::remote_layer_path;
-
     use super::*;
-    use std::sync::{
-        atomic::{AtomicUsize, Ordering},
-        Arc,
-    };
+    use std::sync::atomic::{AtomicUsize, Ordering};
 
     pub struct ConsumerState {
         rx: tokio::sync::mpsc::UnboundedReceiver<ListWriterQueueMessage>,

pageserver/src/disk_usage_eviction_task.rs

@@ -58,6 +58,7 @@ use utils::{completion, id::TimelineId};
 
 use crate::{
     config::PageServerConf,
+    metrics::disk_usage_based_eviction::METRICS,
     task_mgr::{self, TaskKind, BACKGROUND_RUNTIME},
     tenant::{
         self,
@@ -65,7 +66,6 @@ use crate::{
         remote_timeline_client::LayerFileMetadata,
         secondary::SecondaryTenant,
         storage_layer::{AsLayerDesc, EvictionError, Layer, LayerFileName},
-        Timeline,
     },
 };
 
@@ -409,13 +409,23 @@ pub(crate) async fn disk_usage_eviction_task_iteration_impl<U: Usage>(
         "running disk usage based eviction due to pressure"
     );
 
-    let candidates =
+    let (candidates, collection_time) = {
+        let started_at = std::time::Instant::now();
         match collect_eviction_candidates(tenant_manager, eviction_order, cancel).await? {
             EvictionCandidates::Cancelled => {
                 return Ok(IterationOutcome::Cancelled);
             }
-            EvictionCandidates::Finished(partitioned) => partitioned,
-        };
+            EvictionCandidates::Finished(partitioned) => (partitioned, started_at.elapsed()),
+        }
+    };
+
+    METRICS.layers_collected.inc_by(candidates.len() as u64);
+
+    tracing::info!(
+        elapsed_ms = collection_time.as_millis(),
+        total_layers = candidates.len(),
+        "collection completed"
+    );
 
     // Debug-log the list of candidates
     let now = SystemTime::now();
@@ -446,9 +456,10 @@ pub(crate) async fn disk_usage_eviction_task_iteration_impl<U: Usage>(
     // the tenant's min-resident-size threshold, print a warning, and memorize the disk
     // usage at that point, in 'usage_planned_min_resident_size_respecting'.
 
-    let selection = select_victims(&candidates, usage_pre);
+    let (evicted_amount, usage_planned) =
+        select_victims(&candidates, usage_pre).into_amount_and_planned();
 
-    let (evicted_amount, usage_planned) = selection.into_amount_and_planned();
+    METRICS.layers_selected.inc_by(evicted_amount as u64);
 
     // phase2: evict layers
 
@@ -477,9 +488,15 @@ pub(crate) async fn disk_usage_eviction_task_iteration_impl<U: Usage>(
             if let Some(next) = next {
                 match next {
                     Ok(Ok(file_size)) => {
+                        METRICS.layers_evicted.inc();
                         usage_assumed.add_available_bytes(file_size);
                     }
-                    Ok(Err((file_size, EvictionError::NotFound | EvictionError::Downloaded))) => {
+                    Ok(Err((
+                        file_size,
+                        EvictionError::NotFound
+                        | EvictionError::Downloaded
+                        | EvictionError::Timeout,
+                    ))) => {
                         evictions_failed.file_sizes += file_size;
                         evictions_failed.count += 1;
                     }
@@ -495,7 +512,10 @@ pub(crate) async fn disk_usage_eviction_task_iteration_impl<U: Usage>(
 
             // calling again when consumed_all is fine as evicted is fused.
             let Some((_partition, candidate)) = evicted.next() else {
-                consumed_all = true;
+                if !consumed_all {
+                    tracing::info!("all evictions started, waiting");
+                    consumed_all = true;
+                }
                 continue;
             };
 
@@ -503,11 +523,15 @@ pub(crate) async fn disk_usage_eviction_task_iteration_impl<U: Usage>(
                 EvictionLayer::Attached(layer) => {
                     let file_size = layer.layer_desc().file_size;
                     js.spawn(async move {
-                        layer
-                            .evict_and_wait()
-                            .await
-                            .map(|()| file_size)
-                            .map_err(|e| (file_size, e))
+                        // have a low eviction waiting timeout because our LRU calculations go stale fast;
+                        // also individual layer evictions could hang because of bugs and we do not want to
+                        // pause disk_usage_based_eviction for such.
+                        let timeout = std::time::Duration::from_secs(5);
+
+                        match layer.evict_and_wait(timeout).await {
+                            Ok(()) => Ok(file_size),
+                            Err(e) => Err((file_size, e)),
+                        }
                     });
                 }
                 EvictionLayer::Secondary(layer) => {
@@ -529,6 +553,30 @@ pub(crate) async fn disk_usage_eviction_task_iteration_impl<U: Usage>(
         (usage_assumed, evictions_failed)
     };
 
+    let started_at = std::time::Instant::now();
+
+    let evict_layers = async move {
+        let mut evict_layers = std::pin::pin!(evict_layers);
+
+        let maximum_expected = std::time::Duration::from_secs(10);
+
+        let res = tokio::time::timeout(maximum_expected, &mut evict_layers).await;
+        let tuple = if let Ok(tuple) = res {
+            tuple
+        } else {
+            let elapsed = started_at.elapsed();
+            tracing::info!(elapsed_ms = elapsed.as_millis(), "still ongoing");
+            evict_layers.await
+        };
+
+        let elapsed = started_at.elapsed();
+        tracing::info!(elapsed_ms = elapsed.as_millis(), "completed");
+        tuple
+    };
+
+    let evict_layers =
+        evict_layers.instrument(tracing::info_span!("evict_layers", layers=%evicted_amount));
+
     let (usage_assumed, evictions_failed) = tokio::select! {
         tuple = evict_layers => { tuple },
         _ = cancel.cancelled() => {
@@ -763,6 +811,8 @@ async fn collect_eviction_candidates(
     eviction_order: EvictionOrder,
     cancel: &CancellationToken,
 ) -> anyhow::Result<EvictionCandidates> {
+    const LOG_DURATION_THRESHOLD: std::time::Duration = std::time::Duration::from_secs(10);
+
     // get a snapshot of the list of tenants
     let tenants = tenant::mgr::list_tenants()
         .await
@@ -791,6 +841,8 @@ async fn collect_eviction_candidates(
             continue;
         }
 
+        let started_at = std::time::Instant::now();
+
         // collect layers from all timelines in this tenant
         //
         // If one of the timelines becomes `!is_active()` during the iteration,
@@ -805,6 +857,7 @@ async fn collect_eviction_candidates(
             }
             let info = tl.get_local_layers_for_disk_usage_eviction().await;
             debug!(tenant_id=%tl.tenant_shard_id.tenant_id, shard_id=%tl.tenant_shard_id.shard_slug(), timeline_id=%tl.timeline_id, "timeline resident layers count: {}", info.resident_layers.len());
+
             tenant_candidates.extend(info.resident_layers.into_iter());
             max_layer_size = max_layer_size.max(info.max_layer_size.unwrap_or(0));
 
@@ -870,7 +923,25 @@ async fn collect_eviction_candidates(
                     (partition, candidate)
                 });
 
+        METRICS
+            .tenant_layer_count
+            .observe(tenant_candidates.len() as f64);
+
         candidates.extend(tenant_candidates);
+
+        let elapsed = started_at.elapsed();
+        METRICS
+            .tenant_collection_time
+            .observe(elapsed.as_secs_f64());
+
+        if elapsed > LOG_DURATION_THRESHOLD {
+            tracing::info!(
+                tenant_id=%tenant.tenant_shard_id().tenant_id,
+                shard_id=%tenant.tenant_shard_id().shard_slug(),
+                elapsed_ms = elapsed.as_millis(),
+                "collection took longer than threshold"
+            );
+        }
     }
 
     // Note: the same tenant ID might be hit twice, if it transitions from attached to
@@ -885,11 +956,11 @@ async fn collect_eviction_candidates(
         },
     );
 
-    for secondary_tenant in secondary_tenants {
+    for tenant in secondary_tenants {
         // for secondary tenants we use a sum of on_disk layers and already evicted layers. this is
         // to prevent repeated disk usage based evictions from completely draining less often
         // updating secondaries.
-        let (mut layer_info, total_layers) = secondary_tenant.get_layers_for_eviction();
+        let (mut layer_info, total_layers) = tenant.get_layers_for_eviction();
 
         debug_assert!(
             total_layers >= layer_info.resident_layers.len(),
@@ -897,6 +968,8 @@ async fn collect_eviction_candidates(
             layer_info.resident_layers.len()
         );
 
+        let started_at = std::time::Instant::now();
+
         layer_info
             .resident_layers
             .sort_unstable_by_key(|layer_info| std::cmp::Reverse(layer_info.last_activity_ts));
@@ -918,9 +991,27 @@ async fn collect_eviction_candidates(
                     )
                 });
 
+        METRICS
+            .tenant_layer_count
+            .observe(tenant_candidates.len() as f64);
         candidates.extend(tenant_candidates);
 
         tokio::task::yield_now().await;
+
+        let elapsed = started_at.elapsed();
+
+        METRICS
+            .tenant_collection_time
+            .observe(elapsed.as_secs_f64());
+
+        if elapsed > LOG_DURATION_THRESHOLD {
+            tracing::info!(
+                tenant_id=%tenant.tenant_shard_id().tenant_id,
+                shard_id=%tenant.tenant_shard_id().shard_slug(),
+                elapsed_ms = elapsed.as_millis(),
+                "collection took longer than threshold"
+            );
+        }
     }
 
     debug_assert!(MinResidentSizePartition::Above < MinResidentSizePartition::Below,
@@ -997,30 +1088,6 @@ impl<U: Usage> VictimSelection<U> {
     }
 }
 
-struct TimelineKey(Arc<Timeline>);
-
-impl PartialEq for TimelineKey {
-    fn eq(&self, other: &Self) -> bool {
-        Arc::ptr_eq(&self.0, &other.0)
-    }
-}
-
-impl Eq for TimelineKey {}
-
-impl std::hash::Hash for TimelineKey {
-    fn hash<H: std::hash::Hasher>(&self, state: &mut H) {
-        Arc::as_ptr(&self.0).hash(state);
-    }
-}
-
-impl std::ops::Deref for TimelineKey {
-    type Target = Timeline;
-
-    fn deref(&self) -> &Self::Target {
-        self.0.as_ref()
-    }
-}
-
 /// A totally ordered f32 subset we can use with sorting functions.
 pub(crate) mod finite_f32 {
 

pageserver/src/http/openapi_spec.yml

@@ -1339,6 +1339,10 @@ components:
           type: array
           items:
             $ref: "#/components/schemas/TenantShardLocation"
+        stripe_size:
+          description: If multiple shards are present, this field contains the sharding stripe size, else it is null.
+          type: integer
+          nullable: true
     TenantShardLocation:
       type: object
       required:

pageserver/src/http/routes.rs

@@ -1451,11 +1451,12 @@ async fn put_tenant_location_config_handler(
         tenant::SpawnMode::Eager
     };
 
-    let attached = state
+    let tenant = state
         .tenant_manager
         .upsert_location(tenant_shard_id, location_conf, flush, spawn_mode, &ctx)
-        .await?
-        .is_some();
+        .await?;
+    let stripe_size = tenant.as_ref().map(|t| t.get_shard_stripe_size());
+    let attached = tenant.is_some();
 
     if let Some(_flush_ms) = flush {
         match state
@@ -1477,12 +1478,20 @@ async fn put_tenant_location_config_handler(
     // This API returns a vector of pageservers where the tenant is attached: this is
     // primarily for use in the sharding service.  For compatibilty, we also return this
     // when called directly on a pageserver, but the payload is always zero or one shards.
-    let mut response = TenantLocationConfigResponse { shards: Vec::new() };
+    let mut response = TenantLocationConfigResponse {
+        shards: Vec::new(),
+        stripe_size: None,
+    };
     if attached {
         response.shards.push(TenantShardLocation {
             shard_id: tenant_shard_id,
             node_id: state.conf.id,
-        })
+        });
+        if tenant_shard_id.shard_count.count() > 1 {
+            // Stripe size should be set if we are attached
+            debug_assert!(stripe_size.is_some());
+            response.stripe_size = stripe_size;
+        }
     }
 
     json_response(StatusCode::OK, response)

pageserver/src/metrics.rs

@@ -11,6 +11,7 @@ use once_cell::sync::Lazy;
 use pageserver_api::shard::TenantShardId;
 use strum::{EnumCount, IntoEnumIterator, VariantNames};
 use strum_macros::{EnumVariantNames, IntoStaticStr};
+use tracing::warn;
 use utils::id::TimelineId;
 
 /// Prometheus histogram buckets (in seconds) for operations in the critical
@@ -1005,15 +1006,39 @@ impl GlobalAndPerTimelineHistogram {
     }
 }
 
-struct GlobalAndPerTimelineHistogramTimer<'a> {
+struct GlobalAndPerTimelineHistogramTimer<'a, 'c> {
     h: &'a GlobalAndPerTimelineHistogram,
+    ctx: &'c RequestContext,
     start: std::time::Instant,
+    op: SmgrQueryType,
 }
 
-impl<'a> Drop for GlobalAndPerTimelineHistogramTimer<'a> {
+impl<'a, 'c> Drop for GlobalAndPerTimelineHistogramTimer<'a, 'c> {
     fn drop(&mut self) {
         let elapsed = self.start.elapsed();
-        self.h.observe(elapsed.as_secs_f64());
+        let ex_throttled = self
+            .ctx
+            .micros_spent_throttled
+            .close_and_checked_sub_from(elapsed);
+        let ex_throttled = match ex_throttled {
+            Ok(res) => res,
+            Err(error) => {
+                use utils::rate_limit::RateLimit;
+                static LOGGED: Lazy<Mutex<enum_map::EnumMap<SmgrQueryType, RateLimit>>> =
+                    Lazy::new(|| {
+                        Mutex::new(enum_map::EnumMap::from_array(std::array::from_fn(|_| {
+                            RateLimit::new(Duration::from_secs(10))
+                        })))
+                    });
+                let mut guard = LOGGED.lock().unwrap();
+                let rate_limit = &mut guard[self.op];
+                rate_limit.call(|| {
+                    warn!(op=?self.op, error, "error deducting time spent throttled; this message is logged at a global rate limit");
+                });
+                elapsed
+            }
+        };
+        self.h.observe(ex_throttled.as_secs_f64());
     }
 }
 
@@ -1025,6 +1050,7 @@ impl<'a> Drop for GlobalAndPerTimelineHistogramTimer<'a> {
     strum_macros::EnumCount,
     strum_macros::EnumIter,
     strum_macros::FromRepr,
+    enum_map::Enum,
 )]
 #[strum(serialize_all = "snake_case")]
 pub enum SmgrQueryType {
@@ -1130,11 +1156,35 @@ impl SmgrQueryTimePerTimeline {
         });
         Self { metrics }
     }
-    pub(crate) fn start_timer(&self, op: SmgrQueryType) -> impl Drop + '_ {
+    pub(crate) fn start_timer<'c: 'a, 'a>(
+        &'a self,
+        op: SmgrQueryType,
+        ctx: &'c RequestContext,
+    ) -> impl Drop + '_ {
         let metric = &self.metrics[op as usize];
+        let start = Instant::now();
+        match ctx.micros_spent_throttled.open() {
+            Ok(()) => (),
+            Err(error) => {
+                use utils::rate_limit::RateLimit;
+                static LOGGED: Lazy<Mutex<enum_map::EnumMap<SmgrQueryType, RateLimit>>> =
+                    Lazy::new(|| {
+                        Mutex::new(enum_map::EnumMap::from_array(std::array::from_fn(|_| {
+                            RateLimit::new(Duration::from_secs(10))
+                        })))
+                    });
+                let mut guard = LOGGED.lock().unwrap();
+                let rate_limit = &mut guard[op];
+                rate_limit.call(|| {
+                    warn!(?op, error, "error opening micros_spent_throttled; this message is logged at a global rate limit");
+                });
+            }
+        }
         GlobalAndPerTimelineHistogramTimer {
             h: metric,
-            start: std::time::Instant::now(),
+            ctx,
+            start,
+            op,
         }
     }
 }
@@ -1145,6 +1195,11 @@ mod smgr_query_time_tests {
     use strum::IntoEnumIterator;
     use utils::id::{TenantId, TimelineId};
 
+    use crate::{
+        context::{DownloadBehavior, RequestContext},
+        task_mgr::TaskKind,
+    };
+
     // Regression test, we used hard-coded string constants before using an enum.
     #[test]
     fn op_label_name() {
@@ -1193,7 +1248,8 @@ mod smgr_query_time_tests {
             let (pre_global, pre_per_tenant_timeline) = get_counts();
             assert_eq!(pre_per_tenant_timeline, 0);
 
-            let timer = metrics.start_timer(*op);
+            let ctx = RequestContext::new(TaskKind::UnitTest, DownloadBehavior::Download);
+            let timer = metrics.start_timer(*op, &ctx);
             drop(timer);
 
             let (post_global, post_per_tenant_timeline) = get_counts();
@@ -1915,17 +1971,16 @@ impl Drop for TimelineMetrics {
         let tenant_id = &self.tenant_id;
         let timeline_id = &self.timeline_id;
         let shard_id = &self.shard_id;
-        let _ = LAST_RECORD_LSN.remove_label_values(&[tenant_id, &shard_id, timeline_id]);
+        let _ = LAST_RECORD_LSN.remove_label_values(&[tenant_id, shard_id, timeline_id]);
         {
             RESIDENT_PHYSICAL_SIZE_GLOBAL.sub(self.resident_physical_size_get());
-            let _ =
-                RESIDENT_PHYSICAL_SIZE.remove_label_values(&[tenant_id, &shard_id, timeline_id]);
+            let _ = RESIDENT_PHYSICAL_SIZE.remove_label_values(&[tenant_id, shard_id, timeline_id]);
         }
-        let _ = CURRENT_LOGICAL_SIZE.remove_label_values(&[tenant_id, &shard_id, timeline_id]);
+        let _ = CURRENT_LOGICAL_SIZE.remove_label_values(&[tenant_id, shard_id, timeline_id]);
         if let Some(metric) = Lazy::get(&DIRECTORY_ENTRIES_COUNT) {
-            let _ = metric.remove_label_values(&[tenant_id, &shard_id, timeline_id]);
+            let _ = metric.remove_label_values(&[tenant_id, shard_id, timeline_id]);
         }
-        let _ = EVICTIONS.remove_label_values(&[tenant_id, &shard_id, timeline_id]);
+        let _ = EVICTIONS.remove_label_values(&[tenant_id, shard_id, timeline_id]);
 
         self.evictions_with_low_residence_duration
             .write()
@@ -2474,6 +2529,64 @@ pub(crate) mod tenant_throttling {
     }
 }
 
+pub(crate) mod disk_usage_based_eviction {
+    use super::*;
+
+    pub(crate) struct Metrics {
+        pub(crate) tenant_collection_time: Histogram,
+        pub(crate) tenant_layer_count: Histogram,
+        pub(crate) layers_collected: IntCounter,
+        pub(crate) layers_selected: IntCounter,
+        pub(crate) layers_evicted: IntCounter,
+    }
+
+    impl Default for Metrics {
+        fn default() -> Self {
+            let tenant_collection_time = register_histogram!(
+                "pageserver_disk_usage_based_eviction_tenant_collection_seconds",
+                "Time spent collecting layers from a tenant -- not normalized by collected layer amount",
+                vec![0.001, 0.005, 0.01, 0.05, 0.1, 0.5, 1.0, 5.0, 10.0]
+            )
+            .unwrap();
+
+            let tenant_layer_count = register_histogram!(
+                "pageserver_disk_usage_based_eviction_tenant_collected_layers",
+                "Amount of layers gathered from a tenant",
+                vec![5.0, 50.0, 500.0, 5000.0, 50000.0]
+            )
+            .unwrap();
+
+            let layers_collected = register_int_counter!(
+                "pageserver_disk_usage_based_eviction_collected_layers_total",
+                "Amount of layers collected"
+            )
+            .unwrap();
+
+            let layers_selected = register_int_counter!(
+                "pageserver_disk_usage_based_eviction_select_layers_total",
+                "Amount of layers selected"
+            )
+            .unwrap();
+
+            let layers_evicted = register_int_counter!(
+                "pageserver_disk_usage_based_eviction_evicted_layers_total",
+                "Amount of layers successfully evicted"
+            )
+            .unwrap();
+
+            Self {
+                tenant_collection_time,
+                tenant_layer_count,
+                layers_collected,
+                layers_selected,
+                layers_evicted,
+            }
+        }
+    }
+
+    pub(crate) static METRICS: Lazy<Metrics> = Lazy::new(Metrics::default);
+}
+
 pub fn preinitialize_metrics() {
     // Python tests need these and on some we do alerting.
     //
@@ -2508,6 +2621,7 @@ pub fn preinitialize_metrics() {
     Lazy::force(&TENANT_MANAGER);
 
     Lazy::force(&crate::tenant::storage_layer::layer::LAYER_IMPL_METRICS);
+    Lazy::force(&disk_usage_based_eviction::METRICS);
 
     // countervecs
     [&BACKGROUND_LOOP_PERIOD_OVERRUN_COUNT]

pageserver/src/page_cache.rs

@@ -73,7 +73,6 @@
 
 use std::{
     collections::{hash_map::Entry, HashMap},
-    convert::TryInto,
     sync::{
         atomic::{AtomicU64, AtomicU8, AtomicUsize, Ordering},
         Arc, Weak,
@@ -262,7 +261,9 @@ pub struct PageCache {
     size_metrics: &'static PageCacheSizeMetrics,
 }
 
-struct PinnedSlotsPermit(tokio::sync::OwnedSemaphorePermit);
+struct PinnedSlotsPermit {
+    _permit: tokio::sync::OwnedSemaphorePermit,
+}
 
 ///
 /// PageReadGuard is a "lease" on a buffer, for reading. The page is kept locked
@@ -558,9 +559,9 @@ impl PageCache {
         )
         .await
         {
-            Ok(res) => Ok(PinnedSlotsPermit(
-                res.expect("this semaphore is never closed"),
-            )),
+            Ok(res) => Ok(PinnedSlotsPermit {
+                _permit: res.expect("this semaphore is never closed"),
+            }),
             Err(_timeout) => {
                 crate::metrics::page_cache_errors_inc(
                     crate::metrics::PageCacheErrorKind::AcquirePinnedSlotTimeout,

pageserver/src/page_service.rs

@@ -27,7 +27,7 @@ use pageserver_api::models::{
 };
 use pageserver_api::shard::ShardIndex;
 use pageserver_api::shard::ShardNumber;
-use postgres_backend::{self, is_expected_io_error, AuthType, PostgresBackend, QueryError};
+use postgres_backend::{is_expected_io_error, AuthType, PostgresBackend, QueryError};
 use pq_proto::framed::ConnectionError;
 use pq_proto::FeStartupPacket;
 use pq_proto::{BeMessage, FeMessage, RowDescriptor};
@@ -44,7 +44,6 @@ use tokio::io::AsyncWriteExt;
 use tokio::io::{AsyncRead, AsyncWrite};
 use tokio_util::io::StreamReader;
 use tokio_util::sync::CancellationToken;
-use tracing::field;
 use tracing::*;
 use utils::id::ConnectionId;
 use utils::sync::gate::GateGuard;
@@ -911,7 +910,7 @@ impl PageServerHandler {
         let timeline = self.get_timeline_shard_zero(tenant_id, timeline_id).await?;
         let _timer = timeline
             .query_metrics
-            .start_timer(metrics::SmgrQueryType::GetRelExists);
+            .start_timer(metrics::SmgrQueryType::GetRelExists, ctx);
 
         let latest_gc_cutoff_lsn = timeline.get_latest_gc_cutoff_lsn();
         let lsn =
@@ -939,7 +938,7 @@ impl PageServerHandler {
 
         let _timer = timeline
             .query_metrics
-            .start_timer(metrics::SmgrQueryType::GetRelSize);
+            .start_timer(metrics::SmgrQueryType::GetRelSize, ctx);
 
         let latest_gc_cutoff_lsn = timeline.get_latest_gc_cutoff_lsn();
         let lsn =
@@ -967,7 +966,7 @@ impl PageServerHandler {
 
         let _timer = timeline
             .query_metrics
-            .start_timer(metrics::SmgrQueryType::GetDbSize);
+            .start_timer(metrics::SmgrQueryType::GetDbSize, ctx);
 
         let latest_gc_cutoff_lsn = timeline.get_latest_gc_cutoff_lsn();
         let lsn =
@@ -1115,7 +1114,10 @@ impl PageServerHandler {
         ctx: &RequestContext,
     ) -> Result<PagestreamBeMessage, PageStreamError> {
         let timeline = match self.get_cached_timeline_for_page(req) {
-            Ok(tl) => tl,
+            Ok(tl) => {
+                set_tracing_field_shard_id(tl);
+                tl
+            }
             Err(key) => {
                 match self
                     .load_timeline_for_page(tenant_id, timeline_id, key)
@@ -1140,12 +1142,9 @@ impl PageServerHandler {
             }
         };
 
-        // load_timeline_for_page sets shard_id, but get_cached_timeline_for_page doesn't
-        set_tracing_field_shard_id(timeline);
-
         let _timer = timeline
             .query_metrics
-            .start_timer(metrics::SmgrQueryType::GetPageAtLsn);
+            .start_timer(metrics::SmgrQueryType::GetPageAtLsn, ctx);
 
         let latest_gc_cutoff_lsn = timeline.get_latest_gc_cutoff_lsn();
         let lsn =
@@ -1173,7 +1172,7 @@ impl PageServerHandler {
 
         let _timer = timeline
             .query_metrics
-            .start_timer(metrics::SmgrQueryType::GetSlruSegment);
+            .start_timer(metrics::SmgrQueryType::GetSlruSegment, ctx);
 
         let latest_gc_cutoff_lsn = timeline.get_latest_gc_cutoff_lsn();
         let lsn =

pageserver/src/pgdatadir_mapping.rs

@@ -1677,7 +1677,7 @@ struct RelDirectory {
     rels: HashSet<(Oid, u8)>,
 }
 
-#[derive(Debug, Serialize, Deserialize, Default)]
+#[derive(Debug, Serialize, Deserialize, Default, PartialEq)]
 pub(crate) struct AuxFilesDirectory {
     pub(crate) files: HashMap<String, Bytes>,
 }

pageserver/src/repository.rs

@@ -37,7 +37,6 @@ impl Value {
 mod test {
     use super::*;
 
-    use bytes::Bytes;
     use utils::bin_ser::BeSer;
 
     macro_rules! roundtrip {

pageserver/src/tenant.rs

@@ -22,6 +22,7 @@ use pageserver_api::models;
 use pageserver_api::models::TimelineState;
 use pageserver_api::models::WalRedoManagerStatus;
 use pageserver_api::shard::ShardIdentity;
+use pageserver_api::shard::ShardStripeSize;
 use pageserver_api::shard::TenantShardId;
 use remote_storage::DownloadError;
 use remote_storage::GenericRemoteStorage;
@@ -109,7 +110,6 @@ pub use pageserver_api::models::TenantState;
 use tokio::sync::Semaphore;
 
 static INIT_DB_SEMAPHORE: Lazy<Semaphore> = Lazy::new(|| Semaphore::new(8));
-use toml_edit;
 use utils::{
     crashsafe,
     generation::Generation,
@@ -152,7 +152,6 @@ pub(crate) mod ephemeral_file;
 pub mod layer_map;
 
 pub mod metadata;
-mod par_fsync;
 pub mod remote_timeline_client;
 pub mod storage_layer;
 
@@ -2088,6 +2087,10 @@ impl Tenant {
         &self.tenant_shard_id
     }
 
+    pub(crate) fn get_shard_stripe_size(&self) -> ShardStripeSize {
+        self.shard_identity.stripe_size
+    }
+
     pub(crate) fn get_generation(&self) -> Generation {
         self.generation
     }
@@ -2384,7 +2387,7 @@ impl Tenant {
             self.tenant_shard_id,
             self.generation,
             self.shard_identity,
-            self.walredo_mgr.as_ref().map(Arc::clone),
+            self.walredo_mgr.clone(),
             resources,
             pg_version,
             state,
@@ -3593,25 +3596,18 @@ pub async fn dump_layerfile_from_path(
 #[cfg(test)]
 pub(crate) mod harness {
     use bytes::{Bytes, BytesMut};
-    use camino::Utf8PathBuf;
     use once_cell::sync::OnceCell;
     use pageserver_api::models::ShardParameters;
     use pageserver_api::shard::ShardIndex;
-    use std::fs;
-    use std::sync::Arc;
     use utils::logging;
-    use utils::lsn::Lsn;
 
     use crate::deletion_queue::mock::MockDeletionQueue;
     use crate::walredo::apply_neon;
-    use crate::{
-        config::PageServerConf, repository::Key, tenant::Tenant, walrecord::NeonWalRecord,
-    };
+    use crate::{repository::Key, walrecord::NeonWalRecord};
 
     use super::*;
-    use crate::tenant::config::{TenantConf, TenantConfOpt};
     use hex_literal::hex;
-    use utils::id::{TenantId, TimelineId};
+    use utils::id::TenantId;
 
     pub const TIMELINE_ID: TimelineId =
         TimelineId::from_array(hex!("11223344556677881122334455667788"));
@@ -3840,10 +3836,8 @@ mod tests {
     use crate::DEFAULT_PG_VERSION;
     use bytes::BytesMut;
     use hex_literal::hex;
-    use once_cell::sync::Lazy;
     use pageserver_api::keyspace::KeySpace;
     use rand::{thread_rng, Rng};
-    use tokio_util::sync::CancellationToken;
 
     static TEST_KEY: Lazy<Key> =
         Lazy::new(|| Key::from_slice(&hex!("010000000033333333444444445500000001")));

pageserver/src/tenant/blob_io.rs

@@ -12,7 +12,7 @@
 //! len >= 128: 1XXXXXXX XXXXXXXX XXXXXXXX XXXXXXXX
 //!
 use bytes::{BufMut, BytesMut};
-use tokio_epoll_uring::{BoundedBuf, Slice};
+use tokio_epoll_uring::{BoundedBuf, IoBuf, Slice};
 
 use crate::context::RequestContext;
 use crate::page_cache::PAGE_SZ;
@@ -127,7 +127,7 @@ impl<const BUFFERED: bool> BlobWriter<BUFFERED> {
     /// You need to make sure that the internal buffer is empty, otherwise
     /// data will be written in wrong order.
     #[inline(always)]
-    async fn write_all_unbuffered<B: BoundedBuf>(
+    async fn write_all_unbuffered<B: BoundedBuf<Buf = Buf>, Buf: IoBuf + Send>(
         &mut self,
         src_buf: B,
     ) -> (B::Buf, Result<(), Error>) {
@@ -162,7 +162,10 @@ impl<const BUFFERED: bool> BlobWriter<BUFFERED> {
     }
 
     /// Internal, possibly buffered, write function
-    async fn write_all<B: BoundedBuf>(&mut self, src_buf: B) -> (B::Buf, Result<(), Error>) {
+    async fn write_all<B: BoundedBuf<Buf = Buf>, Buf: IoBuf + Send>(
+        &mut self,
+        src_buf: B,
+    ) -> (B::Buf, Result<(), Error>) {
         if !BUFFERED {
             assert!(self.buf.is_empty());
             return self.write_all_unbuffered(src_buf).await;
@@ -210,7 +213,10 @@ impl<const BUFFERED: bool> BlobWriter<BUFFERED> {
 
     /// Write a blob of data. Returns the offset that it was written to,
     /// which can be used to retrieve the data later.
-    pub async fn write_blob<B: BoundedBuf>(&mut self, srcbuf: B) -> (B::Buf, Result<u64, Error>) {
+    pub async fn write_blob<B: BoundedBuf<Buf = Buf>, Buf: IoBuf + Send>(
+        &mut self,
+        srcbuf: B,
+    ) -> (B::Buf, Result<u64, Error>) {
         let offset = self.offset;
 
         let len = srcbuf.bytes_init();

pageserver/src/tenant/config.rs

@@ -52,7 +52,10 @@ pub mod defaults {
     pub const DEFAULT_PITR_INTERVAL: &str = "7 days";
     pub const DEFAULT_WALRECEIVER_CONNECT_TIMEOUT: &str = "10 seconds";
     pub const DEFAULT_WALRECEIVER_LAGGING_WAL_TIMEOUT: &str = "10 seconds";
-    pub const DEFAULT_MAX_WALRECEIVER_LSN_WAL_LAG: u64 = 10 * 1024 * 1024;
+    // The default limit on WAL lag should be set to avoid causing disconnects under high throughput
+    // scenarios: since the broker stats are updated ~1/s, a value of 1GiB should be sufficient for
+    // throughputs up to 1GiB/s per timeline.
+    pub const DEFAULT_MAX_WALRECEIVER_LSN_WAL_LAG: u64 = 1024 * 1024 * 1024;
     pub const DEFAULT_EVICTIONS_LOW_RESIDENCE_DURATION_METRIC_THRESHOLD: &str = "24 hour";
 
     pub const DEFAULT_INGEST_BATCH_SIZE: u64 = 100;

pageserver/src/tenant/disk_btree.rs

@@ -18,11 +18,19 @@
 //! - An Iterator interface would be more convenient for the callers than the
 //!   'visit' function
 //!
+use async_stream::try_stream;
 use byteorder::{ReadBytesExt, BE};
 use bytes::{BufMut, Bytes, BytesMut};
 use either::Either;
+use futures::Stream;
 use hex;
-use std::{cmp::Ordering, io, result};
+use std::{
+    cmp::Ordering,
+    io,
+    iter::Rev,
+    ops::{Range, RangeInclusive},
+    result,
+};
 use thiserror::Error;
 use tracing::error;
 
@@ -251,6 +259,90 @@ where
         Ok(result)
     }
 
+    /// Return a stream which yields all key, value pairs from the index
+    /// starting from the first key greater or equal to `start_key`.
+    ///
+    /// Note that this is a copy of [`Self::visit`].
+    /// TODO: Once the sequential read path is removed this will become
+    /// the only index traversal method.
+    pub fn get_stream_from<'a>(
+        &'a self,
+        start_key: &'a [u8; L],
+        ctx: &'a RequestContext,
+    ) -> impl Stream<Item = std::result::Result<(Vec<u8>, u64), DiskBtreeError>> + 'a {
+        try_stream! {
+            let mut stack = Vec::new();
+            stack.push((self.root_blk, None));
+            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, ctx)
+                    .await?;
+
+                let node = OnDiskNode::deparse(node_buf.as_ref())?;
+                let prefix_len = node.prefix_len as usize;
+                let suffix_len = node.suffix_len as usize;
+
+                assert!(node.num_children > 0);
+
+                let mut keybuf = Vec::new();
+                keybuf.extend(node.prefix);
+                keybuf.resize(prefix_len + suffix_len, 0);
+
+                let mut iter: Either<Range<usize>, Rev<RangeInclusive<usize>>> = if let Some(iter) = opt_iter {
+                    iter
+                } else {
+                    // Locate the first match
+                    let idx = match node.binary_search(start_key, keybuf.as_mut_slice()) {
+                        Ok(idx) => idx,
+                        Err(idx) => {
+                            if node.level == 0 {
+                                // Imagine that the node contains the following keys:
+                                //
+                                // 1
+                                // 3  <-- idx
+                                // 5
+                                //
+                                // If the search key is '2' and there is exact match,
+                                // the binary search would return the index of key
+                                // '3'. That's cool, '3' is the first key to return.
+                                idx
+                            } else {
+                                // This is an internal page, so each key represents a lower
+                                // bound for what's in the child page. If there is no exact
+                                // match, we have to return the *previous* entry.
+                                //
+                                // 1  <-- return this
+                                // 3  <-- idx
+                                // 5
+                                idx.saturating_sub(1)
+                            }
+                        }
+                    };
+                    Either::Left(idx..node.num_children.into())
+                };
+
+                // idx points to the first match now. Keep going from there
+                while let Some(idx) = iter.next() {
+                    let key_off = idx * suffix_len;
+                    let suffix = &node.keys[key_off..key_off + suffix_len];
+                    keybuf[prefix_len..].copy_from_slice(suffix);
+                    let value = node.value(idx);
+                    #[allow(clippy::collapsible_if)]
+                    if node.level == 0 {
+                        // leaf
+                        yield (keybuf.clone(), value.to_u64());
+                    } else {
+                        stack.push((node_blknum, Some(iter)));
+                        stack.push((value.to_blknum(), None));
+                        break;
+                    }
+                }
+            }
+        }
+    }
+
     ///
     /// Scan the tree, starting from 'search_key', in the given direction. 'visitor'
     /// will be called for every key >= 'search_key' (or <= 'search_key', if scanning
@@ -700,8 +792,6 @@ impl<const L: usize> BuildNode<L> {
 #[cfg(test)]
 pub(crate) mod tests {
     use super::*;
-    use crate::context::DownloadBehavior;
-    use crate::task_mgr::TaskKind;
     use crate::tenant::block_io::{BlockCursor, BlockLease, BlockReaderRef};
     use rand::Rng;
     use std::collections::BTreeMap;

pageserver/src/tenant/ephemeral_file.rs

@@ -300,7 +300,7 @@ mod tests {
     use super::*;
     use crate::context::DownloadBehavior;
     use crate::task_mgr::TaskKind;
-    use crate::tenant::block_io::{BlockCursor, BlockReaderRef};
+    use crate::tenant::block_io::BlockReaderRef;
     use rand::{thread_rng, RngCore};
     use std::fs;
     use std::str::FromStr;

pageserver/src/tenant/mgr.rs

@@ -2648,7 +2648,7 @@ pub(crate) async fn immediate_gc(
 
     let tenant = guard
         .get(&tenant_shard_id)
-        .map(Arc::clone)
+        .cloned()
         .with_context(|| format!("tenant {tenant_shard_id}"))
         .map_err(|e| ApiError::NotFound(e.into()))?;
 

pageserver/src/tenant/par_fsync.rs

@@ -1,84 +0,0 @@
-use std::{
-    io,
-    sync::atomic::{AtomicUsize, Ordering},
-};
-
-use camino::{Utf8Path, Utf8PathBuf};
-
-fn fsync_path(path: &Utf8Path) -> io::Result<()> {
-    // TODO use VirtualFile::fsync_all once we fully go async.
-    let file = std::fs::File::open(path)?;
-    file.sync_all()
-}
-
-fn parallel_worker(paths: &[Utf8PathBuf], next_path_idx: &AtomicUsize) -> io::Result<()> {
-    while let Some(path) = paths.get(next_path_idx.fetch_add(1, Ordering::Relaxed)) {
-        fsync_path(path)?;
-    }
-
-    Ok(())
-}
-
-fn fsync_in_thread_pool(paths: &[Utf8PathBuf]) -> io::Result<()> {
-    // TODO: remove this function in favor of `par_fsync_async` once we asyncify everything.
-
-    /// Use at most this number of threads.
-    /// Increasing this limit will
-    /// - use more memory
-    /// - increase the cost of spawn/join latency
-    const MAX_NUM_THREADS: usize = 64;
-    let num_threads = paths.len().min(MAX_NUM_THREADS);
-    let next_path_idx = AtomicUsize::new(0);
-
-    std::thread::scope(|s| -> io::Result<()> {
-        let mut handles = vec![];
-        // Spawn `num_threads - 1`, as the current thread is also a worker.
-        for _ in 1..num_threads {
-            handles.push(s.spawn(|| parallel_worker(paths, &next_path_idx)));
-        }
-
-        parallel_worker(paths, &next_path_idx)?;
-
-        for handle in handles {
-            handle.join().unwrap()?;
-        }
-
-        Ok(())
-    })
-}
-
-/// Parallel fsync all files. Can be used in non-async context as it is using rayon thread pool.
-pub fn par_fsync(paths: &[Utf8PathBuf]) -> io::Result<()> {
-    if paths.len() == 1 {
-        fsync_path(&paths[0])?;
-        return Ok(());
-    }
-
-    fsync_in_thread_pool(paths)
-}
-
-/// Parallel fsync asynchronously.
-pub async fn par_fsync_async(paths: &[Utf8PathBuf]) -> io::Result<()> {
-    const MAX_CONCURRENT_FSYNC: usize = 64;
-    let mut next = paths.iter().peekable();
-    let mut js = tokio::task::JoinSet::new();
-    loop {
-        while js.len() < MAX_CONCURRENT_FSYNC && next.peek().is_some() {
-            let next = next.next().expect("just peeked");
-            let next = next.to_owned();
-            js.spawn_blocking(move || fsync_path(&next));
-        }
-
-        // now the joinset has been filled up, wait for next to complete
-        if let Some(res) = js.join_next().await {
-            res??;
-        } else {
-            // last item had already completed
-            assert!(
-                next.peek().is_none(),
-                "joinset emptied, we shouldn't have more work"
-            );
-            return Ok(());
-        }
-    }
-}

pageserver/src/tenant/remote_timeline_client.rs

@@ -1791,14 +1791,12 @@ mod tests {
         context::RequestContext,
         tenant::{
             harness::{TenantHarness, TIMELINE_ID},
-            storage_layer::Layer,
-            Generation, Tenant, Timeline,
+            Tenant, Timeline,
         },
         DEFAULT_PG_VERSION,
     };
 
     use std::collections::HashSet;
-    use utils::lsn::Lsn;
 
     pub(super) fn dummy_contents(name: &str) -> Vec<u8> {
         format!("contents for {name}").into()

pageserver/src/tenant/remote_timeline_client/download.rs

@@ -14,14 +14,14 @@ use tokio::io::{AsyncSeekExt, AsyncWriteExt};
 use tokio_util::io::StreamReader;
 use tokio_util::sync::CancellationToken;
 use tracing::warn;
-use utils::{backoff, crashsafe};
+use utils::backoff;
 
 use crate::config::PageServerConf;
 use crate::span::debug_assert_current_span_has_tenant_and_timeline_id;
 use crate::tenant::remote_timeline_client::{remote_layer_path, remote_timelines_path};
 use crate::tenant::storage_layer::LayerFileName;
 use crate::tenant::Generation;
-use crate::virtual_file::on_fatal_io_error;
+use crate::virtual_file::{on_fatal_io_error, MaybeFatalIo, VirtualFile};
 use crate::TEMP_FILE_SUFFIX;
 use remote_storage::{DownloadError, GenericRemoteStorage, ListingMode};
 use utils::crashsafe::path_with_suffix_extension;
@@ -50,9 +50,8 @@ pub async fn download_layer_file<'a>(
 ) -> Result<u64, DownloadError> {
     debug_assert_current_span_has_tenant_and_timeline_id();
 
-    let local_path = conf
-        .timeline_path(&tenant_shard_id, &timeline_id)
-        .join(layer_file_name.file_name());
+    let timeline_path = conf.timeline_path(&tenant_shard_id, &timeline_id);
+    let local_path = timeline_path.join(layer_file_name.file_name());
 
     let remote_path = remote_layer_path(
         &tenant_shard_id.tenant_id,
@@ -149,10 +148,21 @@ pub async fn download_layer_file<'a>(
         .with_context(|| format!("rename download layer file to {local_path}"))
         .map_err(DownloadError::Other)?;
 
-    crashsafe::fsync_async(&local_path)
-        .await
-        .with_context(|| format!("fsync layer file {local_path}"))
-        .map_err(DownloadError::Other)?;
+    // We use fatal_err() below because the after the rename above,
+    // the in-memory state of the filesystem already has the layer file in its final place,
+    // and subsequent pageserver code could think it's durable while it really isn't.
+    let work = async move {
+        let timeline_dir = VirtualFile::open(&timeline_path)
+            .await
+            .fatal_err("VirtualFile::open for timeline dir fsync");
+        timeline_dir
+            .sync_all()
+            .await
+            .fatal_err("VirtualFile::sync_all timeline dir");
+    };
+    crate::virtual_file::io_engine::get()
+        .spawn_blocking_and_block_on_if_std(work)
+        .await;
 
     tracing::debug!("download complete: {local_path}");
 
@@ -161,7 +171,7 @@ pub async fn download_layer_file<'a>(
 
 const TEMP_DOWNLOAD_EXTENSION: &str = "temp_download";
 
-pub fn is_temp_download_file(path: &Utf8Path) -> bool {
+pub(crate) fn is_temp_download_file(path: &Utf8Path) -> bool {
     let extension = path.extension();
     match extension {
         Some(TEMP_DOWNLOAD_EXTENSION) => true,

pageserver/src/tenant/secondary.rs

@@ -32,7 +32,7 @@ use remote_storage::GenericRemoteStorage;
 
 use tokio_util::sync::CancellationToken;
 use tracing::instrument;
-use utils::{completion::Barrier, fs_ext, id::TimelineId, sync::gate::Gate};
+use utils::{completion::Barrier, id::TimelineId, sync::gate::Gate};
 
 enum DownloadCommand {
     Download(TenantShardId),
@@ -121,6 +121,10 @@ impl SecondaryTenant {
         })
     }
 
+    pub(crate) fn tenant_shard_id(&self) -> TenantShardId {
+        self.tenant_shard_id
+    }
+
     pub(crate) async fn shutdown(&self) {
         self.cancel.cancel();
 
@@ -164,16 +168,17 @@ impl SecondaryTenant {
         self.detail.lock().unwrap().get_layers_for_eviction(self)
     }
 
+    /// Cancellation safe, but on cancellation the eviction will go through
     #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug(), timeline_id=%timeline_id, name=%name))]
     pub(crate) async fn evict_layer(
-        &self,
+        self: &Arc<Self>,
         conf: &PageServerConf,
         timeline_id: TimelineId,
         name: LayerFileName,
     ) {
         debug_assert_current_span_has_tenant_id();
 
-        let _guard = match self.gate.enter() {
+        let guard = match self.gate.enter() {
             Ok(g) => g,
             Err(_) => {
                 tracing::debug!("Dropping layer evictions, secondary tenant shutting down",);
@@ -187,35 +192,57 @@ impl SecondaryTenant {
             .timeline_path(&self.tenant_shard_id, &timeline_id)
             .join(name.file_name());
 
-        // We tolerate ENOENT, because between planning eviction and executing
-        // it, the secondary downloader could have seen an updated heatmap that
-        // resulted in a layer being deleted.
-        // Other local I/O errors are process-fatal: these should never happen.
-        tokio::fs::remove_file(path)
-            .await
-            .or_else(fs_ext::ignore_not_found)
-            .fatal_err("Deleting layer during eviction");
+        let this = self.clone();
 
-        // Update the timeline's state.  This does not have to be synchronized with
-        // the download process, because:
-        // - If downloader is racing with us to remove a file (e.g. because it is
-        //   removed from heatmap), then our mutual .remove() operations will both
-        //   succeed.
-        // - If downloader is racing with us to download the object (this would require
-        //   multiple eviction iterations to race with multiple download iterations), then
-        //   if we remove it from the state, the worst that happens is the downloader
-        //   downloads it again before re-inserting, or we delete the file but it remains
-        //   in the state map (in which case it will be downloaded if this secondary
-        //   tenant transitions to attached and tries to access it)
-        //
-        // The important assumption here is that the secondary timeline state does not
-        // have to 100% match what is on disk, because it's a best-effort warming
-        // of the cache.
-        let mut detail = self.detail.lock().unwrap();
-        if let Some(timeline_detail) = detail.timelines.get_mut(&timeline_id) {
-            timeline_detail.on_disk_layers.remove(&name);
-            timeline_detail.evicted_at.insert(name, now);
-        }
+        // spawn it to be cancellation safe
+        tokio::task::spawn_blocking(move || {
+            let _guard = guard;
+            // We tolerate ENOENT, because between planning eviction and executing
+            // it, the secondary downloader could have seen an updated heatmap that
+            // resulted in a layer being deleted.
+            // Other local I/O errors are process-fatal: these should never happen.
+            let deleted = std::fs::remove_file(path);
+
+            let not_found = deleted
+                .as_ref()
+                .is_err_and(|x| x.kind() == std::io::ErrorKind::NotFound);
+
+            let deleted = if not_found {
+                false
+            } else {
+                deleted
+                    .map(|()| true)
+                    .fatal_err("Deleting layer during eviction")
+            };
+
+            if !deleted {
+                // skip updating accounting and putting perhaps later timestamp
+                return;
+            }
+
+            // Update the timeline's state.  This does not have to be synchronized with
+            // the download process, because:
+            // - If downloader is racing with us to remove a file (e.g. because it is
+            //   removed from heatmap), then our mutual .remove() operations will both
+            //   succeed.
+            // - If downloader is racing with us to download the object (this would require
+            //   multiple eviction iterations to race with multiple download iterations), then
+            //   if we remove it from the state, the worst that happens is the downloader
+            //   downloads it again before re-inserting, or we delete the file but it remains
+            //   in the state map (in which case it will be downloaded if this secondary
+            //   tenant transitions to attached and tries to access it)
+            //
+            // The important assumption here is that the secondary timeline state does not
+            // have to 100% match what is on disk, because it's a best-effort warming
+            // of the cache.
+            let mut detail = this.detail.lock().unwrap();
+            if let Some(timeline_detail) = detail.timelines.get_mut(&timeline_id) {
+                timeline_detail.on_disk_layers.remove(&name);
+                timeline_detail.evicted_at.insert(name, now);
+            }
+        })
+        .await
+        .expect("secondary eviction should not have panicked");
     }
 }
 

pageserver/src/tenant/secondary/downloader.rs

@@ -16,7 +16,8 @@ use crate::{
         config::SecondaryLocationConfig,
         debug_assert_current_span_has_tenant_and_timeline_id,
         remote_timeline_client::{
-            index::LayerFileMetadata, FAILED_DOWNLOAD_WARN_THRESHOLD, FAILED_REMOTE_OP_RETRIES,
+            index::LayerFileMetadata, is_temp_download_file, FAILED_DOWNLOAD_WARN_THRESHOLD,
+            FAILED_REMOTE_OP_RETRIES,
         },
         span::debug_assert_current_span_has_tenant_id,
         storage_layer::LayerFileName,
@@ -788,7 +789,7 @@ async fn init_timeline_state(
             // Secondary mode doesn't use local metadata files, but they might have been left behind by an attached tenant.
             warn!(path=?dentry.path(), "found legacy metadata file, these should have been removed in load_tenant_config");
             continue;
-        } else if crate::is_temporary(&file_path) {
+        } else if crate::is_temporary(&file_path) || is_temp_download_file(&file_path) {
             // Temporary files are frequently left behind from restarting during downloads
             tracing::info!("Cleaning up temporary file {file_path}");
             if let Err(e) = tokio::fs::remove_file(&file_path)

pageserver/src/tenant/secondary/heatmap_uploader.rs

@@ -18,7 +18,6 @@ use crate::{
 };
 
 use futures::Future;
-use md5;
 use pageserver_api::shard::TenantShardId;
 use rand::Rng;
 use remote_storage::{GenericRemoteStorage, TimeoutOrCancel};

pageserver/src/tenant/storage_layer.rs

@@ -72,7 +72,7 @@ where
 /// the same ValueReconstructState struct in the next 'get_value_reconstruct_data'
 /// call, to collect more records.
 ///
-#[derive(Debug)]
+#[derive(Debug, Default)]
 pub struct ValueReconstructState {
     pub records: Vec<(Lsn, NeonWalRecord)>,
     pub img: Option<(Lsn, Bytes)>,

pageserver/src/tenant/storage_layer/delta_layer.rs

@@ -46,6 +46,7 @@ use crate::{DELTA_FILE_MAGIC, STORAGE_FORMAT_VERSION};
 use anyhow::{anyhow, bail, ensure, Context, Result};
 use bytes::BytesMut;
 use camino::{Utf8Path, Utf8PathBuf};
+use futures::StreamExt;
 use pageserver_api::keyspace::KeySpace;
 use pageserver_api::models::LayerAccessKind;
 use pageserver_api::shard::TenantShardId;
@@ -847,10 +848,33 @@ impl DeltaLayerInner {
         reconstruct_state: &mut ValuesReconstructState,
         ctx: &RequestContext,
     ) -> Result<(), GetVectoredError> {
-        let reads = self
-            .plan_reads(keyspace, lsn_range, reconstruct_state, ctx)
-            .await
-            .map_err(GetVectoredError::Other)?;
+        let block_reader = FileBlockReader::new(&self.file, self.file_id);
+        let index_reader = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(
+            self.index_start_blk,
+            self.index_root_blk,
+            block_reader,
+        );
+
+        let planner = VectoredReadPlanner::new(
+            self.max_vectored_read_bytes
+                .expect("Layer is loaded with max vectored bytes config")
+                .0
+                .into(),
+        );
+
+        let data_end_offset = self.index_start_blk as u64 * PAGE_SZ as u64;
+
+        let reads = Self::plan_reads(
+            keyspace,
+            lsn_range,
+            data_end_offset,
+            index_reader,
+            planner,
+            reconstruct_state,
+            ctx,
+        )
+        .await
+        .map_err(GetVectoredError::Other)?;
 
         self.do_reads_and_update_state(reads, reconstruct_state)
             .await;
@@ -858,73 +882,64 @@ impl DeltaLayerInner {
         Ok(())
     }
 
-    async fn plan_reads(
-        &self,
+    async fn plan_reads<Reader>(
         keyspace: KeySpace,
         lsn_range: Range<Lsn>,
+        data_end_offset: u64,
+        index_reader: DiskBtreeReader<Reader, DELTA_KEY_SIZE>,
+        mut planner: VectoredReadPlanner,
         reconstruct_state: &mut ValuesReconstructState,
         ctx: &RequestContext,
-    ) -> anyhow::Result<Vec<VectoredRead>> {
-        let mut planner = VectoredReadPlanner::new(
-            self.max_vectored_read_bytes
-                .expect("Layer is loaded with max vectored bytes config")
-                .0
-                .into(),
-        );
-
-        let block_reader = FileBlockReader::new(&self.file, self.file_id);
-        let tree_reader = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(
-            self.index_start_blk,
-            self.index_root_blk,
-            block_reader,
-        );
+    ) -> anyhow::Result<Vec<VectoredRead>>
+    where
+        Reader: BlockReader,
+    {
+        let ctx = RequestContextBuilder::extend(ctx)
+            .page_content_kind(PageContentKind::DeltaLayerBtreeNode)
+            .build();
 
         for range in keyspace.ranges.iter() {
             let mut range_end_handled = false;
 
             let start_key = DeltaKey::from_key_lsn(&range.start, lsn_range.start);
-            tree_reader
-                .visit(
-                    &start_key.0,
-                    VisitDirection::Forwards,
-                    |raw_key, value| {
-                        let key = Key::from_slice(&raw_key[..KEY_SIZE]);
-                        let lsn = DeltaKey::extract_lsn_from_buf(raw_key);
-                        let blob_ref = BlobRef(value);
+            let index_stream = index_reader.get_stream_from(&start_key.0, &ctx);
+            let mut index_stream = std::pin::pin!(index_stream);
 
-                        assert!(key >= range.start && lsn >= lsn_range.start);
+            while let Some(index_entry) = index_stream.next().await {
+                let (raw_key, value) = index_entry?;
+                let key = Key::from_slice(&raw_key[..KEY_SIZE]);
+                let lsn = DeltaKey::extract_lsn_from_buf(&raw_key);
+                let blob_ref = BlobRef(value);
 
-                        let cached_lsn = reconstruct_state.get_cached_lsn(&key);
-                        let flag = {
-                            if cached_lsn >= Some(lsn) {
-                                BlobFlag::Ignore
-                            } else if blob_ref.will_init() {
-                                BlobFlag::Replaces
-                            } else {
-                                BlobFlag::None
-                            }
-                        };
+                // Lsns are not monotonically increasing across keys, so we don't assert on them.
+                assert!(key >= range.start);
 
-                        if key >= range.end || (key.next() == range.end && lsn >= lsn_range.end) {
-                            planner.handle_range_end(blob_ref.pos());
-                            range_end_handled = true;
-                            false
-                        } else {
-                            planner.handle(key, lsn, blob_ref.pos(), flag);
-                            true
-                        }
-                    },
-                    &RequestContextBuilder::extend(ctx)
-                        .page_content_kind(PageContentKind::DeltaLayerBtreeNode)
-                        .build(),
-                )
-                .await
-                .map_err(|err| anyhow!(err))?;
+                let outside_lsn_range = !lsn_range.contains(&lsn);
+                let below_cached_lsn = reconstruct_state.get_cached_lsn(&key) >= Some(lsn);
+
+                let flag = {
+                    if outside_lsn_range || below_cached_lsn {
+                        BlobFlag::Ignore
+                    } else if blob_ref.will_init() {
+                        BlobFlag::ReplaceAll
+                    } else {
+                        // Usual path: add blob to the read
+                        BlobFlag::None
+                    }
+                };
+
+                if key >= range.end || (key.next() == range.end && lsn >= lsn_range.end) {
+                    planner.handle_range_end(blob_ref.pos());
+                    range_end_handled = true;
+                    break;
+                } else {
+                    planner.handle(key, lsn, blob_ref.pos(), flag);
+                }
+            }
 
             if !range_end_handled {
-                let payload_end = self.index_start_blk as u64 * PAGE_SZ as u64;
-                tracing::info!("Handling range end fallback at {}", payload_end);
-                planner.handle_range_end(payload_end);
+                tracing::info!("Handling range end fallback at {}", data_end_offset);
+                planner.handle_range_end(data_end_offset);
             }
         }
 
@@ -1190,3 +1205,131 @@ impl<'a> pageserver_compaction::interface::CompactionDeltaEntry<'a, Key> for Del
         self.size
     }
 }
+
+#[cfg(test)]
+mod test {
+    use std::collections::BTreeMap;
+
+    use super::*;
+    use crate::{
+        context::DownloadBehavior, task_mgr::TaskKind, tenant::disk_btree::tests::TestDisk,
+    };
+
+    /// Construct an index for a fictional delta layer and and then
+    /// traverse in order to plan vectored reads for a query. Finally,
+    /// verify that the traversal fed the right index key and value
+    /// pairs into the planner.
+    #[tokio::test]
+    async fn test_delta_layer_index_traversal() {
+        let base_key = Key {
+            field1: 0,
+            field2: 1663,
+            field3: 12972,
+            field4: 16396,
+            field5: 0,
+            field6: 246080,
+        };
+
+        // Populate the index with some entries
+        let entries: BTreeMap<Key, Vec<Lsn>> = BTreeMap::from([
+            (base_key, vec![Lsn(1), Lsn(5), Lsn(25), Lsn(26), Lsn(28)]),
+            (base_key.add(1), vec![Lsn(2), Lsn(5), Lsn(10), Lsn(50)]),
+            (base_key.add(2), vec![Lsn(2), Lsn(5), Lsn(10), Lsn(50)]),
+            (base_key.add(5), vec![Lsn(10), Lsn(15), Lsn(16), Lsn(20)]),
+        ]);
+
+        let mut disk = TestDisk::default();
+        let mut writer = DiskBtreeBuilder::<_, DELTA_KEY_SIZE>::new(&mut disk);
+
+        let mut disk_offset = 0;
+        for (key, lsns) in &entries {
+            for lsn in lsns {
+                let index_key = DeltaKey::from_key_lsn(key, *lsn);
+                let blob_ref = BlobRef::new(disk_offset, false);
+                writer
+                    .append(&index_key.0, blob_ref.0)
+                    .expect("In memory disk append should never fail");
+
+                disk_offset += 1;
+            }
+        }
+
+        // Prepare all the arguments for the call into `plan_reads` below
+        let (root_offset, _writer) = writer
+            .finish()
+            .expect("In memory disk finish should never fail");
+        let reader = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(0, root_offset, disk);
+        let planner = VectoredReadPlanner::new(100);
+        let mut reconstruct_state = ValuesReconstructState::new();
+        let ctx = RequestContext::new(TaskKind::UnitTest, DownloadBehavior::Error);
+
+        let keyspace = KeySpace {
+            ranges: vec![
+                base_key..base_key.add(3),
+                base_key.add(3)..base_key.add(100),
+            ],
+        };
+        let lsn_range = Lsn(2)..Lsn(40);
+
+        // Plan and validate
+        let vectored_reads = DeltaLayerInner::plan_reads(
+            keyspace.clone(),
+            lsn_range.clone(),
+            disk_offset,
+            reader,
+            planner,
+            &mut reconstruct_state,
+            &ctx,
+        )
+        .await
+        .expect("Read planning should not fail");
+
+        validate(keyspace, lsn_range, vectored_reads, entries);
+    }
+
+    fn validate(
+        keyspace: KeySpace,
+        lsn_range: Range<Lsn>,
+        vectored_reads: Vec<VectoredRead>,
+        index_entries: BTreeMap<Key, Vec<Lsn>>,
+    ) {
+        #[derive(Debug, PartialEq, Eq)]
+        struct BlobSpec {
+            key: Key,
+            lsn: Lsn,
+            at: u64,
+        }
+
+        let mut planned_blobs = Vec::new();
+        for read in vectored_reads {
+            for (at, meta) in read.blobs_at.as_slice() {
+                planned_blobs.push(BlobSpec {
+                    key: meta.key,
+                    lsn: meta.lsn,
+                    at: *at,
+                });
+            }
+        }
+
+        let mut expected_blobs = Vec::new();
+        let mut disk_offset = 0;
+        for (key, lsns) in index_entries {
+            for lsn in lsns {
+                let key_included = keyspace.ranges.iter().any(|range| range.contains(&key));
+                let lsn_included = lsn_range.contains(&lsn);
+
+                if key_included && lsn_included {
+                    expected_blobs.push(BlobSpec {
+                        key,
+                        lsn,
+                        at: disk_offset,
+                    });
+                }
+
+                disk_offset += 1;
+            }
+        }
+
+        assert_eq!(planned_blobs, expected_blobs);
+    }
+}

pageserver/src/tenant/storage_layer/image_layer.rs

@@ -55,6 +55,7 @@ use std::ops::Range;
 use std::os::unix::prelude::FileExt;
 use std::sync::Arc;
 use tokio::sync::OnceCell;
+use tokio_stream::StreamExt;
 use tracing::*;
 
 use utils::{
@@ -489,35 +490,33 @@ impl ImageLayerInner {
         let tree_reader =
             DiskBtreeReader::new(self.index_start_blk, self.index_root_blk, block_reader);
 
+        let ctx = RequestContextBuilder::extend(ctx)
+            .page_content_kind(PageContentKind::ImageLayerBtreeNode)
+            .build();
+
         for range in keyspace.ranges.iter() {
             let mut range_end_handled = false;
 
             let mut search_key: [u8; KEY_SIZE] = [0u8; KEY_SIZE];
             range.start.write_to_byte_slice(&mut search_key);
 
-            tree_reader
-                .visit(
-                    &search_key,
-                    VisitDirection::Forwards,
-                    |raw_key, offset| {
-                        let key = Key::from_slice(&raw_key[..KEY_SIZE]);
-                        assert!(key >= range.start);
+            let index_stream = tree_reader.get_stream_from(&search_key, &ctx);
+            let mut index_stream = std::pin::pin!(index_stream);
 
-                        if key >= range.end {
-                            planner.handle_range_end(offset);
-                            range_end_handled = true;
-                            false
-                        } else {
-                            planner.handle(key, self.lsn, offset, BlobFlag::None);
-                            true
-                        }
-                    },
-                    &RequestContextBuilder::extend(ctx)
-                        .page_content_kind(PageContentKind::ImageLayerBtreeNode)
-                        .build(),
-                )
-                .await
-                .map_err(|err| GetVectoredError::Other(anyhow!(err)))?;
+            while let Some(index_entry) = index_stream.next().await {
+                let (raw_key, offset) = index_entry?;
+
+                let key = Key::from_slice(&raw_key[..KEY_SIZE]);
+                assert!(key >= range.start);
+
+                if key >= range.end {
+                    planner.handle_range_end(offset);
+                    range_end_handled = true;
+                    break;
+                } else {
+                    planner.handle(key, self.lsn, offset, BlobFlag::None);
+                }
+            }
 
             if !range_end_handled {
                 let payload_end = self.index_start_blk as u64 * PAGE_SZ as u64;

pageserver/src/tenant/storage_layer/layer.rs

@@ -8,7 +8,7 @@ use pageserver_api::shard::ShardIndex;
 use std::ops::Range;
 use std::sync::atomic::{AtomicBool, AtomicUsize, Ordering};
 use std::sync::{Arc, Weak};
-use std::time::SystemTime;
+use std::time::{Duration, SystemTime};
 use tracing::Instrument;
 use utils::lsn::Lsn;
 use utils::sync::heavier_once_cell;
@@ -195,6 +195,7 @@ impl Layer {
         let downloaded = resident.expect("just initialized");
 
         // if the rename works, the path is as expected
+        // TODO: sync system call
         std::fs::rename(temp_path, owner.local_path())
             .with_context(|| format!("rename temporary file as correct path for {owner}"))?;
 
@@ -208,10 +209,15 @@ impl Layer {
     /// If for a bad luck or blocking of the executor, we miss the actual eviction and the layer is
     /// re-downloaded, [`EvictionError::Downloaded`] is returned.
     ///
+    /// Timeout is mandatory, because waiting for eviction is only needed for our tests; eviction
+    /// will happen regardless the future returned by this method completing unless there is a
+    /// read access (currently including [`Layer::keep_resident`]) before eviction gets to
+    /// complete.
+    ///
     /// Technically cancellation safe, but cancelling might shift the viewpoint of what generation
     /// of download-evict cycle on retry.
-    pub(crate) async fn evict_and_wait(&self) -> Result<(), EvictionError> {
-        self.0.evict_and_wait().await
+    pub(crate) async fn evict_and_wait(&self, timeout: Duration) -> Result<(), EvictionError> {
+        self.0.evict_and_wait(timeout).await
     }
 
     /// Delete the layer file when the `self` gets dropped, also try to schedule a remote index upload
@@ -363,7 +369,7 @@ impl Layer {
     ///
     /// Does not start local deletion, use [`Self::delete_on_drop`] for that
     /// separatedly.
-    #[cfg(feature = "testing")]
+    #[cfg(any(feature = "testing", test))]
     pub(crate) fn wait_drop(&self) -> impl std::future::Future<Output = ()> + 'static {
         let mut rx = self.0.status.subscribe();
 
@@ -632,7 +638,7 @@ impl LayerInner {
 
     /// Cancellation safe, however dropping the future and calling this method again might result
     /// in a new attempt to evict OR join the previously started attempt.
-    pub(crate) async fn evict_and_wait(&self) -> Result<(), EvictionError> {
+    pub(crate) async fn evict_and_wait(&self, timeout: Duration) -> Result<(), EvictionError> {
         use tokio::sync::broadcast::error::RecvError;
 
         assert!(self.have_remote_client);
@@ -652,16 +658,22 @@ impl LayerInner {
         if strong.is_some() {
             // drop the DownloadedLayer outside of the holding the guard
             drop(strong);
+
+            // idea here is that only one evicter should ever get to witness a strong reference,
+            // which means whenever get_or_maybe_download upgrades a weak, it must mark up a
+            // cancelled eviction and signal us, like it currently does.
+            //
+            // a second concurrent evict_and_wait will not see a strong reference.
             LAYER_IMPL_METRICS.inc_started_evictions();
         }
 
-        match rx.recv().await {
-            Ok(Status::Evicted) => Ok(()),
-            Ok(Status::Downloaded) => Err(EvictionError::Downloaded),
-            Err(RecvError::Closed) => {
+        match tokio::time::timeout(timeout, rx.recv()).await {
+            Ok(Ok(Status::Evicted)) => Ok(()),
+            Ok(Ok(Status::Downloaded)) => Err(EvictionError::Downloaded),
+            Ok(Err(RecvError::Closed)) => {
                 unreachable!("sender cannot be dropped while we are in &self method")
             }
-            Err(RecvError::Lagged(_)) => {
+            Ok(Err(RecvError::Lagged(_))) => {
                 // this is quite unlikely, but we are blocking a lot in the async context, so
                 // we might be missing this because we are stuck on a LIFO slot on a thread
                 // which is busy blocking for a 1TB database create_image_layers.
@@ -674,6 +686,7 @@ impl LayerInner {
                     None => Ok(()),
                 }
             }
+            Err(_timeout) => Err(EvictionError::Timeout),
         }
     }
 
@@ -963,7 +976,7 @@ impl LayerInner {
                 }
 
                 self.consecutive_failures.store(0, Ordering::Relaxed);
-                tracing::info!("on-demand download successful");
+                tracing::info!(size=%self.desc.file_size, "on-demand download successful");
 
                 Ok(permit)
             }
@@ -1195,6 +1208,9 @@ pub(crate) enum EvictionError {
     /// Evictions must always lose to downloads in races, and this time it happened.
     #[error("layer was downloaded instead")]
     Downloaded,
+
+    #[error("eviction did not happen within timeout")]
+    Timeout,
 }
 
 /// Error internal to the [`LayerInner::get_or_maybe_download`]

pageserver/src/tenant/storage_layer/layer/tests.rs

@@ -1,13 +1,173 @@
 use futures::StreamExt;
+use pageserver_api::key::CONTROLFILE_KEY;
 use tokio::task::JoinSet;
+use tracing::Instrument;
 use utils::{
     completion::{self, Completion},
     id::TimelineId,
 };
 
 use super::*;
-use crate::task_mgr::BACKGROUND_RUNTIME;
-use crate::tenant::harness::TenantHarness;
+use crate::{context::DownloadBehavior, task_mgr::BACKGROUND_RUNTIME};
+use crate::{task_mgr::TaskKind, tenant::harness::TenantHarness};
+
+/// Used in tests to advance a future to wanted await point, and not futher.
+const ADVANCE: std::time::Duration = std::time::Duration::from_secs(3600);
+
+/// Used in tests to indicate forever long timeout; has to be longer than the amount of ADVANCE
+/// timeout uses to advance futures.
+const FOREVER: std::time::Duration = std::time::Duration::from_secs(ADVANCE.as_secs() * 24 * 7);
+
+/// Demonstrate the API and resident -> evicted -> resident -> deleted transitions.
+#[tokio::test]
+async fn smoke_test() {
+    let handle = BACKGROUND_RUNTIME.handle();
+
+    let h = TenantHarness::create("smoke_test").unwrap();
+    let span = h.span();
+    let download_span = span.in_scope(|| tracing::info_span!("downloading", timeline_id = 1));
+    let (tenant, _) = h.load().await;
+
+    let ctx = RequestContext::new(TaskKind::UnitTest, DownloadBehavior::Download);
+
+    let timeline = tenant
+        .create_test_timeline(TimelineId::generate(), Lsn(0x10), 14, &ctx)
+        .await
+        .unwrap();
+
+    let layer = {
+        let mut layers = {
+            let layers = timeline.layers.read().await;
+            layers.resident_layers().collect::<Vec<_>>().await
+        };
+
+        assert_eq!(layers.len(), 1);
+
+        layers.swap_remove(0)
+    };
+
+    // all layers created at pageserver are like `layer`, initialized with strong
+    // Arc<DownloadedLayer>.
+
+    let img_before = {
+        let mut data = ValueReconstructState::default();
+        layer
+            .get_value_reconstruct_data(CONTROLFILE_KEY, Lsn(0x10)..Lsn(0x11), &mut data, &ctx)
+            .await
+            .unwrap();
+        data.img
+            .take()
+            .expect("tenant harness writes the control file")
+    };
+
+    // important part is evicting the layer, which can be done when there are no more ResidentLayer
+    // instances -- there currently are none, only two `Layer` values, one in the layermap and on
+    // in scope.
+    layer.evict_and_wait(FOREVER).await.unwrap();
+
+    // double-evict returns an error, which is valid if both eviction_task and disk usage based
+    // eviction would both evict the same layer at the same time.
+
+    let e = layer.evict_and_wait(FOREVER).await.unwrap_err();
+    assert!(matches!(e, EvictionError::NotFound));
+
+    // on accesses when the layer is evicted, it will automatically be downloaded.
+    let img_after = {
+        let mut data = ValueReconstructState::default();
+        layer
+            .get_value_reconstruct_data(CONTROLFILE_KEY, Lsn(0x10)..Lsn(0x11), &mut data, &ctx)
+            .instrument(download_span.clone())
+            .await
+            .unwrap();
+        data.img.take().unwrap()
+    };
+
+    assert_eq!(img_before, img_after);
+
+    // evict_and_wait can timeout, but it doesn't cancel the evicting itself
+    //
+    // ZERO for timeout does not work reliably, so first take up all spawn_blocking slots to
+    // artificially slow it down.
+    let helper = SpawnBlockingPoolHelper::consume_all_spawn_blocking_threads(handle).await;
+
+    match layer
+        .evict_and_wait(std::time::Duration::ZERO)
+        .await
+        .unwrap_err()
+    {
+        EvictionError::Timeout => {
+            // expected, but note that the eviction is "still ongoing"
+            helper.release().await;
+            // exhaust spawn_blocking pool to ensure it is now complete
+            SpawnBlockingPoolHelper::consume_and_release_all_of_spawn_blocking_threads(handle)
+                .await;
+        }
+        other => unreachable!("{other:?}"),
+    }
+
+    // only way to query if a layer is resident is to acquire a ResidentLayer instance.
+    // Layer::keep_resident never downloads, but it might initialize if the layer file is found
+    // downloaded locally.
+    let none = layer.keep_resident().await.unwrap();
+    assert!(
+        none.is_none(),
+        "Expected none, because eviction removed the local file, found: {none:?}"
+    );
+
+    // plain downloading is rarely needed
+    layer
+        .download_and_keep_resident()
+        .instrument(download_span)
+        .await
+        .unwrap();
+
+    // last important part is deletion on drop: gc and compaction use it for compacted L0 layers
+    // or fully garbage collected layers. deletion means deleting the local file, and scheduling a
+    // deletion of the already unlinked from index_part.json remote file.
+    //
+    // marking a layer to be deleted on drop is irreversible; there is no technical reason against
+    // reversiblity, but currently it is not needed so it is not provided.
+    layer.delete_on_drop();
+
+    let path = layer.local_path().to_owned();
+
+    // wait_drop produces an unconnected to Layer future which will resolve when the
+    // LayerInner::drop has completed.
+    let mut wait_drop = std::pin::pin!(layer.wait_drop());
+
+    // paused time doesn't really work well with timeouts and evict_and_wait, so delay pausing
+    // until here
+    tokio::time::pause();
+    tokio::time::timeout(ADVANCE, &mut wait_drop)
+        .await
+        .expect_err("should had timed out because two strong references exist");
+
+    tokio::fs::metadata(&path)
+        .await
+        .expect("the local layer file still exists");
+
+    let rtc = timeline.remote_client.as_ref().unwrap();
+
+    {
+        let layers = &[layer];
+        let mut g = timeline.layers.write().await;
+        g.finish_gc_timeline(layers);
+        // this just updates the remote_physical_size for demonstration purposes
+        rtc.schedule_gc_update(layers).unwrap();
+    }
+
+    // when strong references are dropped, the file is deleted and remote deletion is scheduled
+    wait_drop.await;
+
+    let e = tokio::fs::metadata(&path)
+        .await
+        .expect_err("the local file is deleted");
+    assert_eq!(e.kind(), std::io::ErrorKind::NotFound);
+
+    rtc.wait_completion().await.unwrap();
+
+    assert_eq!(rtc.get_remote_physical_size(), 0);
+}
 
 /// This test demonstrates a previous hang when a eviction and deletion were requested at the same
 /// time. Now both of them complete per Arc drop semantics.
@@ -41,10 +201,10 @@ async fn evict_and_wait_on_wanted_deleted() {
     let resident = layer.keep_resident().await.unwrap();
 
     {
-        let mut evict_and_wait = std::pin::pin!(layer.evict_and_wait());
+        let mut evict_and_wait = std::pin::pin!(layer.evict_and_wait(FOREVER));
 
         // drive the future to await on the status channel
-        tokio::time::timeout(std::time::Duration::from_secs(3600), &mut evict_and_wait)
+        tokio::time::timeout(ADVANCE, &mut evict_and_wait)
             .await
             .expect_err("should had been a timeout since we are holding the layer resident");
 
@@ -115,10 +275,10 @@ async fn residency_check_while_evict_and_wait_on_clogged_spawn_blocking() {
 
     let resident = layer.keep_resident().await.unwrap();
 
-    let mut evict_and_wait = std::pin::pin!(layer.evict_and_wait());
+    let mut evict_and_wait = std::pin::pin!(layer.evict_and_wait(FOREVER));
 
     // drive the future to await on the status channel
-    tokio::time::timeout(std::time::Duration::from_secs(3600), &mut evict_and_wait)
+    tokio::time::timeout(ADVANCE, &mut evict_and_wait)
         .await
         .expect_err("should had been a timeout since we are holding the layer resident");
     assert_eq!(1, LAYER_IMPL_METRICS.started_evictions.get());
@@ -138,7 +298,7 @@ async fn residency_check_while_evict_and_wait_on_clogged_spawn_blocking() {
 
     // because the keep_resident check alters wanted evicted without sending a message, we will
     // never get completed
-    let e = tokio::time::timeout(std::time::Duration::from_secs(3600), &mut evict_and_wait)
+    let e = tokio::time::timeout(ADVANCE, &mut evict_and_wait)
         .await
         .expect("no timeout, because keep_resident re-initialized")
         .expect_err("eviction should not have succeeded because re-initialized");
@@ -158,9 +318,10 @@ async fn residency_check_while_evict_and_wait_on_clogged_spawn_blocking() {
             .sum::<u64>()
     );
 
-    let mut second_eviction = std::pin::pin!(layer.evict_and_wait());
+    let mut second_eviction = std::pin::pin!(layer.evict_and_wait(FOREVER));
 
-    tokio::time::timeout(std::time::Duration::from_secs(3600), &mut second_eviction)
+    // advance to the wait on the queue
+    tokio::time::timeout(ADVANCE, &mut second_eviction)
         .await
         .expect_err("timeout because spawn_blocking is clogged");
 
@@ -171,7 +332,12 @@ async fn residency_check_while_evict_and_wait_on_clogged_spawn_blocking() {
 
     helper.release().await;
 
-    tokio::time::timeout(std::time::Duration::from_secs(3600), &mut second_eviction)
+    // the second_eviction gets to run here
+    //
+    // synchronize to be *strictly* after the second_eviction spawn_blocking run
+    SpawnBlockingPoolHelper::consume_and_release_all_of_spawn_blocking_threads(handle).await;
+
+    tokio::time::timeout(ADVANCE, &mut second_eviction)
         .await
         .expect("eviction goes through now that spawn_blocking is unclogged")
         .expect("eviction should succeed, because version matches");
@@ -261,3 +427,49 @@ impl SpawnBlockingPoolHelper {
             .await
     }
 }
+
+#[test]
+fn spawn_blocking_pool_helper_actually_works() {
+    // create a custom runtime for which we know and control how many blocking threads it has
+    //
+    // because the amount is not configurable for our helper, expect the same amount as
+    // BACKGROUND_RUNTIME using the tokio defaults would have.
+    let rt = tokio::runtime::Builder::new_current_thread()
+        .max_blocking_threads(512)
+        .enable_all()
+        .build()
+        .unwrap();
+
+    let handle = rt.handle();
+
+    rt.block_on(async move {
+        // this will not return until all threads are spun up and actually executing the code
+        // waiting on `consumed` to be `SpawnBlockingPoolHelper::release`'d.
+        let consumed = SpawnBlockingPoolHelper::consume_all_spawn_blocking_threads(handle).await;
+
+        println!("consumed");
+
+        let mut jh = std::pin::pin!(tokio::task::spawn_blocking(move || {
+            // this will not get to run before we release
+        }));
+
+        println!("spawned");
+
+        tokio::time::timeout(std::time::Duration::from_secs(1), &mut jh)
+            .await
+            .expect_err("the task should not have gotten to run yet");
+
+        println!("tried to join");
+
+        consumed.release().await;
+
+        println!("released");
+
+        tokio::time::timeout(std::time::Duration::from_secs(1), jh)
+            .await
+            .expect("no timeout")
+            .expect("no join error");
+
+        println!("joined");
+    });
+}

pageserver/src/tenant/tasks.rs

@@ -217,7 +217,7 @@ async fn compaction_loop(tenant: Arc<Tenant>, cancel: CancellationToken) {
                 }
                 let allowed_rps = tenant.timeline_get_throttle.steady_rps();
                 let delta = now - prev;
-                warn!(
+                info!(
                     n_seconds=%format_args!("{:.3}",
                     delta.as_secs_f64()),
                     count_accounted,

pageserver/src/tenant/throttle.rs

@@ -2,14 +2,14 @@ use std::{
     str::FromStr,
     sync::{
         atomic::{AtomicU64, Ordering},
-        Arc,
+        Arc, Mutex,
     },
     time::{Duration, Instant},
 };
 
 use arc_swap::ArcSwap;
 use enumset::EnumSet;
-use tracing::error;
+use tracing::{error, warn};
 
 use crate::{context::RequestContext, task_mgr::TaskKind};
 
@@ -157,6 +157,19 @@ where
                 .fetch_add(wait_time.as_micros() as u64, Ordering::Relaxed);
             let observation = Observation { wait_time };
             self.metric.observe_throttling(&observation);
+            match ctx.micros_spent_throttled.add(wait_time) {
+                Ok(res) => res,
+                Err(error) => {
+                    use once_cell::sync::Lazy;
+                    use utils::rate_limit::RateLimit;
+                    static WARN_RATE_LIMIT: Lazy<Mutex<RateLimit>> =
+                        Lazy::new(|| Mutex::new(RateLimit::new(Duration::from_secs(10))));
+                    let mut guard = WARN_RATE_LIMIT.lock().unwrap();
+                    guard.call(move || {
+                        warn!(error, "error adding time spent throttled; this message is logged at a global rate limit");
+                    });
+                }
+            }
         }
     }
 }

pageserver/src/tenant/timeline/compaction.rs

@@ -4,24 +4,32 @@
 //!
 //! The old legacy algorithm is implemented directly in `timeline.rs`.
 
+use std::collections::BinaryHeap;
 use std::ops::{Deref, Range};
 use std::sync::Arc;
 
-use super::Timeline;
+use super::layer_manager::LayerManager;
+use super::{CompactFlags, DurationRecorder, RecordedDuration, Timeline};
 
+use anyhow::{anyhow, Context};
 use async_trait::async_trait;
+use enumset::EnumSet;
 use fail::fail_point;
+use itertools::Itertools;
+use pageserver_api::shard::TenantShardId;
 use tokio_util::sync::CancellationToken;
-use tracing::{debug, trace, warn};
+use tracing::{debug, info, info_span, trace, warn, Instrument};
+use utils::id::TimelineId;
 
-use crate::context::RequestContext;
+use crate::context::{AccessStatsBehavior, RequestContext, RequestContextBuilder};
 use crate::tenant::storage_layer::{AsLayerDesc, PersistentLayerDesc};
-use crate::tenant::timeline::{is_rel_fsm_block_key, is_rel_vm_block_key};
+use crate::tenant::timeline::{drop_rlock, is_rel_fsm_block_key, is_rel_vm_block_key, Hole};
 use crate::tenant::timeline::{DeltaLayerWriter, ImageLayerWriter};
 use crate::tenant::timeline::{Layer, ResidentLayer};
 use crate::tenant::DeltaLayer;
 use crate::tenant::PageReconstructError;
-use crate::ZERO_PAGE;
+use crate::virtual_file::{MaybeFatalIo, VirtualFile};
+use crate::{page_cache, ZERO_PAGE};
 
 use crate::keyspace::KeySpace;
 use crate::repository::Key;
@@ -33,6 +41,694 @@ use pageserver_compaction::interface::*;
 
 use super::CompactionError;
 
+impl Timeline {
+    /// TODO: cancellation
+    pub(crate) async fn compact_legacy(
+        self: &Arc<Self>,
+        _cancel: &CancellationToken,
+        flags: EnumSet<CompactFlags>,
+        ctx: &RequestContext,
+    ) -> Result<(), CompactionError> {
+        // High level strategy for compaction / image creation:
+        //
+        // 1. First, calculate the desired "partitioning" of the
+        // currently in-use key space. The goal is to partition the
+        // key space into roughly fixed-size chunks, but also take into
+        // account any existing image layers, and try to align the
+        // chunk boundaries with the existing image layers to avoid
+        // too much churn. Also try to align chunk boundaries with
+        // relation boundaries.  In principle, we don't know about
+        // relation boundaries here, we just deal with key-value
+        // pairs, and the code in pgdatadir_mapping.rs knows how to
+        // map relations into key-value pairs. But in practice we know
+        // that 'field6' is the block number, and the fields 1-5
+        // identify a relation. This is just an optimization,
+        // though.
+        //
+        // 2. Once we know the partitioning, for each partition,
+        // decide if it's time to create a new image layer. The
+        // criteria is: there has been too much "churn" since the last
+        // image layer? The "churn" is fuzzy concept, it's a
+        // combination of too many delta files, or too much WAL in
+        // total in the delta file. Or perhaps: if creating an image
+        // file would allow to delete some older files.
+        //
+        // 3. After that, we compact all level0 delta files if there
+        // are too many of them.  While compacting, we also garbage
+        // collect any page versions that are no longer needed because
+        // of the new image layers we created in step 2.
+        //
+        // TODO: This high level strategy hasn't been implemented yet.
+        // Below are functions compact_level0() and create_image_layers()
+        // but they are a bit ad hoc and don't quite work like it's explained
+        // above. Rewrite it.
+
+        // Is the timeline being deleted?
+        if self.is_stopping() {
+            trace!("Dropping out of compaction on timeline shutdown");
+            return Err(CompactionError::ShuttingDown);
+        }
+
+        let target_file_size = self.get_checkpoint_distance();
+
+        // Define partitioning schema if needed
+
+        // FIXME: the match should only cover repartitioning, not the next steps
+        match self
+            .repartition(
+                self.get_last_record_lsn(),
+                self.get_compaction_target_size(),
+                flags,
+                ctx,
+            )
+            .await
+        {
+            Ok((partitioning, lsn)) => {
+                // Disables access_stats updates, so that the files we read remain candidates for eviction after we're done with them
+                let image_ctx = RequestContextBuilder::extend(ctx)
+                    .access_stats_behavior(AccessStatsBehavior::Skip)
+                    .build();
+
+                // 2. Compact
+                let timer = self.metrics.compact_time_histo.start_timer();
+                self.compact_level0(target_file_size, ctx).await?;
+                timer.stop_and_record();
+
+                // 3. Create new image layers for partitions that have been modified
+                // "enough".
+                let layers = self
+                    .create_image_layers(
+                        &partitioning,
+                        lsn,
+                        flags.contains(CompactFlags::ForceImageLayerCreation),
+                        &image_ctx,
+                    )
+                    .await
+                    .map_err(anyhow::Error::from)?;
+                if let Some(remote_client) = &self.remote_client {
+                    for layer in layers {
+                        remote_client.schedule_layer_file_upload(layer)?;
+                    }
+                }
+
+                if let Some(remote_client) = &self.remote_client {
+                    // should any new image layer been created, not uploading index_part will
+                    // result in a mismatch between remote_physical_size and layermap calculated
+                    // size, which will fail some tests, but should not be an issue otherwise.
+                    remote_client.schedule_index_upload_for_file_changes()?;
+                }
+            }
+            Err(err) => {
+                // no partitioning? This is normal, if the timeline was just created
+                // as an empty timeline. Also in unit tests, when we use the timeline
+                // as a simple key-value store, ignoring the datadir layout. Log the
+                // error but continue.
+                //
+                // Suppress error when it's due to cancellation
+                if !self.cancel.is_cancelled() {
+                    tracing::error!("could not compact, repartitioning keyspace failed: {err:?}");
+                }
+            }
+        };
+
+        Ok(())
+    }
+
+    /// Collect a bunch of Level 0 layer files, and compact and reshuffle them as
+    /// as Level 1 files.
+    async fn compact_level0(
+        self: &Arc<Self>,
+        target_file_size: u64,
+        ctx: &RequestContext,
+    ) -> Result<(), CompactionError> {
+        let CompactLevel0Phase1Result {
+            new_layers,
+            deltas_to_compact,
+        } = {
+            let phase1_span = info_span!("compact_level0_phase1");
+            let ctx = ctx.attached_child();
+            let mut stats = CompactLevel0Phase1StatsBuilder {
+                version: Some(2),
+                tenant_id: Some(self.tenant_shard_id),
+                timeline_id: Some(self.timeline_id),
+                ..Default::default()
+            };
+
+            let begin = tokio::time::Instant::now();
+            let phase1_layers_locked = Arc::clone(&self.layers).read_owned().await;
+            let now = tokio::time::Instant::now();
+            stats.read_lock_acquisition_micros =
+                DurationRecorder::Recorded(RecordedDuration(now - begin), now);
+            self.compact_level0_phase1(phase1_layers_locked, stats, target_file_size, &ctx)
+                .instrument(phase1_span)
+                .await?
+        };
+
+        if new_layers.is_empty() && deltas_to_compact.is_empty() {
+            // nothing to do
+            return Ok(());
+        }
+
+        self.finish_compact_batch(&new_layers, &Vec::new(), &deltas_to_compact)
+            .await?;
+        Ok(())
+    }
+
+    /// Level0 files first phase of compaction, explained in the [`Self::compact_legacy`] comment.
+    async fn compact_level0_phase1(
+        self: &Arc<Self>,
+        guard: tokio::sync::OwnedRwLockReadGuard<LayerManager>,
+        mut stats: CompactLevel0Phase1StatsBuilder,
+        target_file_size: u64,
+        ctx: &RequestContext,
+    ) -> Result<CompactLevel0Phase1Result, CompactionError> {
+        stats.read_lock_held_spawn_blocking_startup_micros =
+            stats.read_lock_acquisition_micros.till_now(); // set by caller
+        let layers = guard.layer_map();
+        let level0_deltas = layers.get_level0_deltas()?;
+        let mut level0_deltas = level0_deltas
+            .into_iter()
+            .map(|x| guard.get_from_desc(&x))
+            .collect_vec();
+        stats.level0_deltas_count = Some(level0_deltas.len());
+        // Only compact if enough layers have accumulated.
+        let threshold = self.get_compaction_threshold();
+        if level0_deltas.is_empty() || level0_deltas.len() < threshold {
+            debug!(
+                level0_deltas = level0_deltas.len(),
+                threshold, "too few deltas to compact"
+            );
+            return Ok(CompactLevel0Phase1Result::default());
+        }
+
+        // This failpoint is used together with `test_duplicate_layers` integration test.
+        // It returns the compaction result exactly the same layers as input to compaction.
+        // We want to ensure that this will not cause any problem when updating the layer map
+        // after the compaction is finished.
+        //
+        // Currently, there are two rare edge cases that will cause duplicated layers being
+        // inserted.
+        // 1. The compaction job is inturrupted / did not finish successfully. Assume we have file 1, 2, 3, 4, which
+        //    is compacted to 5, but the page server is shut down, next time we start page server we will get a layer
+        //    map containing 1, 2, 3, 4, and 5, whereas 5 has the same content as 4. If we trigger L0 compation at this
+        //    point again, it is likely that we will get a file 6 which has the same content and the key range as 5,
+        //    and this causes an overwrite. This is acceptable because the content is the same, and we should do a
+        //    layer replace instead of the normal remove / upload process.
+        // 2. The input workload pattern creates exactly n files that are sorted, non-overlapping and is of target file
+        //    size length. Compaction will likely create the same set of n files afterwards.
+        //
+        // This failpoint is a superset of both of the cases.
+        if cfg!(feature = "testing") {
+            let active = (|| {
+                ::fail::fail_point!("compact-level0-phase1-return-same", |_| true);
+                false
+            })();
+
+            if active {
+                let mut new_layers = Vec::with_capacity(level0_deltas.len());
+                for delta in &level0_deltas {
+                    // we are just faking these layers as being produced again for this failpoint
+                    new_layers.push(
+                        delta
+                            .download_and_keep_resident()
+                            .await
+                            .context("download layer for failpoint")?,
+                    );
+                }
+                tracing::info!("compact-level0-phase1-return-same"); // so that we can check if we hit the failpoint
+                return Ok(CompactLevel0Phase1Result {
+                    new_layers,
+                    deltas_to_compact: level0_deltas,
+                });
+            }
+        }
+
+        // Gather the files to compact in this iteration.
+        //
+        // Start with the oldest Level 0 delta file, and collect any other
+        // level 0 files that form a contiguous sequence, such that the end
+        // LSN of previous file matches the start LSN of the next file.
+        //
+        // Note that if the files don't form such a sequence, we might
+        // "compact" just a single file. That's a bit pointless, but it allows
+        // us to get rid of the level 0 file, and compact the other files on
+        // the next iteration. This could probably made smarter, but such
+        // "gaps" in the sequence of level 0 files should only happen in case
+        // of a crash, partial download from cloud storage, or something like
+        // that, so it's not a big deal in practice.
+        level0_deltas.sort_by_key(|l| l.layer_desc().lsn_range.start);
+        let mut level0_deltas_iter = level0_deltas.iter();
+
+        let first_level0_delta = level0_deltas_iter.next().unwrap();
+        let mut prev_lsn_end = first_level0_delta.layer_desc().lsn_range.end;
+        let mut deltas_to_compact = Vec::with_capacity(level0_deltas.len());
+
+        deltas_to_compact.push(first_level0_delta.download_and_keep_resident().await?);
+        for l in level0_deltas_iter {
+            let lsn_range = &l.layer_desc().lsn_range;
+
+            if lsn_range.start != prev_lsn_end {
+                break;
+            }
+            deltas_to_compact.push(l.download_and_keep_resident().await?);
+            prev_lsn_end = lsn_range.end;
+        }
+        let lsn_range = Range {
+            start: deltas_to_compact
+                .first()
+                .unwrap()
+                .layer_desc()
+                .lsn_range
+                .start,
+            end: deltas_to_compact.last().unwrap().layer_desc().lsn_range.end,
+        };
+
+        info!(
+            "Starting Level0 compaction in LSN range {}-{} for {} layers ({} deltas in total)",
+            lsn_range.start,
+            lsn_range.end,
+            deltas_to_compact.len(),
+            level0_deltas.len()
+        );
+
+        for l in deltas_to_compact.iter() {
+            info!("compact includes {l}");
+        }
+
+        // We don't need the original list of layers anymore. Drop it so that
+        // we don't accidentally use it later in the function.
+        drop(level0_deltas);
+
+        stats.read_lock_held_prerequisites_micros = stats
+            .read_lock_held_spawn_blocking_startup_micros
+            .till_now();
+
+        // Determine N largest holes where N is number of compacted layers.
+        let max_holes = deltas_to_compact.len();
+        let last_record_lsn = self.get_last_record_lsn();
+        let min_hole_range = (target_file_size / page_cache::PAGE_SZ as u64) as i128;
+        let min_hole_coverage_size = 3; // TODO: something more flexible?
+
+        // min-heap (reserve space for one more element added before eviction)
+        let mut heap: BinaryHeap<Hole> = BinaryHeap::with_capacity(max_holes + 1);
+        let mut prev: Option<Key> = None;
+
+        let mut all_keys = Vec::new();
+
+        for l in deltas_to_compact.iter() {
+            all_keys.extend(l.load_keys(ctx).await?);
+        }
+
+        // FIXME: should spawn_blocking the rest of this function
+
+        // The current stdlib sorting implementation is designed in a way where it is
+        // particularly fast where the slice is made up of sorted sub-ranges.
+        all_keys.sort_by_key(|DeltaEntry { key, lsn, .. }| (*key, *lsn));
+
+        stats.read_lock_held_key_sort_micros = stats.read_lock_held_prerequisites_micros.till_now();
+
+        for &DeltaEntry { key: next_key, .. } in all_keys.iter() {
+            if let Some(prev_key) = prev {
+                // just first fast filter
+                if next_key.to_i128() - prev_key.to_i128() >= min_hole_range {
+                    let key_range = prev_key..next_key;
+                    // Measuring hole by just subtraction of i128 representation of key range boundaries
+                    // has not so much sense, because largest holes will corresponds field1/field2 changes.
+                    // But we are mostly interested to eliminate holes which cause generation of excessive image layers.
+                    // That is why it is better to measure size of hole as number of covering image layers.
+                    let coverage_size = layers.image_coverage(&key_range, last_record_lsn).len();
+                    if coverage_size >= min_hole_coverage_size {
+                        heap.push(Hole {
+                            key_range,
+                            coverage_size,
+                        });
+                        if heap.len() > max_holes {
+                            heap.pop(); // remove smallest hole
+                        }
+                    }
+                }
+            }
+            prev = Some(next_key.next());
+        }
+        stats.read_lock_held_compute_holes_micros = stats.read_lock_held_key_sort_micros.till_now();
+        drop_rlock(guard);
+        stats.read_lock_drop_micros = stats.read_lock_held_compute_holes_micros.till_now();
+        let mut holes = heap.into_vec();
+        holes.sort_unstable_by_key(|hole| hole.key_range.start);
+        let mut next_hole = 0; // index of next hole in holes vector
+
+        // This iterator walks through all key-value pairs from all the layers
+        // we're compacting, in key, LSN order.
+        let all_values_iter = all_keys.iter();
+
+        // This iterator walks through all keys and is needed to calculate size used by each key
+        let mut all_keys_iter = all_keys
+            .iter()
+            .map(|DeltaEntry { key, lsn, size, .. }| (*key, *lsn, *size))
+            .coalesce(|mut prev, cur| {
+                // Coalesce keys that belong to the same key pair.
+                // This ensures that compaction doesn't put them
+                // into different layer files.
+                // Still limit this by the target file size,
+                // so that we keep the size of the files in
+                // check.
+                if prev.0 == cur.0 && prev.2 < target_file_size {
+                    prev.2 += cur.2;
+                    Ok(prev)
+                } else {
+                    Err((prev, cur))
+                }
+            });
+
+        // Merge the contents of all the input delta layers into a new set
+        // of delta layers, based on the current partitioning.
+        //
+        // We split the new delta layers on the key dimension. We iterate through the key space, and for each key, check if including the next key to the current output layer we're building would cause the layer to become too large. If so, dump the current output layer and start new one.
+        // It's possible that there is a single key with so many page versions that storing all of them in a single layer file
+        // would be too large. In that case, we also split on the LSN dimension.
+        //
+        // LSN
+        //  ^
+        //  |
+        //  | +-----------+            +--+--+--+--+
+        //  | |           |            |  |  |  |  |
+        //  | +-----------+            |  |  |  |  |
+        //  | |           |            |  |  |  |  |
+        //  | +-----------+     ==>    |  |  |  |  |
+        //  | |           |            |  |  |  |  |
+        //  | +-----------+            |  |  |  |  |
+        //  | |           |            |  |  |  |  |
+        //  | +-----------+            +--+--+--+--+
+        //  |
+        //  +--------------> key
+        //
+        //
+        // If one key (X) has a lot of page versions:
+        //
+        // LSN
+        //  ^
+        //  |                                 (X)
+        //  | +-----------+            +--+--+--+--+
+        //  | |           |            |  |  |  |  |
+        //  | +-----------+            |  |  +--+  |
+        //  | |           |            |  |  |  |  |
+        //  | +-----------+     ==>    |  |  |  |  |
+        //  | |           |            |  |  +--+  |
+        //  | +-----------+            |  |  |  |  |
+        //  | |           |            |  |  |  |  |
+        //  | +-----------+            +--+--+--+--+
+        //  |
+        //  +--------------> key
+        // TODO: this actually divides the layers into fixed-size chunks, not
+        // based on the partitioning.
+        //
+        // TODO: we should also opportunistically materialize and
+        // garbage collect what we can.
+        let mut new_layers = Vec::new();
+        let mut prev_key: Option<Key> = None;
+        let mut writer: Option<DeltaLayerWriter> = None;
+        let mut key_values_total_size = 0u64;
+        let mut dup_start_lsn: Lsn = Lsn::INVALID; // start LSN of layer containing values of the single key
+        let mut dup_end_lsn: Lsn = Lsn::INVALID; // end LSN of layer containing values of the single key
+
+        for &DeltaEntry {
+            key, lsn, ref val, ..
+        } in all_values_iter
+        {
+            let value = val.load(ctx).await?;
+            let same_key = prev_key.map_or(false, |prev_key| prev_key == key);
+            // We need to check key boundaries once we reach next key or end of layer with the same key
+            if !same_key || lsn == dup_end_lsn {
+                let mut next_key_size = 0u64;
+                let is_dup_layer = dup_end_lsn.is_valid();
+                dup_start_lsn = Lsn::INVALID;
+                if !same_key {
+                    dup_end_lsn = Lsn::INVALID;
+                }
+                // Determine size occupied by this key. We stop at next key or when size becomes larger than target_file_size
+                for (next_key, next_lsn, next_size) in all_keys_iter.by_ref() {
+                    next_key_size = next_size;
+                    if key != next_key {
+                        if dup_end_lsn.is_valid() {
+                            // We are writting segment with duplicates:
+                            // place all remaining values of this key in separate segment
+                            dup_start_lsn = dup_end_lsn; // new segments starts where old stops
+                            dup_end_lsn = lsn_range.end; // there are no more values of this key till end of LSN range
+                        }
+                        break;
+                    }
+                    key_values_total_size += next_size;
+                    // Check if it is time to split segment: if total keys size is larger than target file size.
+                    // We need to avoid generation of empty segments if next_size > target_file_size.
+                    if key_values_total_size > target_file_size && lsn != next_lsn {
+                        // Split key between multiple layers: such layer can contain only single key
+                        dup_start_lsn = if dup_end_lsn.is_valid() {
+                            dup_end_lsn // new segment with duplicates starts where old one stops
+                        } else {
+                            lsn // start with the first LSN for this key
+                        };
+                        dup_end_lsn = next_lsn; // upper LSN boundary is exclusive
+                        break;
+                    }
+                }
+                // handle case when loop reaches last key: in this case dup_end is non-zero but dup_start is not set.
+                if dup_end_lsn.is_valid() && !dup_start_lsn.is_valid() {
+                    dup_start_lsn = dup_end_lsn;
+                    dup_end_lsn = lsn_range.end;
+                }
+                if writer.is_some() {
+                    let written_size = writer.as_mut().unwrap().size();
+                    let contains_hole =
+                        next_hole < holes.len() && key >= holes[next_hole].key_range.end;
+                    // check if key cause layer overflow or contains hole...
+                    if is_dup_layer
+                        || dup_end_lsn.is_valid()
+                        || written_size + key_values_total_size > target_file_size
+                        || contains_hole
+                    {
+                        // ... if so, flush previous layer and prepare to write new one
+                        new_layers.push(
+                            writer
+                                .take()
+                                .unwrap()
+                                .finish(prev_key.unwrap().next(), self)
+                                .await?,
+                        );
+                        writer = None;
+
+                        if contains_hole {
+                            // skip hole
+                            next_hole += 1;
+                        }
+                    }
+                }
+                // Remember size of key value because at next iteration we will access next item
+                key_values_total_size = next_key_size;
+            }
+            fail_point!("delta-layer-writer-fail-before-finish", |_| {
+                Err(CompactionError::Other(anyhow::anyhow!(
+                    "failpoint delta-layer-writer-fail-before-finish"
+                )))
+            });
+
+            if !self.shard_identity.is_key_disposable(&key) {
+                if writer.is_none() {
+                    // Create writer if not initiaized yet
+                    writer = Some(
+                        DeltaLayerWriter::new(
+                            self.conf,
+                            self.timeline_id,
+                            self.tenant_shard_id,
+                            key,
+                            if dup_end_lsn.is_valid() {
+                                // this is a layer containing slice of values of the same key
+                                debug!("Create new dup layer {}..{}", dup_start_lsn, dup_end_lsn);
+                                dup_start_lsn..dup_end_lsn
+                            } else {
+                                debug!("Create new layer {}..{}", lsn_range.start, lsn_range.end);
+                                lsn_range.clone()
+                            },
+                        )
+                        .await?,
+                    );
+                }
+
+                writer.as_mut().unwrap().put_value(key, lsn, value).await?;
+            } else {
+                debug!(
+                    "Dropping key {} during compaction (it belongs on shard {:?})",
+                    key,
+                    self.shard_identity.get_shard_number(&key)
+                );
+            }
+
+            if !new_layers.is_empty() {
+                fail_point!("after-timeline-compacted-first-L1");
+            }
+
+            prev_key = Some(key);
+        }
+        if let Some(writer) = writer {
+            new_layers.push(writer.finish(prev_key.unwrap().next(), self).await?);
+        }
+
+        // Sync layers
+        if !new_layers.is_empty() {
+            // Print a warning if the created layer is larger than double the target size
+            // Add two pages for potential overhead. This should in theory be already
+            // accounted for in the target calculation, but for very small targets,
+            // we still might easily hit the limit otherwise.
+            let warn_limit = target_file_size * 2 + page_cache::PAGE_SZ as u64 * 2;
+            for layer in new_layers.iter() {
+                if layer.layer_desc().file_size > warn_limit {
+                    warn!(
+                        %layer,
+                        "created delta file of size {} larger than double of target of {target_file_size}", layer.layer_desc().file_size
+                    );
+                }
+            }
+
+            // The writer.finish() above already did the fsync of the inodes.
+            // We just need to fsync the directory in which these inodes are linked,
+            // which we know to be the timeline directory.
+            //
+            // We use fatal_err() below because the after writer.finish() returns with success,
+            // the in-memory state of the filesystem already has the layer file in its final place,
+            // and subsequent pageserver code could think it's durable while it really isn't.
+            let timeline_dir = VirtualFile::open(
+                &self
+                    .conf
+                    .timeline_path(&self.tenant_shard_id, &self.timeline_id),
+            )
+            .await
+            .fatal_err("VirtualFile::open for timeline dir fsync");
+            timeline_dir
+                .sync_all()
+                .await
+                .fatal_err("VirtualFile::sync_all timeline dir");
+        }
+
+        stats.write_layer_files_micros = stats.read_lock_drop_micros.till_now();
+        stats.new_deltas_count = Some(new_layers.len());
+        stats.new_deltas_size = Some(new_layers.iter().map(|l| l.layer_desc().file_size).sum());
+
+        match TryInto::<CompactLevel0Phase1Stats>::try_into(stats)
+            .and_then(|stats| serde_json::to_string(&stats).context("serde_json::to_string"))
+        {
+            Ok(stats_json) => {
+                info!(
+                    stats_json = stats_json.as_str(),
+                    "compact_level0_phase1 stats available"
+                )
+            }
+            Err(e) => {
+                warn!("compact_level0_phase1 stats failed to serialize: {:#}", e);
+            }
+        }
+
+        Ok(CompactLevel0Phase1Result {
+            new_layers,
+            deltas_to_compact: deltas_to_compact
+                .into_iter()
+                .map(|x| x.drop_eviction_guard())
+                .collect::<Vec<_>>(),
+        })
+    }
+}
+
+#[derive(Default)]
+struct CompactLevel0Phase1Result {
+    new_layers: Vec<ResidentLayer>,
+    deltas_to_compact: Vec<Layer>,
+}
+
+#[derive(Default)]
+struct CompactLevel0Phase1StatsBuilder {
+    version: Option<u64>,
+    tenant_id: Option<TenantShardId>,
+    timeline_id: Option<TimelineId>,
+    read_lock_acquisition_micros: DurationRecorder,
+    read_lock_held_spawn_blocking_startup_micros: DurationRecorder,
+    read_lock_held_key_sort_micros: DurationRecorder,
+    read_lock_held_prerequisites_micros: DurationRecorder,
+    read_lock_held_compute_holes_micros: DurationRecorder,
+    read_lock_drop_micros: DurationRecorder,
+    write_layer_files_micros: DurationRecorder,
+    level0_deltas_count: Option<usize>,
+    new_deltas_count: Option<usize>,
+    new_deltas_size: Option<u64>,
+}
+
+#[derive(serde::Serialize)]
+struct CompactLevel0Phase1Stats {
+    version: u64,
+    tenant_id: TenantShardId,
+    timeline_id: TimelineId,
+    read_lock_acquisition_micros: RecordedDuration,
+    read_lock_held_spawn_blocking_startup_micros: RecordedDuration,
+    read_lock_held_key_sort_micros: RecordedDuration,
+    read_lock_held_prerequisites_micros: RecordedDuration,
+    read_lock_held_compute_holes_micros: RecordedDuration,
+    read_lock_drop_micros: RecordedDuration,
+    write_layer_files_micros: RecordedDuration,
+    level0_deltas_count: usize,
+    new_deltas_count: usize,
+    new_deltas_size: u64,
+}
+
+impl TryFrom<CompactLevel0Phase1StatsBuilder> for CompactLevel0Phase1Stats {
+    type Error = anyhow::Error;
+
+    fn try_from(value: CompactLevel0Phase1StatsBuilder) -> Result<Self, Self::Error> {
+        Ok(Self {
+            version: value.version.ok_or_else(|| anyhow!("version not set"))?,
+            tenant_id: value
+                .tenant_id
+                .ok_or_else(|| anyhow!("tenant_id not set"))?,
+            timeline_id: value
+                .timeline_id
+                .ok_or_else(|| anyhow!("timeline_id not set"))?,
+            read_lock_acquisition_micros: value
+                .read_lock_acquisition_micros
+                .into_recorded()
+                .ok_or_else(|| anyhow!("read_lock_acquisition_micros not set"))?,
+            read_lock_held_spawn_blocking_startup_micros: value
+                .read_lock_held_spawn_blocking_startup_micros
+                .into_recorded()
+                .ok_or_else(|| anyhow!("read_lock_held_spawn_blocking_startup_micros not set"))?,
+            read_lock_held_key_sort_micros: value
+                .read_lock_held_key_sort_micros
+                .into_recorded()
+                .ok_or_else(|| anyhow!("read_lock_held_key_sort_micros not set"))?,
+            read_lock_held_prerequisites_micros: value
+                .read_lock_held_prerequisites_micros
+                .into_recorded()
+                .ok_or_else(|| anyhow!("read_lock_held_prerequisites_micros not set"))?,
+            read_lock_held_compute_holes_micros: value
+                .read_lock_held_compute_holes_micros
+                .into_recorded()
+                .ok_or_else(|| anyhow!("read_lock_held_compute_holes_micros not set"))?,
+            read_lock_drop_micros: value
+                .read_lock_drop_micros
+                .into_recorded()
+                .ok_or_else(|| anyhow!("read_lock_drop_micros not set"))?,
+            write_layer_files_micros: value
+                .write_layer_files_micros
+                .into_recorded()
+                .ok_or_else(|| anyhow!("write_layer_files_micros not set"))?,
+            level0_deltas_count: value
+                .level0_deltas_count
+                .ok_or_else(|| anyhow!("level0_deltas_count not set"))?,
+            new_deltas_count: value
+                .new_deltas_count
+                .ok_or_else(|| anyhow!("new_deltas_count not set"))?,
+            new_deltas_size: value
+                .new_deltas_size
+                .ok_or_else(|| anyhow!("new_deltas_size not set"))?,
+        })
+    }
+}
+
 impl Timeline {
     /// Entry point for new tiered compaction algorithm.
     ///
@@ -75,14 +771,13 @@ impl Timeline {
 
         let keyspace = self.collect_keyspace(end_lsn, ctx).await?;
         let mut adaptor = TimelineAdaptor::new(self, (end_lsn, keyspace));
-        let ctx_adaptor = RequestContextAdaptor(ctx.clone());
 
         pageserver_compaction::compact_tiered::compact_tiered(
             &mut adaptor,
             end_lsn,
             target_file_size,
             fanout,
-            &ctx_adaptor,
+            ctx,
         )
         .await?;
 
@@ -135,7 +830,6 @@ struct ResidentDeltaLayer(ResidentLayer);
 #[derive(Clone)]
 struct ResidentImageLayer(ResidentLayer);
 
-#[async_trait]
 impl CompactionJobExecutor for TimelineAdaptor {
     type Key = crate::repository::Key;
 
@@ -143,13 +837,13 @@ impl CompactionJobExecutor for TimelineAdaptor {
     type DeltaLayer = ResidentDeltaLayer;
     type ImageLayer = ResidentImageLayer;
 
-    type RequestContext = RequestContextAdaptor;
+    type RequestContext = crate::context::RequestContext;
 
     async fn get_layers(
         &mut self,
         key_range: &Range<Key>,
         lsn_range: &Range<Lsn>,
-        _ctx: &RequestContextAdaptor,
+        _ctx: &RequestContext,
     ) -> anyhow::Result<Vec<OwnArc<PersistentLayerDesc>>> {
         self.flush_updates().await?;
 
@@ -170,7 +864,7 @@ impl CompactionJobExecutor for TimelineAdaptor {
         &mut self,
         key_range: &Range<Key>,
         lsn: Lsn,
-        _ctx: &RequestContextAdaptor,
+        _ctx: &RequestContext,
     ) -> anyhow::Result<Vec<Range<Key>>> {
         if lsn == self.keyspace.0 {
             Ok(pageserver_compaction::helpers::intersect_keyspace(
@@ -206,7 +900,7 @@ impl CompactionJobExecutor for TimelineAdaptor {
         &mut self,
         lsn: Lsn,
         key_range: &Range<Key>,
-        ctx: &RequestContextAdaptor,
+        ctx: &RequestContext,
     ) -> anyhow::Result<()> {
         Ok(self.create_image_impl(lsn, key_range, ctx).await?)
     }
@@ -216,7 +910,7 @@ impl CompactionJobExecutor for TimelineAdaptor {
         lsn_range: &Range<Lsn>,
         key_range: &Range<Key>,
         input_layers: &[ResidentDeltaLayer],
-        ctx: &RequestContextAdaptor,
+        ctx: &RequestContext,
     ) -> anyhow::Result<()> {
         debug!("Create new layer {}..{}", lsn_range.start, lsn_range.end);
 
@@ -287,7 +981,7 @@ impl CompactionJobExecutor for TimelineAdaptor {
     async fn delete_layer(
         &mut self,
         layer: &OwnArc<PersistentLayerDesc>,
-        _ctx: &RequestContextAdaptor,
+        _ctx: &RequestContext,
     ) -> anyhow::Result<()> {
         self.layers_to_delete.push(layer.clone().0);
         Ok(())
@@ -299,7 +993,7 @@ impl TimelineAdaptor {
         &mut self,
         lsn: Lsn,
         key_range: &Range<Key>,
-        ctx: &RequestContextAdaptor,
+        ctx: &RequestContext,
     ) -> Result<(), PageReconstructError> {
         let timer = self.timeline.metrics.create_images_time_histo.start_timer();
 
@@ -361,17 +1055,7 @@ impl TimelineAdaptor {
     }
 }
 
-pub struct RequestContextAdaptor(pub RequestContext);
-
-impl std::ops::Deref for RequestContextAdaptor {
-    type Target = RequestContext;
-
-    fn deref(&self) -> &Self::Target {
-        &self.0
-    }
-}
-
-impl CompactionRequestContext for RequestContextAdaptor {}
+impl CompactionRequestContext for crate::context::RequestContext {}
 
 #[derive(Debug, Clone)]
 pub struct OwnArc<T>(pub Arc<T>);
@@ -449,10 +1133,7 @@ impl CompactionLayer<Key> for ResidentDeltaLayer {
 impl CompactionDeltaLayer<TimelineAdaptor> for ResidentDeltaLayer {
     type DeltaEntry<'a> = DeltaEntry<'a>;
 
-    async fn load_keys<'a>(
-        &self,
-        ctx: &RequestContextAdaptor,
-    ) -> anyhow::Result<Vec<DeltaEntry<'_>>> {
+    async fn load_keys<'a>(&self, ctx: &RequestContext) -> anyhow::Result<Vec<DeltaEntry<'_>>> {
         self.0.load_keys(ctx).await
     }
 }

pageserver/src/tenant/timeline/eviction_task.rs

@@ -204,6 +204,7 @@ impl Timeline {
             evicted: usize,
             errors: usize,
             not_evictable: usize,
+            timeouts: usize,
             #[allow(dead_code)]
             skipped_for_shutdown: usize,
         }
@@ -267,7 +268,11 @@ impl Timeline {
                 let layer = guard.drop_eviction_guard();
                 if no_activity_for > p.threshold {
                     // this could cause a lot of allocations in some cases
-                    js.spawn(async move { layer.evict_and_wait().await });
+                    js.spawn(async move {
+                        layer
+                            .evict_and_wait(std::time::Duration::from_secs(5))
+                            .await
+                    });
                     stats.candidates += 1;
                 }
             }
@@ -280,6 +285,9 @@ impl Timeline {
                     Ok(Err(EvictionError::NotFound | EvictionError::Downloaded)) => {
                         stats.not_evictable += 1;
                     }
+                    Ok(Err(EvictionError::Timeout)) => {
+                        stats.timeouts += 1;
+                    }
                     Err(je) if je.is_cancelled() => unreachable!("not used"),
                     Err(je) if je.is_panic() => {
                         /* already logged */
@@ -295,7 +303,8 @@ impl Timeline {
             stats = join_all => {
                 if stats.candidates == stats.not_evictable {
                     debug!(stats=?stats, "eviction iteration complete");
-                } else if stats.errors > 0 || stats.not_evictable > 0 {
+                } else if stats.errors > 0 || stats.not_evictable > 0 || stats.timeouts > 0 {
+                    // reminder: timeouts are not eviction cancellations
                     warn!(stats=?stats, "eviction iteration complete");
                 } else {
                     info!(stats=?stats, "eviction iteration complete");

pageserver/src/tenant/vectored_blob_io.rs

@@ -128,7 +128,7 @@ impl VectoredReadBuilder {
 pub enum BlobFlag {
     None,
     Ignore,
-    Replaces,
+    ReplaceAll,
 }
 
 /// Planner for vectored blob reads.
@@ -170,7 +170,7 @@ impl VectoredReadPlanner {
     /// incorrect data to the user.
     ///
     /// The `flag` argument has two interesting values:
-    /// * [`BlobFlag::Replaces`]: The blob for this key should replace all existing blobs.
+    /// * [`BlobFlag::ReplaceAll`]: The blob for this key should replace all existing blobs.
     /// This is used for WAL records that `will_init`.
     /// * [`BlobFlag::Ignore`]: This blob should not be included in the read. This happens
     /// if the blob is cached.
@@ -204,7 +204,7 @@ impl VectoredReadPlanner {
                 let blobs_for_key = self.blobs.entry(key).or_default();
                 blobs_for_key.push((lsn, start_offset, end_offset));
             }
-            BlobFlag::Replaces => {
+            BlobFlag::ReplaceAll => {
                 let blobs_for_key = self.blobs.entry(key).or_default();
                 blobs_for_key.clear();
                 blobs_for_key.push((lsn, start_offset, end_offset));
@@ -411,10 +411,10 @@ mod tests {
         let blob_descriptions = vec![
             (first_key, lsn, 0, BlobFlag::None),    // First in read 1
             (first_key, lsn, 1024, BlobFlag::None), // Last in read 1
-            (second_key, lsn, 2 * 1024, BlobFlag::Replaces),
+            (second_key, lsn, 2 * 1024, BlobFlag::ReplaceAll),
             (second_key, lsn, 3 * 1024, BlobFlag::None),
-            (second_key, lsn, 4 * 1024, BlobFlag::Replaces), // First in read 2
-            (second_key, lsn, 5 * 1024, BlobFlag::None),     // Last in read 2
+            (second_key, lsn, 4 * 1024, BlobFlag::ReplaceAll), // First in read 2
+            (second_key, lsn, 5 * 1024, BlobFlag::None),       // Last in read 2
         ];
 
         let ranges = [&blob_descriptions[0..2], &blob_descriptions[4..]];

pageserver/src/virtual_file.rs

@@ -17,20 +17,21 @@ use crate::tenant::TENANTS_SEGMENT_NAME;
 use camino::{Utf8Path, Utf8PathBuf};
 use once_cell::sync::OnceCell;
 use pageserver_api::shard::TenantShardId;
-use std::fs::{self, File};
+use std::fs::File;
 use std::io::{Error, ErrorKind, Seek, SeekFrom};
 use tokio_epoll_uring::{BoundedBuf, IoBuf, IoBufMut, Slice};
 
 use std::os::fd::{AsRawFd, FromRawFd, IntoRawFd, OwnedFd, RawFd};
-use std::os::unix::fs::FileExt;
 use std::sync::atomic::{AtomicBool, AtomicUsize, Ordering};
 use tokio::sync::{RwLock, RwLockReadGuard, RwLockWriteGuard};
 use tokio::time::Instant;
 
 pub use pageserver_api::models::virtual_file as api;
 pub(crate) mod io_engine;
+mod metadata;
 mod open_options;
 pub(crate) use io_engine::IoEngineKind;
+pub(crate) use metadata::Metadata;
 pub(crate) use open_options::*;
 
 ///
@@ -435,13 +436,25 @@ impl VirtualFile {
 
     /// Call File::sync_all() on the underlying File.
     pub async fn sync_all(&self) -> Result<(), Error> {
-        with_file!(self, StorageIoOperation::Fsync, |file_guard| file_guard
-            .with_std_file(|std_file| std_file.sync_all()))
+        with_file!(self, StorageIoOperation::Fsync, |file_guard| {
+            let (_file_guard, res) = io_engine::get().sync_all(file_guard).await;
+            res
+        })
     }
 
-    pub async fn metadata(&self) -> Result<fs::Metadata, Error> {
-        with_file!(self, StorageIoOperation::Metadata, |file_guard| file_guard
-            .with_std_file(|std_file| std_file.metadata()))
+    /// Call File::sync_data() on the underlying File.
+    pub async fn sync_data(&self) -> Result<(), Error> {
+        with_file!(self, StorageIoOperation::Fsync, |file_guard| {
+            let (_file_guard, res) = io_engine::get().sync_data(file_guard).await;
+            res
+        })
+    }
+
+    pub async fn metadata(&self) -> Result<Metadata, Error> {
+        with_file!(self, StorageIoOperation::Metadata, |file_guard| {
+            let (_file_guard, res) = io_engine::get().metadata(file_guard).await;
+            res
+        })
     }
 
     /// Helper function internal to `VirtualFile` that looks up the underlying File,
@@ -579,7 +592,7 @@ impl VirtualFile {
     }
 
     // Copied from https://doc.rust-lang.org/1.72.0/src/std/os/unix/fs.rs.html#219-235
-    pub async fn write_all_at<B: BoundedBuf>(
+    pub async fn write_all_at<B: BoundedBuf<Buf = Buf>, Buf: IoBuf + Send>(
         &self,
         buf: B,
         mut offset: u64,
@@ -590,8 +603,9 @@ impl VirtualFile {
         }
         let mut buf = buf.slice(0..buf_len);
         while !buf.is_empty() {
-            // TODO: push `buf` further down
-            match self.write_at(&buf, offset).await {
+            let res;
+            (buf, res) = self.write_at(buf, offset).await;
+            match res {
                 Ok(0) => {
                     return (
                         Slice::into_inner(buf),
@@ -605,7 +619,7 @@ impl VirtualFile {
                     buf = buf.slice(n..);
                     offset += n as u64;
                 }
-                Err(ref e) if e.kind() == std::io::ErrorKind::Interrupted => {}
+                Err(e) if e.kind() == std::io::ErrorKind::Interrupted => {}
                 Err(e) => return (Slice::into_inner(buf), Err(e)),
             }
         }
@@ -616,15 +630,19 @@ impl VirtualFile {
     /// Returns the IoBuf that is underlying the BoundedBuf `buf`.
     /// I.e., the returned value's `bytes_init()` method returns something different than the `bytes_init()` that was passed in.
     /// It's quite brittle and easy to mis-use, so, we return the size in the Ok() variant.
-    pub async fn write_all<B: BoundedBuf>(&mut self, buf: B) -> (B::Buf, Result<usize, Error>) {
+    pub async fn write_all<B: BoundedBuf<Buf = Buf>, Buf: IoBuf + Send>(
+        &mut self,
+        buf: B,
+    ) -> (B::Buf, Result<usize, Error>) {
         let nbytes = buf.bytes_init();
         if nbytes == 0 {
             return (Slice::into_inner(buf.slice_full()), Ok(0));
         }
         let mut buf = buf.slice(0..nbytes);
         while !buf.is_empty() {
-            // TODO: push `Slice` further down
-            match self.write(&buf).await {
+            let res;
+            (buf, res) = self.write(buf).await;
+            match res {
                 Ok(0) => {
                     return (
                         Slice::into_inner(buf),
@@ -644,11 +662,18 @@ impl VirtualFile {
         (Slice::into_inner(buf), Ok(nbytes))
     }
 
-    async fn write(&mut self, buf: &[u8]) -> Result<usize, std::io::Error> {
+    async fn write<B: IoBuf + Send>(
+        &mut self,
+        buf: Slice<B>,
+    ) -> (Slice<B>, Result<usize, std::io::Error>) {
         let pos = self.pos;
-        let n = self.write_at(buf, pos).await?;
+        let (buf, res) = self.write_at(buf, pos).await;
+        let n = match res {
+            Ok(n) => n,
+            Err(e) => return (buf, Err(e)),
+        };
         self.pos += n as u64;
-        Ok(n)
+        (buf, Ok(n))
     }
 
     pub(crate) async fn read_at<B>(&self, buf: B, offset: u64) -> (B, Result<usize, Error>)
@@ -676,16 +701,30 @@ impl VirtualFile {
         })
     }
 
-    async fn write_at(&self, buf: &[u8], offset: u64) -> Result<usize, Error> {
-        let result = with_file!(self, StorageIoOperation::Write, |file_guard| {
-            file_guard.with_std_file(|std_file| std_file.write_at(buf, offset))
-        });
-        if let Ok(size) = result {
-            STORAGE_IO_SIZE
-                .with_label_values(&["write", &self.tenant_id, &self.shard_id, &self.timeline_id])
-                .add(size as i64);
-        }
-        result
+    async fn write_at<B: IoBuf + Send>(
+        &self,
+        buf: Slice<B>,
+        offset: u64,
+    ) -> (Slice<B>, Result<usize, Error>) {
+        let file_guard = match self.lock_file().await {
+            Ok(file_guard) => file_guard,
+            Err(e) => return (buf, Err(e)),
+        };
+        observe_duration!(StorageIoOperation::Write, {
+            let ((_file_guard, buf), result) =
+                io_engine::get().write_at(file_guard, offset, buf).await;
+            if let Ok(size) = result {
+                STORAGE_IO_SIZE
+                    .with_label_values(&[
+                        "write",
+                        &self.tenant_id,
+                        &self.shard_id,
+                        &self.timeline_id,
+                    ])
+                    .add(size as i64);
+            }
+            (buf, result)
+        })
     }
 }
 
@@ -1083,6 +1122,7 @@ mod tests {
     use rand::Rng;
     use std::future::Future;
     use std::io::Write;
+    use std::os::unix::fs::FileExt;
     use std::sync::Arc;
 
     enum MaybeVirtualFile {
@@ -1103,7 +1143,11 @@ mod tests {
                 MaybeVirtualFile::File(file) => file.read_exact_at(&mut buf, offset).map(|()| buf),
             }
         }
-        async fn write_all_at<B: BoundedBuf>(&self, buf: B, offset: u64) -> Result<(), Error> {
+        async fn write_all_at<B: BoundedBuf<Buf = Buf>, Buf: IoBuf + Send>(
+            &self,
+            buf: B,
+            offset: u64,
+        ) -> Result<(), Error> {
             match self {
                 MaybeVirtualFile::VirtualFile(file) => {
                     let (_buf, res) = file.write_all_at(buf, offset).await;
@@ -1124,7 +1168,10 @@ mod tests {
                 MaybeVirtualFile::File(file) => file.seek(pos),
             }
         }
-        async fn write_all<B: BoundedBuf>(&mut self, buf: B) -> Result<(), Error> {
+        async fn write_all<B: BoundedBuf<Buf = Buf>, Buf: IoBuf + Send>(
+            &mut self,
+            buf: B,
+        ) -> Result<(), Error> {
             match self {
                 MaybeVirtualFile::VirtualFile(file) => {
                     let (_buf, res) = file.write_all(buf).await;

pageserver/src/virtual_file/io_engine.rs

@@ -7,6 +7,9 @@
 //!
 //! Then use [`get`] and  [`super::OpenOptions`].
 
+use tokio_epoll_uring::{IoBuf, Slice};
+use tracing::Instrument;
+
 pub(crate) use super::api::IoEngineKind;
 #[derive(Clone, Copy)]
 #[repr(u8)]
@@ -61,7 +64,8 @@ pub(super) fn init(engine_kind: IoEngineKind) {
     set(engine_kind);
 }
 
-pub(super) fn get() -> IoEngine {
+/// Longer-term, this API should only be used by [`super::VirtualFile`].
+pub(crate) fn get() -> IoEngine {
     let cur = IoEngine::try_from(IO_ENGINE.load(Ordering::Relaxed)).unwrap();
     if cfg!(test) {
         let env_var_name = "NEON_PAGESERVER_UNIT_TEST_VIRTUAL_FILE_IOENGINE";
@@ -98,7 +102,17 @@ use std::{
     sync::atomic::{AtomicU8, Ordering},
 };
 
-use super::FileGuard;
+use super::{FileGuard, Metadata};
+
+#[cfg(target_os = "linux")]
+fn epoll_uring_error_to_std(e: tokio_epoll_uring::Error<std::io::Error>) -> std::io::Error {
+    match e {
+        tokio_epoll_uring::Error::Op(e) => e,
+        tokio_epoll_uring::Error::System(system) => {
+            std::io::Error::new(std::io::ErrorKind::Other, system)
+        }
+    }
+}
 
 impl IoEngine {
     pub(super) async fn read_at<B>(
@@ -133,16 +147,109 @@ impl IoEngine {
             IoEngine::TokioEpollUring => {
                 let system = tokio_epoll_uring::thread_local_system().await;
                 let (resources, res) = system.read(file_guard, offset, buf).await;
+                (resources, res.map_err(epoll_uring_error_to_std))
+            }
+        }
+    }
+    pub(super) async fn sync_all(&self, file_guard: FileGuard) -> (FileGuard, std::io::Result<()>) {
+        match self {
+            IoEngine::NotSet => panic!("not initialized"),
+            IoEngine::StdFs => {
+                let res = file_guard.with_std_file(|std_file| std_file.sync_all());
+                (file_guard, res)
+            }
+            #[cfg(target_os = "linux")]
+            IoEngine::TokioEpollUring => {
+                let system = tokio_epoll_uring::thread_local_system().await;
+                let (resources, res) = system.fsync(file_guard).await;
+                (resources, res.map_err(epoll_uring_error_to_std))
+            }
+        }
+    }
+    pub(super) async fn sync_data(
+        &self,
+        file_guard: FileGuard,
+    ) -> (FileGuard, std::io::Result<()>) {
+        match self {
+            IoEngine::NotSet => panic!("not initialized"),
+            IoEngine::StdFs => {
+                let res = file_guard.with_std_file(|std_file| std_file.sync_data());
+                (file_guard, res)
+            }
+            #[cfg(target_os = "linux")]
+            IoEngine::TokioEpollUring => {
+                let system = tokio_epoll_uring::thread_local_system().await;
+                let (resources, res) = system.fdatasync(file_guard).await;
+                (resources, res.map_err(epoll_uring_error_to_std))
+            }
+        }
+    }
+    pub(super) async fn metadata(
+        &self,
+        file_guard: FileGuard,
+    ) -> (FileGuard, std::io::Result<Metadata>) {
+        match self {
+            IoEngine::NotSet => panic!("not initialized"),
+            IoEngine::StdFs => {
+                let res =
+                    file_guard.with_std_file(|std_file| std_file.metadata().map(Metadata::from));
+                (file_guard, res)
+            }
+            #[cfg(target_os = "linux")]
+            IoEngine::TokioEpollUring => {
+                let system = tokio_epoll_uring::thread_local_system().await;
+                let (resources, res) = system.statx(file_guard).await;
                 (
                     resources,
-                    res.map_err(|e| match e {
-                        tokio_epoll_uring::Error::Op(e) => e,
-                        tokio_epoll_uring::Error::System(system) => {
-                            std::io::Error::new(std::io::ErrorKind::Other, system)
-                        }
-                    }),
+                    res.map_err(epoll_uring_error_to_std).map(Metadata::from),
                 )
             }
         }
     }
+    pub(super) async fn write_at<B: IoBuf + Send>(
+        &self,
+        file_guard: FileGuard,
+        offset: u64,
+        buf: Slice<B>,
+    ) -> ((FileGuard, Slice<B>), std::io::Result<usize>) {
+        match self {
+            IoEngine::NotSet => panic!("not initialized"),
+            IoEngine::StdFs => {
+                let result = file_guard.with_std_file(|std_file| std_file.write_at(&buf, offset));
+                ((file_guard, buf), result)
+            }
+            #[cfg(target_os = "linux")]
+            IoEngine::TokioEpollUring => {
+                let system = tokio_epoll_uring::thread_local_system().await;
+                let (resources, res) = system.write(file_guard, offset, buf).await;
+                (resources, res.map_err(epoll_uring_error_to_std))
+            }
+        }
+    }
+
+    /// If we switch a user of [`tokio::fs`] to use [`super::io_engine`],
+    /// they'd start blocking the executor thread if [`IoEngine::StdFs`] is configured
+    /// whereas before the switch to [`super::io_engine`], that wasn't the case.
+    /// This method helps avoid such a regression.
+    ///
+    /// Panics if the `spawn_blocking` fails, see [`tokio::task::JoinError`] for reasons why that can happen.
+    pub(crate) async fn spawn_blocking_and_block_on_if_std<Fut, R>(&self, work: Fut) -> R
+    where
+        Fut: 'static + Send + std::future::Future<Output = R>,
+        R: 'static + Send,
+    {
+        match self {
+            IoEngine::NotSet => panic!("not initialized"),
+            IoEngine::StdFs => {
+                let span = tracing::info_span!("spawn_blocking_block_on_if_std");
+                tokio::task::spawn_blocking({
+                    move || tokio::runtime::Handle::current().block_on(work.instrument(span))
+                })
+                .await
+                .expect("failed to join blocking code most likely it panicked, panicking as well")
+            }
+            #[cfg(target_os = "linux")]
+            IoEngine::TokioEpollUring => work.await,
+        }
+    }
 }

pageserver/src/virtual_file/metadata.rs

@@ -0,0 +1,30 @@
+use std::fs;
+
+pub enum Metadata {
+    StdFs(fs::Metadata),
+    #[cfg(target_os = "linux")]
+    TokioEpollUring(Box<tokio_epoll_uring::ops::statx::statx>),
+}
+
+#[cfg(target_os = "linux")]
+impl From<Box<tokio_epoll_uring::ops::statx::statx>> for Metadata {
+    fn from(value: Box<tokio_epoll_uring::ops::statx::statx>) -> Self {
+        Metadata::TokioEpollUring(value)
+    }
+}
+
+impl From<std::fs::Metadata> for Metadata {
+    fn from(value: std::fs::Metadata) -> Self {
+        Metadata::StdFs(value)
+    }
+}
+
+impl Metadata {
+    pub fn len(&self) -> u64 {
+        match self {
+            Metadata::StdFs(metadata) => metadata.len(),
+            #[cfg(target_os = "linux")]
+            Metadata::TokioEpollUring(statx) => statx.stx_size,
+        }
+    }
+}

pageserver/src/walingest.rs

@@ -1667,8 +1667,6 @@ mod tests {
     use super::*;
     use crate::tenant::harness::*;
     use crate::tenant::remote_timeline_client::{remote_initdb_archive_path, INITDB_PATH};
-    use crate::tenant::Timeline;
-    use postgres_ffi::v14::xlog_utils::SIZEOF_CHECKPOINT;
     use postgres_ffi::RELSEG_SIZE;
 
     use crate::DEFAULT_PG_VERSION;

pageserver/src/walredo.rs

@@ -262,7 +262,7 @@ impl PostgresRedoManager {
             // next request will launch a new one.
             if let Err(e) = result.as_ref() {
                 error!(
-                    "error applying {} WAL records {}..{} ({} bytes) to base image with LSN {} to reconstruct page image at LSN {} n_attempts={}: {:?}",
+                    "error applying {} WAL records {}..{} ({} bytes) to key {key}, from base image with LSN {} to reconstruct page image at LSN {} n_attempts={}: {:?}",
                     records.len(),
                     records.first().map(|p| p.0).unwrap_or(Lsn(0)),
                     records.last().map(|p| p.0).unwrap_or(Lsn(0)),