test: write out a walredo file in test

fix: time the reconstruction, not future creation (#5459 )
`pageserver_getpage_reconstruct_seconds` histogram had been only recording the time it takes to create a future, not await on it. Since: eb0a698adc.
2026-01-25 06:10:37 +00:00 · 2023-10-04 11:48:07 +00:00 · 2023-10-04 11:01:07 +01:00 · 2023-10-04 11:24:30 +03:00 · 2023-10-04 02:38:42 +02:00 · 2023-10-03 18:42:39 +01:00
56 changed files with 1947 additions and 900 deletions
--- a/.github/workflows/build_and_test.yml
+++ b/.github/workflows/build_and_test.yml
@@ -834,7 +834,7 @@ jobs:
      run:
        shell: sh -eu {0}
    env:
-      VM_BUILDER_VERSION: v0.17.11
+      VM_BUILDER_VERSION: v0.17.12

    steps:
      - name: Checkout
--- a/2
+++ b/2
@@ -5,7 +5,7 @@
 /libs/remote_storage/ @neondatabase/storage
 /libs/safekeeper_api/ @neondatabase/safekeepers
 /libs/vm_monitor/ @neondatabase/autoscaling @neondatabase/compute
-/pageserver/ @neondatabase/compute @neondatabase/storage
+/pageserver/ @neondatabase/storage
 /pgxn/ @neondatabase/compute
 /proxy/ @neondatabase/proxy
 /safekeeper/ @neondatabase/safekeepers
--- a/Cargo.lock
+++ b/Cargo.lock
@@ -1780,18 +1780,9 @@ checksum = "95505c38b4572b2d910cecb0281560f54b440a19336cbbcb27bf6ce6adc6f5a8"

 [[package]]
 name = "hermit-abi"
-version = "0.2.6"
+version = "0.3.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "ee512640fe35acbfb4bb779db6f0d80704c2cacfa2e39b601ef3e3f47d1ae4c7"
-dependencies = [
- "libc",
-]
-
-[[package]]
-name = "hermit-abi"
-version = "0.3.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "fed44880c466736ef9a5c5b5facefb5ed0785676d0c02d612db14e54f0d84286"
+checksum = "d77f7ec81a6d05a3abb01ab6eb7590f6083d08449fe5a1c8b1e620283546ccb7"

 [[package]]
 name = "hex"
@@ -2053,7 +2044,7 @@ version = "1.0.11"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "eae7b9aee968036d54dce06cebaefd919e4472e753296daccd6d344e3e2df0c2"
 dependencies = [
- "hermit-abi 0.3.1",
+ "hermit-abi",
 "libc",
 "windows-sys 0.48.0",
 ]
@@ -2070,7 +2061,7 @@ version = "0.4.7"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "adcf93614601c8129ddf72e2d5633df827ba6551541c6d8c59520a371475be1f"
 dependencies = [
- "hermit-abi 0.3.1",
+ "hermit-abi",
 "io-lifetimes",
 "rustix 0.37.19",
 "windows-sys 0.48.0",
@@ -2444,11 +2435,11 @@ dependencies = [

 [[package]]
 name = "num_cpus"
-version = "1.15.0"
+version = "1.16.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0fac9e2da13b5eb447a6ce3d392f23a29d8694bff781bf03a16cd9ac8697593b"
+checksum = "4161fcb6d602d4d2081af7c3a45852d875a03dd337a6bfdd6e06407b61342a43"
 dependencies = [
- "hermit-abi 0.2.6",
+ "hermit-abi",
 "libc",
 ]

@@ -3246,6 +3237,7 @@ dependencies = [
 "reqwest-tracing",
 "routerify",
 "rstest",
+ "rustc-hash",
 "rustls",
 "rustls-pemfile",
 "scopeguard",
--- a/Cargo.toml
+++ b/Cargo.toml
@@ -107,6 +107,7 @@ reqwest-middleware = "0.2.0"
 reqwest-retry = "0.2.2"
 routerify = "3"
 rpds = "0.13"
+rustc-hash = "1.1.0"
 rustls = "0.21"
 rustls-pemfile = "1"
 rustls-split = "0.3"
--- a/Dockerfile.compute-node
+++ b/Dockerfile.compute-node
@@ -615,11 +615,7 @@ RUN wget https://gitlab.com/dalibo/postgresql_anonymizer/-/archive/1.1.0/postgre
 #########################################################################################
 #
 # Layer "rust extensions"
-# This layer is used to build `pgx` deps
-#
-# FIXME: This needs to be updated to latest version of 'pgrx' (it was renamed from
-# 'pgx' to 'pgrx') for PostgreSQL 16. And that in turn requires bumping the pgx
-# dependency on all the rust extension that depend on it, too.
+# This layer is used to build `pgrx` deps
 #
 #########################################################################################
 FROM build-deps AS rust-extensions-build
@@ -635,22 +631,12 @@ USER nonroot
 WORKDIR /home/nonroot
 ARG PG_VERSION

-RUN case "${PG_VERSION}" in \
-      "v14" | "v15") \
-        ;; \
-      "v16") \
-        echo "TODO: Not yet supported for PostgreSQL 16. Need to update pgrx dependencies" && exit 0 \
-	;; \
-      *) \
-        echo "unexpected PostgreSQL version ${PG_VERSION}" && exit 1 \
-        ;; \
-    esac && \
-    curl -sSO https://static.rust-lang.org/rustup/dist/$(uname -m)-unknown-linux-gnu/rustup-init && \
+RUN curl -sSO https://static.rust-lang.org/rustup/dist/$(uname -m)-unknown-linux-gnu/rustup-init && \
    chmod +x rustup-init && \
    ./rustup-init -y --no-modify-path --profile minimal --default-toolchain stable && \
    rm rustup-init && \
-    cargo install --locked --version 0.7.3 cargo-pgx && \
-    /bin/bash -c 'cargo pgx init --pg${PG_VERSION:1}=/usr/local/pgsql/bin/pg_config'
+    cargo install --locked --version 0.10.2 cargo-pgrx && \
+    /bin/bash -c 'cargo pgrx init --pg${PG_VERSION:1}=/usr/local/pgsql/bin/pg_config'

 USER root

@@ -664,23 +650,11 @@ USER root
 FROM rust-extensions-build AS pg-jsonschema-pg-build
 ARG PG_VERSION

-# caeab60d70b2fd3ae421ec66466a3abbb37b7ee6 made on 06/03/2023
-# there is no release tag yet, but we need it due to the superuser fix in the control file, switch to git tag after release >= 0.1.5
-RUN case "${PG_VERSION}" in \
-      "v14" | "v15") \
-        ;; \
-      "v16") \
-        echo "TODO: Not yet supported for PostgreSQL 16. Need to update pgrx dependencies" && exit 0 \
-	;; \
-      *) \
-        echo "unexpected PostgreSQL version \"${PG_VERSION}\"" && exit 1 \
-        ;; \
-    esac && \
-    wget https://github.com/supabase/pg_jsonschema/archive/caeab60d70b2fd3ae421ec66466a3abbb37b7ee6.tar.gz -O pg_jsonschema.tar.gz && \
-    echo "54129ce2e7ee7a585648dbb4cef6d73f795d94fe72f248ac01119992518469a4 pg_jsonschema.tar.gz" | sha256sum --check && \
+RUN wget https://github.com/supabase/pg_jsonschema/archive/refs/tags/v0.2.0.tar.gz -O pg_jsonschema.tar.gz && \
+    echo "9118fc508a6e231e7a39acaa6f066fcd79af17a5db757b47d2eefbe14f7794f0 pg_jsonschema.tar.gz" | sha256sum --check && \
    mkdir pg_jsonschema-src && cd pg_jsonschema-src && tar xvzf ../pg_jsonschema.tar.gz --strip-components=1 -C . && \
-    sed -i 's/pgx = "0.7.1"/pgx = { version = "0.7.3", features = [ "unsafe-postgres" ] }/g' Cargo.toml && \
-    cargo pgx install --release && \
+    sed -i 's/pgrx = "0.10.2"/pgrx = { version = "0.10.2", features = [ "unsafe-postgres" ] }/g' Cargo.toml && \
+    cargo pgrx install --release && \
    echo "trusted = true" >> /usr/local/pgsql/share/extension/pg_jsonschema.control

 #########################################################################################
@@ -693,26 +667,11 @@ RUN case "${PG_VERSION}" in \
 FROM rust-extensions-build AS pg-graphql-pg-build
 ARG PG_VERSION

-# b4988843647450a153439be367168ed09971af85 made on 22/02/2023 (from remove-pgx-contrib-spiext branch)
-# Currently pgx version bump to >= 0.7.2  causes "call to unsafe function" compliation errors in
-# pgx-contrib-spiext. There is a branch that removes that dependency, so use it. It is on the
-# same 1.1 version we've used before.
-RUN case "${PG_VERSION}" in \
-      "v14" | "v15") \
-        ;; \
-      "v16") \
-        echo "TODO: Not yet supported for PostgreSQL 16. Need to update pgrx dependencies" && exit 0 \
-	;; \
-      *) \
-        echo "unexpected PostgreSQL version" && exit 1 \
-        ;; \
-    esac && \
-    wget https://github.com/yrashk/pg_graphql/archive/b4988843647450a153439be367168ed09971af85.tar.gz -O pg_graphql.tar.gz && \
-    echo "0c7b0e746441b2ec24187d0e03555faf935c2159e2839bddd14df6dafbc8c9bd pg_graphql.tar.gz" | sha256sum --check && \
+RUN wget https://github.com/supabase/pg_graphql/archive/refs/tags/v1.4.0.tar.gz -O pg_graphql.tar.gz && \
+    echo "bd8dc7230282b3efa9ae5baf053a54151ed0e66881c7c53750e2d0c765776edc pg_graphql.tar.gz" | sha256sum --check && \
    mkdir pg_graphql-src && cd pg_graphql-src && tar xvzf ../pg_graphql.tar.gz --strip-components=1 -C . && \
-    sed -i 's/pgx = "~0.7.1"/pgx = { version = "0.7.3", features = [ "unsafe-postgres" ] }/g' Cargo.toml && \
-    sed -i 's/pgx-tests = "~0.7.1"/pgx-tests = "0.7.3"/g' Cargo.toml && \
-    cargo pgx install --release && \
+    sed -i 's/pgrx = "=0.10.2"/pgrx = { version = "0.10.2", features = [ "unsafe-postgres" ] }/g' Cargo.toml && \
+    cargo pgrx install --release && \
    # it's needed to enable extension because it uses untrusted C language
    sed -i 's/superuser = false/superuser = true/g' /usr/local/pgsql/share/extension/pg_graphql.control && \
    echo "trusted = true" >> /usr/local/pgsql/share/extension/pg_graphql.control
@@ -727,21 +686,11 @@ RUN case "${PG_VERSION}" in \
 FROM rust-extensions-build AS pg-tiktoken-pg-build
 ARG PG_VERSION

-# 801f84f08c6881c8aa30f405fafbf00eec386a72 made on 10/03/2023
-RUN case "${PG_VERSION}" in \
-      "v14" | "v15") \
-        ;; \
-      "v16") \
-        echo "TODO: Not yet supported for PostgreSQL 16. Need to update pgrx dependencies" && exit 0 \
-	;; \
-      *) \
-        echo "unexpected PostgreSQL version" && exit 1 \
-        ;; \
-    esac && \
-    wget https://github.com/kelvich/pg_tiktoken/archive/801f84f08c6881c8aa30f405fafbf00eec386a72.tar.gz -O pg_tiktoken.tar.gz && \
-    echo "52f60ac800993a49aa8c609961842b611b6b1949717b69ce2ec9117117e16e4a pg_tiktoken.tar.gz" | sha256sum --check && \
+# 26806147b17b60763039c6a6878884c41a262318 made on 26/09/2023
+RUN wget https://github.com/kelvich/pg_tiktoken/archive/26806147b17b60763039c6a6878884c41a262318.tar.gz -O pg_tiktoken.tar.gz && \
+    echo "e64e55aaa38c259512d3e27c572da22c4637418cf124caba904cd50944e5004e pg_tiktoken.tar.gz" | sha256sum --check && \
    mkdir pg_tiktoken-src && cd pg_tiktoken-src && tar xvzf ../pg_tiktoken.tar.gz --strip-components=1 -C . && \
-    cargo pgx install --release && \
+    cargo pgrx install --release && \
    echo "trusted = true" >> /usr/local/pgsql/share/extension/pg_tiktoken.control

 #########################################################################################
@@ -754,21 +703,15 @@ RUN case "${PG_VERSION}" in \
 FROM rust-extensions-build AS pg-pgx-ulid-build
 ARG PG_VERSION

-RUN case "${PG_VERSION}" in \
-      "v14" | "v15") \
-        ;; \
-      "v16") \
-        echo "TODO: Not yet supported for PostgreSQL 16. Need to update pgrx dependencies" && exit 0 \
-	;; \
-      *) \
-        echo "unexpected PostgreSQL version" && exit 1 \
-        ;; \
-    esac && \
-    wget https://github.com/pksunkara/pgx_ulid/archive/refs/tags/v0.1.0.tar.gz -O pgx_ulid.tar.gz && \
-    echo "908b7358e6f846e87db508ae5349fb56a88ee6305519074b12f3d5b0ff09f791 pgx_ulid.tar.gz" | sha256sum --check && \
+RUN wget https://github.com/pksunkara/pgx_ulid/archive/refs/tags/v0.1.3.tar.gz -O pgx_ulid.tar.gz && \
+    echo "ee5db82945d2d9f2d15597a80cf32de9dca67b897f605beb830561705f12683c pgx_ulid.tar.gz" | sha256sum --check && \
    mkdir pgx_ulid-src && cd pgx_ulid-src && tar xvzf ../pgx_ulid.tar.gz --strip-components=1 -C . && \
-    sed -i 's/pgx        = "=0.7.3"/pgx = { version = "0.7.3", features = [ "unsafe-postgres" ] }/g' Cargo.toml && \
-    cargo pgx install --release && \
+    echo "******************* Apply a patch for Postgres 16 support; delete in the next release ******************" && \
+    wget https://github.com/pksunkara/pgx_ulid/commit/f84954cf63fc8c80d964ac970d9eceed3c791196.patch && \
+    patch -p1 < f84954cf63fc8c80d964ac970d9eceed3c791196.patch && \
+    echo "********************************************************************************************************" && \
+    sed -i 's/pgrx       = "=0.10.2"/pgrx = { version = "=0.10.2", features = [ "unsafe-postgres" ] }/g' Cargo.toml && \
+    cargo pgrx install --release && \
    echo "trusted = true" >> /usr/local/pgsql/share/extension/ulid.control

 #########################################################################################
--- a/compute_tools/src/monitor.rs
+++ b/compute_tools/src/monitor.rs
@@ -1,5 +1,5 @@
 use std::sync::Arc;
-use std::{thread, time};
+use std::{thread, time::Duration};

 use chrono::{DateTime, Utc};
 use postgres::{Client, NoTls};
@@ -7,7 +7,7 @@ use tracing::{debug, info};

 use crate::compute::ComputeNode;

-const MONITOR_CHECK_INTERVAL: u64 = 500; // milliseconds
+const MONITOR_CHECK_INTERVAL: Duration = Duration::from_millis(500);

 // Spin in a loop and figure out the last activity time in the Postgres.
 // Then update it in the shared state. This function never errors out.
@@ -17,13 +17,12 @@ fn watch_compute_activity(compute: &ComputeNode) {
    let connstr = compute.connstr.as_str();
    // Define `client` outside of the loop to reuse existing connection if it's active.
    let mut client = Client::connect(connstr, NoTls);
-    let timeout = time::Duration::from_millis(MONITOR_CHECK_INTERVAL);

    info!("watching Postgres activity at {}", connstr);

    loop {
        // Should be outside of the write lock to allow others to read while we sleep.
-        thread::sleep(timeout);
+        thread::sleep(MONITOR_CHECK_INTERVAL);

        match &mut client {
            Ok(cli) => {
--- a/deny.toml
+++ b/deny.toml
@@ -23,7 +23,7 @@ vulnerability = "deny"
 unmaintained = "warn"
 yanked = "warn"
 notice = "warn"
-ignore = ["RUSTSEC-2023-0052"]
+ignore = []

 # This section is considered when running `cargo deny check licenses`
 # More documentation for the licenses section can be found here:
--- a/docs/rfcs/028-pageserver-migration.md
+++ b/docs/rfcs/028-pageserver-migration.md
@@ -0,0 +1,599 @@
+# Seamless tenant migration
+
+- Author: john@neon.tech
+- Created on 2023-08-11
+- Implemented on ..
+
+## Summary
+
+The preceding [generation numbers RFC](025-generation-numbers.md) may be thought of as "making tenant
+migration safe". Following that,
+this RFC is about how those migrations are to be done:
+
+1. Seamlessly (without interruption to client availability)
+2. Quickly (enabling faster operations)
+3. Efficiently (minimizing I/O and $ cost)
+
+These points are in priority order: if we have to sacrifice
+efficiency to make a migration seamless for clients, we will
+do so, etc.
+
+This is accomplished by introducing two high level changes:
+
+- A dual-attached state for tenants, used in a control-plane-orchestrated
+  migration procedure that preserves availability during a migration.
+- Warm secondary locations for tenants, where on-disk content is primed
+  for a fast migration of the tenant from its current attachment to this
+  secondary location.
+
+## Motivation
+
+Migrating tenants between pageservers is essential to operating a service
+at scale, in several contexts:
+
+1. Responding to a pageserver node failure by migrating tenants to other pageservers
+2. Balancing load and capacity across pageservers, for example when a user expands their
+   database and they need to migrate to a pageserver with more capacity.
+3. Restarting pageservers for upgrades and maintenance
+
+The current situation steps for migration are:
+
+- detach from old node; skip if old node is dead; (the [skip part is still WIP](https://github.com/neondatabase/cloud/issues/5426)).
+- attach to new node
+- re-configure endpoints to use the new node
+
+Once [generation numbers](025-generation-numbers.md) are implemented,
+the detach step is no longer critical for correctness. So, we can
+
+- attach to a new node,
+- re-configure endpoints to use the new node, and then
+- detach from the old node.
+
+However, this still does not meet our seamless/fast/efficient goals:
+
+- Not fast: The new node will have to download potentially large amounts
+  of data from S3, which may take many minutes.
+- Not seamless: If we attach to a new pageserver before detaching an old one,
+  the new one might delete some objects that interrupt availability of reads on the old one.
+- Not efficient: the old pageserver will continue uploading
+  S3 content during the migration that will never be read.
+
+The user expectations for availability are:
+
+- For planned maintenance, there should be zero availability
+  gap. This expectation is fulfilled by this RFC.
+- For unplanned changes (e.g. node failures), there should be
+  minimal availability gap. This RFC provides the _mechanism_
+  to fail over quickly, but does not provide the failure _detection_
+  nor failover _policy_.
+
+## Non Goals
+
+- Defining service tiers with different storage strategies: the same
+  level of HA & overhead will apply to all tenants. This doesn't rule out
+  adding such tiers in future.
+- Enabling pageserver failover in the absence of a control plane: the control
+  plane will remain the source of truth for what should be attached where.
+- Totally avoiding availability gaps on unplanned migrations during
+  a failure (we expect a small, bounded window of
+  read unavailability of very recent LSNs)
+- Workload balancing: this RFC defines the mechanism for moving tenants
+  around, not the higher level logic for deciding who goes where.
+- Defining all possible configuration flows for tenants: the migration process
+  defined in this RFC demonstrates the sufficiency of the pageserver API, but
+  is not the only kind of configuration change the control plane will ever do.
+  The APIs defined here should let the control plane move tenants around in
+  whatever way is needed while preserving data safety and read availability.
+
+## Impacted components
+
+Pageserver, control plane
+
+## Terminology
+
+- **Attachment**: a tenant is _attached_ to a pageserver if it has
+  been issued a generation number, and is running an instance of
+  the `Tenant` type, ingesting the WAL, and available to serve
+  page reads.
+- **Location**: locations are a superset of attachments. A location
+  is a combination of a tenant and a pageserver. We may _attach_ at a _location_.
+
+- **Secondary location**: a location which is not currently attached.
+- **Warm secondary location**: a location which is not currently attached, but is endeavoring to maintain a warm local cache of layers. We avoid calling this a _warm standby_ to avoid confusion with similar postgres features.
+
+## Implementation (high level)
+
+### Warm secondary locations
+
+To enable faster migrations, we will identify at least one _secondary location_
+for each tenant. This secondary location will keep a warm cache of layers
+for the tenant, so that if it is later attached, it can catch up with the
+latest LSN quickly: rather than downloading everything, it only has to replay
+the recent part of the WAL to advance from the remote_consistent_offset to the
+most recent LSN in the WAL.
+
+The control plane is responsible for selecting secondary locations, and
+calling into pageservers to configure tenants into a secondary mode at this
+new location, as well as attaching the tenant in its existing primary location.
+
+The attached pageserver for a tenant will publish a [layer heatmap](#layer-heatmap)
+to advise secondaries of which layers should be downloaded.
+
+### Location modes
+
+Currently, we consider a tenant to be in one of two states on a pageserver:
+
+- Attached: active `Tenant` object, and layers on local disk
+- Detached: no layers on local disk, no runtime state.
+
+We will extend this with finer-grained modes, whose purpose will become
+clear in later sections:
+
+- **AttachedSingle**: equivalent the existing attached state.
+- **AttachedMulti**: like AttachedSingle, holds an up to date generation, but
+  does not do deletions.
+- **AttachedStale**: like AttachedSingle, holds a stale generation,
+  do not do any remote storage operations.
+- **Secondary**: keep local state on disk, periodically update from S3.
+- **Detached**: equivalent to existing detached state.
+
+To control these finer grained states, a new pageserver API endpoint will be added.
+
+### Cutover procedure
+
+Define old location and new location as "Node A" and "Node B". Consider
+the case where both nodes are available, and Node B was previously configured
+as a secondary location for the tenant we are migrating.
+
+The cutover procedure is orchestrated by the control plane, calling into
+the pageservers' APIs:
+
+1. Call to Node A requesting it to flush to S3 and enter AttachedStale state
+2. Increment generation, and call to Node B requesting it to enter AttachedMulti
+   state with the new generation.
+3. Call to Node B, requesting it to download the latest hot layers from remote storage,
+   according to the latest heatmap flushed by Node A.
+4. Wait for Node B's WAL ingestion to catch up with node A's
+5. Update endpoints to use node B instead of node A
+6. Call to node B requesting it to enter state AttachedSingle.
+7. Call to node A requesting it to enter state Secondary
+
+The following table summarizes how the state of the system advances:
+
+|     Step      |     Node A     |     Node B     | Node used by endpoints |
+| :-----------: | :------------: | :------------: | :--------------------: |
+| 1 (_initial_) | AttachedSingle |   Secondary    |           A            |
+|       2       | AttachedStale  | AttachedMulti  |           A            |
+|       3       | AttachedStale  | AttachedMulti  |           A            |
+|       4       | AttachedStale  | AttachedMulti  |           A            |
+| 5 (_cutover_) | AttachedStale  | AttachedMulti  |           B            |
+|       6       | AttachedStale  | AttachedSingle |           B            |
+|  7 (_final_)  |   Secondary    | AttachedSingle |           B            |
+
+The procedure described for a clean handover from a live node to a secondary
+is also used for failure cases and for migrations to a location that is not
+configured as a secondary, by simply skipping irrelevant steps, as described in
+the following sections.
+
+#### Migration from an unresponsive node
+
+If node A is unavailable, then all calls into
+node A are skipped and we don't wait for B to catch up before
+switching updating the endpoints to use B.
+
+#### Migration to a location that is not a secondary
+
+If node B is initially in Detached state, the procedure is identical. Since Node B
+is coming from a Detached state rather than Secondary, the download of layers and
+catch up with WAL will take much longer.
+
+We might do this if:
+
+- Attached and secondary locations are both critically low on disk, and we need
+  to migrate to a third node with more resources available.
+- We are migrating a tenant which does not use secondary locations to save on cost.
+
+#### Permanent migration away from a node
+
+In the final step of the migration, we generally request the original node to enter a Secondary
+state. This is typical if we are doing a planned migration during maintenance, or to
+balance CPU/network load away from a node.
+
+One might also want to permanently migrate away: this can be done by simply removing the secondary
+location after the migration is complete, or as an optimization by substituting the Detached state
+for the Secondary state in the final step.
+
+#### Cutover diagram
+
+```mermaid
+sequenceDiagram
+participant CP as Control plane
+participant A as Node A
+participant B as Node B
+participant E as Endpoint
+
+CP->>A: PUT Flush & go to AttachedStale
+note right of A: A continues to ingest WAL
+CP->>B: PUT AttachedMulti
+CP->>B: PUT Download layers from latest heatmap
+note right of B: B downloads from S3
+loop Poll until download complete
+CP->>B: GET download status
+end
+activate B
+note right of B: B ingests WAL
+loop Poll until catch up
+CP->>B: GET visible WAL
+CP->>A: GET visible WAL
+end
+deactivate B
+CP->>E: Configure to use Node B
+E->>B: Connect for reads
+CP->>B: PUT AttachedSingle
+CP->>A: PUT Secondary
+```
+
+#### Cutover from an unavailable pageserver
+
+This case is far simpler: we may skip straight to our intended
+end state.
+
+```mermaid
+sequenceDiagram
+participant A as Node A
+participant CP as Control plane
+participant B as Node B
+participant E as Endpoint
+
+note right of A: Node A offline
+activate A
+CP->>B: PUT AttachedSingle
+CP->>E: Configure to use Node B
+E->>B: Connect for reads
+deactivate A
+```
+
+## Implementation (detail)
+
+### Purpose of AttachedMulti, AttachedStale
+
+#### AttachedMulti
+
+Ordinarily, an attached pageserver whose generation is the latest may delete
+layers at will (e.g. during compaction). If a previous generation pageserver
+is also still attached, and in use by endpoints, then this layer deletion could
+lead to a loss of availability for the endpoint when reading from the previous
+generation pageserver.
+
+The _AttachedMulti_ state simply disables deletions. These will be enqueued
+in `RemoteTimelineClient` until the control plane transitions the
+node into AttachedSingle, which unblocks deletions.  Other remote storage operations
+such as uploads are not blocked.
+
+AttachedMulti is not required for data safety, only to preserve availability
+on pageservers running with stale generations.
+
+A node enters AttachedMulti only when explicitly asked to by the control plane. It should
+only remain in this state for the duration of a migration.
+
+If a control plane bug leaves
+the node in AttachedMulti for a long time, then we must avoid unbounded memory use from enqueued
+deletions. This may be accomplished simply, by dropping enqueued deletions when some modest
+threshold of delayed deletions (e.g. 10k layers per tenant) is reached. As with all deletions,
+it is safe to skip them, and the leaked objects will be eventually cleaned up by scrub or
+by timeline deletion.
+
+During AttachedMulti, the Tenant is free to drop layers from local disk in response to
+disk pressure: only the deletion of remote layers is blocked.
+
+#### AttachedStale
+
+Currently, a pageserver with a stale generation number will continue to
+upload layers, but be prevented from completing deletions. This is safe, but inefficient: layers uploaded by this stale generation
+will not be read back by future generations of pageservers.
+
+The _AttachedStale_ state disables S3 uploads. The stale pageserver
+will continue to ingest the WAL and write layers to local disk, but not to
+do any uploads to S3.
+
+A node may enter AttachedStale in two ways:
+
+- Explicitly, when control plane calls into the node at the start of a migration.
+- Implicitly, when the node tries to validate some deletions and discovers
+  that its generation is stale.
+
+The AttachedStale state also disables sending consumption metrics from
+that location: it is interpreted as an indication that some other pageserver
+is already attached or is about to be attached, and that new pageserver will
+be responsible for sending consumption metrics.
+
+#### Disk Pressure & AttachedStale
+
+Over long periods of time, a tenant location in AttachedStale will accumulate data
+on local disk, as it cannot evict any layers written since it entered the
+AttachStale state. We rely on the control plane to revert the location to
+Secondary or Detached at the end of a migration.
+
+This scenario is particularly noteworthy when evacuating all tenants on a pageserver:
+since _all_ the attached tenants will go into AttachedStale, we will be doing no
+uploads at all, therefore ingested data will cause disk usage to increase continuously.
+Under nominal conditions, the available disk space on pageservers should be sufficient
+to complete the evacuation before this becomes a problem, but we must also handle
+the case where we hit a low disk situation while in this state.
+
+The concept of disk pressure already exists in the pageserver: the `disk_usage_eviction_task`
+touches each Tenant when it determines that a low-disk condition requires
+some layer eviction. Having selected layers for eviction, the eviction
+task calls `Timeline::evict_layers`.
+
+**Safety**: If evict_layers is called while in AttachedStale state, and some of the to-be-evicted
+layers are not yet uploaded to S3, then the block on uploads will be lifted. This
+will result in leaking some objects once a migration is complete, but will enable
+the node to manage its disk space properly: if a node is left with some tenants
+in AttachedStale indefinitely due to a network partition or control plane bug,
+these tenants will not cause a full disk condition.
+
+### Warm secondary updates
+
+#### Layer heatmap
+
+The secondary location's job is to serve reads **with the same quality of service as the original location
+was serving them around the time of a migration**. This does not mean the secondary
+location needs the whole set of layers: inactive layers that might soon
+be evicted on the attached pageserver need not be downloaded by the
+secondary. A totally idle tenant only needs to maintain enough on-disk
+state to enable a fast cold start (i.e. the most recent image layers are
+typically sufficient).
+
+To enable this, we introduce the concept of a _layer heatmap_, which
+acts as an advisory input to secondary locations to decide which
+layers to download from S3.
+
+#### Attached pageserver
+
+The attached pageserver, if in state AttachedSingle, periodically
+uploads a serialized heat map to S3. It may skip this if there
+is no change since the last time it uploaded (e.g. if the tenant
+is totally idle).
+
+Additionally, when the tenant is flushed to remote storage prior to a migration
+(the first step in [cutover procedure](#cutover-procedure)), 
+the heatmap is written out. This enables a future attached pageserver
+to get an up to date view when deciding which layers to download.
+
+#### Secondary location behavior
+
+Secondary warm locations run a simple loop, implemented separately from
+the main `Tenant` type, which represents attached tenants:
+
+- Download the layer heatmap
+- Select any "hot enough" layers to download, if there is sufficient
+  free disk space.
+- Download layers, if they were not previously evicted (see below)
+- Download the latest index_part.json
+- Check if any layers currently on disk are no longer referenced by
+  IndexPart & delete them
+
+Note that the heatmap is only advisory: if a secondary location has plenty
+of disk space, it may choose to retain layers that aren't referenced
+by the heatmap, as long as they are still referenced by the IndexPart. Conversely,
+if a node is very low on disk space, it might opt to raise the heat threshold required
+to both downloading a layer, until more disk space is available.
+
+#### Secondary locations & disk pressure
+
+Secondary locations are subject to eviction on disk pressure, just as
+attached locations are.  For eviction purposes, the access time of a
+layer in a secondary location will be the access time given in the heatmap,
+rather than the literal time at which the local layer file was accessed.
+
+The heatmap will indicate which layers are in local storage on the attached
+location.  The secondary will always attempt to get back to having that
+set of layers on disk, but to avoid flapping, it will remember the access
+time of the layer it was most recently asked to evict, and layers whose
+access time is below that will not be re-downloaded.
+
+The resulting behavior is that after a layer is evicted from a secondary
+location, it is only re-downloaded once the attached pageserver accesses
+the layer and uploads a heatmap reflecting that access time.  On a pageserver
+restart, the secondary location will attempt to download all layers in
+the heatmap again, if they are not on local disk.
+
+This behavior will be slightly different when secondary locations are
+used for "low energy tenants", but that is beyond the scope of this RFC.
+
+### Location configuration API
+
+Currently, the `/tenant/<tenant_id>/config` API defines various
+tunables like compaction settings, which apply to the tenant irrespective
+of which pageserver it is running on.
+
+A new "location config" structure will be introduced, which defines
+configuration which is per-tenant, but local to a particular pageserver,
+such as the attachment mode and whether it is a secondary.
+
+The pageserver will expose a new per-tenant API for setting
+the state: `/tenant/<tenant_id>/location/config`.
+
+Body content:
+
+```
+{
+  state: 'enum{Detached, Secondary, AttachedSingle, AttachedMulti, AttachedStale}',
+  generation: Option<u32>,
+  configuration: `Option<TenantConfig>`
+  flush: bool
+}
+```
+
+Existing `/attach` and `/detach` endpoint will have the same
+behavior as calling `/location/config` with `AttachedSingle` and `Detached`
+states respectively. These endpoints will be deprecated and later
+removed.
+
+The generation attribute is mandatory for entering `AttachedSingle` or
+`AttachedMulti`.
+
+The configuration attribute is mandatory when entering any state other
+than `Detached`. This configuration is the same as the body for
+the existing `/tenant/<tenant_id>/config` endpoint.
+
+The `flush` argument indicates whether the pageservers should flush
+to S3 before proceeding: this only has any effect if the node is
+currently in AttachedSingle or AttachedMulti. This is used
+during the first phase of migration, when transitioning the
+old pageserver to AttachedSingle.
+
+The `/re-attach` API response will be extended to include a `state` as
+well as a `generation`, enabling the pageserver to enter the
+correct state for each tenant on startup.
+
+### Database schema for locations
+
+A new table `ProjectLocation`:
+
+- pageserver_id: int
+- tenant_id: TenantId
+- generation: Option<int>
+- state: `enum(Secondary, AttachedSingle, AttachedMulti)`
+
+Notes:
+
+- It is legacy for a Project to have zero `ProjectLocation`s
+- The `pageserver` column in `Project` now means "to which pageserver should
+  endpoints connect", rather than simply which pageserver is attached.
+- The `generation` column in `Project` remains, and is incremented and used
+  to set the generation of `ProjectLocation` rows when they are set into
+  an attached state.
+- The `Detached` state is implicitly represented as the absence of
+  a `ProjectLocation`.
+
+### Executing migrations
+
+Migrations will be implemented as Go functions, within the
+existing `Operation` framework in the control plane. These
+operations are persistent, such that they will always keep
+trying until completion: this property is important to avoid
+leaving garbage behind on pageservers, such as AttachedStale
+locations.
+
+### Recovery from failures during migration
+
+During migration, the control plane may encounter failures of either
+the original or new pageserver, or both:
+
+- If the original fails, skip past waiting for the new pageserver
+  to catch up, and put it into AttachedSingle immediately.
+- If the new node fails, put the old pageserver into Secondary
+  and then back into AttachedSingle (this has the effect of
+  retaining on-disk state and granting it a fresh generation number).
+- If both nodes fail, keep trying until one of them is available
+  again.
+
+### Control plane -> Pageserver reconciliation
+
+A migration may be done while the old node is unavailable,
+in which case the old node may still be running in an AttachedStale
+state.
+
+In this case, it is undesirable to have the migration `Operation`
+stay alive until the old node eventually comes back online
+and can be cleaned up. To handle this, the control plane
+should run a background reconciliation process to compare
+a pageserver's attachments with the database, and clean up
+any that shouldn't be there any more.
+
+Note that there will be no work to do if the old node was really
+offline, as during startup it will call into `/re-attach` and
+be updated that way. The reconciliation will only be needed
+if the node was unavailable but still running.
+
+## Alternatives considered
+
+### Only enabling secondary locations for tenants on a higher service tier
+
+This will make sense in future, especially for tiny databases that may be
+downloaded from S3 in milliseconds when needed.
+
+However, it is not wise to do it immediately, because pageservers contain
+a mixture of higher and lower tier workloads. If we had 1 tenant with
+a secondary location and 9 without, then those other 9 tenants will do
+a lot of I/O as they try to recover from S3, which may degrade the
+service of the tenant which had a secondary location.
+
+Until we segregate tenant on different service tiers on different pageserver
+nodes, or implement & test QoS to ensure that tenants with secondaries are
+not harmed by tenants without, we should use the same failover approach
+for all the tenants.
+
+### Hot secondary locations (continuous WAL replay)
+
+Instead of secondary locations populating their caches from S3, we could
+have them consume the WAL from safekeepers. The downsides of this would be:
+
+- Double load on safekeepers, which are a less scalable service than S3
+- Secondary locations' on-disk state would end up subtly different to
+  the remote state, which would make synchronizing with S3 more complex/expensive
+  when going into attached state.
+
+The downside of only updating secondary locations from S3 is that we will
+have a delay during migration from replaying the LSN range between what's
+in S3 and what's in the pageserver. This range will be very small on
+planned migrations, as we have the old pageserver flush to S3 immediately
+before attaching the new pageserver. On unplanned migrations (old pageserver
+is unavailable), the range of LSNs to replay is bounded by the flush frequency
+on the old pageserver. However, the migration doesn't have to wait for the
+replay: it's just that not-yet-replayed LSNs will be unavailable for read
+until the new pageserver catches up.
+
+We expect that pageserver reads of the most recent LSNs will be relatively
+rare, as for an active endpoint those pages will usually still be in the postgres
+page cache: this leads us to prefer synchronizing from S3 on secondary
+locations, rather than consuming the WAL from safekeepers.
+
+### Cold secondary locations
+
+It is not functionally necessary to keep warm caches on secondary locations at all. However, if we do not, then
+we would experience a de-facto availability loss in unplanned migrations, as reads to the new node would take an extremely long time (many seconds, perhaps minutes).
+
+Warm caches on secondary locations are necessary to meet
+our availability goals.
+
+### Pageserver-granularity failover
+
+Instead of migrating tenants individually, we could have entire spare nodes,
+and on a node death, move all its work to one of these spares.
+
+This approach is avoided for several reasons:
+
+- we would still need fine-grained tenant migration for other
+  purposes such as balancing load
+- by sharing the spare capacity over many peers rather than one spare node,
+  these peers may use the capacity for other purposes, until it is needed
+  to handle migrated tenants. e.g. for keeping a deeper cache of their
+  attached tenants.
+
+### Readonly during migration
+
+We could simplify migrations by making both previous and new nodes go into a
+readonly state, then flush remote content from the previous node, then activate
+attachment on the secondary node.
+
+The downside to this approach is a potentially large gap in readability of
+recent LSNs while loading data onto the new node. To avoid this, it is worthwhile
+to incur the extra cost of double-replaying the WAL onto old and new nodes' local
+storage during a migration.
+
+### Peer-to-peer pageserver communication
+
+Rather than uploading the heatmap to S3, attached pageservers could make it
+available to peers.
+
+Currently, pageservers have no peer to peer communication, so adding this
+for heatmaps would incur significant overhead in deployment and configuration
+of the service, and ensuring that when a new pageserver is deployed, other
+pageservers are updated to be aware of it.
+
+As well as simplifying implementation, putting heatmaps in S3 will be useful
+for future analytics purposes -- gathering aggregated statistics on activity
+pattersn across many tenants may be done directly from data in S3.
--- a/libs/consumption_metrics/src/lib.rs
+++ b/libs/consumption_metrics/src/lib.rs
@@ -107,7 +107,7 @@ pub const CHUNK_SIZE: usize = 1000;

 // Just a wrapper around a slice of events
 // to serialize it as `{"events" : [ ] }
-#[derive(serde::Serialize)]
+#[derive(serde::Serialize, serde::Deserialize)]
 pub struct EventChunk<'a, T: Clone> {
    pub events: std::borrow::Cow<'a, [T]>,
 }
--- a/libs/remote_storage/src/lib.rs
+++ b/libs/remote_storage/src/lib.rs
@@ -25,11 +25,7 @@ use tokio::io;
 use toml_edit::Item;
 use tracing::info;

-pub use self::{
-    local_fs::LocalFs,
-    s3_bucket::S3Bucket,
-    simulate_failures::{SimulatedError, UnreliableWrapper},
-};
+pub use self::{local_fs::LocalFs, s3_bucket::S3Bucket, simulate_failures::UnreliableWrapper};

 /// How many different timelines can be processed simultaneously when synchronizing layers with the remote storage.
 /// During regular work, pageserver produces one layer file per timeline checkpoint, with bursts of concurrency
@@ -194,44 +190,26 @@ impl Debug for Download {
 #[derive(Debug)]
 pub enum DownloadError {
    /// Validation or other error happened due to user input.
-    ///
-    /// This is only used by LOCAL_FS.
    BadInput(anyhow::Error),
-
    /// The file was not found in the remote storage.
-    ///
-    /// This can only happen during download, never during delete.
    NotFound,
-
-    /// The file was found in the remote storage, but the operation failed.
-    ///
-    /// The error should have context already describing the real failed operation.
+    /// The file was found in the remote storage, but the download failed.
    Other(anyhow::Error),
 }

 impl std::fmt::Display for DownloadError {
    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
-        use DownloadError::*;
        match self {
-            NotFound => write!(f, "No file found for the remote object id given"),
-            // this is same as thiserror error(transparent); it handles {} and {:#}
-            Other(e) | BadInput(e) => std::fmt::Display::fmt(e, f),
+            DownloadError::BadInput(e) => {
+                write!(f, "Failed to download a remote file due to user input: {e}")
+            }
+            DownloadError::NotFound => write!(f, "No file found for the remote object id given"),
+            DownloadError::Other(e) => write!(f, "Failed to download a remote file: {e:?}"),
        }
    }
 }

-impl std::error::Error for DownloadError {
-    fn source(&self) -> Option<&(dyn std::error::Error + 'static)> {
-        use DownloadError::*;
-        match self {
-            NotFound => None,
-            Other(_) | BadInput(_) => {
-                // TODO: these are anyhow, cannot return here
-                None
-            }
-        }
-    }
-}
+impl std::error::Error for DownloadError {}

 /// Every storage, currently supported.
 /// Serves as a simple way to pass around the [`RemoteStorage`] without dealing with generics.
--- a/libs/remote_storage/src/s3_bucket.rs
+++ b/libs/remote_storage/src/s3_bucket.rs
@@ -47,10 +47,47 @@ pub struct S3Bucket {
    bucket_name: String,
    prefix_in_bucket: Option<String>,
    max_keys_per_list_response: Option<i32>,
+    concurrency_limiter: ConcurrencyLimiter,
+}
+
+struct ConcurrencyLimiter {
    // Every request to S3 can be throttled or cancelled, if a certain number of requests per second is exceeded.
    // Same goes to IAM, which is queried before every S3 request, if enabled. IAM has even lower RPS threshold.
    // The helps to ensure we don't exceed the thresholds.
-    concurrency_limiter: Arc<Semaphore>,
+    write: Arc<Semaphore>,
+    read: Arc<Semaphore>,
+}
+
+impl ConcurrencyLimiter {
+    fn for_kind(&self, kind: RequestKind) -> &Arc<Semaphore> {
+        match kind {
+            RequestKind::Get => &self.read,
+            RequestKind::Put => &self.write,
+            RequestKind::List => &self.read,
+            RequestKind::Delete => &self.write,
+        }
+    }
+
+    async fn acquire(
+        &self,
+        kind: RequestKind,
+    ) -> Result<tokio::sync::SemaphorePermit<'_>, tokio::sync::AcquireError> {
+        self.for_kind(kind).acquire().await
+    }
+
+    async fn acquire_owned(
+        &self,
+        kind: RequestKind,
+    ) -> Result<tokio::sync::OwnedSemaphorePermit, tokio::sync::AcquireError> {
+        Arc::clone(self.for_kind(kind)).acquire_owned().await
+    }
+
+    fn new(limit: usize) -> ConcurrencyLimiter {
+        Self {
+            read: Arc::new(Semaphore::new(limit)),
+            write: Arc::new(Semaphore::new(limit)),
+        }
+    }
 }

 #[derive(Default)]
@@ -117,7 +154,7 @@ impl S3Bucket {
            bucket_name: aws_config.bucket_name.clone(),
            max_keys_per_list_response: aws_config.max_keys_per_list_response,
            prefix_in_bucket,
-            concurrency_limiter: Arc::new(Semaphore::new(aws_config.concurrency_limit.get())),
+            concurrency_limiter: ConcurrencyLimiter::new(aws_config.concurrency_limit.get()),
        })
    }

@@ -156,7 +193,7 @@ impl S3Bucket {
        let started_at = start_counting_cancelled_wait(kind);
        let permit = self
            .concurrency_limiter
-            .acquire()
+            .acquire(kind)
            .await
            .expect("semaphore is never closed");

@@ -172,8 +209,7 @@ impl S3Bucket {
        let started_at = start_counting_cancelled_wait(kind);
        let permit = self
            .concurrency_limiter
-            .clone()
-            .acquire_owned()
+            .acquire_owned(kind)
            .await
            .expect("semaphore is never closed");

@@ -345,7 +381,7 @@ impl RemoteStorage for S3Bucket {
                .set_max_keys(self.max_keys_per_list_response)
                .send()
                .await
-                .context("list S3 prefixes")
+                .context("Failed to list S3 prefixes")
                .map_err(DownloadError::Other);

            let started_at = ScopeGuard::into_inner(started_at);
@@ -397,7 +433,7 @@ impl RemoteStorage for S3Bucket {
                .set_max_keys(self.max_keys_per_list_response)
                .send()
                .await
-                .context("list files in S3 bucket");
+                .context("Failed to list files in S3 bucket");

            let started_at = ScopeGuard::into_inner(started_at);
            metrics::BUCKET_METRICS
@@ -521,7 +557,10 @@ impl RemoteStorage for S3Bucket {
                        .deleted_objects_total
                        .inc_by(chunk.len() as u64);
                    if let Some(errors) = resp.errors {
-                        return Err(anyhow::anyhow!("delete {} objects", errors.len()));
+                        return Err(anyhow::format_err!(
+                            "Failed to delete {} objects",
+                            errors.len()
+                        ));
                    }
                }
                Err(e) => {
--- a/libs/remote_storage/src/simulate_failures.rs
+++ b/libs/remote_storage/src/simulate_failures.rs
@@ -18,7 +18,7 @@ pub struct UnreliableWrapper {
 }

 /// Used to identify retries of different unique operation.
-#[derive(Hash, Eq, PartialEq)]
+#[derive(Debug, Hash, Eq, PartialEq)]
 enum RemoteOp {
    ListPrefixes(Option<RemotePath>),
    Upload(RemotePath),
@@ -27,22 +27,6 @@ enum RemoteOp {
    DeleteObjects(Vec<RemotePath>),
 }

-impl std::fmt::Debug for RemoteOp {
-    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
-        use RemoteOp::*;
-        match self {
-            ListPrefixes(arg0) => f.debug_tuple("ListPrefixes").field(arg0).finish(),
-            Upload(arg0) => f.debug_tuple("Upload").field(arg0).finish(),
-            Download(arg0) => f.debug_tuple("Download").field(arg0).finish(),
-            Delete(arg0) => f.debug_tuple("Delete").field(arg0).finish(),
-            DeleteObjects(many) if many.len() > 3 => {
-                write!(f, "DeleteObjects({} paths)", many.len())
-            }
-            DeleteObjects(few) => f.debug_tuple("DeleteObjects").field(few).finish(),
-        }
-    }
-}
-
 impl UnreliableWrapper {
    pub fn new(inner: crate::GenericRemoteStorage, attempts_to_fail: u64) -> Self {
        assert!(attempts_to_fail > 0);
@@ -75,12 +59,13 @@ impl UnreliableWrapper {
                    e.remove();
                    Ok(attempts_before_this)
                } else {
-                    let error = anyhow::anyhow!(SimulatedError::from(e.key()));
+                    let error =
+                        anyhow::anyhow!("simulated failure of remote operation {:?}", e.key());
                    Err(DownloadError::Other(error))
                }
            }
            Entry::Vacant(e) => {
-                let error = anyhow::anyhow!(SimulatedError::from(e.key()));
+                let error = anyhow::anyhow!("simulated failure of remote operation {:?}", e.key());
                e.insert(1);
                Err(DownloadError::Other(error))
            }
@@ -95,26 +80,6 @@ impl UnreliableWrapper {
    }
 }

-/// `pub` type for checking if this is the root cause around logging.
-///
-/// This is just a string to avoid cloning a huge number of paths a second time.
-#[derive(Debug)]
-pub struct SimulatedError(String);
-
-impl<'a> From<&'a RemoteOp> for SimulatedError {
-    fn from(value: &'_ RemoteOp) -> Self {
-        SimulatedError(format!("simulated failure of remote operation {:?}", value))
-    }
-}
-
-impl std::fmt::Display for SimulatedError {
-    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
-        f.write_str(&self.0)
-    }
-}
-
-impl std::error::Error for SimulatedError {}
-
 #[async_trait::async_trait]
 impl RemoteStorage for UnreliableWrapper {
    async fn list_prefixes(
--- a/libs/utils/src/http/error.rs
+++ b/libs/utils/src/http/error.rs
@@ -24,6 +24,9 @@ pub enum ApiError {
    #[error("Precondition failed: {0}")]
    PreconditionFailed(Box<str>),

+    #[error("Resource temporarily unavailable: {0}")]
+    ResourceUnavailable(String),
+
    #[error("Shutting down")]
    ShuttingDown,

@@ -59,6 +62,10 @@ impl ApiError {
                "Shutting down".to_string(),
                StatusCode::SERVICE_UNAVAILABLE,
            ),
+            ApiError::ResourceUnavailable(err) => HttpErrorBody::response_from_msg_and_status(
+                err.to_string(),
+                StatusCode::SERVICE_UNAVAILABLE,
+            ),
            ApiError::InternalServerError(err) => HttpErrorBody::response_from_msg_and_status(
                err.to_string(),
                StatusCode::INTERNAL_SERVER_ERROR,
--- a/libs/utils/src/logging.rs
+++ b/libs/utils/src/logging.rs
@@ -216,6 +216,24 @@ impl std::fmt::Debug for PrettyLocation<'_, '_> {
    }
 }

+/// When you will store a secret but want to make sure it won't
+/// be accidentally logged, wrap it in a SecretString, whose Debug
+/// implementation does not expose the contents.
+#[derive(Clone, Eq, PartialEq)]
+pub struct SecretString(String);
+
+impl SecretString {
+    pub fn get_contents(&self) -> &str {
+        self.0.as_str()
+    }
+}
+
+impl std::fmt::Debug for SecretString {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        write!(f, "[SECRET]")
+    }
+}
+
 #[cfg(test)]
 mod tests {
    use metrics::{core::Opts, IntCounterVec};
--- a/libs/vm_monitor/src/cgroup.rs
+++ b/libs/vm_monitor/src/cgroup.rs
@@ -431,14 +431,14 @@ impl CgroupWatcher {
                            .context("failed to request upscale")?;

                        let memory_high =
-                            self.get_high_bytes().context("failed to get memory.high")?;
+                            self.get_memory_high_bytes().context("failed to get memory.high")?;
                        let new_high = memory_high + self.config.memory_high_increase_by_bytes;
                        info!(
                            current_high_bytes = memory_high,
                            new_high_bytes = new_high,
                            "updating memory.high"
                        );
-                        self.set_high_bytes(new_high)
+                        self.set_memory_high_bytes(new_high)
                            .context("failed to set memory.high")?;
                        last_memory_high_increase_at = Some(Instant::now());
                        continue;
@@ -556,14 +556,6 @@ impl CgroupWatcher {
    }
 }

-/// Represents a set of limits we apply to a cgroup to control memory usage.
-///
-/// Setting these values also affects the thresholds for receiving usage alerts.
-#[derive(Debug)]
-pub struct MemoryLimits {
-    pub high: u64,
-}
-
 // Methods for manipulating the actual cgroup
 impl CgroupWatcher {
    /// Get a handle on the freezer subsystem.
@@ -624,50 +616,29 @@ impl CgroupWatcher {
    }

    /// Set cgroup memory.high threshold.
-    pub fn set_high_bytes(&self, bytes: u64) -> anyhow::Result<()> {
+    pub fn set_memory_high_bytes(&self, bytes: u64) -> anyhow::Result<()> {
+        self.set_memory_high_internal(MaxValue::Value(u64::min(bytes, i64::MAX as u64) as i64))
+    }
+
+    /// Set the cgroup's memory.high to 'max', disabling it.
+    pub fn unset_memory_high(&self) -> anyhow::Result<()> {
+        self.set_memory_high_internal(MaxValue::Max)
+    }
+
+    fn set_memory_high_internal(&self, value: MaxValue) -> anyhow::Result<()> {
        self.memory()
            .context("failed to get memory subsystem")?
            .set_mem(cgroups_rs::memory::SetMemory {
                low: None,
-                high: Some(MaxValue::Value(u64::min(bytes, i64::MAX as u64) as i64)),
+                high: Some(value),
                min: None,
                max: None,
            })
-            .context("failed to set memory.high")
-    }
-
-    /// Set cgroup memory.high and memory.max.
-    pub fn set_limits(&self, limits: &MemoryLimits) -> anyhow::Result<()> {
-        info!(limits.high, path = self.path(), "writing new memory limits",);
-        self.memory()
-            .context("failed to get memory subsystem while setting memory limits")?
-            .set_mem(cgroups_rs::memory::SetMemory {
-                min: None,
-                low: None,
-                high: Some(MaxValue::Value(
-                    u64::min(limits.high, i64::MAX as u64) as i64
-                )),
-                max: None,
-            })
-            .context("failed to set memory limits")
-    }
-
-    /// Given some amount of available memory, set the desired cgroup memory limits
-    pub fn set_memory_limits(&mut self, available_memory: u64) -> anyhow::Result<()> {
-        let new_high = self.config.calculate_memory_high_value(available_memory);
-        let limits = MemoryLimits { high: new_high };
-        info!(
-            path = self.path(),
-            memory = ?limits,
-            "setting cgroup memory",
-        );
-        self.set_limits(&limits)
-            .context("failed to set cgroup memory limits")?;
-        Ok(())
+            .map_err(anyhow::Error::from)
    }

    /// Get memory.high threshold.
-    pub fn get_high_bytes(&self) -> anyhow::Result<u64> {
+    pub fn get_memory_high_bytes(&self) -> anyhow::Result<u64> {
        let high = self
            .memory()
            .context("failed to get memory subsystem while getting memory statistics")?
--- a/libs/vm_monitor/src/runner.rs
+++ b/libs/vm_monitor/src/runner.rs
@@ -16,7 +16,7 @@ use tokio::sync::mpsc;
 use tokio_util::sync::CancellationToken;
 use tracing::{error, info, warn};

-use crate::cgroup::{CgroupWatcher, MemoryLimits, Sequenced};
+use crate::cgroup::{CgroupWatcher, Sequenced};
 use crate::dispatcher::Dispatcher;
 use crate::filecache::{FileCacheConfig, FileCacheState};
 use crate::protocol::{InboundMsg, InboundMsgKind, OutboundMsg, OutboundMsgKind, Resources};
@@ -106,6 +106,51 @@ impl Runner {
            kill,
        };

+        // If we have both the cgroup and file cache integrations enabled, it's possible for
+        // temporary failures to result in cgroup throttling (from memory.high), that in turn makes
+        // it near-impossible to connect to the file cache (because it times out). Unfortunately,
+        // we *do* still want to determine the file cache size before setting the cgroup's
+        // memory.high, so it's not as simple as just swapping the order.
+        //
+        // Instead, the resolution here is that on vm-monitor startup (note: happens on each
+        // connection from autoscaler-agent, possibly multiple times per compute_ctl lifecycle), we
+        // temporarily unset memory.high, to allow any existing throttling to dissipate. It's a bit
+        // of a hacky solution, but helps with reliability.
+        if let Some(name) = &args.cgroup {
+            // Best not to set up cgroup stuff more than once, so we'll initialize cgroup state
+            // now, and then set limits later.
+            info!("initializing cgroup");
+
+            let (cgroup, cgroup_event_stream) = CgroupWatcher::new(name.clone(), requesting_send)
+                .context("failed to create cgroup manager")?;
+
+            info!("temporarily unsetting memory.high");
+
+            // Temporarily un-set cgroup memory.high; see above.
+            cgroup
+                .unset_memory_high()
+                .context("failed to unset memory.high")?;
+
+            let cgroup = Arc::new(cgroup);
+
+            let cgroup_clone = Arc::clone(&cgroup);
+            spawn_with_cancel(
+                token.clone(),
+                |_| error!("cgroup watcher terminated"),
+                async move { cgroup_clone.watch(notified_recv, cgroup_event_stream).await },
+            );
+
+            state.cgroup = Some(cgroup);
+        } else {
+            // *NOTE*: We need to forget the sender so that its drop impl does not get ran.
+            // This allows us to poll it in `Monitor::run` regardless of whether we
+            // are managing a cgroup or not. If we don't forget it, all receives will
+            // immediately return an error because the sender is droped and it will
+            // claim all select! statements, effectively turning `Monitor::run` into
+            // `loop { fail to receive }`.
+            mem::forget(requesting_send);
+        }
+
        let mut file_cache_reserved_bytes = 0;
        let mem = get_total_system_memory();

@@ -119,7 +164,7 @@ impl Runner {
                false => FileCacheConfig::default_in_memory(),
            };

-            let mut file_cache = FileCacheState::new(connstr, config, token.clone())
+            let mut file_cache = FileCacheState::new(connstr, config, token)
                .await
                .context("failed to create file cache")?;

@@ -152,35 +197,15 @@ impl Runner {
            state.filecache = Some(file_cache);
        }

-        if let Some(name) = &args.cgroup {
-            let (mut cgroup, cgroup_event_stream) =
-                CgroupWatcher::new(name.clone(), requesting_send)
-                    .context("failed to create cgroup manager")?;
-
+        if let Some(cgroup) = &state.cgroup {
            let available = mem - file_cache_reserved_bytes;
+            let value = cgroup.config.calculate_memory_high_value(available);
+
+            info!(value, "setting memory.high");

            cgroup
-                .set_memory_limits(available)
-                .context("failed to set cgroup memory limits")?;
-
-            let cgroup = Arc::new(cgroup);
-
-            // Some might call this . . . cgroup v2
-            let cgroup_clone = Arc::clone(&cgroup);
-
-            spawn_with_cancel(token, |_| error!("cgroup watcher terminated"), async move {
-                cgroup_clone.watch(notified_recv, cgroup_event_stream).await
-            });
-
-            state.cgroup = Some(cgroup);
-        } else {
-            // *NOTE*: We need to forget the sender so that its drop impl does not get ran.
-            // This allows us to poll it in `Monitor::run` regardless of whether we
-            // are managing a cgroup or not. If we don't forget it, all receives will
-            // immediately return an error because the sender is droped and it will
-            // claim all select! statements, effectively turning `Monitor::run` into
-            // `loop { fail to receive }`.
-            mem::forget(requesting_send);
+                .set_memory_high_bytes(value)
+                .context("failed to set cgroup memory.high")?;
        }

        Ok(state)
@@ -257,14 +282,11 @@ impl Runner {
                new_cgroup_mem_high = cgroup.config.calculate_memory_high_value(available_memory);
            }

-            let limits = MemoryLimits {
-                // new_cgroup_mem_high is initialized to 0 but it is guarancontextd to not be here
-                // since it is properly initialized in the previous cgroup if let block
-                high: new_cgroup_mem_high,
-            };
+            // new_cgroup_mem_high is initialized to 0 but it is guaranteed to not be here
+            // since it is properly initialized in the previous cgroup if let block
            cgroup
-                .set_limits(&limits)
-                .context("failed to set cgroup memory limits")?;
+                .set_memory_high_bytes(new_cgroup_mem_high)
+                .context("failed to set cgroup memory.high")?;

            let message = format!(
                "set cgroup memory.high to {} MiB, of new max {} MiB",
@@ -327,12 +349,9 @@ impl Runner {
                name = cgroup.path(),
                "updating cgroup memory.high",
            );
-            let limits = MemoryLimits {
-                high: new_cgroup_mem_high,
-            };
            cgroup
-                .set_limits(&limits)
-                .context("failed to set file cache size")?;
+                .set_memory_high_bytes(new_cgroup_mem_high)
+                .context("failed to set cgroup memory.high")?;
        }

        Ok(())
--- a/pageserver/src/bin/pageserver.rs
+++ b/pageserver/src/bin/pageserver.rs
@@ -644,7 +644,7 @@ fn create_remote_storage_client(
    let config = if let Some(config) = &conf.remote_storage_config {
        config
    } else {
-        // No remote storage configured.
+        tracing::warn!("no remote storage configured, this is a deprecated configuration");
        return Ok(None);
    };

--- a/pageserver/src/config.rs
+++ b/pageserver/src/config.rs
@@ -11,6 +11,7 @@ use std::env;
 use storage_broker::Uri;
 use utils::crashsafe::path_with_suffix_extension;
 use utils::id::ConnectionId;
+use utils::logging::SecretString;

 use once_cell::sync::OnceCell;
 use reqwest::Url;
@@ -207,6 +208,9 @@ pub struct PageServerConf {
    pub background_task_maximum_delay: Duration,

    pub control_plane_api: Option<Url>,
+
+    /// JWT token for use with the control plane API.
+    pub control_plane_api_token: Option<SecretString>,
 }

 /// We do not want to store this in a PageServerConf because the latter may be logged
@@ -283,6 +287,7 @@ struct PageServerConfigBuilder {
    background_task_maximum_delay: BuilderValue<Duration>,

    control_plane_api: BuilderValue<Option<Url>>,
+    control_plane_api_token: BuilderValue<Option<SecretString>>,
 }

 impl Default for PageServerConfigBuilder {
@@ -347,6 +352,7 @@ impl Default for PageServerConfigBuilder {
            .unwrap()),

            control_plane_api: Set(None),
+            control_plane_api_token: Set(None),
        }
    }
 }
@@ -567,6 +573,9 @@ impl PageServerConfigBuilder {
            control_plane_api: self
                .control_plane_api
                .ok_or(anyhow!("missing control_plane_api"))?,
+            control_plane_api_token: self
+                .control_plane_api_token
+                .ok_or(anyhow!("missing control_plane_api_token"))?,
        })
    }
 }
@@ -945,6 +954,7 @@ impl PageServerConf {
            ondemand_download_behavior_treat_error_as_warn: false,
            background_task_maximum_delay: Duration::ZERO,
            control_plane_api: None,
+            control_plane_api_token: None,
        }
    }
 }
@@ -1168,7 +1178,8 @@ background_task_maximum_delay = '334 s'
                background_task_maximum_delay: humantime::parse_duration(
                    defaults::DEFAULT_BACKGROUND_TASK_MAXIMUM_DELAY
                )?,
-                control_plane_api: None
+                control_plane_api: None,
+                control_plane_api_token: None
            },
            "Correct defaults should be used when no config values are provided"
        );
@@ -1224,7 +1235,8 @@ background_task_maximum_delay = '334 s'
                test_remote_failures: 0,
                ondemand_download_behavior_treat_error_as_warn: false,
                background_task_maximum_delay: Duration::from_secs(334),
-                control_plane_api: None
+                control_plane_api: None,
+                control_plane_api_token: None
            },
            "Should be able to parse all basic config values correctly"
        );
--- a/pageserver/src/control_plane_client.rs
+++ b/pageserver/src/control_plane_client.rs
@@ -53,12 +53,16 @@ impl ControlPlaneClient {
            segs.pop_if_empty().push("");
        }

-        let client = reqwest::ClientBuilder::new()
-            .build()
-            .expect("Failed to construct http client");
+        let mut client = reqwest::ClientBuilder::new();
+
+        if let Some(jwt) = &conf.control_plane_api_token {
+            let mut headers = hyper::HeaderMap::new();
+            headers.insert("Authorization", jwt.get_contents().parse().unwrap());
+            client = client.default_headers(headers);
+        }

        Some(Self {
-            http_client: client,
+            http_client: client.build().expect("Failed to construct HTTP client"),
            base_url: url,
            node_id: conf.id,
            cancel: cancel.clone(),
--- a/pageserver/src/http/routes.rs
+++ b/pageserver/src/http/routes.rs
@@ -132,7 +132,7 @@ impl From<PageReconstructError> for ApiError {
                ApiError::InternalServerError(anyhow::anyhow!("request was cancelled"))
            }
            PageReconstructError::AncestorStopping(_) => {
-                ApiError::InternalServerError(anyhow::Error::new(pre))
+                ApiError::ResourceUnavailable(format!("{pre}"))
            }
            PageReconstructError::WalRedo(pre) => {
                ApiError::InternalServerError(anyhow::Error::new(pre))
@@ -145,7 +145,7 @@ impl From<TenantMapInsertError> for ApiError {
    fn from(tmie: TenantMapInsertError) -> ApiError {
        match tmie {
            TenantMapInsertError::StillInitializing | TenantMapInsertError::ShuttingDown => {
-                ApiError::InternalServerError(anyhow::Error::new(tmie))
+                ApiError::ResourceUnavailable(format!("{tmie}"))
            }
            TenantMapInsertError::TenantAlreadyExists(id, state) => {
                ApiError::Conflict(format!("tenant {id} already exists, state: {state:?}"))
@@ -159,6 +159,12 @@ impl From<TenantStateError> for ApiError {
    fn from(tse: TenantStateError) -> ApiError {
        match tse {
            TenantStateError::NotFound(tid) => ApiError::NotFound(anyhow!("tenant {}", tid).into()),
+            TenantStateError::NotActive(_) => {
+                ApiError::ResourceUnavailable("Tenant not yet active".into())
+            }
+            TenantStateError::IsStopping(_) => {
+                ApiError::ResourceUnavailable("Tenant is stopping".into())
+            }
            _ => ApiError::InternalServerError(anyhow::Error::new(tse)),
        }
    }
@@ -168,14 +174,17 @@ impl From<GetTenantError> for ApiError {
    fn from(tse: GetTenantError) -> ApiError {
        match tse {
            GetTenantError::NotFound(tid) => ApiError::NotFound(anyhow!("tenant {}", tid).into()),
-            e @ GetTenantError::NotActive(_) => {
+            GetTenantError::Broken(reason) => {
+                ApiError::InternalServerError(anyhow!("tenant is broken: {}", reason))
+            }
+            GetTenantError::NotActive(_) => {
                // Why is this not `ApiError::NotFound`?
                // Because we must be careful to never return 404 for a tenant if it does
                // in fact exist locally. If we did, the caller could draw the conclusion
                // that it can attach the tenant to another PS and we'd be in split-brain.
                //
                // (We can produce this variant only in `mgr::get_tenant(..., active=true)` calls).
-                ApiError::InternalServerError(anyhow::Error::new(e))
+                ApiError::ResourceUnavailable("Tenant not yet active".into())
            }
        }
    }
@@ -622,8 +631,9 @@ async fn tenant_list_handler(
    let response_data = mgr::list_tenants()
        .instrument(info_span!("tenant_list"))
        .await
-        .map_err(anyhow::Error::new)
-        .map_err(ApiError::InternalServerError)?
+        .map_err(|_| {
+            ApiError::ResourceUnavailable("Tenant map is initializing or shutting down".to_string())
+        })?
        .iter()
        .map(|(id, state)| TenantInfo {
            id: *id,
--- a/pageserver/src/metrics.rs
+++ b/pageserver/src/metrics.rs
@@ -264,6 +264,46 @@ pub static PAGE_CACHE_SIZE: Lazy<PageCacheSizeMetrics> = Lazy::new(|| PageCacheS
    },
 });

+pub(crate) static PAGE_CACHE_ACQUIRE_PINNED_SLOT_TIME: Lazy<Histogram> = Lazy::new(|| {
+    register_histogram!(
+        "pageserver_page_cache_acquire_pinned_slot_seconds",
+        "Time spent acquiring a pinned slot in the page cache",
+        CRITICAL_OP_BUCKETS.into(),
+    )
+    .expect("failed to define a metric")
+});
+
+pub(crate) static PAGE_CACHE_FIND_VICTIMS_ITERS_TOTAL: Lazy<IntCounter> = Lazy::new(|| {
+    register_int_counter!(
+        "pageserver_page_cache_find_victim_iters_total",
+        "Counter for the number of iterations in the find_victim loop",
+    )
+    .expect("failed to define a metric")
+});
+
+static PAGE_CACHE_ERRORS: Lazy<IntCounterVec> = Lazy::new(|| {
+    register_int_counter_vec!(
+        "page_cache_errors_total",
+        "Number of timeouts while acquiring a pinned slot in the page cache",
+        &["error_kind"]
+    )
+    .expect("failed to define a metric")
+});
+
+#[derive(IntoStaticStr)]
+#[strum(serialize_all = "kebab_case")]
+pub(crate) enum PageCacheErrorKind {
+    AcquirePinnedSlotTimeout,
+    EvictIterLimit,
+}
+
+pub(crate) fn page_cache_errors_inc(error_kind: PageCacheErrorKind) {
+    PAGE_CACHE_ERRORS
+        .get_metric_with_label_values(&[error_kind.into()])
+        .unwrap()
+        .inc();
+}
+
 pub(crate) static WAIT_LSN_TIME: Lazy<Histogram> = Lazy::new(|| {
    register_histogram!(
        "pageserver_wait_lsn_seconds",
@@ -291,6 +331,14 @@ static RESIDENT_PHYSICAL_SIZE: Lazy<UIntGaugeVec> = Lazy::new(|| {
    .expect("failed to define a metric")
 });

+pub(crate) static RESIDENT_PHYSICAL_SIZE_GLOBAL: Lazy<UIntGauge> = Lazy::new(|| {
+    register_uint_gauge!(
+        "pageserver_resident_physical_size_global",
+        "Like `pageserver_resident_physical_size`, but without tenant/timeline dimensions."
+    )
+    .expect("failed to define a metric")
+});
+
 static REMOTE_PHYSICAL_SIZE: Lazy<UIntGaugeVec> = Lazy::new(|| {
    register_uint_gauge_vec!(
        "pageserver_remote_physical_size",
@@ -301,6 +349,14 @@ static REMOTE_PHYSICAL_SIZE: Lazy<UIntGaugeVec> = Lazy::new(|| {
    .expect("failed to define a metric")
 });

+static REMOTE_PHYSICAL_SIZE_GLOBAL: Lazy<UIntGauge> = Lazy::new(|| {
+    register_uint_gauge!(
+        "pageserver_remote_physical_size_global",
+        "Like `pageserver_remote_physical_size`, but without tenant/timeline dimensions."
+    )
+    .expect("failed to define a metric")
+});
+
 pub(crate) static REMOTE_ONDEMAND_DOWNLOADED_LAYERS: Lazy<IntCounter> = Lazy::new(|| {
    register_int_counter!(
        "pageserver_remote_ondemand_downloaded_layers_total",
@@ -1209,7 +1265,7 @@ pub struct TimelineMetrics {
    pub load_layer_map_histo: StorageTimeMetrics,
    pub garbage_collect_histo: StorageTimeMetrics,
    pub last_record_gauge: IntGauge,
-    pub resident_physical_size_gauge: UIntGauge,
+    resident_physical_size_gauge: UIntGauge,
    /// copy of LayeredTimeline.current_logical_size
    pub current_logical_size_gauge: UIntGauge,
    pub num_persistent_files_created: IntCounter,
@@ -1287,10 +1343,29 @@ impl TimelineMetrics {
    }

    pub fn record_new_file_metrics(&self, sz: u64) {
-        self.resident_physical_size_gauge.add(sz);
+        self.resident_physical_size_add(sz);
        self.num_persistent_files_created.inc_by(1);
        self.persistent_bytes_written.inc_by(sz);
    }
+
+    pub fn resident_physical_size_sub(&self, sz: u64) {
+        self.resident_physical_size_gauge.sub(sz);
+        crate::metrics::RESIDENT_PHYSICAL_SIZE_GLOBAL.sub(sz);
+    }
+
+    pub fn resident_physical_size_add(&self, sz: u64) {
+        self.resident_physical_size_gauge.add(sz);
+        crate::metrics::RESIDENT_PHYSICAL_SIZE_GLOBAL.add(sz);
+    }
+
+    pub fn resident_physical_size_set(&self, sz: u64) {
+        self.resident_physical_size_gauge.set(sz);
+        crate::metrics::RESIDENT_PHYSICAL_SIZE_GLOBAL.set(sz);
+    }
+
+    pub fn resident_physical_size_get(&self) -> u64 {
+        self.resident_physical_size_gauge.get()
+    }
 }

 impl Drop for TimelineMetrics {
@@ -1298,7 +1373,10 @@ impl Drop for TimelineMetrics {
        let tenant_id = &self.tenant_id;
        let timeline_id = &self.timeline_id;
        let _ = LAST_RECORD_LSN.remove_label_values(&[tenant_id, timeline_id]);
-        let _ = RESIDENT_PHYSICAL_SIZE.remove_label_values(&[tenant_id, timeline_id]);
+        {
+            RESIDENT_PHYSICAL_SIZE_GLOBAL.sub(self.resident_physical_size_get());
+            let _ = RESIDENT_PHYSICAL_SIZE.remove_label_values(&[tenant_id, timeline_id]);
+        }
        let _ = CURRENT_LOGICAL_SIZE.remove_label_values(&[tenant_id, timeline_id]);
        let _ = NUM_PERSISTENT_FILES_CREATED.remove_label_values(&[tenant_id, timeline_id]);
        let _ = PERSISTENT_BYTES_WRITTEN.remove_label_values(&[tenant_id, timeline_id]);
@@ -1352,10 +1430,43 @@ use std::time::{Duration, Instant};
 use crate::context::{PageContentKind, RequestContext};
 use crate::task_mgr::TaskKind;

+/// Maintain a per timeline gauge in addition to the global gauge.
+struct PerTimelineRemotePhysicalSizeGauge {
+    last_set: u64,
+    gauge: UIntGauge,
+}
+
+impl PerTimelineRemotePhysicalSizeGauge {
+    fn new(per_timeline_gauge: UIntGauge) -> Self {
+        Self {
+            last_set: per_timeline_gauge.get(),
+            gauge: per_timeline_gauge,
+        }
+    }
+    fn set(&mut self, sz: u64) {
+        self.gauge.set(sz);
+        if sz < self.last_set {
+            REMOTE_PHYSICAL_SIZE_GLOBAL.sub(self.last_set - sz);
+        } else {
+            REMOTE_PHYSICAL_SIZE_GLOBAL.add(sz - self.last_set);
+        };
+        self.last_set = sz;
+    }
+    fn get(&self) -> u64 {
+        self.gauge.get()
+    }
+}
+
+impl Drop for PerTimelineRemotePhysicalSizeGauge {
+    fn drop(&mut self) {
+        REMOTE_PHYSICAL_SIZE_GLOBAL.sub(self.last_set);
+    }
+}
+
 pub struct RemoteTimelineClientMetrics {
    tenant_id: String,
    timeline_id: String,
-    remote_physical_size_gauge: Mutex<Option<UIntGauge>>,
+    remote_physical_size_gauge: Mutex<Option<PerTimelineRemotePhysicalSizeGauge>>,
    calls_unfinished_gauge: Mutex<HashMap<(&'static str, &'static str), IntGauge>>,
    bytes_started_counter: Mutex<HashMap<(&'static str, &'static str), IntCounter>>,
    bytes_finished_counter: Mutex<HashMap<(&'static str, &'static str), IntCounter>>,
@@ -1373,18 +1484,24 @@ impl RemoteTimelineClientMetrics {
        }
    }

-    pub fn remote_physical_size_gauge(&self) -> UIntGauge {
+    pub(crate) fn remote_physical_size_set(&self, sz: u64) {
        let mut guard = self.remote_physical_size_gauge.lock().unwrap();
-        guard
-            .get_or_insert_with(|| {
+        let gauge = guard.get_or_insert_with(|| {
+            PerTimelineRemotePhysicalSizeGauge::new(
                REMOTE_PHYSICAL_SIZE
                    .get_metric_with_label_values(&[
                        &self.tenant_id.to_string(),
                        &self.timeline_id.to_string(),
                    ])
-                    .unwrap()
-            })
-            .clone()
+                    .unwrap(),
+            )
+        });
+        gauge.set(sz);
+    }
+
+    pub(crate) fn remote_physical_size_get(&self) -> u64 {
+        let guard = self.remote_physical_size_gauge.lock().unwrap();
+        guard.as_ref().map(|gauge| gauge.get()).unwrap_or(0)
    }

    pub fn remote_operation_time(
--- a/pageserver/src/page_cache.rs
+++ b/pageserver/src/page_cache.rs
@@ -75,7 +75,11 @@
 use std::{
    collections::{hash_map::Entry, HashMap},
    convert::TryInto,
-    sync::atomic::{AtomicU64, AtomicU8, AtomicUsize, Ordering},
+    sync::{
+        atomic::{AtomicU64, AtomicU8, AtomicUsize, Ordering},
+        Arc, Weak,
+    },
+    time::Duration,
 };

 use anyhow::Context;
@@ -165,6 +169,8 @@ struct Slot {

 struct SlotInner {
    key: Option<CacheKey>,
+    // for `coalesce_readers_permit`
+    permit: std::sync::Mutex<Weak<PinnedSlotsPermit>>,
    buf: &'static mut [u8; PAGE_SZ],
 }

@@ -207,6 +213,22 @@ impl Slot {
    }
 }

+impl SlotInner {
+    /// If there is aready a reader, drop our permit and share its permit, just like we share read access.
+    fn coalesce_readers_permit(&self, permit: PinnedSlotsPermit) -> Arc<PinnedSlotsPermit> {
+        let mut guard = self.permit.lock().unwrap();
+        if let Some(existing_permit) = guard.upgrade() {
+            drop(guard);
+            drop(permit);
+            existing_permit
+        } else {
+            let permit = Arc::new(permit);
+            *guard = Arc::downgrade(&permit);
+            permit
+        }
+    }
+}
+
 pub struct PageCache {
    /// This contains the mapping from the cache key to buffer slot that currently
    /// contains the page, if any.
@@ -224,6 +246,8 @@ pub struct PageCache {
    /// The actual buffers with their metadata.
    slots: Box<[Slot]>,

+    pinned_slots: Arc<tokio::sync::Semaphore>,
+
    /// Index of the next candidate to evict, for the Clock replacement algorithm.
    /// This is interpreted modulo the page cache size.
    next_evict_slot: AtomicUsize,
@@ -231,23 +255,28 @@ pub struct PageCache {
    size_metrics: &'static PageCacheSizeMetrics,
 }

+struct PinnedSlotsPermit(tokio::sync::OwnedSemaphorePermit);
+
 ///
 /// PageReadGuard is a "lease" on a buffer, for reading. The page is kept locked
 /// until the guard is dropped.
 ///
-pub struct PageReadGuard<'i>(tokio::sync::RwLockReadGuard<'i, SlotInner>);
+pub struct PageReadGuard<'i> {
+    _permit: Arc<PinnedSlotsPermit>,
+    slot_guard: tokio::sync::RwLockReadGuard<'i, SlotInner>,
+}

 impl std::ops::Deref for PageReadGuard<'_> {
    type Target = [u8; PAGE_SZ];

    fn deref(&self) -> &Self::Target {
-        self.0.buf
+        self.slot_guard.buf
    }
 }

 impl AsRef<[u8; PAGE_SZ]> for PageReadGuard<'_> {
    fn as_ref(&self) -> &[u8; PAGE_SZ] {
-        self.0.buf
+        self.slot_guard.buf
    }
 }

@@ -264,6 +293,8 @@ impl AsRef<[u8; PAGE_SZ]> for PageReadGuard<'_> {
 pub struct PageWriteGuard<'i> {
    inner: tokio::sync::RwLockWriteGuard<'i, SlotInner>,

+    _permit: PinnedSlotsPermit,
+
    // Are the page contents currently valid?
    // Used to mark pages as invalid that are assigned but not yet filled with data.
    valid: bool,
@@ -348,6 +379,10 @@ impl PageCache {
        lsn: Lsn,
        ctx: &RequestContext,
    ) -> Option<(Lsn, PageReadGuard)> {
+        let Ok(permit) = self.try_get_pinned_slot_permit().await else {
+            return None;
+        };
+
        crate::metrics::PAGE_CACHE
            .for_ctx(ctx)
            .read_accesses_materialized_page
@@ -362,7 +397,10 @@ impl PageCache {
            lsn,
        };

-        if let Some(guard) = self.try_lock_for_read(&mut cache_key).await {
+        if let Some(guard) = self
+            .try_lock_for_read(&mut cache_key, &mut Some(permit))
+            .await
+        {
            if let CacheKey::MaterializedPage {
                hash_key: _,
                lsn: available_lsn,
@@ -445,6 +483,29 @@ impl PageCache {
    // "mappings" after this section. But the routines in this section should
    // not require changes.

+    async fn try_get_pinned_slot_permit(&self) -> anyhow::Result<PinnedSlotsPermit> {
+        let timer = crate::metrics::PAGE_CACHE_ACQUIRE_PINNED_SLOT_TIME.start_timer();
+        match tokio::time::timeout(
+            // Choose small timeout, neon_smgr does its own retries.
+            // https://neondb.slack.com/archives/C04DGM6SMTM/p1694786876476869
+            Duration::from_secs(10),
+            Arc::clone(&self.pinned_slots).acquire_owned(),
+        )
+        .await
+        {
+            Ok(res) => Ok(PinnedSlotsPermit(
+                res.expect("this semaphore is never closed"),
+            )),
+            Err(_timeout) => {
+                timer.stop_and_discard();
+                crate::metrics::page_cache_errors_inc(
+                    crate::metrics::PageCacheErrorKind::AcquirePinnedSlotTimeout,
+                );
+                anyhow::bail!("timeout: there were page guards alive for all page cache slots")
+            }
+        }
+    }
+
    /// Look up a page in the cache.
    ///
    /// If the search criteria is not exact, *cache_key is updated with the key
@@ -454,7 +515,11 @@ impl PageCache {
    ///
    /// If no page is found, returns None and *cache_key is left unmodified.
    ///
-    async fn try_lock_for_read(&self, cache_key: &mut CacheKey) -> Option<PageReadGuard> {
+    async fn try_lock_for_read(
+        &self,
+        cache_key: &mut CacheKey,
+        permit: &mut Option<PinnedSlotsPermit>,
+    ) -> Option<PageReadGuard> {
        let cache_key_orig = cache_key.clone();
        if let Some(slot_idx) = self.search_mapping(cache_key) {
            // The page was found in the mapping. Lock the slot, and re-check
@@ -464,7 +529,10 @@ impl PageCache {
            let inner = slot.inner.read().await;
            if inner.key.as_ref() == Some(cache_key) {
                slot.inc_usage_count();
-                return Some(PageReadGuard(inner));
+                return Some(PageReadGuard {
+                    _permit: inner.coalesce_readers_permit(permit.take().unwrap()),
+                    slot_guard: inner,
+                });
            } else {
                // search_mapping might have modified the search key; restore it.
                *cache_key = cache_key_orig;
@@ -507,6 +575,8 @@ impl PageCache {
        cache_key: &mut CacheKey,
        ctx: &RequestContext,
    ) -> anyhow::Result<ReadBufResult> {
+        let mut permit = Some(self.try_get_pinned_slot_permit().await?);
+
        let (read_access, hit) = match cache_key {
            CacheKey::MaterializedPage { .. } => {
                unreachable!("Materialized pages use lookup_materialized_page")
@@ -523,17 +593,21 @@ impl PageCache {
        let mut is_first_iteration = true;
        loop {
            // First check if the key already exists in the cache.
-            if let Some(read_guard) = self.try_lock_for_read(cache_key).await {
+            if let Some(read_guard) = self.try_lock_for_read(cache_key, &mut permit).await {
+                debug_assert!(permit.is_none());
                if is_first_iteration {
                    hit.inc();
                }
                return Ok(ReadBufResult::Found(read_guard));
            }
+            debug_assert!(permit.is_some());
            is_first_iteration = false;

            // Not found. Find a victim buffer
-            let (slot_idx, mut inner) =
-                self.find_victim().context("Failed to find evict victim")?;
+            let (slot_idx, mut inner) = self
+                .find_victim(permit.as_ref().unwrap())
+                .await
+                .context("Failed to find evict victim")?;

            // Insert mapping for this. At this point, we may find that another
            // thread did the same thing concurrently. In that case, we evicted
@@ -555,7 +629,16 @@ impl PageCache {
            inner.key = Some(cache_key.clone());
            slot.set_usage_count(1);

+            debug_assert!(
+                {
+                    let guard = inner.permit.lock().unwrap();
+                    guard.upgrade().is_none()
+                },
+                "we hold a write lock, so, no one else should have a permit"
+            );
+
            return Ok(ReadBufResult::NotFound(PageWriteGuard {
+                _permit: permit.take().unwrap(),
                inner,
                valid: false,
            }));
@@ -566,7 +649,11 @@ impl PageCache {
    /// found, returns None.
    ///
    /// When locking a page for writing, the search criteria is always "exact".
-    async fn try_lock_for_write(&self, cache_key: &CacheKey) -> Option<PageWriteGuard> {
+    async fn try_lock_for_write(
+        &self,
+        cache_key: &CacheKey,
+        permit: &mut Option<PinnedSlotsPermit>,
+    ) -> Option<PageWriteGuard> {
        if let Some(slot_idx) = self.search_mapping_for_write(cache_key) {
            // The page was found in the mapping. Lock the slot, and re-check
            // that it's still what we expected (because we don't released the mapping
@@ -575,7 +662,18 @@ impl PageCache {
            let inner = slot.inner.write().await;
            if inner.key.as_ref() == Some(cache_key) {
                slot.inc_usage_count();
-                return Some(PageWriteGuard { inner, valid: true });
+                debug_assert!(
+                    {
+                        let guard = inner.permit.lock().unwrap();
+                        guard.upgrade().is_none()
+                    },
+                    "we hold a write lock, so, no one else should have a permit"
+                );
+                return Some(PageWriteGuard {
+                    _permit: permit.take().unwrap(),
+                    inner,
+                    valid: true,
+                });
            }
        }
        None
@@ -586,15 +684,20 @@ impl PageCache {
    /// Similar to lock_for_read(), but the returned buffer is write-locked and
    /// may be modified by the caller even if it's already found in the cache.
    async fn lock_for_write(&self, cache_key: &CacheKey) -> anyhow::Result<WriteBufResult> {
+        let mut permit = Some(self.try_get_pinned_slot_permit().await?);
        loop {
            // First check if the key already exists in the cache.
-            if let Some(write_guard) = self.try_lock_for_write(cache_key).await {
+            if let Some(write_guard) = self.try_lock_for_write(cache_key, &mut permit).await {
+                debug_assert!(permit.is_none());
                return Ok(WriteBufResult::Found(write_guard));
            }
+            debug_assert!(permit.is_some());

            // Not found. Find a victim buffer
-            let (slot_idx, mut inner) =
-                self.find_victim().context("Failed to find evict victim")?;
+            let (slot_idx, mut inner) = self
+                .find_victim(permit.as_ref().unwrap())
+                .await
+                .context("Failed to find evict victim")?;

            // Insert mapping for this. At this point, we may find that another
            // thread did the same thing concurrently. In that case, we evicted
@@ -616,7 +719,16 @@ impl PageCache {
            inner.key = Some(cache_key.clone());
            slot.set_usage_count(1);

+            debug_assert!(
+                {
+                    let guard = inner.permit.lock().unwrap();
+                    guard.upgrade().is_none()
+                },
+                "we hold a write lock, so, no one else should have a permit"
+            );
+
            return Ok(WriteBufResult::NotFound(PageWriteGuard {
+                _permit: permit.take().unwrap(),
                inner,
                valid: false,
            }));
@@ -769,7 +881,10 @@ impl PageCache {
    /// Find a slot to evict.
    ///
    /// On return, the slot is empty and write-locked.
-    fn find_victim(&self) -> anyhow::Result<(usize, tokio::sync::RwLockWriteGuard<SlotInner>)> {
+    async fn find_victim(
+        &self,
+        _permit_witness: &PinnedSlotsPermit,
+    ) -> anyhow::Result<(usize, tokio::sync::RwLockWriteGuard<SlotInner>)> {
        let iter_limit = self.slots.len() * 10;
        let mut iters = 0;
        loop {
@@ -782,13 +897,40 @@ impl PageCache {
                let mut inner = match slot.inner.try_write() {
                    Ok(inner) => inner,
                    Err(_err) => {
-                        // If we have looped through the whole buffer pool 10 times
-                        // and still haven't found a victim buffer, something's wrong.
-                        // Maybe all the buffers were in locked. That could happen in
-                        // theory, if you have more threads holding buffers locked than
-                        // there are buffers in the pool. In practice, with a reasonably
-                        // large buffer pool it really shouldn't happen.
                        if iters > iter_limit {
+                            // NB: Even with the permits, there's no hard guarantee that we will find a slot with
+                            // any particular number of iterations: other threads might race ahead and acquire and
+                            // release pins just as we're scanning the array.
+                            //
+                            // Imagine that nslots is 2, and as starting point, usage_count==1 on all
+                            // slots. There are two threads running concurrently, A and B. A has just
+                            // acquired the permit from the semaphore.
+                            //
+                            //   A: Look at slot 1. Its usage_count == 1, so decrement it to zero, and continue the search
+                            //   B: Acquire permit.
+                            //   B: Look at slot 2, decrement its usage_count to zero and continue the search
+                            //   B: Look at slot 1. Its usage_count is zero, so pin it and bump up its usage_count to 1.
+                            //   B: Release pin and permit again
+                            //   B: Acquire permit.
+                            //   B: Look at slot 2. Its usage_count is zero, so pin it and bump up its usage_count to 1.
+                            //   B: Release pin and permit again
+                            //
+                            // Now we're back in the starting situation that both slots have
+                            // usage_count 1, but A has now been through one iteration of the
+                            // find_victim() loop. This can repeat indefinitely and on each
+                            // iteration, A's iteration count increases by one.
+                            //
+                            // So, even though the semaphore for the permits is fair, the victim search
+                            // itself happens in parallel and is not fair.
+                            // Hence even with a permit, a task can theoretically be starved.
+                            // To avoid this, we'd need tokio to give priority to tasks that are holding
+                            // permits for longer.
+                            // Note that just yielding to tokio during iteration without such
+                            // priority boosting is likely counter-productive. We'd just give more opportunities
+                            // for B to bump usage count, further starving A.
+                            crate::metrics::page_cache_errors_inc(
+                                crate::metrics::PageCacheErrorKind::EvictIterLimit,
+                            );
                            anyhow::bail!("exceeded evict iter limit");
                        }
                        continue;
@@ -799,6 +941,7 @@ impl PageCache {
                    self.remove_mapping(old_key);
                    inner.key = None;
                }
+                crate::metrics::PAGE_CACHE_FIND_VICTIMS_ITERS_TOTAL.inc_by(iters as u64);
                return Ok((slot_idx, inner));
            }
        }
@@ -826,7 +969,11 @@ impl PageCache {
                let buf: &mut [u8; PAGE_SZ] = chunk.try_into().unwrap();

                Slot {
-                    inner: tokio::sync::RwLock::new(SlotInner { key: None, buf }),
+                    inner: tokio::sync::RwLock::new(SlotInner {
+                        key: None,
+                        buf,
+                        permit: std::sync::Mutex::new(Weak::new()),
+                    }),
                    usage_count: AtomicU8::new(0),
                }
            })
@@ -838,6 +985,7 @@ impl PageCache {
            slots,
            next_evict_slot: AtomicUsize::new(0),
            size_metrics,
+            pinned_slots: Arc::new(tokio::sync::Semaphore::new(num_pages)),
        }
    }
 }
--- a/pageserver/src/page_service.rs
+++ b/pageserver/src/page_service.rs
@@ -1265,7 +1265,10 @@ async fn get_active_tenant_with_timeout(
        Ok(tenant) => tenant,
        Err(e @ GetTenantError::NotFound(_)) => return Err(GetActiveTenantError::NotFound(e)),
        Err(GetTenantError::NotActive(_)) => {
-            unreachable!("we're calling get_tenant with active=false")
+            unreachable!("we're calling get_tenant with active_only=false")
+        }
+        Err(GetTenantError::Broken(_)) => {
+            unreachable!("we're calling get_tenant with active_only=false")
        }
    };
    let wait_time = Duration::from_secs(30);
--- a/pageserver/src/tenant/mgr.rs
+++ b/pageserver/src/tenant/mgr.rs
@@ -130,10 +130,15 @@ pub async fn init_tenant_mgr(
        // deletion list entries may still be valid.  We provide that by pushing a recovery operation into
        // the queue. Sequential processing of te queue ensures that recovery is done before any new tenant deletions
        // are processed, even though we don't block on recovery completing here.
-        resources
-            .deletion_queue_client
-            .recover(result.clone())
-            .await?;
+        //
+        // Must only do this if remote storage is enabled, otherwise deletion queue
+        // is not running and channel push will fail.
+        if resources.remote_storage.is_some() {
+            resources
+                .deletion_queue_client
+                .recover(result.clone())
+                .await?;
+        }

        Some(result)
    } else {
@@ -505,6 +510,11 @@ pub enum GetTenantError {
    NotFound(TenantId),
    #[error("Tenant {0} is not active")]
    NotActive(TenantId),
+    /// Broken is logically a subset of NotActive, but a distinct error is useful as
+    /// NotActive is usually a retryable state for API purposes, whereas Broken
+    /// is a stuck error state
+    #[error("Tenant is broken: {0}")]
+    Broken(String),
 }

 /// Gets the tenant from the in-memory data, erroring if it's absent or is not fitting to the query.
@@ -519,10 +529,20 @@ pub async fn get_tenant(
    let tenant = m
        .get(&tenant_id)
        .ok_or(GetTenantError::NotFound(tenant_id))?;
-    if active_only && !tenant.is_active() {
-        Err(GetTenantError::NotActive(tenant_id))
-    } else {
-        Ok(Arc::clone(tenant))
+
+    match tenant.current_state() {
+        TenantState::Broken {
+            reason,
+            backtrace: _,
+        } if active_only => Err(GetTenantError::Broken(reason)),
+        TenantState::Active => Ok(Arc::clone(tenant)),
+        _ => {
+            if active_only {
+                Err(GetTenantError::NotActive(tenant_id))
+            } else {
+                Ok(Arc::clone(tenant))
+            }
+        }
    }
 }

--- a/pageserver/src/tenant/remote_timeline_client.rs
+++ b/pageserver/src/tenant/remote_timeline_client.rs
@@ -453,11 +453,11 @@ impl RemoteTimelineClient {
        } else {
            0
        };
-        self.metrics.remote_physical_size_gauge().set(size);
+        self.metrics.remote_physical_size_set(size);
    }

    pub fn get_remote_physical_size(&self) -> u64 {
-        self.metrics.remote_physical_size_gauge().get()
+        self.metrics.remote_physical_size_get()
    }

    //
@@ -1161,11 +1161,7 @@ impl RemoteTimelineClient {
                    // at info level at first, and only WARN if the operation fails repeatedly.
                    //
                    // (See similar logic for downloads in `download::download_retry`)
-
-                    let is_simulated = cfg!(feature = "testing")
-                        && e.root_cause().is::<remote_storage::SimulatedError>();
-
-                    if retries < FAILED_UPLOAD_WARN_THRESHOLD || is_simulated {
+                    if retries < FAILED_UPLOAD_WARN_THRESHOLD {
                        info!(
                            "failed to perform remote task {}, will retry (attempt {}): {:#}",
                            task.op, retries, e
--- a/pageserver/src/tenant/timeline.rs
+++ b/pageserver/src/tenant/timeline.rs
@@ -496,13 +496,36 @@ impl Timeline {
        };

        let timer = crate::metrics::GET_RECONSTRUCT_DATA_TIME.start_timer();
-        self.get_reconstruct_data(key, lsn, &mut reconstruct_state, ctx)
+        let path = self
+            .get_reconstruct_data(key, lsn, &mut reconstruct_state, ctx)
            .await?;
        timer.stop_and_record();

-        RECONSTRUCT_TIME
-            .observe_closure_duration(|| self.reconstruct_value(key, lsn, reconstruct_state))
-            .await
+        let timer = RECONSTRUCT_TIME.start_timer();
+        let res = self.reconstruct_value(key, lsn, reconstruct_state).await;
+        timer.stop_and_record();
+
+        if cfg!(feature = "testing") && res.is_err() {
+            // it can only be walredo issue
+            use std::fmt::Write;
+
+            let mut msg = String::new();
+
+            path.into_iter().for_each(|(res, cont_lsn, layer)| {
+                writeln!(
+                    msg,
+                    "- layer traversal: result {res:?}, cont_lsn {cont_lsn}, layer: {}",
+                    layer(),
+                )
+                .expect("string grows")
+            });
+
+            // this is to rule out or provide evidence that we could in some cases read a duplicate
+            // walrecord
+            tracing::info!("walredo failed, path:\n{msg}");
+        }
+
+        res
    }

    /// Get last or prev record separately. Same as get_last_record_rlsn().last/prev.
@@ -559,7 +582,7 @@ impl Timeline {
    }

    pub fn resident_physical_size(&self) -> u64 {
-        self.metrics.resident_physical_size_gauge.get()
+        self.metrics.resident_physical_size_get()
    }

    ///
@@ -1309,10 +1332,7 @@ impl Timeline {
        // will treat the file as a local layer again, count it towards resident size,
        // and it'll be like the layer removal never happened.
        // The bump in resident size is perhaps unexpected but overall a robust behavior.
-        self.metrics
-            .resident_physical_size_gauge
-            .sub(layer_file_size);
-
+        self.metrics.resident_physical_size_sub(layer_file_size);
        self.metrics.evictions.inc();

        if let Some(delta) = local_layer_residence_duration {
@@ -1846,9 +1866,7 @@ impl Timeline {
            "loaded layer map with {} layers at {}, total physical size: {}",
            num_layers, disk_consistent_lsn, total_physical_size
        );
-        self.metrics
-            .resident_physical_size_gauge
-            .set(total_physical_size);
+        self.metrics.resident_physical_size_set(total_physical_size);

        timer.stop_and_record();
        Ok(())
@@ -2229,7 +2247,7 @@ impl Timeline {
        request_lsn: Lsn,
        reconstruct_state: &mut ValueReconstructState,
        ctx: &RequestContext,
-    ) -> Result<(), PageReconstructError> {
+    ) -> Result<Vec<TraversalPathItem>, PageReconstructError> {
        // Start from the current timeline.
        let mut timeline_owned;
        let mut timeline = self;
@@ -2260,12 +2278,12 @@ impl Timeline {
            // The function should have updated 'state'
            //info!("CALLED for {} at {}: {:?} with {} records, cached {}", key, cont_lsn, result, reconstruct_state.records.len(), cached_lsn);
            match result {
-                ValueReconstructResult::Complete => return Ok(()),
+                ValueReconstructResult::Complete => return Ok(traversal_path),
                ValueReconstructResult::Continue => {
                    // If we reached an earlier cached page image, we're done.
                    if cont_lsn == cached_lsn + 1 {
                        MATERIALIZED_PAGE_CACHE_HIT.inc_by(1);
-                        return Ok(());
+                        return Ok(traversal_path);
                    }
                    if prev_lsn <= cont_lsn {
                        // Didn't make any progress in last iteration. Error out to avoid
@@ -4398,7 +4416,7 @@ impl Timeline {

                    // XXX the temp file is still around in Err() case
                    // and consumes space until we clean up upon pageserver restart.
-                    self_clone.metrics.resident_physical_size_gauge.add(*size);
+                    self_clone.metrics.resident_physical_size_add(*size);

                    // Download complete. Replace the RemoteLayer with the corresponding
                    // Delta- or ImageLayer in the layer map.
--- a/pageserver/src/tenant/timeline/layer_manager.rs
+++ b/pageserver/src/tenant/timeline/layer_manager.rs
@@ -263,7 +263,7 @@ impl LayerManager {
        let desc = layer.layer_desc();
        if !layer.is_remote_layer() {
            layer.delete_resident_layer_file()?;
-            metrics.resident_physical_size_gauge.sub(desc.file_size);
+            metrics.resident_physical_size_sub(desc.file_size);
        }

        // TODO Removing from the bottom of the layer map is expensive.
--- a/pageserver/src/virtual_file.rs
+++ b/pageserver/src/virtual_file.rs
@@ -177,19 +177,19 @@ impl OpenFiles {
 pub enum CrashsafeOverwriteError {
    #[error("final path has no parent dir")]
    FinalPathHasNoParentDir,
-    #[error("remove tempfile: {0}")]
+    #[error("remove tempfile")]
    RemovePreviousTempfile(#[source] std::io::Error),
-    #[error("create tempfile: {0}")]
+    #[error("create tempfile")]
    CreateTempfile(#[source] std::io::Error),
-    #[error("write tempfile: {0}")]
+    #[error("write tempfile")]
    WriteContents(#[source] std::io::Error),
-    #[error("sync tempfile: {0}")]
+    #[error("sync tempfile")]
    SyncTempfile(#[source] std::io::Error),
-    #[error("rename tempfile to final path: {0}")]
+    #[error("rename tempfile to final path")]
    RenameTempfileToFinalPath(#[source] std::io::Error),
-    #[error("open final path parent dir: {0}")]
+    #[error("open final path parent dir")]
    OpenFinalPathParentDir(#[source] std::io::Error),
-    #[error("sync final path parent dir: {0}")]
+    #[error("sync final path parent dir")]
    SyncFinalPathParentDir(#[source] std::io::Error),
 }
 impl CrashsafeOverwriteError {
--- a/pageserver/src/walredo.rs
+++ b/pageserver/src/walredo.rs
@@ -38,6 +38,9 @@ use tracing::*;
 use utils::crashsafe::path_with_suffix_extension;
 use utils::{bin_ser::BeSer, id::TenantId, lsn::Lsn, nonblock::set_nonblock};

+#[cfg(feature = "testing")]
+use std::sync::atomic::{AtomicUsize, Ordering};
+
 use crate::metrics::{
    WAL_REDO_BYTES_HISTOGRAM, WAL_REDO_RECORDS_HISTOGRAM, WAL_REDO_RECORD_COUNTER, WAL_REDO_TIME,
    WAL_REDO_WAIT_TIME,
@@ -113,6 +116,9 @@ struct ProcessOutput {
 pub struct PostgresRedoManager {
    tenant_id: TenantId,
    conf: &'static PageServerConf,
+    /// Counter to separate same sized walredo inputs failing at the same millisecond.
+    #[cfg(feature = "testing")]
+    dump_sequence: AtomicUsize,

    stdout: Mutex<Option<ProcessOutput>>,
    stdin: Mutex<Option<ProcessInput>>,
@@ -224,6 +230,8 @@ impl PostgresRedoManager {
        PostgresRedoManager {
            tenant_id,
            conf,
+            #[cfg(feature = "testing")]
+            dump_sequence: AtomicUsize::default(),
            stdin: Mutex::new(None),
            stdout: Mutex::new(None),
            stderr: Mutex::new(None),
@@ -290,25 +298,25 @@ impl PostgresRedoManager {
            WAL_REDO_BYTES_HISTOGRAM.observe(nbytes as f64);

            debug!(
-				"postgres applied {} WAL records ({} bytes) in {} us to reconstruct page image at LSN {}",
-				len,
-				nbytes,
-				duration.as_micros(),
-				lsn
-			);
+                "postgres applied {} WAL records ({} bytes) in {} us to reconstruct page image at LSN {}",
+                len,
+                nbytes,
+                duration.as_micros(),
+                lsn
+            );

            // If something went wrong, don't try to reuse the process. Kill it, and
            // next request will launch a new one.
            if result.is_err() {
                error!(
-                "error applying {} WAL records {}..{} ({} bytes) to base image with LSN {} to reconstruct page image at LSN {}",
-                records.len(),
-				records.first().map(|p| p.0).unwrap_or(Lsn(0)),
-				records.last().map(|p| p.0).unwrap_or(Lsn(0)),
-                nbytes,
-				base_img_lsn,
-                lsn
-            );
+                    "error applying {} WAL records {}..{} ({} bytes) to base image with LSN {} to reconstruct page image at LSN {}",
+                    records.len(),
+                    records.first().map(|p| p.0).unwrap_or(Lsn(0)),
+                    records.last().map(|p| p.0).unwrap_or(Lsn(0)),
+                    nbytes,
+                    base_img_lsn,
+                    lsn
+                );
                // self.stdin only holds stdin & stderr as_raw_fd().
                // Dropping it as part of take() doesn't close them.
                // The owning objects (ChildStdout and ChildStderr) are stored in
@@ -742,7 +750,7 @@ impl PostgresRedoManager {
    #[instrument(skip_all, fields(tenant_id=%self.tenant_id, pid=%input.as_ref().unwrap().child.id()))]
    fn apply_wal_records(
        &self,
-        mut input: MutexGuard<Option<ProcessInput>>,
+        input: MutexGuard<Option<ProcessInput>>,
        tag: BufferTag,
        base_img: &Option<Bytes>,
        records: &[(Lsn, NeonWalRecord)],
@@ -779,6 +787,23 @@ impl PostgresRedoManager {
        build_get_page_msg(tag, &mut writebuf);
        WAL_REDO_RECORD_COUNTER.inc_by(records.len() as u64);

+        let res = self.apply_wal_records0(&writebuf, input, wal_redo_timeout);
+
+        if res.is_err() {
+            // not all of these can be caused by this particular input, however these are so rare
+            // in tests so capture all.
+            self.record_and_log(&writebuf);
+        }
+
+        res
+    }
+
+    fn apply_wal_records0(
+        &self,
+        writebuf: &[u8],
+        mut input: MutexGuard<Option<ProcessInput>>,
+        wal_redo_timeout: Duration,
+    ) -> Result<Bytes, std::io::Error> {
        let proc = input.as_mut().unwrap();
        let mut nwrite = 0usize;
        let stdout_fd = proc.stdout_fd;
@@ -984,6 +1009,38 @@ impl PostgresRedoManager {
        }
        Ok(res)
    }
+
+    #[cfg(feature = "testing")]
+    fn record_and_log(&self, writebuf: &[u8]) {
+        let millis = std::time::SystemTime::now()
+            .duration_since(std::time::SystemTime::UNIX_EPOCH)
+            .unwrap()
+            .as_millis();
+
+        let seq = self.dump_sequence.fetch_add(1, Ordering::Relaxed);
+
+        // these files will be collected to an allure report
+        let filename = format!("walredo-{millis}-{}-{seq}.walredo", writebuf.len());
+
+        let path = self.conf.tenant_path(&self.tenant_id).join(&filename);
+
+        let res = std::fs::OpenOptions::new()
+            .write(true)
+            .create_new(true)
+            .read(true)
+            .open(path)
+            .and_then(|mut f| f.write_all(writebuf));
+
+        // trip up allowed_errors
+        if let Err(e) = res {
+            tracing::error!(target=%filename, length=writebuf.len(), "failed to write out the walredo errored input: {e}");
+        } else {
+            tracing::error!(filename, "erroring walredo input saved");
+        }
+    }
+
+    #[cfg(not(feature = "testing"))]
+    fn record_and_log(&self, _: &[u8]) {}
 }

 /// Wrapper type around `std::process::Child` which guarantees that the child
--- a/pgxn/neon/control_plane_connector.c
+++ b/pgxn/neon/control_plane_connector.c
@@ -741,6 +741,13 @@ NeonProcessUtility(
 			break;
 		case T_DropdbStmt:
 			HandleDropDb(castNode(DropdbStmt, parseTree));
+			/*
+			 * We do this here to hack around the fact that Postgres performs the drop
+			 * INSIDE of standard_ProcessUtility, which means that if we try to
+			 * abort the drop normally it'll be too late. DROP DATABASE can't be inside
+			 * of a transaction block anyway, so this should be fine to do.
+			 */
+			NeonXactCallback(XACT_EVENT_PRE_COMMIT, NULL);
 			break;
 		case T_CreateRoleStmt:
 			HandleCreateRole(castNode(CreateRoleStmt, parseTree));
--- a/pgxn/neon/file_cache.c
+++ b/pgxn/neon/file_cache.c
@@ -14,7 +14,6 @@
 */

 #include <sys/file.h>
-#include <sys/statvfs.h>
 #include <unistd.h>
 #include <fcntl.h>

@@ -38,9 +37,6 @@
 #include "storage/fd.h"
 #include "storage/pg_shmem.h"
 #include "storage/buf_internals.h"
-#include "storage/procsignal.h"
-#include "postmaster/bgworker.h"
-#include "postmaster/interrupt.h"

 /*
 * Local file cache is used to temporary store relations pages in local file system.
@@ -66,9 +62,6 @@

 #define SIZE_MB_TO_CHUNKS(size) ((uint32)((size) * MB / BLCKSZ / BLOCKS_PER_CHUNK))

-#define MAX_MONITOR_INTERVAL_USEC 1000000 /* 1 second */
-#define MAX_DISK_WRITE_RATE       1000 /* MB/sec */
-
 typedef struct FileCacheEntry
 {
 	BufferTag	key;
@@ -91,14 +84,12 @@ static int   lfc_desc = 0;
 static LWLockId lfc_lock;
 static int   lfc_max_size;
 static int   lfc_size_limit;
-static int   lfc_free_space_watermark;
 static char* lfc_path;
 static  FileCacheControl* lfc_ctl;
 static shmem_startup_hook_type prev_shmem_startup_hook;
 #if PG_VERSION_NUM>=150000
 static shmem_request_hook_type prev_shmem_request_hook;
 #endif
-static int   lfc_shrinking_factor; /* power of two by which local cache size will be shrinked when lfc_free_space_watermark is reached */

 void FileCacheMonitorMain(Datum main_arg);

@@ -254,80 +245,6 @@ lfc_change_limit_hook(int newval, void *extra)
 	LWLockRelease(lfc_lock);
 }

-/*
- * Local file system state monitor check available free space.
- * If it is lower than lfc_free_space_watermark then we shrink size of local cache
- * but throwing away least recently accessed chunks.
- * First time low space watermark is reached cache size is divided by two,
- * second time by four,... Finally we remove all chunks from local cache.
- *
- * Please notice that we are not changing lfc_cache_size: it is used to be adjusted by autoscaler.
- * We only throw away cached chunks but do not prevent from filling cache by new chunks.
- *
- * Interval of poooling cache state is calculated as minimal time needed to consume lfc_free_space_watermark
- * disk space with maximal possible disk write speed (1Gb/sec). But not larger than 1 second.
- * Calling statvfs each second should not add any noticeable overhead.
- */
-void
-FileCacheMonitorMain(Datum main_arg)
-{
-	/*
-	 * Choose file system state monitor interval so that space can not be exosted
-	 * during this period but not longer than  MAX_MONITOR_INTERVAL (10 sec)
-	 */
-	uint64 monitor_interval = Min(MAX_MONITOR_INTERVAL_USEC, lfc_free_space_watermark*MB/MAX_DISK_WRITE_RATE);
-
-	/* Establish signal handlers. */
-	pqsignal(SIGUSR1, procsignal_sigusr1_handler);
-	pqsignal(SIGHUP, SignalHandlerForConfigReload);
-	pqsignal(SIGTERM, SignalHandlerForShutdownRequest);
-	BackgroundWorkerUnblockSignals();
-
-	/* Periodically dump buffers until terminated. */
-	while (!ShutdownRequestPending)
-	{
-		if (lfc_size_limit != 0)
-		{
-			struct statvfs sfs;
-			if (statvfs(lfc_path, &sfs) < 0)
-			{
-				elog(WARNING, "Failed to obtain status of %s: %m", lfc_path);
-			}
-			else
-			{
-				if (sfs.f_bavail*sfs.f_bsize < lfc_free_space_watermark*MB)
-				{
-					if (lfc_shrinking_factor < 31) {
-						lfc_shrinking_factor += 1;
-					}
-					lfc_change_limit_hook(lfc_size_limit >> lfc_shrinking_factor, NULL);
-				}
-				else
-					lfc_shrinking_factor = 0; /* reset to initial value */
-			}
-		}
-		pg_usleep(monitor_interval);
-	}
-}
-
-static void
-lfc_register_free_space_monitor(void)
-{
-	BackgroundWorker bgw;
-	memset(&bgw, 0, sizeof(bgw));
-	bgw.bgw_flags = BGWORKER_SHMEM_ACCESS;
-	bgw.bgw_start_time = BgWorkerStart_RecoveryFinished;
-	snprintf(bgw.bgw_library_name, BGW_MAXLEN, "neon");
-	snprintf(bgw.bgw_function_name, BGW_MAXLEN, "FileCacheMonitorMain");
-	snprintf(bgw.bgw_name, BGW_MAXLEN, "Local free space monitor");
-	snprintf(bgw.bgw_type, BGW_MAXLEN, "Local free space monitor");
-	bgw.bgw_restart_time = 5;
-	bgw.bgw_notify_pid = 0;
-	bgw.bgw_main_arg = (Datum) 0;
-
-	RegisterBackgroundWorker(&bgw);
-}
-
 void
 lfc_init(void)
 {
@@ -364,19 +281,6 @@ lfc_init(void)
 							lfc_change_limit_hook,
 							NULL);

-	DefineCustomIntVariable("neon.free_space_watermark",
-							"Minimal free space in local file system after reaching which local file cache will be truncated",
-							NULL,
-							&lfc_free_space_watermark,
-							1024, /* 1GB */
-							0,
-							INT_MAX,
-							PGC_SIGHUP,
-							GUC_UNIT_MB,
-							NULL,
-							NULL,
-							NULL);
-
 	DefineCustomStringVariable("neon.file_cache_path",
 							   "Path to local file cache (can be raw device)",
 							   NULL,
@@ -391,9 +295,6 @@ lfc_init(void)
 	if (lfc_max_size == 0)
 		return;

-	if (lfc_free_space_watermark != 0)
-		lfc_register_free_space_monitor();
-
 	prev_shmem_startup_hook = shmem_startup_hook;
 	shmem_startup_hook = lfc_shmem_startup;
 #if PG_VERSION_NUM>=150000
--- a/pgxn/neon/pagestore_smgr.c
+++ b/pgxn/neon/pagestore_smgr.c
@@ -1790,6 +1790,14 @@ neon_zeroextend(SMgrRelation reln, ForkNumber forkNum, BlockNumber blocknum,
 	if (!XLogInsertAllowed())
 		return;

+	/* ensure we have enough xlog buffers to log max-sized records */
+	XLogEnsureRecordSpace(Min(remblocks, (XLR_MAX_BLOCK_ID - 1)), 0);
+
+	/*
+	 * Iterate over all the pages. They are collected into batches of
+	 * XLR_MAX_BLOCK_ID pages, and a single WAL-record is written for each
+	 * batch.
+	 */
 	while (remblocks > 0)
 	{
 		int			count = Min(remblocks, XLR_MAX_BLOCK_ID);
--- a/poetry.lock
+++ b/poetry.lock
@@ -2415,18 +2415,18 @@ files = [

 [[package]]
 name = "urllib3"
-version = "1.26.11"
+version = "1.26.17"
 description = "HTTP library with thread-safe connection pooling, file post, and more."
 optional = false
-python-versions = ">=2.7, !=3.0.*, !=3.1.*, !=3.2.*, !=3.3.*, !=3.4.*, !=3.5.*, <4"
+python-versions = ">=2.7, !=3.0.*, !=3.1.*, !=3.2.*, !=3.3.*, !=3.4.*, !=3.5.*"
 files = [
-    {file = "urllib3-1.26.11-py2.py3-none-any.whl", hash = "sha256:c33ccba33c819596124764c23a97d25f32b28433ba0dedeb77d873a38722c9bc"},
-    {file = "urllib3-1.26.11.tar.gz", hash = "sha256:ea6e8fb210b19d950fab93b60c9009226c63a28808bc8386e05301e25883ac0a"},
+    {file = "urllib3-1.26.17-py2.py3-none-any.whl", hash = "sha256:94a757d178c9be92ef5539b8840d48dc9cf1b2709c9d6b588232a055c524458b"},
+    {file = "urllib3-1.26.17.tar.gz", hash = "sha256:24d6a242c28d29af46c3fae832c36db3bbebcc533dd1bb549172cd739c82df21"},
 ]

 [package.extras]
-brotli = ["brotli (>=1.0.9)", "brotlicffi (>=0.8.0)", "brotlipy (>=0.6.0)"]
-secure = ["certifi", "cryptography (>=1.3.4)", "idna (>=2.0.0)", "ipaddress", "pyOpenSSL (>=0.14)"]
+brotli = ["brotli (==1.0.9)", "brotli (>=1.0.9)", "brotlicffi (>=0.8.0)", "brotlipy (>=0.6.0)"]
+secure = ["certifi", "cryptography (>=1.3.4)", "idna (>=2.0.0)", "ipaddress", "pyOpenSSL (>=0.14)", "urllib3-secure-extra"]
 socks = ["PySocks (>=1.5.6,!=1.5.7,<2.0)"]

 [[package]]
--- a/proxy/Cargo.toml
+++ b/proxy/Cargo.toml
@@ -42,6 +42,7 @@ reqwest-middleware.workspace = true
 reqwest-retry.workspace = true
 reqwest-tracing.workspace = true
 routerify.workspace = true
+rustc-hash.workspace = true
 rustls-pemfile.workspace = true
 rustls.workspace = true
 scopeguard.workspace = true
--- a/proxy/src/auth/backend.rs
+++ b/proxy/src/auth/backend.rs
@@ -160,6 +160,19 @@ impl BackendType<'_, ClientCredentials<'_>> {
            Test(_) => Some("test".to_owned()),
        }
    }
+
+    /// Get username from the credentials.
+    pub fn get_user(&self) -> &str {
+        use BackendType::*;
+
+        match self {
+            Console(_, creds) => creds.user,
+            Postgres(_, creds) => creds.user,
+            Link(_) => "link",
+            Test(_) => "test",
+        }
+    }
+
    /// Authenticate the client via the requested backend, possibly using credentials.
    #[tracing::instrument(fields(allow_cleartext = allow_cleartext), skip_all)]
    pub async fn authenticate(
--- a/proxy/src/http/conn_pool.rs
+++ b/proxy/src/http/conn_pool.rs
@@ -17,11 +17,12 @@ use std::{
 use tokio::time;
 use tokio_postgres::AsyncMessage;

-use crate::{auth, console};
+use crate::{
+    auth, console,
+    metrics::{Ids, MetricCounter, USAGE_METRICS},
+};
 use crate::{compute, config};

-use super::sql_over_http::MAX_RESPONSE_SIZE;
-
 use crate::proxy::ConnectMechanism;

 use tracing::{error, warn};
@@ -400,7 +401,6 @@ async fn connect_to_compute_once(
        .user(&conn_info.username)
        .password(&conn_info.password)
        .dbname(&conn_info.dbname)
-        .max_backend_message_size(MAX_RESPONSE_SIZE)
        .connect_timeout(timeout)
        .connect(tokio_postgres::NoTls)
        .await?;
@@ -412,6 +412,10 @@ async fn connect_to_compute_once(
    span.in_scope(|| {
        info!(%conn_info, %session, "new connection");
    });
+    let ids = Ids {
+        endpoint_id: node_info.aux.endpoint_id.to_string(),
+        branch_id: node_info.aux.branch_id.to_string(),
+    };

    tokio::spawn(
        poll_fn(move |cx| {
@@ -450,10 +454,18 @@ async fn connect_to_compute_once(
    Ok(Client {
        inner: client,
        session: tx,
+        ids,
    })
 }

 pub struct Client {
    pub inner: tokio_postgres::Client,
    session: tokio::sync::watch::Sender<uuid::Uuid>,
+    ids: Ids,
+}
+
+impl Client {
+    pub fn metrics(&self) -> Arc<MetricCounter> {
+        USAGE_METRICS.register(self.ids.clone())
+    }
 }
--- a/proxy/src/http/sql_over_http.rs
+++ b/proxy/src/http/sql_over_http.rs
@@ -3,10 +3,12 @@ use std::sync::Arc;
 use anyhow::bail;
 use futures::pin_mut;
 use futures::StreamExt;
-use hashbrown::HashMap;
 use hyper::body::HttpBody;
+use hyper::header;
 use hyper::http::HeaderName;
 use hyper::http::HeaderValue;
+use hyper::Response;
+use hyper::StatusCode;
 use hyper::{Body, HeaderMap, Request};
 use serde_json::json;
 use serde_json::Map;
@@ -16,7 +18,11 @@ use tokio_postgres::types::Type;
 use tokio_postgres::GenericClient;
 use tokio_postgres::IsolationLevel;
 use tokio_postgres::Row;
+use tracing::error;
+use tracing::instrument;
 use url::Url;
+use utils::http::error::ApiError;
+use utils::http::json::json_response;

 use super::conn_pool::ConnInfo;
 use super::conn_pool::GlobalConnPool;
@@ -39,7 +45,6 @@ enum Payload {
    Batch(BatchQueryData),
 }

-pub const MAX_RESPONSE_SIZE: usize = 10 * 1024 * 1024; // 10 MB
 const MAX_REQUEST_SIZE: u64 = 1024 * 1024; // 1 MB

 static RAW_TEXT_OUTPUT: HeaderName = HeaderName::from_static("neon-raw-text-output");
@@ -182,7 +187,45 @@ pub async fn handle(
    sni_hostname: Option<String>,
    conn_pool: Arc<GlobalConnPool>,
    session_id: uuid::Uuid,
-) -> anyhow::Result<(Value, HashMap<HeaderName, HeaderValue>)> {
+) -> Result<Response<Body>, ApiError> {
+    let result = handle_inner(request, sni_hostname, conn_pool, session_id).await;
+
+    let mut response = match result {
+        Ok(r) => r,
+        Err(e) => {
+            let message = format!("{:?}", e);
+            let code = match e.downcast_ref::<tokio_postgres::Error>() {
+                Some(e) => match e.code() {
+                    Some(e) => serde_json::to_value(e.code()).unwrap(),
+                    None => Value::Null,
+                },
+                None => Value::Null,
+            };
+            error!(
+                ?code,
+                "sql-over-http per-client task finished with an error: {e:#}"
+            );
+            // TODO: this shouldn't always be bad request.
+            json_response(
+                StatusCode::BAD_REQUEST,
+                json!({ "message": message, "code": code }),
+            )?
+        }
+    };
+    response.headers_mut().insert(
+        "Access-Control-Allow-Origin",
+        hyper::http::HeaderValue::from_static("*"),
+    );
+    Ok(response)
+}
+
+#[instrument(name = "sql-over-http", skip_all)]
+async fn handle_inner(
+    request: Request<Body>,
+    sni_hostname: Option<String>,
+    conn_pool: Arc<GlobalConnPool>,
+    session_id: uuid::Uuid,
+) -> anyhow::Result<Response<Body>> {
    //
    // Determine the destination and connection params
    //
@@ -233,13 +276,18 @@ pub async fn handle(

    let mut client = conn_pool.get(&conn_info, !allow_pool, session_id).await?;

+    let mut response = Response::builder()
+        .status(StatusCode::OK)
+        .header(header::CONTENT_TYPE, "application/json");
+
    //
    // Now execute the query and return the result
    //
+    let mut size = 0;
    let result = match payload {
-        Payload::Single(query) => query_to_json(&client.inner, query, raw_output, array_mode)
-            .await
-            .map(|x| (x, HashMap::default())),
+        Payload::Single(query) => {
+            query_to_json(&client.inner, query, &mut size, raw_output, array_mode).await
+        }
        Payload::Batch(batch_query) => {
            let mut results = Vec::new();
            let mut builder = client.inner.build_transaction();
@@ -254,7 +302,8 @@ pub async fn handle(
            }
            let transaction = builder.start().await?;
            for query in batch_query.queries {
-                let result = query_to_json(&transaction, query, raw_output, array_mode).await;
+                let result =
+                    query_to_json(&transaction, query, &mut size, raw_output, array_mode).await;
                match result {
                    Ok(r) => results.push(r),
                    Err(e) => {
@@ -264,26 +313,27 @@ pub async fn handle(
                }
            }
            transaction.commit().await?;
-            let mut headers = HashMap::default();
            if txn_read_only {
-                headers.insert(
+                response = response.header(
                    TXN_READ_ONLY.clone(),
                    HeaderValue::try_from(txn_read_only.to_string())?,
                );
            }
            if txn_deferrable {
-                headers.insert(
+                response = response.header(
                    TXN_DEFERRABLE.clone(),
                    HeaderValue::try_from(txn_deferrable.to_string())?,
                );
            }
            if let Some(txn_isolation_level) = txn_isolation_level_raw {
-                headers.insert(TXN_ISOLATION_LEVEL.clone(), txn_isolation_level);
+                response = response.header(TXN_ISOLATION_LEVEL.clone(), txn_isolation_level);
            }
-            Ok((json!({ "results": results }), headers))
+            Ok(json!({ "results": results }))
        }
    };

+    let metrics = client.metrics();
+
    if allow_pool {
        let current_span = tracing::Span::current();
        // return connection to the pool
@@ -293,12 +343,30 @@ pub async fn handle(
        });
    }

-    result
+    match result {
+        Ok(value) => {
+            // how could this possibly fail
+            let body = serde_json::to_string(&value).expect("json serialization should not fail");
+            let len = body.len();
+            let response = response
+                .body(Body::from(body))
+                // only fails if invalid status code or invalid header/values are given.
+                // these are not user configurable so it cannot fail dynamically
+                .expect("building response payload should not fail");
+
+            // count the egress bytes - we miss the TLS and header overhead but oh well...
+            // moving this later in the stack is going to be a lot of effort and ehhhh
+            metrics.record_egress(len as u64);
+            Ok(response)
+        }
+        Err(e) => Err(e),
+    }
 }

 async fn query_to_json<T: GenericClient>(
    client: &T,
    data: QueryData,
+    current_size: &mut usize,
    raw_output: bool,
    array_mode: bool,
 ) -> anyhow::Result<Value> {
@@ -312,16 +380,10 @@ async fn query_to_json<T: GenericClient>(
    // big.
    pin_mut!(row_stream);
    let mut rows: Vec<tokio_postgres::Row> = Vec::new();
-    let mut current_size = 0;
    while let Some(row) = row_stream.next().await {
        let row = row?;
-        current_size += row.body_len();
+        *current_size += row.body_len();
        rows.push(row);
-        if current_size > MAX_RESPONSE_SIZE {
-            return Err(anyhow::anyhow!(
-                "response is too large (max is {MAX_RESPONSE_SIZE} bytes)"
-            ));
-        }
    }

    // grab the command tag and number of rows affected
--- a/proxy/src/http/websocket.rs
+++ b/proxy/src/http/websocket.rs
@@ -7,7 +7,6 @@ use crate::{
 };
 use bytes::{Buf, Bytes};
 use futures::{Sink, Stream, StreamExt};
-use hashbrown::HashMap;
 use hyper::{
    server::{
        accept,
@@ -18,7 +17,6 @@ use hyper::{
 };
 use hyper_tungstenite::{tungstenite::Message, HyperWebsocket, WebSocketStream};
 use pin_project_lite::pin_project;
-use serde_json::{json, Value};

 use std::{
    convert::Infallible,
@@ -204,44 +202,7 @@ async fn ws_handler(
    // TODO: that deserves a refactor as now this function also handles http json client besides websockets.
    // Right now I don't want to blow up sql-over-http patch with file renames and do that as a follow up instead.
    } else if request.uri().path() == "/sql" && request.method() == Method::POST {
-        let result = sql_over_http::handle(request, sni_hostname, conn_pool, session_id)
-            .instrument(info_span!("sql-over-http"))
-            .await;
-        let status_code = match result {
-            Ok(_) => StatusCode::OK,
-            Err(_) => StatusCode::BAD_REQUEST,
-        };
-        let (json, headers) = match result {
-            Ok(r) => r,
-            Err(e) => {
-                let message = format!("{:?}", e);
-                let code = match e.downcast_ref::<tokio_postgres::Error>() {
-                    Some(e) => match e.code() {
-                        Some(e) => serde_json::to_value(e.code()).unwrap(),
-                        None => Value::Null,
-                    },
-                    None => Value::Null,
-                };
-                error!(
-                    ?code,
-                    "sql-over-http per-client task finished with an error: {e:#}"
-                );
-                (
-                    json!({ "message": message, "code": code }),
-                    HashMap::default(),
-                )
-            }
-        };
-        json_response(status_code, json).map(|mut r| {
-            r.headers_mut().insert(
-                "Access-Control-Allow-Origin",
-                hyper::http::HeaderValue::from_static("*"),
-            );
-            for (k, v) in headers {
-                r.headers_mut().insert(k, v);
-            }
-            r
-        })
+        sql_over_http::handle(request, sni_hostname, conn_pool, session_id).await
    } else if request.uri().path() == "/sql" && request.method() == Method::OPTIONS {
        Response::builder()
            .header("Allow", "OPTIONS, POST")
@@ -253,7 +214,7 @@ async fn ws_handler(
            .header("Access-Control-Max-Age", "86400" /* 24 hours */)
            .status(StatusCode::OK) // 204 is also valid, but see: https://developer.mozilla.org/en-US/docs/Web/HTTP/Methods/OPTIONS#status_code
            .body(Body::empty())
-            .map_err(|e| ApiError::BadRequest(e.into()))
+            .map_err(|e| ApiError::InternalServerError(e.into()))
    } else {
        json_response(StatusCode::BAD_REQUEST, "query is not supported")
    }
--- a/proxy/src/metrics.rs
+++ b/proxy/src/metrics.rs
@@ -3,9 +3,18 @@
 use crate::{config::MetricCollectionConfig, http};
 use chrono::{DateTime, Utc};
 use consumption_metrics::{idempotency_key, Event, EventChunk, EventType, CHUNK_SIZE};
-use serde::Serialize;
-use std::{collections::HashMap, convert::Infallible, time::Duration};
-use tracing::{error, info, instrument, trace, warn};
+use dashmap::{mapref::entry::Entry, DashMap};
+use once_cell::sync::Lazy;
+use serde::{Deserialize, Serialize};
+use std::{
+    convert::Infallible,
+    sync::{
+        atomic::{AtomicU64, AtomicUsize, Ordering},
+        Arc,
+    },
+    time::Duration,
+};
+use tracing::{error, info, instrument, trace};

 const PROXY_IO_BYTES_PER_CLIENT: &str = "proxy_io_bytes_per_client";

@@ -18,12 +27,95 @@ const DEFAULT_HTTP_REPORTING_TIMEOUT: Duration = Duration::from_secs(60);
 /// Both the proxy and the ingestion endpoint will live in the same region (or cell)
 /// so while the project-id is unique across regions the whole pipeline will work correctly
 /// because we enrich the event with project_id in the control-plane endpoint.
-#[derive(Eq, Hash, PartialEq, Serialize, Debug, Clone)]
+#[derive(Eq, Hash, PartialEq, Serialize, Deserialize, Debug, Clone)]
 pub struct Ids {
    pub endpoint_id: String,
    pub branch_id: String,
 }

+#[derive(Debug)]
+pub struct MetricCounter {
+    transmitted: AtomicU64,
+    opened_connections: AtomicUsize,
+}
+
+impl MetricCounter {
+    /// Record that some bytes were sent from the proxy to the client
+    pub fn record_egress(&self, bytes: u64) {
+        self.transmitted.fetch_add(bytes, Ordering::AcqRel);
+    }
+
+    /// extract the value that should be reported
+    fn should_report(self: &Arc<Self>) -> Option<u64> {
+        // heuristic to see if the branch is still open
+        // if a clone happens while we are observing, the heuristic will be incorrect.
+        //
+        // Worst case is that we won't report an event for this endpoint.
+        // However, for the strong count to be 1 it must have occured that at one instant
+        // all the endpoints were closed, so missing a report because the endpoints are closed is valid.
+        let is_open = Arc::strong_count(self) > 1;
+        let opened = self.opened_connections.swap(0, Ordering::AcqRel);
+
+        // update cached metrics eagerly, even if they can't get sent
+        // (to avoid sending the same metrics twice)
+        // see the relevant discussion on why to do so even if the status is not success:
+        // https://github.com/neondatabase/neon/pull/4563#discussion_r1246710956
+        let value = self.transmitted.swap(0, Ordering::AcqRel);
+
+        // Our only requirement is that we report in every interval if there was an open connection
+        // if there were no opened connections since, then we don't need to report
+        if value == 0 && !is_open && opened == 0 {
+            None
+        } else {
+            Some(value)
+        }
+    }
+
+    /// Determine whether the counter should be cleared from the global map.
+    fn should_clear(self: &mut Arc<Self>) -> bool {
+        // we can't clear this entry if it's acquired elsewhere
+        let Some(counter) = Arc::get_mut(self) else {
+            return false;
+        };
+        let opened = *counter.opened_connections.get_mut();
+        let value = *counter.transmitted.get_mut();
+        // clear if there's no data to report
+        value == 0 && opened == 0
+    }
+}
+
+// endpoint and branch IDs are not user generated so we don't run the risk of hash-dos
+type FastHasher = std::hash::BuildHasherDefault<rustc_hash::FxHasher>;
+
+#[derive(Default)]
+pub struct Metrics {
+    endpoints: DashMap<Ids, Arc<MetricCounter>, FastHasher>,
+}
+
+impl Metrics {
+    /// Register a new byte metrics counter for this endpoint
+    pub fn register(&self, ids: Ids) -> Arc<MetricCounter> {
+        let entry = if let Some(entry) = self.endpoints.get(&ids) {
+            entry.clone()
+        } else {
+            self.endpoints
+                .entry(ids)
+                .or_insert_with(|| {
+                    Arc::new(MetricCounter {
+                        transmitted: AtomicU64::new(0),
+                        opened_connections: AtomicUsize::new(0),
+                    })
+                })
+                .clone()
+        };
+
+        entry.opened_connections.fetch_add(1, Ordering::AcqRel);
+        entry
+    }
+}
+
+pub static USAGE_METRICS: Lazy<Metrics> = Lazy::new(Metrics::default);
+
 pub async fn task_main(config: &MetricCollectionConfig) -> anyhow::Result<Infallible> {
    info!("metrics collector config: {config:?}");
    scopeguard::defer! {
@@ -31,145 +123,83 @@ pub async fn task_main(config: &MetricCollectionConfig) -> anyhow::Result<Infall
    }

    let http_client = http::new_client_with_timeout(DEFAULT_HTTP_REPORTING_TIMEOUT);
-    let mut cached_metrics: HashMap<Ids, (u64, DateTime<Utc>)> = HashMap::new();
    let hostname = hostname::get()?.as_os_str().to_string_lossy().into_owned();

+    let mut prev = Utc::now();
    let mut ticker = tokio::time::interval(config.interval);
    loop {
        ticker.tick().await;

-        let res = collect_metrics_iteration(
+        let now = Utc::now();
+        collect_metrics_iteration(
+            &USAGE_METRICS,
            &http_client,
-            &mut cached_metrics,
            &config.endpoint,
            &hostname,
+            prev,
+            now,
        )
        .await;
-
-        match res {
-            Err(e) => error!("failed to send consumption metrics: {e} "),
-            Ok(_) => trace!("periodic metrics collection completed successfully"),
-        }
+        prev = now;
    }
 }

-fn gather_proxy_io_bytes_per_client() -> Vec<(Ids, (u64, DateTime<Utc>))> {
-    let mut current_metrics: Vec<(Ids, (u64, DateTime<Utc>))> = Vec::new();
-    let metrics = prometheus::default_registry().gather();
-
-    for m in metrics {
-        if m.get_name() == "proxy_io_bytes_per_client" {
-            for ms in m.get_metric() {
-                let direction = ms
-                    .get_label()
-                    .iter()
-                    .find(|l| l.get_name() == "direction")
-                    .unwrap()
-                    .get_value();
-
-                // Only collect metric for outbound traffic
-                if direction == "tx" {
-                    let endpoint_id = ms
-                        .get_label()
-                        .iter()
-                        .find(|l| l.get_name() == "endpoint_id")
-                        .unwrap()
-                        .get_value();
-                    let branch_id = ms
-                        .get_label()
-                        .iter()
-                        .find(|l| l.get_name() == "branch_id")
-                        .unwrap()
-                        .get_value();
-
-                    let value = ms.get_counter().get_value() as u64;
-
-                    // Report if the metric value is suspiciously large
-                    if value > (1u64 << 40) {
-                        warn!(
-                            "potentially abnormal counter value: branch_id {} endpoint_id {} val: {}",
-                            branch_id, endpoint_id, value
-                        );
-                    }
-
-                    current_metrics.push((
-                        Ids {
-                            endpoint_id: endpoint_id.to_string(),
-                            branch_id: branch_id.to_string(),
-                        },
-                        (value, Utc::now()),
-                    ));
-                }
-            }
-        }
-    }
-
-    current_metrics
-}
-
 #[instrument(skip_all)]
 async fn collect_metrics_iteration(
+    metrics: &Metrics,
    client: &http::ClientWithMiddleware,
-    cached_metrics: &mut HashMap<Ids, (u64, DateTime<Utc>)>,
    metric_collection_endpoint: &reqwest::Url,
    hostname: &str,
-) -> anyhow::Result<()> {
+    prev: DateTime<Utc>,
+    now: DateTime<Utc>,
+) {
    info!(
        "starting collect_metrics_iteration. metric_collection_endpoint: {}",
        metric_collection_endpoint
    );

-    let current_metrics = gather_proxy_io_bytes_per_client();
+    let mut metrics_to_clear = Vec::new();

-    let metrics_to_send: Vec<Event<Ids, &'static str>> = current_metrics
+    let metrics_to_send: Vec<(Ids, u64)> = metrics
+        .endpoints
        .iter()
-        .filter_map(|(curr_key, (curr_val, curr_time))| {
-            let mut start_time = *curr_time;
-            let mut value = *curr_val;
-
-            if let Some((prev_val, prev_time)) = cached_metrics.get(curr_key) {
-                // Only send metrics updates if the metric has increased
-                if curr_val > prev_val {
-                    value = curr_val - prev_val;
-                    start_time = *prev_time;
-                } else {
-                    if curr_val < prev_val {
-                        error!("proxy_io_bytes_per_client metric value decreased from {} to {} for key {:?}",
-                        prev_val, curr_val, curr_key);
-                    }
-                    return None;
-                }
+        .filter_map(|counter| {
+            let key = counter.key().clone();
+            let Some(value) = counter.should_report() else {
+                metrics_to_clear.push(key);
+                return None;
            };
-
-            Some(Event {
-                kind: EventType::Incremental {
-                    start_time,
-                    stop_time: *curr_time,
-                },
-                metric: PROXY_IO_BYTES_PER_CLIENT,
-                idempotency_key: idempotency_key(hostname),
-                value,
-                extra: Ids {
-                    endpoint_id: curr_key.endpoint_id.clone(),
-                    branch_id: curr_key.branch_id.clone(),
-                },
-            })
+            Some((key, value))
        })
        .collect();

    if metrics_to_send.is_empty() {
        trace!("no new metrics to send");
-        return Ok(());
    }

    // Send metrics.
    // Split into chunks of 1000 metrics to avoid exceeding the max request size
    for chunk in metrics_to_send.chunks(CHUNK_SIZE) {
+        let events = chunk
+            .iter()
+            .map(|(ids, value)| Event {
+                kind: EventType::Incremental {
+                    start_time: prev,
+                    stop_time: now,
+                },
+                metric: PROXY_IO_BYTES_PER_CLIENT,
+                idempotency_key: idempotency_key(hostname),
+                value: *value,
+                extra: Ids {
+                    endpoint_id: ids.endpoint_id.clone(),
+                    branch_id: ids.branch_id.clone(),
+                },
+            })
+            .collect();
+
        let res = client
            .post(metric_collection_endpoint.clone())
-            .json(&EventChunk {
-                events: chunk.into(),
-            })
+            .json(&EventChunk { events })
            .send()
            .await;

@@ -183,34 +213,113 @@ async fn collect_metrics_iteration(

        if !res.status().is_success() {
            error!("metrics endpoint refused the sent metrics: {:?}", res);
-            for metric in chunk.iter().filter(|metric| metric.value > (1u64 << 40)) {
+            for metric in chunk.iter().filter(|(_, value)| *value > (1u64 << 40)) {
                // Report if the metric value is suspiciously large
                error!("potentially abnormal metric value: {:?}", metric);
            }
        }
-        // update cached metrics after they were sent
-        // (to avoid sending the same metrics twice)
-        // see the relevant discussion on why to do so even if the status is not success:
-        // https://github.com/neondatabase/neon/pull/4563#discussion_r1246710956
-        for send_metric in chunk {
-            let stop_time = match send_metric.kind {
-                EventType::Incremental { stop_time, .. } => stop_time,
-                _ => unreachable!(),
-            };
+    }

-            cached_metrics
-                .entry(Ids {
-                    endpoint_id: send_metric.extra.endpoint_id.clone(),
-                    branch_id: send_metric.extra.branch_id.clone(),
-                })
-                // update cached value (add delta) and time
-                .and_modify(|e| {
-                    e.0 = e.0.saturating_add(send_metric.value);
-                    e.1 = stop_time
-                })
-                // cache new metric
-                .or_insert((send_metric.value, stop_time));
+    for metric in metrics_to_clear {
+        match metrics.endpoints.entry(metric) {
+            Entry::Occupied(mut counter) => {
+                if counter.get_mut().should_clear() {
+                    counter.remove_entry();
+                }
+            }
+            Entry::Vacant(_) => {}
        }
    }
-    Ok(())
+}
+
+#[cfg(test)]
+mod tests {
+    use std::{
+        net::TcpListener,
+        sync::{Arc, Mutex},
+    };
+
+    use anyhow::Error;
+    use chrono::Utc;
+    use consumption_metrics::{Event, EventChunk};
+    use hyper::{
+        service::{make_service_fn, service_fn},
+        Body, Response,
+    };
+    use url::Url;
+
+    use super::{collect_metrics_iteration, Ids, Metrics};
+    use crate::http;
+
+    #[tokio::test]
+    async fn metrics() {
+        let listener = TcpListener::bind("0.0.0.0:0").unwrap();
+
+        let reports = Arc::new(Mutex::new(vec![]));
+        let reports2 = reports.clone();
+
+        let server = hyper::server::Server::from_tcp(listener)
+            .unwrap()
+            .serve(make_service_fn(move |_| {
+                let reports = reports.clone();
+                async move {
+                    Ok::<_, Error>(service_fn(move |req| {
+                        let reports = reports.clone();
+                        async move {
+                            let bytes = hyper::body::to_bytes(req.into_body()).await?;
+                            let events: EventChunk<'static, Event<Ids, String>> =
+                                serde_json::from_slice(&bytes)?;
+                            reports.lock().unwrap().push(events);
+                            Ok::<_, Error>(Response::new(Body::from(vec![])))
+                        }
+                    }))
+                }
+            }));
+        let addr = server.local_addr();
+        tokio::spawn(server);
+
+        let metrics = Metrics::default();
+        let client = http::new_client();
+        let endpoint = Url::parse(&format!("http://{addr}")).unwrap();
+        let now = Utc::now();
+
+        // no counters have been registered
+        collect_metrics_iteration(&metrics, &client, &endpoint, "foo", now, now).await;
+        let r = std::mem::take(&mut *reports2.lock().unwrap());
+        assert!(r.is_empty());
+
+        // register a new counter
+        let counter = metrics.register(Ids {
+            endpoint_id: "e1".to_string(),
+            branch_id: "b1".to_string(),
+        });
+
+        // the counter should be observed despite 0 egress
+        collect_metrics_iteration(&metrics, &client, &endpoint, "foo", now, now).await;
+        let r = std::mem::take(&mut *reports2.lock().unwrap());
+        assert_eq!(r.len(), 1);
+        assert_eq!(r[0].events.len(), 1);
+        assert_eq!(r[0].events[0].value, 0);
+
+        // record egress
+        counter.record_egress(1);
+
+        // egress should be observered
+        collect_metrics_iteration(&metrics, &client, &endpoint, "foo", now, now).await;
+        let r = std::mem::take(&mut *reports2.lock().unwrap());
+        assert_eq!(r.len(), 1);
+        assert_eq!(r[0].events.len(), 1);
+        assert_eq!(r[0].events[0].value, 1);
+
+        // release counter
+        drop(counter);
+
+        // we do not observe the counter
+        collect_metrics_iteration(&metrics, &client, &endpoint, "foo", now, now).await;
+        let r = std::mem::take(&mut *reports2.lock().unwrap());
+        assert!(r.is_empty());
+
+        // counter is unregistered
+        assert!(metrics.endpoints.is_empty());
+    }
 }
--- a/proxy/src/proxy.rs
+++ b/proxy/src/proxy.rs
@@ -7,6 +7,7 @@ use crate::{
    compute::{self, PostgresConnection},
    config::{ProxyConfig, TlsConfig},
    console::{self, errors::WakeComputeError, messages::MetricsAuxInfo, Api},
+    metrics::{Ids, USAGE_METRICS},
    protocol2::WithClientIp,
    stream::{PqStream, Stream},
 };
@@ -602,6 +603,11 @@ pub async fn proxy_pass(
    compute: impl AsyncRead + AsyncWrite + Unpin,
    aux: &MetricsAuxInfo,
 ) -> anyhow::Result<()> {
+    let usage = USAGE_METRICS.register(Ids {
+        endpoint_id: aux.endpoint_id.to_string(),
+        branch_id: aux.branch_id.to_string(),
+    });
+
    let m_sent = NUM_BYTES_PROXIED_COUNTER.with_label_values(&aux.traffic_labels("tx"));
    let mut client = MeasuredStream::new(
        client,
@@ -609,6 +615,7 @@ pub async fn proxy_pass(
        |cnt| {
            // Number of bytes we sent to the client (outbound).
            m_sent.inc_by(cnt as u64);
+            usage.record_egress(cnt as u64);
        },
    );

@@ -690,7 +697,14 @@ impl<S: AsyncRead + AsyncWrite + Unpin> Client<'_, S> {
            .await
        {
            Ok(auth_result) => auth_result,
-            Err(e) => return stream.throw_error(e).await,
+            Err(e) => {
+                let user = creds.get_user();
+                let db = params.get("database");
+                let app = params.get("application_name");
+                let params_span = tracing::info_span!("", ?user, ?db, ?app);
+
+                return stream.throw_error(e).instrument(params_span).await;
+            }
        };

        let AuthSuccess {
--- a/test_runner/fixtures/compare_fixtures.py
+++ b/test_runner/fixtures/compare_fixtures.py
@@ -105,6 +105,8 @@ class NeonCompare(PgCompare):
        self._pg_bin = pg_bin
        self.pageserver_http_client = self.env.pageserver.http_client()

+        # note that neon_simple_env now uses LOCAL_FS remote storage
+
        # Create tenant
        tenant_conf: Dict[str, str] = {}
        if False:  # TODO add pytest setting for this
--- a/test_runner/fixtures/neon_fixtures.py
+++ b/test_runner/fixtures/neon_fixtures.py
@@ -37,6 +37,7 @@ from psycopg2.extensions import connection as PgConnection
 from psycopg2.extensions import cursor as PgCursor
 from psycopg2.extensions import make_dsn, parse_dsn
 from typing_extensions import Literal
+from urllib3.util.retry import Retry

 from fixtures.broker import NeonBroker
 from fixtures.log_helper import log
@@ -460,9 +461,11 @@ class NeonEnvBuilder:
        ), "Unexpectedly instantiated from outside a test function"
        self.test_name = test_name

-    def init_configs(self) -> NeonEnv:
+    def init_configs(self, default_remote_storage_if_missing: bool = True) -> NeonEnv:
        # Cannot create more than one environment from one builder
        assert self.env is None, "environment already initialized"
+        if default_remote_storage_if_missing and self.pageserver_remote_storage is None:
+            self.enable_pageserver_remote_storage(RemoteStorageKind.LOCAL_FS)
        self.env = NeonEnv(self)
        return self.env

@@ -470,8 +473,19 @@ class NeonEnvBuilder:
        assert self.env is not None, "environment is not already initialized, call init() first"
        self.env.start()

-    def init_start(self, initial_tenant_conf: Optional[Dict[str, str]] = None) -> NeonEnv:
-        env = self.init_configs()
+    def init_start(
+        self,
+        initial_tenant_conf: Optional[Dict[str, str]] = None,
+        default_remote_storage_if_missing: bool = True,
+    ) -> NeonEnv:
+        """
+        Default way to create and start NeonEnv. Also creates the initial_tenant with root initial_timeline.
+
+        To avoid creating initial_tenant, call init_configs to setup the environment.
+
+        Configuring pageserver with remote storage is now the default. There will be a warning if pageserver is created without one.
+        """
+        env = self.init_configs(default_remote_storage_if_missing=default_remote_storage_if_missing)
        self.start()

        # Prepare the default branch to start the postgres on later.
@@ -546,7 +560,7 @@ class NeonEnvBuilder:
        user: RemoteStorageUser,
        bucket_name: Optional[str] = None,
        bucket_region: Optional[str] = None,
-    ) -> Optional[RemoteStorage]:
+    ) -> RemoteStorage:
        ret = kind.configure(
            self.repo_dir,
            self.mock_s3_server,
@@ -889,6 +903,8 @@ def _shared_simple_env(
    """
    # Internal fixture backing the `neon_simple_env` fixture. If TEST_SHARED_FIXTURES
     is set, this is shared by all tests using `neon_simple_env`.
+
+    This fixture will use RemoteStorageKind.LOCAL_FS with pageserver.
    """

    if os.environ.get("TEST_SHARED_FIXTURES") is None:
@@ -1636,11 +1652,14 @@ class NeonPageserver(PgProtocol):
        if '"testing"' not in self.version:
            pytest.skip("pageserver was built without 'testing' feature")

-    def http_client(self, auth_token: Optional[str] = None) -> PageserverHttpClient:
+    def http_client(
+        self, auth_token: Optional[str] = None, retries: Optional[Retry] = None
+    ) -> PageserverHttpClient:
        return PageserverHttpClient(
            port=self.service_port.http,
            auth_token=auth_token,
            is_testing_enabled_or_skip=self.is_testing_enabled_or_skip,
+            retries=retries,
        )

    @property
--- a/test_runner/fixtures/pageserver/http.py
+++ b/test_runner/fixtures/pageserver/http.py
@@ -7,6 +7,8 @@ from dataclasses import dataclass
 from typing import Any, Dict, List, Optional, Tuple

 import requests
+from requests.adapters import HTTPAdapter
+from urllib3.util.retry import Retry

 from fixtures.log_helper import log
 from fixtures.metrics import Metrics, parse_metrics
@@ -113,12 +115,40 @@ class TenantConfig:


 class PageserverHttpClient(requests.Session):
-    def __init__(self, port: int, is_testing_enabled_or_skip: Fn, auth_token: Optional[str] = None):
+    def __init__(
+        self,
+        port: int,
+        is_testing_enabled_or_skip: Fn,
+        auth_token: Optional[str] = None,
+        retries: Optional[Retry] = None,
+    ):
        super().__init__()
        self.port = port
        self.auth_token = auth_token
        self.is_testing_enabled_or_skip = is_testing_enabled_or_skip

+        if retries is None:
+            # We apply a retry policy that is different to the default `requests` behavior,
+            # because the pageserver has various transiently unavailable states that benefit
+            # from a client retrying on 503
+
+            retries = Retry(
+                # Status retries are for retrying on 503 while e.g. waiting for tenants to activate
+                status=5,
+                # Connection retries are for waiting for the pageserver to come up and listen
+                connect=5,
+                # No read retries: if a request hangs that is not expected behavior
+                # (this may change in future if we do fault injection of a kind that causes
+                #  requests TCP flows to stick)
+                read=False,
+                backoff_factor=0,
+                status_forcelist=[503],
+                allowed_methods=None,
+                remove_headers_on_redirect=[],
+            )
+
+        self.mount("http://", HTTPAdapter(max_retries=retries))
+
        if auth_token is not None:
            self.headers["Authorization"] = f"Bearer {auth_token}"

--- a/test_runner/fixtures/remote_storage.py
+++ b/test_runner/fixtures/remote_storage.py
@@ -202,9 +202,6 @@ class RemoteStorageKind(str, enum.Enum):
    LOCAL_FS = "local_fs"
    MOCK_S3 = "mock_s3"
    REAL_S3 = "real_s3"
-    # Pass to tests that are generic to remote storage
-    # to ensure the test pass with or without the remote storage
-    NOOP = "noop"

    def configure(
        self,
@@ -215,10 +212,7 @@ class RemoteStorageKind(str, enum.Enum):
        user: RemoteStorageUser,
        bucket_name: Optional[str] = None,
        bucket_region: Optional[str] = None,
-    ) -> Optional[RemoteStorage]:
-        if self == RemoteStorageKind.NOOP:
-            return None
-
+    ) -> RemoteStorage:
        if self == RemoteStorageKind.LOCAL_FS:
            return LocalFsStorage(LocalFsStorage.component_path(repo_dir, user))

--- a/test_runner/fixtures/utils.py
+++ b/test_runner/fixtures/utils.py
@@ -222,7 +222,7 @@ def get_scale_for_db(size_mb: int) -> int:


 ATTACHMENT_NAME_REGEX: re.Pattern = re.compile(  # type: ignore[type-arg]
-    r"regression\.diffs|.+\.(?:log|stderr|stdout|filediff|metrics|html)"
+    r"regression\.diffs|.+\.(?:log|stderr|stdout|filediff|metrics|html|walredo)"
 )


@@ -250,6 +250,9 @@ def allure_attach_from_dir(dir: Path):
            elif source.endswith(".html"):
                attachment_type = "text/html"
                extension = "html"
+            elif source.endswith(".walredo"):
+                attachment_type = "application/octet-stream"
+                extension = "walredo"
            else:
                attachment_type = "text/plain"
                extension = attachment.suffix.removeprefix(".")
--- a/test_runner/regress/test_broken_timeline.py
+++ b/test_runner/regress/test_broken_timeline.py
@@ -4,7 +4,12 @@ from typing import List, Tuple

 import pytest
 from fixtures.log_helper import log
-from fixtures.neon_fixtures import Endpoint, NeonEnv, NeonEnvBuilder
+from fixtures.neon_fixtures import (
+    Endpoint,
+    NeonEnv,
+    NeonEnvBuilder,
+    wait_for_last_flush_lsn,
+)
 from fixtures.types import TenantId, TimelineId


@@ -26,17 +31,18 @@ def test_broken_timeline(neon_env_builder: NeonEnvBuilder):

    tenant_timelines: List[Tuple[TenantId, TimelineId, Endpoint]] = []

-    for _ in range(4):
+    for _ in range(3):
        tenant_id, timeline_id = env.neon_cli.create_tenant()

        endpoint = env.endpoints.create_start("main", tenant_id=tenant_id)
        with endpoint.cursor() as cur:
            cur.execute("CREATE TABLE t(key int primary key, value text)")
            cur.execute("INSERT INTO t SELECT generate_series(1,100), 'payload'")
+            wait_for_last_flush_lsn(env, endpoint, tenant_id, timeline_id)
        endpoint.stop()
        tenant_timelines.append((tenant_id, timeline_id, endpoint))

-    # Stop the pageserver
+    # Stop the pageserver -- this has to be not immediate or we need to wait for uploads
    env.pageserver.stop()

    # Leave the first timeline alone, but corrupt the others in different ways
@@ -45,30 +51,21 @@ def test_broken_timeline(neon_env_builder: NeonEnvBuilder):

    (tenant1, timeline1, pg1) = tenant_timelines[1]
    metadata_path = f"{env.pageserver.workdir}/tenants/{tenant1}/timelines/{timeline1}/metadata"
-    f = open(metadata_path, "w")
-    f.write("overwritten with garbage!")
-    f.close()
+    with open(metadata_path, "w") as f:
+        f.write("overwritten with garbage!")
    log.info(f"Timeline {tenant1}/{timeline1} got its metadata spoiled")

    (tenant2, timeline2, pg2) = tenant_timelines[2]
    timeline_path = f"{env.pageserver.workdir}/tenants/{tenant2}/timelines/{timeline2}/"
-    for filename in os.listdir(timeline_path):
-        if filename.startswith("00000"):
-            # Looks like a layer file. Remove it
-            os.remove(f"{timeline_path}/{filename}")
-    log.info(
-        f"Timeline {tenant2}/{timeline2} got its layer files removed (no remote storage enabled)"
-    )
-
-    (tenant3, timeline3, pg3) = tenant_timelines[3]
-    timeline_path = f"{env.pageserver.workdir}/tenants/{tenant3}/timelines/{timeline3}/"
    for filename in os.listdir(timeline_path):
        if filename.startswith("00000"):
            # Looks like a layer file. Corrupt it
-            f = open(f"{timeline_path}/{filename}", "w")
-            f.write("overwritten with garbage!")
-            f.close()
-    log.info(f"Timeline {tenant3}/{timeline3} got its layer files spoiled")
+            p = f"{timeline_path}/{filename}"
+            size = os.path.getsize(p)
+            with open(p, "wb") as f:
+                f.truncate(0)
+                f.truncate(size)
+    log.info(f"Timeline {tenant2}/{timeline2} got its local layer files spoiled")

    env.pageserver.start()

@@ -87,22 +84,13 @@ def test_broken_timeline(neon_env_builder: NeonEnvBuilder):
        f"As expected, compute startup failed eagerly for timeline with corrupt metadata: {err}"
    )

-    # Second timeline has no ancestors, only the metadata file and no layer files locally,
-    # and we don't have the remote storage enabled. It is loaded into memory, but getting
-    # the basebackup from it will fail.
-    with pytest.raises(
-        Exception, match=f"Tenant {tenant2} will not become active. Current state: Broken"
-    ) as err:
-        pg2.start()
-    log.info(f"As expected, compute startup failed for timeline with missing layers: {err}")
-
-    # Third timeline will also fail during basebackup, because the layer file is corrupt.
+    # Second timeline will fail during basebackup, because the local layer file is corrupt.
    # It will fail when we try to read (and reconstruct) a page from it, ergo the error message.
    # (We don't check layer file contents on startup, when loading the timeline)
    with pytest.raises(Exception, match="Failed to load delta layer") as err:
-        pg3.start()
+        pg2.start()
    log.info(
-        f"As expected, compute startup failed for timeline {tenant3}/{timeline3} with corrupt layers: {err}"
+        f"As expected, compute startup failed for timeline {tenant2}/{timeline2} with corrupt layers: {err}"
    )


--- a/test_runner/regress/test_ddl_forwarding.py
+++ b/test_runner/regress/test_ddl_forwarding.py
@@ -211,4 +211,12 @@ def test_ddl_forwarding(ddl: DdlForwardingContext):
        ddl.wait()

    ddl.failures(False)
+    cur.execute("CREATE DATABASE failure WITH OWNER=cork")
+    ddl.wait()
+    with pytest.raises(psycopg2.InternalError):
+        ddl.failures(True)
+        cur.execute("DROP DATABASE failure")
+        ddl.wait()
+    ddl.pg.connect(dbname="failure")  # Ensure we can connect after a failed drop
+
    conn.close()
--- a/test_runner/regress/test_disk_usage_eviction.py
+++ b/test_runner/regress/test_disk_usage_eviction.py
@@ -74,11 +74,13 @@ class EvictionEnv:
    pgbench_init_lsns: Dict[TenantId, Lsn]

    def timelines_du(self) -> Tuple[int, int, int]:
-        return poor_mans_du(self.neon_env, [(tid, tlid) for tid, tlid in self.timelines])
+        return poor_mans_du(
+            self.neon_env, [(tid, tlid) for tid, tlid in self.timelines], verbose=False
+        )

    def du_by_timeline(self) -> Dict[Tuple[TenantId, TimelineId], int]:
        return {
-            (tid, tlid): poor_mans_du(self.neon_env, [(tid, tlid)])[0]
+            (tid, tlid): poor_mans_du(self.neon_env, [(tid, tlid)], verbose=True)[0]
            for tid, tlid in self.timelines
        }

@@ -89,7 +91,21 @@ class EvictionEnv:
        """
        lsn = self.pgbench_init_lsns[tenant_id]
        with self.neon_env.endpoints.create_start("main", tenant_id=tenant_id, lsn=lsn) as endpoint:
-            self.pg_bin.run(["pgbench", "-S", endpoint.connstr()])
+            # instead of using pgbench --select-only which does point selects,
+            # run full table scans for all tables
+            with endpoint.connect() as conn:
+                cur = conn.cursor()
+
+                tables_cols = {
+                    "pgbench_accounts": "abalance",
+                    "pgbench_tellers": "tbalance",
+                    "pgbench_branches": "bbalance",
+                    "pgbench_history": "delta",
+                }
+
+                for table, column in tables_cols.items():
+                    cur.execute(f"select avg({column}) from {table}")
+                    _avg = cur.fetchone()

    def pageserver_start_with_disk_usage_eviction(
        self, period, max_usage_pct, min_avail_bytes, mock_behavior
@@ -127,6 +143,19 @@ class EvictionEnv:
        self.neon_env.pageserver.allowed_errors.append(".*WARN.* disk usage still high.*")


+def human_bytes(amt: float) -> str:
+    suffixes = ["", "Ki", "Mi", "Gi"]
+
+    last = suffixes[-1]
+
+    for name in suffixes:
+        if amt < 1024 or name == last:
+            return f"{int(round(amt))} {name}B"
+        amt = amt / 1024
+
+    raise RuntimeError("unreachable")
+
+
@pytest.fixture
 def eviction_env(request, neon_env_builder: NeonEnvBuilder, pg_bin: PgBin) -> EvictionEnv:
    """
@@ -215,8 +244,12 @@ def test_broken_tenants_are_skipped(eviction_env: EvictionEnv):

    healthy_tenant_id, healthy_timeline_id = env.timelines[1]

-    broken_size_pre, _, _ = poor_mans_du(env.neon_env, [(broken_tenant_id, broken_timeline_id)])
-    healthy_size_pre, _, _ = poor_mans_du(env.neon_env, [(healthy_tenant_id, healthy_timeline_id)])
+    broken_size_pre, _, _ = poor_mans_du(
+        env.neon_env, [(broken_tenant_id, broken_timeline_id)], verbose=True
+    )
+    healthy_size_pre, _, _ = poor_mans_du(
+        env.neon_env, [(healthy_tenant_id, healthy_timeline_id)], verbose=True
+    )

    # try to evict everything, then validate that broken tenant wasn't touched
    target = broken_size_pre + healthy_size_pre
@@ -224,8 +257,12 @@ def test_broken_tenants_are_skipped(eviction_env: EvictionEnv):
    response = env.pageserver_http.disk_usage_eviction_run({"evict_bytes": target})
    log.info(f"{response}")

-    broken_size_post, _, _ = poor_mans_du(env.neon_env, [(broken_tenant_id, broken_timeline_id)])
-    healthy_size_post, _, _ = poor_mans_du(env.neon_env, [(healthy_tenant_id, healthy_timeline_id)])
+    broken_size_post, _, _ = poor_mans_du(
+        env.neon_env, [(broken_tenant_id, broken_timeline_id)], verbose=True
+    )
+    healthy_size_post, _, _ = poor_mans_du(
+        env.neon_env, [(healthy_tenant_id, healthy_timeline_id)], verbose=True
+    )

    assert broken_size_pre == broken_size_post, "broken tenant should not be touched"
    assert healthy_size_post < healthy_size_pre
@@ -366,18 +403,16 @@ def test_partial_evict_tenant(eviction_env: EvictionEnv):
    du_by_timeline = env.du_by_timeline()

    # pick any tenant
-    [our_tenant, other_tenant] = list(du_by_timeline.keys())
-    (tenant_id, timeline_id) = our_tenant
+    [warm, cold] = list(du_by_timeline.keys())
+    (tenant_id, timeline_id) = warm

-    # make our tenant more recently used than the other one
+    # make picked tenant more recently used than the other one
    env.warm_up_tenant(tenant_id)

    # Build up enough pressure to require evictions from both tenants,
    # but not enough to fall into global LRU.
-    # So, set target to all occipied space, except 2*env.layer_size per tenant
-    target = (
-        du_by_timeline[other_tenant] + (du_by_timeline[our_tenant] // 2) - 2 * 2 * env.layer_size
-    )
+    # So, set target to all occupied space, except 2*env.layer_size per tenant
+    target = du_by_timeline[cold] + (du_by_timeline[warm] // 2) - 2 * 2 * env.layer_size
    response = ps_http.disk_usage_eviction_run({"evict_bytes": target})
    log.info(f"{response}")

@@ -392,22 +427,33 @@ def test_partial_evict_tenant(eviction_env: EvictionEnv):
            later_tenant_usage < du_by_timeline[tenant]
        ), "all tenants should have lost some layers"

+    warm_size = later_du_by_timeline[warm]
+
+    # bounds for warmed_size
+    warm_lower = 0.5 * du_by_timeline[warm]
+
+    # We don't know exactly whether the cold tenant needs 2 or just 1 env.layer_size wiggle room.
+    # So, check for up to 3 here.
+    warm_upper = warm_lower + 3 * env.layer_size
+
+    cold_size = later_du_by_timeline[cold]
+    cold_upper = 2 * env.layer_size
+
+    log.info(
+        f"expecting for warm tenant: {human_bytes(warm_lower)} < {human_bytes(warm_size)} < {human_bytes(warm_upper)}"
+    )
+    log.info(f"expecting for cold tenant: {human_bytes(cold_size)} < {human_bytes(cold_upper)}")
+
+    assert warm_size > warm_lower, "warmed up tenant should be at about half size (lower)"
+    assert warm_size < warm_upper, "warmed up tenant should be at about half size (upper)"
+
    assert (
-        later_du_by_timeline[our_tenant] > 0.5 * du_by_timeline[our_tenant]
-    ), "our warmed up tenant should be at about half capacity, part 1"
-    assert (
-        # We don't know exactly whether the cold tenant needs 2 or just 1 env.layer_size wiggle room.
-        # So, check for up to 3 here.
-        later_du_by_timeline[our_tenant]
-        < 0.5 * du_by_timeline[our_tenant] + 3 * env.layer_size
-    ), "our warmed up tenant should be at about half capacity, part 2"
-    assert (
-        later_du_by_timeline[other_tenant] < 2 * env.layer_size
-    ), "the other tenant should be evicted to is min_resident_size, i.e., max layer file size"
+        cold_size < cold_upper
+    ), "the cold tenant should be evicted to its min_resident_size, i.e., max layer file size"


 def poor_mans_du(
-    env: NeonEnv, timelines: list[Tuple[TenantId, TimelineId]]
+    env: NeonEnv, timelines: list[Tuple[TenantId, TimelineId]], verbose: bool = False
 ) -> Tuple[int, int, int]:
    """
    Disk usage, largest, smallest layer for layer files over the given (tenant, timeline) tuples;
@@ -430,9 +476,11 @@ def poor_mans_du(
                smallest_layer = min(smallest_layer, size)
            else:
                smallest_layer = size
-            log.info(f"{tenant_id}/{timeline_id} => {file.name} {size}")
+            if verbose:
+                log.info(f"{tenant_id}/{timeline_id} => {file.name} {size} ({human_bytes(size)})")

-        log.info(f"{tenant_id}/{timeline_id}: sum {total}")
+        if verbose:
+            log.info(f"{tenant_id}/{timeline_id}: sum {total} ({human_bytes(total)})")
        total_on_disk += total

    assert smallest_layer is not None or total_on_disk == 0 and largest_layer == 0
--- a/test_runner/regress/test_pageserver_metric_collection.py
+++ b/test_runner/regress/test_pageserver_metric_collection.py
@@ -5,7 +5,6 @@ from pathlib import Path
 from queue import SimpleQueue
 from typing import Any, Dict, Set

-import pytest
 from fixtures.log_helper import log
 from fixtures.neon_fixtures import (
    NeonEnvBuilder,
@@ -17,15 +16,13 @@ from pytest_httpserver import HTTPServer
 from werkzeug.wrappers.request import Request
 from werkzeug.wrappers.response import Response

+# TODO: collect all of the env setup *AFTER* removal of RemoteStorageKind.NOOP
+

-@pytest.mark.parametrize(
-    "remote_storage_kind", [RemoteStorageKind.NOOP, RemoteStorageKind.LOCAL_FS]
-)
 def test_metric_collection(
    httpserver: HTTPServer,
    neon_env_builder: NeonEnvBuilder,
    httpserver_listen_address,
-    remote_storage_kind: RemoteStorageKind,
 ):
    (host, port) = httpserver_listen_address
    metric_collection_endpoint = f"http://{host}:{port}/billing/api/v1/usage_events"
@@ -55,7 +52,7 @@ def test_metric_collection(
        synthetic_size_calculation_interval="3s"
        """

-    neon_env_builder.enable_pageserver_remote_storage(remote_storage_kind)
+    neon_env_builder.enable_pageserver_remote_storage(RemoteStorageKind.LOCAL_FS)

    log.info(f"test_metric_collection endpoint is {metric_collection_endpoint}")

@@ -68,6 +65,14 @@ def test_metric_collection(
    env = neon_env_builder.init_start(initial_tenant_conf={"pitr_interval": "0 sec"})
    # httpserver is shut down before pageserver during passing run
    env.pageserver.allowed_errors.append(".*metrics endpoint refused the sent metrics*")
+    # we have a fast rate of calculation, these can happen at shutdown
+    env.pageserver.allowed_errors.append(
+        ".*synthetic_size_worker:calculate_synthetic_size.*:gather_size_inputs.*: failed to calculate logical size at .*: cancelled.*"
+    )
+    env.pageserver.allowed_errors.append(
+        ".*synthetic_size_worker: failed to calculate synthetic size for tenant .*: failed to calculate some logical_sizes"
+    )
+
    tenant_id = env.initial_tenant
    timeline_id = env.initial_timeline
    endpoint = env.endpoints.create_start("main", tenant_id=tenant_id)
@@ -98,17 +103,14 @@ def test_metric_collection(
            total += sample[2]
        return int(total)

-    remote_uploaded = 0
-
    # upload some data to remote storage
-    if remote_storage_kind == RemoteStorageKind.LOCAL_FS:
-        wait_for_last_flush_lsn(env, endpoint, tenant_id, timeline_id)
-        pageserver_http = env.pageserver.http_client()
-        pageserver_http.timeline_checkpoint(tenant_id, timeline_id)
-        pageserver_http.timeline_gc(tenant_id, timeline_id, 10000)
+    wait_for_last_flush_lsn(env, endpoint, tenant_id, timeline_id)
+    pageserver_http = env.pageserver.http_client()
+    pageserver_http.timeline_checkpoint(tenant_id, timeline_id)
+    pageserver_http.timeline_gc(tenant_id, timeline_id, 10000)

-        remote_uploaded = get_num_remote_ops("index", "upload")
-        assert remote_uploaded > 0
+    remote_uploaded = get_num_remote_ops("index", "upload")
+    assert remote_uploaded > 0

    # we expect uploads at 1Hz, on busy runners this could be too optimistic,
    # so give 5s we only want to get the following upload after "ready" value.
@@ -211,6 +213,14 @@ def test_metric_collection_cleans_up_tempfile(

    # httpserver is shut down before pageserver during passing run
    env.pageserver.allowed_errors.append(".*metrics endpoint refused the sent metrics*")
+    # we have a fast rate of calculation, these can happen at shutdown
+    env.pageserver.allowed_errors.append(
+        ".*synthetic_size_worker:calculate_synthetic_size.*:gather_size_inputs.*: failed to calculate logical size at .*: cancelled.*"
+    )
+    env.pageserver.allowed_errors.append(
+        ".*synthetic_size_worker: failed to calculate synthetic size for tenant .*: failed to calculate some logical_sizes"
+    )
+
    tenant_id = env.initial_tenant
    timeline_id = env.initial_timeline
    endpoint = env.endpoints.create_start("main", tenant_id=tenant_id)
--- a/test_runner/regress/test_pageserver_restart.py
+++ b/test_runner/regress/test_pageserver_restart.py
@@ -169,3 +169,6 @@ def test_pageserver_chaos(neon_env_builder: NeonEnvBuilder):
        # Check that all the updates are visible
        num_updates = endpoint.safe_psql("SELECT sum(updates) FROM foo")[0][0]
        assert num_updates == i * 100000
+
+    with open(neon_env_builder.test_output_dir / "foobar.walredo", "w") as file:
+        file.write("lets see if this ends in the report")
--- a/test_runner/regress/test_tenant_delete.py
+++ b/test_runner/regress/test_tenant_delete.py
@@ -30,9 +30,7 @@ from fixtures.types import TenantId
 from fixtures.utils import run_pg_bench_small


-@pytest.mark.parametrize(
-    "remote_storage_kind", [RemoteStorageKind.NOOP, *available_remote_storages()]
-)
+@pytest.mark.parametrize("remote_storage_kind", available_remote_storages())
 def test_tenant_delete_smoke(
    neon_env_builder: NeonEnvBuilder,
    remote_storage_kind: RemoteStorageKind,
@@ -144,18 +142,12 @@ FAILPOINTS_BEFORE_BACKGROUND = [
 def combinations():
    result = []

-    remotes = [RemoteStorageKind.NOOP, RemoteStorageKind.MOCK_S3]
+    remotes = [RemoteStorageKind.MOCK_S3]
    if os.getenv("ENABLE_REAL_S3_REMOTE_STORAGE"):
        remotes.append(RemoteStorageKind.REAL_S3)

    for remote_storage_kind in remotes:
        for delete_failpoint in FAILPOINTS:
-            if remote_storage_kind is RemoteStorageKind.NOOP and delete_failpoint in (
-                "timeline-delete-before-index-delete",
-            ):
-                # the above failpoint are not relevant for config without remote storage
-                continue
-
            # Simulate failures for only one type of remote storage
            # to avoid log pollution and make tests run faster
            if remote_storage_kind is RemoteStorageKind.MOCK_S3:
@@ -215,21 +207,18 @@ def test_delete_tenant_exercise_crash_safety_failpoints(
    with env.endpoints.create_start("delete", tenant_id=tenant_id) as endpoint:
        # generate enough layers
        run_pg_bench_small(pg_bin, endpoint.connstr())
-        if remote_storage_kind is RemoteStorageKind.NOOP:
-            wait_for_last_flush_lsn(env, endpoint, tenant_id, timeline_id)
-        else:
-            last_flush_lsn_upload(env, endpoint, tenant_id, timeline_id)
+        last_flush_lsn_upload(env, endpoint, tenant_id, timeline_id)

-            if remote_storage_kind in available_s3_storages():
-                assert_prefix_not_empty(
-                    neon_env_builder,
-                    prefix="/".join(
-                        (
-                            "tenants",
-                            str(tenant_id),
-                        )
-                    ),
-                )
+        if remote_storage_kind in available_s3_storages():
+            assert_prefix_not_empty(
+                neon_env_builder,
+                prefix="/".join(
+                    (
+                        "tenants",
+                        str(tenant_id),
+                    )
+                ),
+            )

    ps_http.configure_failpoints((failpoint, "return"))

@@ -260,12 +249,7 @@ def test_delete_tenant_exercise_crash_safety_failpoints(
        env.pageserver.stop()
        env.pageserver.start()

-        if (
-            remote_storage_kind is RemoteStorageKind.NOOP
-            and failpoint == "tenant-delete-before-create-local-mark"
-        ):
-            tenant_delete_wait_completed(ps_http, tenant_id, iterations=iterations)
-        elif failpoint in (
+        if failpoint in (
            "tenant-delete-before-shutdown",
            "tenant-delete-before-create-remote-mark",
        ):
--- a/test_runner/regress/test_tenant_detach.py
+++ b/test_runner/regress/test_tenant_detach.py
@@ -519,11 +519,8 @@ def test_detach_while_attaching(
 # * restart the pageserver and verify that ignored tenant is still not loaded
 # * `load` the same tenant
 # * ensure that it's status is `Active` and it's present in pageserver's memory with all timelines
-@pytest.mark.parametrize("remote_storage_kind", [RemoteStorageKind.NOOP, RemoteStorageKind.MOCK_S3])
-def test_ignored_tenant_reattach(
-    neon_env_builder: NeonEnvBuilder, remote_storage_kind: RemoteStorageKind
-):
-    neon_env_builder.enable_pageserver_remote_storage(remote_storage_kind)
+def test_ignored_tenant_reattach(neon_env_builder: NeonEnvBuilder):
+    neon_env_builder.enable_pageserver_remote_storage(RemoteStorageKind.MOCK_S3)
    env = neon_env_builder.init_start()
    pageserver_http = env.pageserver.http_client()

--- a/test_runner/regress/test_tenant_size.py
+++ b/test_runner/regress/test_tenant_size.py
@@ -15,7 +15,7 @@ from fixtures.pageserver.utils import (
    timeline_delete_wait_completed,
    wait_until_tenant_active,
 )
-from fixtures.pg_version import PgVersion, xfail_on_postgres
+from fixtures.pg_version import PgVersion
 from fixtures.types import Lsn, TenantId, TimelineId


@@ -532,7 +532,24 @@ def test_single_branch_get_tenant_size_grows(
    assert size_after == prev, "size after restarting pageserver should not have changed"


-@xfail_on_postgres(PgVersion.V15, reason="Test significantly more flaky on Postgres 15")
+def assert_size_approx_equal(size_a, size_b):
+    """
+    Tests that evaluate sizes are checking the pageserver space consumption
+    that sits many layers below the user input.  The exact space needed
+    varies slightly depending on postgres behavior.
+
+    Rather than expecting postgres to be determinstic and occasionally
+    failing the test, we permit sizes for the same data to vary by a few pages.
+    """
+
+    # Determined empirically from examples of equality failures: they differ
+    # by page multiples of 8272, and usually by 1-3 pages.  Tolerate 4 to avoid
+    # failing on outliers from that observed range.
+    threshold = 4 * 8272
+
+    assert size_a == pytest.approx(size_b, abs=threshold)
+
+
 def test_get_tenant_size_with_multiple_branches(
    neon_env_builder: NeonEnvBuilder, test_output_dir: Path
 ):
@@ -573,7 +590,7 @@ def test_get_tenant_size_with_multiple_branches(
    )

    size_after_first_branch = http_client.tenant_size(tenant_id)
-    assert size_after_first_branch == size_at_branch
+    assert_size_approx_equal(size_after_first_branch, size_at_branch)

    first_branch_endpoint = env.endpoints.create_start("first-branch", tenant_id=tenant_id)

@@ -599,7 +616,7 @@ def test_get_tenant_size_with_multiple_branches(
        "second-branch", main_branch_name, tenant_id
    )
    size_after_second_branch = http_client.tenant_size(tenant_id)
-    assert size_after_second_branch == size_after_continuing_on_main
+    assert_size_approx_equal(size_after_second_branch, size_after_continuing_on_main)

    second_branch_endpoint = env.endpoints.create_start("second-branch", tenant_id=tenant_id)

@@ -635,7 +652,7 @@ def test_get_tenant_size_with_multiple_branches(
    # tenant_size but so far this has been reliable, even though at least gc
    # and tenant_size race for the same locks
    size_after = http_client.tenant_size(tenant_id)
-    assert size_after == size_after_thinning_branch
+    assert_size_approx_equal(size_after, size_after_thinning_branch)

    size_debug_file_before = open(test_output_dir / "size_debug_before.html", "w")
    size_debug = http_client.tenant_size_debug(tenant_id)
--- a/test_runner/regress/test_tenants.py
+++ b/test_runner/regress/test_tenants.py
@@ -12,7 +12,6 @@ from fixtures.log_helper import log
 from fixtures.metrics import (
    PAGESERVER_GLOBAL_METRICS,
    PAGESERVER_PER_TENANT_METRICS,
-    PAGESERVER_PER_TENANT_REMOTE_TIMELINE_CLIENT_METRICS,
    parse_metrics,
 )
 from fixtures.neon_fixtures import (
@@ -232,17 +231,10 @@ def test_metrics_normal_work(neon_env_builder: NeonEnvBuilder):
        assert value


-@pytest.mark.parametrize(
-    "remote_storage_kind",
-    # exercise both the code paths where remote_storage=None and remote_storage=Some(...)
-    [RemoteStorageKind.NOOP, RemoteStorageKind.MOCK_S3],
-)
-def test_pageserver_metrics_removed_after_detach(
-    neon_env_builder: NeonEnvBuilder, remote_storage_kind: RemoteStorageKind
-):
+def test_pageserver_metrics_removed_after_detach(neon_env_builder: NeonEnvBuilder):
    """Tests that when a tenant is detached, the tenant specific metrics are not left behind"""

-    neon_env_builder.enable_pageserver_remote_storage(remote_storage_kind)
+    neon_env_builder.enable_pageserver_remote_storage(RemoteStorageKind.MOCK_S3)

    neon_env_builder.num_safekeepers = 3

@@ -282,9 +274,6 @@ def test_pageserver_metrics_removed_after_detach(
    for tenant in [tenant_1, tenant_2]:
        pre_detach_samples = set([x.name for x in get_ps_metric_samples_for_tenant(tenant)])
        expected = set(PAGESERVER_PER_TENANT_METRICS)
-        if remote_storage_kind == RemoteStorageKind.NOOP:
-            # if there's no remote storage configured, we don't expose the remote timeline client metrics
-            expected -= set(PAGESERVER_PER_TENANT_REMOTE_TIMELINE_CLIENT_METRICS)
        assert pre_detach_samples == expected

        env.pageserver.http_client().tenant_detach(tenant)
@@ -294,9 +283,7 @@ def test_pageserver_metrics_removed_after_detach(


 # Check that empty tenants work with or without the remote storage
-@pytest.mark.parametrize(
-    "remote_storage_kind", available_remote_storages() + [RemoteStorageKind.NOOP]
-)
+@pytest.mark.parametrize("remote_storage_kind", available_remote_storages())
 def test_pageserver_with_empty_tenants(
    neon_env_builder: NeonEnvBuilder, remote_storage_kind: RemoteStorageKind
 ):
--- a/test_runner/regress/test_timeline_delete.py
+++ b/test_runner/regress/test_timeline_delete.py
@@ -12,7 +12,6 @@ from fixtures.neon_fixtures import (
    NeonEnvBuilder,
    PgBin,
    last_flush_lsn_upload,
-    wait_for_last_flush_lsn,
 )
 from fixtures.pageserver.http import PageserverApiException
 from fixtures.pageserver.utils import (
@@ -35,6 +34,7 @@ from fixtures.remote_storage import (
 )
 from fixtures.types import Lsn, TenantId, TimelineId
 from fixtures.utils import query_scalar, run_pg_bench_small, wait_until
+from urllib3.util.retry import Retry


 def test_timeline_delete(neon_simple_env: NeonEnv):
@@ -145,19 +145,12 @@ DELETE_FAILPOINTS = [
 def combinations():
    result = []

-    remotes = [RemoteStorageKind.NOOP, RemoteStorageKind.MOCK_S3]
+    remotes = [RemoteStorageKind.MOCK_S3]
    if os.getenv("ENABLE_REAL_S3_REMOTE_STORAGE"):
        remotes.append(RemoteStorageKind.REAL_S3)

    for remote_storage_kind in remotes:
        for delete_failpoint in DELETE_FAILPOINTS:
-            if remote_storage_kind == RemoteStorageKind.NOOP and delete_failpoint in (
-                "timeline-delete-before-index-delete",
-                "timeline-delete-after-index-delete",
-            ):
-                # the above failpoints are not relevant for config without remote storage
-                continue
-
            result.append((remote_storage_kind, delete_failpoint))
    return result

@@ -205,23 +198,21 @@ def test_delete_timeline_exercise_crash_safety_failpoints(
    with env.endpoints.create_start("delete") as endpoint:
        # generate enough layers
        run_pg_bench_small(pg_bin, endpoint.connstr())
-        if remote_storage_kind is RemoteStorageKind.NOOP:
-            wait_for_last_flush_lsn(env, endpoint, env.initial_tenant, timeline_id)
-        else:
-            last_flush_lsn_upload(env, endpoint, env.initial_tenant, timeline_id)

-            if remote_storage_kind in available_s3_storages():
-                assert_prefix_not_empty(
-                    neon_env_builder,
-                    prefix="/".join(
-                        (
-                            "tenants",
-                            str(env.initial_tenant),
-                            "timelines",
-                            str(timeline_id),
-                        )
-                    ),
-                )
+        last_flush_lsn_upload(env, endpoint, env.initial_tenant, timeline_id)
+
+        if remote_storage_kind in available_s3_storages():
+            assert_prefix_not_empty(
+                neon_env_builder,
+                prefix="/".join(
+                    (
+                        "tenants",
+                        str(env.initial_tenant),
+                        "timelines",
+                        str(timeline_id),
+                    )
+                ),
+            )

    env.pageserver.allowed_errors.append(f".*{timeline_id}.*failpoint: {failpoint}")
    # It appears when we stopped flush loop during deletion and then pageserver is stopped
@@ -624,7 +615,7 @@ def test_delete_timeline_client_hangup(neon_env_builder: NeonEnvBuilder):

    child_timeline_id = env.neon_cli.create_branch("child", "main")

-    ps_http = env.pageserver.http_client()
+    ps_http = env.pageserver.http_client(retries=Retry(0, read=False))

    failpoint_name = "persist_deleted_index_part"
    ps_http.configure_failpoints((failpoint_name, "pause"))
@@ -864,7 +855,7 @@ def test_timeline_delete_resumed_on_attach(
            # error from http response is also logged
            ".*InternalServerError\\(Tenant is marked as deleted on remote storage.*",
            # Polling after attach may fail with this
-            f".*InternalServerError\\(Tenant {tenant_id} is not active.*",
+            ".*Resource temporarily unavailable.*Tenant not yet active",
            '.*shutdown_pageserver{exit_code=0}: stopping left-over name="remote upload".*',
        )
    )
--- a/test_runner/regress/test_timeline_size.py
+++ b/test_runner/regress/test_timeline_size.py
@@ -301,12 +301,8 @@ def test_timeline_initial_logical_size_calculation_cancellation(
    # message emitted by the code behind failpoint "timeline-calculate-logical-size-check-dir-exists"


-@pytest.mark.parametrize("remote_storage_kind", [None, RemoteStorageKind.LOCAL_FS])
-def test_timeline_physical_size_init(
-    neon_env_builder: NeonEnvBuilder, remote_storage_kind: Optional[RemoteStorageKind]
-):
-    if remote_storage_kind is not None:
-        neon_env_builder.enable_pageserver_remote_storage(remote_storage_kind)
+def test_timeline_physical_size_init(neon_env_builder: NeonEnvBuilder):
+    neon_env_builder.enable_pageserver_remote_storage(RemoteStorageKind.LOCAL_FS)

    env = neon_env_builder.init_start()

@@ -337,17 +333,12 @@ def test_timeline_physical_size_init(
    )

    assert_physical_size_invariants(
-        get_physical_size_values(env, env.initial_tenant, new_timeline_id, remote_storage_kind),
-        remote_storage_kind,
+        get_physical_size_values(env, env.initial_tenant, new_timeline_id),
    )


-@pytest.mark.parametrize("remote_storage_kind", [None, RemoteStorageKind.LOCAL_FS])
-def test_timeline_physical_size_post_checkpoint(
-    neon_env_builder: NeonEnvBuilder, remote_storage_kind: Optional[RemoteStorageKind]
-):
-    if remote_storage_kind is not None:
-        neon_env_builder.enable_pageserver_remote_storage(remote_storage_kind)
+def test_timeline_physical_size_post_checkpoint(neon_env_builder: NeonEnvBuilder):
+    neon_env_builder.enable_pageserver_remote_storage(RemoteStorageKind.LOCAL_FS)

    env = neon_env_builder.init_start()

@@ -369,19 +360,14 @@ def test_timeline_physical_size_post_checkpoint(

    def check():
        assert_physical_size_invariants(
-            get_physical_size_values(env, env.initial_tenant, new_timeline_id, remote_storage_kind),
-            remote_storage_kind,
+            get_physical_size_values(env, env.initial_tenant, new_timeline_id),
        )

    wait_until(10, 1, check)


-@pytest.mark.parametrize("remote_storage_kind", [None, RemoteStorageKind.LOCAL_FS])
-def test_timeline_physical_size_post_compaction(
-    neon_env_builder: NeonEnvBuilder, remote_storage_kind: Optional[RemoteStorageKind]
-):
-    if remote_storage_kind is not None:
-        neon_env_builder.enable_pageserver_remote_storage(remote_storage_kind)
+def test_timeline_physical_size_post_compaction(neon_env_builder: NeonEnvBuilder):
+    neon_env_builder.enable_pageserver_remote_storage(RemoteStorageKind.LOCAL_FS)

    # Disable background compaction as we don't want it to happen after `get_physical_size` request
    # and before checking the expected size on disk, which makes the assertion failed
@@ -420,21 +406,15 @@ def test_timeline_physical_size_post_compaction(
    pageserver_http.timeline_checkpoint(env.initial_tenant, new_timeline_id)
    pageserver_http.timeline_compact(env.initial_tenant, new_timeline_id)

-    if remote_storage_kind is not None:
-        wait_for_upload_queue_empty(pageserver_http, env.initial_tenant, new_timeline_id)
+    wait_for_upload_queue_empty(pageserver_http, env.initial_tenant, new_timeline_id)

    assert_physical_size_invariants(
-        get_physical_size_values(env, env.initial_tenant, new_timeline_id, remote_storage_kind),
-        remote_storage_kind,
+        get_physical_size_values(env, env.initial_tenant, new_timeline_id),
    )


-@pytest.mark.parametrize("remote_storage_kind", [None, RemoteStorageKind.LOCAL_FS])
-def test_timeline_physical_size_post_gc(
-    neon_env_builder: NeonEnvBuilder, remote_storage_kind: Optional[RemoteStorageKind]
-):
-    if remote_storage_kind is not None:
-        neon_env_builder.enable_pageserver_remote_storage(remote_storage_kind)
+def test_timeline_physical_size_post_gc(neon_env_builder: NeonEnvBuilder):
+    neon_env_builder.enable_pageserver_remote_storage(RemoteStorageKind.LOCAL_FS)

    # Disable background compaction and GC as we don't want it to happen after `get_physical_size` request
    # and before checking the expected size on disk, which makes the assertion failed
@@ -471,12 +451,10 @@ def test_timeline_physical_size_post_gc(
    pageserver_http.timeline_checkpoint(env.initial_tenant, new_timeline_id)
    pageserver_http.timeline_gc(env.initial_tenant, new_timeline_id, gc_horizon=None)

-    if remote_storage_kind is not None:
-        wait_for_upload_queue_empty(pageserver_http, env.initial_tenant, new_timeline_id)
+    wait_for_upload_queue_empty(pageserver_http, env.initial_tenant, new_timeline_id)

    assert_physical_size_invariants(
-        get_physical_size_values(env, env.initial_tenant, new_timeline_id, remote_storage_kind),
-        remote_storage_kind,
+        get_physical_size_values(env, env.initial_tenant, new_timeline_id),
    )


@@ -560,14 +538,10 @@ def test_timeline_size_metrics(
    assert math.isclose(dbsize_sum, tl_logical_size_metric, abs_tol=2 * 1024 * 1024)


-@pytest.mark.parametrize("remote_storage_kind", [None, RemoteStorageKind.LOCAL_FS])
-def test_tenant_physical_size(
-    neon_env_builder: NeonEnvBuilder, remote_storage_kind: Optional[RemoteStorageKind]
-):
+def test_tenant_physical_size(neon_env_builder: NeonEnvBuilder):
    random.seed(100)

-    if remote_storage_kind is not None:
-        neon_env_builder.enable_pageserver_remote_storage(remote_storage_kind)
+    neon_env_builder.enable_pageserver_remote_storage(RemoteStorageKind.LOCAL_FS)

    env = neon_env_builder.init_start()

@@ -575,12 +549,10 @@ def test_tenant_physical_size(
    client = env.pageserver.http_client()

    tenant, timeline = env.neon_cli.create_tenant()
-    if remote_storage_kind is not None:
-        wait_for_upload_queue_empty(pageserver_http, tenant, timeline)

    def get_timeline_resident_physical_size(timeline: TimelineId):
-        sizes = get_physical_size_values(env, tenant, timeline, remote_storage_kind)
-        assert_physical_size_invariants(sizes, remote_storage_kind)
+        sizes = get_physical_size_values(env, tenant, timeline)
+        assert_physical_size_invariants(sizes)
        return sizes.prometheus_resident_physical

    timeline_total_resident_physical_size = get_timeline_resident_physical_size(timeline)
@@ -600,8 +572,7 @@ def test_tenant_physical_size(
        wait_for_last_flush_lsn(env, endpoint, tenant, timeline)
        pageserver_http.timeline_checkpoint(tenant, timeline)

-        if remote_storage_kind is not None:
-            wait_for_upload_queue_empty(pageserver_http, tenant, timeline)
+        wait_for_upload_queue_empty(pageserver_http, tenant, timeline)

        timeline_total_resident_physical_size += get_timeline_resident_physical_size(timeline)

@@ -630,7 +601,6 @@ def get_physical_size_values(
    env: NeonEnv,
    tenant_id: TenantId,
    timeline_id: TimelineId,
-    remote_storage_kind: Optional[RemoteStorageKind],
 ) -> TimelinePhysicalSizeValues:
    res = TimelinePhysicalSizeValues()

@@ -646,12 +616,9 @@ def get_physical_size_values(
    res.prometheus_resident_physical = metrics.query_one(
        "pageserver_resident_physical_size", metrics_filter
    ).value
-    if remote_storage_kind is not None:
-        res.prometheus_remote_physical = metrics.query_one(
-            "pageserver_remote_physical_size", metrics_filter
-        ).value
-    else:
-        res.prometheus_remote_physical = None
+    res.prometheus_remote_physical = metrics.query_one(
+        "pageserver_remote_physical_size", metrics_filter
+    ).value

    detail = client.timeline_detail(
        tenant_id, timeline_id, include_timeline_dir_layer_file_size_sum=True
@@ -664,20 +631,15 @@ def get_physical_size_values(
    return res


-def assert_physical_size_invariants(
-    sizes: TimelinePhysicalSizeValues, remote_storage_kind: Optional[RemoteStorageKind]
-):
+def assert_physical_size_invariants(sizes: TimelinePhysicalSizeValues):
    # resident phyiscal size is defined as
    assert sizes.python_timelinedir_layerfiles_physical == sizes.prometheus_resident_physical
    assert sizes.python_timelinedir_layerfiles_physical == sizes.layer_map_file_size_sum

    # we don't do layer eviction, so, all layers are resident
    assert sizes.api_current_physical == sizes.prometheus_resident_physical
-    if remote_storage_kind is not None:
-        assert sizes.prometheus_resident_physical == sizes.prometheus_remote_physical
-        # XXX would be nice to assert layer file physical storage utilization here as well, but we can only do that for LocalFS
-    else:
-        assert sizes.prometheus_remote_physical is None
+    assert sizes.prometheus_resident_physical == sizes.prometheus_remote_physical
+    # XXX would be nice to assert layer file physical storage utilization here as well, but we can only do that for LocalFS


 # Timeline logical size initialization is an asynchronous background task that runs once,
Author	SHA1	Message	Date
Joonas Koivunen	eb3c8f6653	test: write out a walredo file in test	2023-10-04 11:48:07 +00:00
Joonas Koivunen	dee2bcca44	fix: time the reconstruction, not future creation (#5459 ) `pageserver_getpage_reconstruct_seconds` histogram had been only recording the time it takes to create a future, not await on it. Since: `eb0a698adc`.	2023-10-04 11:01:07 +01:00
Joonas Koivunen	db8ff9d64b	testing: record walredo failures to test reports (#5451 ) We have rare walredo failures with pg16. Let's introduce recording of failing walredo input in `#[cfg(feature = "testing")]`. There is additional logging (the value reconstruction path logging usually shown with not found keys), keeping it for `#[cfg(features = "testing")]`. Cc: #5404.	2023-10-04 11:24:30 +03:00
Rahul Modpur	af6a20dfc2	Improve CrashsafeOverwriteError source printing (#5410 ) ## Problem Duplication of error in log Fixes #5366 ## Summary of changes Removed `{0}` from error description above each enum due to presence of `#[source]` to avoid duplication Signed-off-by: Rahul Modpur <rmodpur2@gmail.com>	2023-10-04 02:38:42 +02:00
Alexander Bayandin	fec94ad5b3	Update checksums for pg_jsonschema & pg_graphql (#5455 ) ## Problem Folks have re-taged releases for `pg_jsonschema` and `pg_graphql` (to increase timeouts on their CI), for us, these are a noop changes, but unfortunately, this will cause our builds to fail due to checksums mismatch (this might not strike right away because of the build cache). - `8ba7c7be9d` - `aa7509370a` ## Summary of changes - `pg_jsonschema` update checksum - `pg_graphql` update checksum	2023-10-03 18:42:39 +01:00
John Spray	ace0c775fc	pageserver: prefer 503 to 500 for transient unavailability (#5439 ) ## Problem The 500 status code should only be used for bugs or unrecoverable failures: situations we did not expect. Currently, the pageserver is misusing this response code for some situations that are totally normal, like requests targeting tenants that are in the process of activating. The 503 response is a convenient catch-all for "I can't right now, but I will be able to". ## Summary of changes - Change some transient availability error conditions to return 503 instead of 500 - Update the HTTP client configuration in integration tests to retry on 503 After these changes, things like creating a tenant and then trying to create a timeline within it will no longer require carefully checking its status first, or retrying on 500s. Instead, a client which is properly configured to retry on 503 can quietly handle such situations.	2023-10-03 17:00:55 +01:00
dependabot[bot]	78dde31827	build(deps): bump urllib3 from 1.26.11 to 1.26.17 (#5442 )	2023-10-03 11:50:27 +01:00
Christian Schwarz	de0e96d2be	remote_storage: separate semaphores for read and write ops (#5440 ) Before this PR, a compaction that queues a lot of uploads could grab all the semaphore permits. Any readers that need on-demand downloads would queue up, causing getpage@lsn outliers. Internal context: https://neondb.slack.com/archives/C05NXJFNRPA/p1696264359425419?thread_ts=1696250393.840899&cid=C05NXJFNRPA	2023-10-03 11:22:11 +03:00
Alexander Bayandin	00369c8c2a	Update pg_jsonschema & pg_grapgql extensions (#5438 ) - Update `pg_jsonschema` to 0.2.0 with Postgres 16 support - Update `pg_grapgql` to 1.4.0 with Postgres 16 support - Remove `pgx` (old name of `pgrx`) layer from Dockerfile	2023-10-02 23:50:27 +01:00
Vadim Kharitonov	c1dcf61ca2	Update pgx-ulid extension (#5382 ) - Update `pgx-ulid` from 0.1.0 to 0.1.3, and add it to Postgres 16 - Add `pg_tiktoken` to Postgres 16 image Closes #5374 --------- Co-authored-by: Alexander Bayandin <alexander@neon.tech>	2023-10-02 15:52:45 +01:00
Sasha Krassovsky	89275f6c1e	Fix invalid database resulting from failed DROP DB (#5423 ) ## Problem If the control plane happened to respond to a DROP DATABASE request with a non-200 response, we'd abort the DROP DATABASE transaction in the usual spot. However, Postgres for some reason actually performs the drop inside of `standard_ProcessUtility`. As such, the database was left in a weird state after aborting the transaction. We had test coverage of a failed CREATE DATABASE but not a failed DROP DATABASE. ## Summary of changes Since DROP DATABASE can't be inside of a transaction block, we can just forward the DDL changes to the control plane inside of `ProcessUtility_hook`, and if we respond with 500 bail out of `ProcessUtility` before we perform the drop. This change also adds a test, which reproduced the invalid database issue before the fix was applied.	2023-09-29 19:39:28 +01:00
Christian Schwarz	c07eef8ea5	page_cache: find_victim: don't spin while there's no chance for a slot (#5319 ) It is wasteful to cycle through the page cache slots trying to find a victim slot if all the slots are currently un-evictable because a read / write guard is alive. We suspect this wasteful cycling to be the root cause for an "indigestion" we observed in staging (#5291). The hypothesis is that we `.await` after we get ahold of a read / write guard, and that tokio actually deschedules us in favor of another future. If that other future then needs a page slot, it can't get ours because we're holding the guard. Repeat this, and eventually, the other future(s) will find themselves doing `find_victim` until they hit `exceeded evict iter limit`. The `find_victim` is wasteful and CPU-starves the futures that are already holding the read/write guard. A `yield` inside `find_victim` could mitigate the starvation, but wouldn't fix the wasting of CPU cycles. So instead, this PR queues waiters behind a tokio semaphore that counts evictable slots. The downside is that this stops the clock page replacement if we have 0 evictable slots. Also, as explained by the big block comment in `find_victims`, the semaphore doesn't fully prevent starvation because because we can't make tokio prioritize those tasks executing `find_victim` that have been trying the longest. Implementation =============== We need to acquire the semaphore permit before locking the slot. Otherwise, we could deadlock / discover that all permits are gone and would have to relinquish the slot, having moved forward the Clock LRU without making progress. The downside is that, we never get full throughput for read-heavy workloads, because, until the reader coalesces onto an existing permit, it'll hold its own permit. Addendum To Root-Cause Analysis In #5291 ======================================== Since merging that PR, @arpad-m pointed out that we couldn't have reached the `slot.write().await` with his patches because the VirtualFile slots can't have all been write-locked, because we only hold them locked while the IO is ongoing, and the IO is still done with synchronous system calls in that patch set, so, we can have had at most $number_of_executor_threads locked at any given time. I count 3 tokio runtimes that do `Timeline::get`, each with 8 executor threads in our deployment => $number_of_executor_threads = 3*8 = 24 . But the virtual file cache has 100 slots. We both agree that nothing changed about the core hypothesis, i.e., additional await points inside VirtualFile caused higher concurrency resulting in exhaustion of page cache slots. But we'll need to reproduce the issue and investigate further to truly understand the root cause, or find out that & why we were indeed using 100 VirtualFile slots. TODO: could it be compaction that needs to hold guards of many VirtualFile's in its iterators?	2023-09-29 20:03:56 +02:00
Alexander Bayandin	86dd28d4fb	Bump hermit-abi & num_cpus packages (#5427 ) ## Problem I've noticed that `hermit-abi` 0.3.1 [1] has been yanked from crates.io (looks like nothing too bad [2]). Also, we have 2 versions of `hermit-api` in dependencies (0.3.* and 0.2.*), update `num-cpus` to use the latest `hermit-api` 0.3.3. - [1] https://crates.io/crates/hermit-abi/0.3.1 - [2] https://github.com/hermit-os/hermit-rs/issues/436 ## Summary of changes - `cargo update -p num-cpus` - `cargo update -p hermit-abi` - Unignore `RUSTSEC-2023-0052` in `deny.toml` (it has been fixed in https://github.com/neondatabase/neon/pull/5069)	2023-09-29 12:57:45 +01:00
Conrad Ludgate	fd20bbc6cb	proxy: log params when no endpoint (#5418 ) ## Problem Our SNI error dashboard features IP addresses but it's not immediately clear who that is still (#5369) ## Summary of changes Log some startup params with this error	2023-09-29 09:40:27 +01:00
John Spray	6a1903987a	tests: use approximate equality in test_get_tenant_size_with_multiple_branches (#5411 ) ## Problem This test has been flaky for a long time. As far as I can tell, the test was simply wrong to expect postgres activity to result in deterministic sizes: making the match fuzzy is not a hack, it's just matching the reality that postgres doesn't promise to write exactly the same number of pages every time it runs a given query. ## Summary of changes Equalities now tolerate up to 4 pages different. This is big enough to tolerate the deltas we've seen in practice. Closes: https://github.com/neondatabase/neon/issues/2962	2023-09-29 09:15:43 +01:00
John Spray	1881373ec4	Update CODEOWNERS (#5421 ) It is usually not intended to notify a random member of the compute team for pageserver PRs. Leaving the notification of the storage team in place, because this serves a purpose when some external contributor opens a PR and isn't sure who to ask.	2023-09-28 17:34:51 +01:00
John Spray	ca3ca2bb9c	pageserver: don't try and recover deletion queue if no remote storage (#5419 ) ## Problem Because `neon_local` by default runs with no remote storage, it was not running the deletion queue workers, and the attempt to call into `recover()` was failing. This is a bogus configuration that will go away when we make remote storage mandatory. ## Summary of changes Don't try and do deletion queue recovery when remote storage is disabled. The reason we don't just unset `control_plane_api` to avoid this is that generations will soon become mandatory, irrespective of when we make remote storage mandatory.	2023-09-28 17:20:34 +01:00
Em Sharnoff	b497d0094e	file cache: Remove free space monitor (#5406 ) This effectively reverts #3832. There's a couple issues we just discovered with the free space monitor, and to my knowledge, the fact we're putting the file cache on a separate filesystem (even when on disk) that's guaranteed to have more room than the maximum size means that this free space monitor should have no effect. More details: 1. The control plane sets the maximum file cache size based on max CU 2. The control plane sets the size of the filesystem underlying the file cache based on the maximum user selectable CU (or, if the endpoint is larger, then that size), so that there's always enough room 3. If postmaster gets SIGKILL'd, then the free space monitor process does not exit 4. If the free space monitor is acting on the cache file but not subject to locking or up-to-date metadata from a newer postgres instance, then this could lead to data corruption. So, in practice I belive the risk of data corruption is low but not nothing, and given the issues we hit because of (3), and given that this the free space monitor shouldn't be necessary because of (1) and (2), it's best to just remove it outright. See also: neondatabase/autoscaling#534, #5405	2023-09-28 06:47:44 -07:00
Conrad Ludgate	528fb1bd81	proxy: metrics2 (#5179 ) ## Problem We need to count metrics always when a connection is open. Not only when the transfer is 0. We also need to count bytes usage for HTTP. ## Summary of changes New structure for usage metrics. A `DashMap<Ids, Arc<Counters>>`. If the arc has 1 owner (the map) then I can conclude that no connections are open. If the counters has "open_connections" non zero, then I can conclude a new connection was opened in the last interval and should be reported on. Also, keep count of how many bytes processed for HTTP and report it here.	2023-09-28 11:38:26 +01:00
Joonas Koivunen	af28362a47	tests: Default to LOCAL_FS for pageserver remote storage (#5402 ) Part of #5172. Builds upon #5243, #5298. Includes the test changes: - no more RemoteStorageKind.NOOP - no more testing of pageserver without remote storage - benchmarks now use LOCAL_FS as well Support for running without RemoteStorage is still kept but in practice, there are no tests and should not be any tests. Co-authored-by: Christian Schwarz <christian@neon.tech>	2023-09-28 12:25:20 +03:00
John Spray	6b4bb91d0a	docs/rfcs: add RFC for fast tenant migration/failover (#5029 ) ## Problem Currently we don't have a way to migrate tenants from one pageserver to another without a risk of gap in availability. ## Summary of changes This follows on from https://github.com/neondatabase/neon/pull/4919 Migrating tenants between pageservers is essential to operating a service at scale, in several contexts: 1. Responding to a pageserver node failure by migrating tenants to other pageservers 2. Balancing load and capacity across pageservers, for example when a user expands their database and they need to migrate to a pageserver with more capacity. 3. Restarting pageservers for upgrades and maintenance Currently, a tenant may migrated by attaching to a new node, re-configuring endpoints to use the new node, and then later detaching from the old node. This is safe once [generation numbers](025-generation-numbers.md) are implemented, but does meet our seamless/fast/efficient goals: Co-authored-by: Christian Schwarz <christian@neon.tech>	2023-09-28 10:07:11 +01:00
Em Sharnoff	5fdc80db03	Bump vm-builder v0.17.11 -> v0.17.12 (#5407 ) Only relevant change is neondatabase/autoscaling#534 - refer there for more details.	2023-09-28 09:52:39 +02:00
Em Sharnoff	48e85460fc	vm-monitor: Unset memory.high on start + refactor cgroup handling (#5348 ) ## Problem Over the past couple days, we've had a couple VMs hit issues with postgres getting hit by memory.high throttling, even after #5303 was supposed to fix that. The tl;dr of those issues is that because vm-monitor startup sets the file cache size first, before interacting with the cgroup, cgroup throttling can mean we timeout connecting to the file cache and never reset the cgroup, even if memory has been upscaled since then. See e.g.: - https://neondb.slack.com/archives/C03F5SM1N02/p1695218132208249 - https://neondb.slack.com/archives/C03F5SM1N02/p1695314613696659 ## Summary of changes This PR adds an additional step into vm-monitor startup, where we first set the cgroup's memory.high value to 'max', removing the capacity for throttling. This preferable to just setting memory.high before the file cache, because it's theoretically possible that the new value of memory.high could still be less than the current memory usage, in which case postgres could continue to be throttled without sufficient memory events to relieve that. Implementing this properly involved adding a method to our internal cgroup interface, and it seemed like there was duplicated functionality there, so this PR unifies that as well, making things a bit more consistent.	2023-09-27 21:27:23 -07:00
Christian Schwarz	090a644392	metrics for resident & remote physical size without tenant/timeline dimension (#5389 ) So that we can compute worst-case /storage size dashboard panel more cheaply.	2023-09-27 13:18:05 +01:00
John Spray	2cced770da	pageserver: add control_plane_api_token config (#5383 ) ## Problem Control plane API calls in prod will need authentication. ## Summary of changes `control_plane_api_token` config is loaded and set as HTTP `Authorization` header. Closes: https://github.com/neondatabase/neon/issues/5139	2023-09-27 13:12:13 +01:00
MMeent	7038ce40ce	Fix neon_zeroextend's WAL logging (#5387 ) When you log more than a few blocks, you need to reserve the space in advance. We didn't do that, so we got errors. Now we do that, and shouldn't get errors.	2023-09-27 13:48:30 +02:00
Joonas Koivunen	ce45fd4cc7	test_pageserver_metric_collection: allowed synthetic size to be cancelled at shutdown (#5398 ) [evidence] of these messages during shutdown. They can happen if we are unlucky enough. [evidence]: https://neon-github-public-dev.s3.amazonaws.com/reports/main/6323709725/index.html#suites/e557ea0d920cfebd45c1921296031273/4120269a64eed172	2023-09-27 12:00:49 +01:00
Joonas Koivunen	6cc8c31fd8	disk_usage_based_eviction: switch warmup to use full table scans (#5384 ) Fixes #3978. `test_partial_evict_tenant` can fail multiple times so even though we retry it as flaky, it will still haunt us. Originally was going to just relax the comparison, then ended up replacing warming up to use full table scans instead of `pgbench --select-only`. This seems to help by producing the expected layer accesses. There might be something off with how many layers pg16 produces compared to pg14 and pg15. Created #5392.	2023-09-27 10:00:21 +01:00