Merge pull request #4803 from neondatabase/releases/2023-07-25

Release 2023-07-25
fix: count broken tenant more correct (#4800 )
2026-01-21 04:12:55 +00:00 · 2023-07-25 14:21:07 +03:00 · 2023-07-25 12:31:24 +03:00 · 2023-07-25 11:15:54 +03:00 · 2023-07-25 10:19:04 +03:00 · 2023-07-25 00:43:27 +03:00
183 changed files with 4799 additions and 2244 deletions
--- a/.cargo/config.toml
+++ b/.cargo/config.toml
@@ -12,6 +12,11 @@ opt-level = 3
 # Turn on a small amount of optimization in Development mode.
 opt-level = 1

+[build]
+# This is only present for local builds, as it will be overridden
+# by the RUSTDOCFLAGS env var in CI.
+rustdocflags = ["-Arustdoc::private_intra_doc_links"]
+
 [alias]
 build_testing = ["build", "--features", "testing"]
 neon = ["run", "--bin", "neon_local"]
--- a/.github/actions/allure-report-generate/action.yml
+++ b/.github/actions/allure-report-generate/action.yml
@@ -105,7 +105,7 @@ runs:
        # Get previously uploaded data for this run
        ZSTD_NBTHREADS=0

-        S3_FILEPATHS=$(aws s3api list-objects-v2 --bucket ${BUCKET} --prefix ${RAW_PREFIX}/ | jq --raw-output  '.Contents[].Key')
+        S3_FILEPATHS=$(aws s3api list-objects-v2 --bucket ${BUCKET} --prefix ${RAW_PREFIX}/ | jq --raw-output '.Contents[]?.Key')
        if [ -z "$S3_FILEPATHS" ]; then
          # There's no previously uploaded data for this $GITHUB_RUN_ID
          exit 0
--- a/.github/actions/run-python-test-set/action.yml
+++ b/.github/actions/run-python-test-set/action.yml
@@ -150,6 +150,14 @@ runs:
          EXTRA_PARAMS="--flaky-tests-json $TEST_OUTPUT/flaky.json $EXTRA_PARAMS"
        fi

+        # We use pytest-split plugin to run benchmarks in parallel on different CI runners
+        if [ "${TEST_SELECTION}" = "test_runner/performance" ] && [ "${{ inputs.build_type }}" != "remote" ]; then
+          mkdir -p $TEST_OUTPUT
+          poetry run ./scripts/benchmark_durations.py "${TEST_RESULT_CONNSTR}" --days 10 --output "$TEST_OUTPUT/benchmark_durations.json"
+
+          EXTRA_PARAMS="--durations-path $TEST_OUTPUT/benchmark_durations.json $EXTRA_PARAMS"
+        fi
+
        if [[ "${{ inputs.build_type }}" == "debug" ]]; then
          cov_prefix=(scripts/coverage "--profraw-prefix=$GITHUB_JOB" --dir=/tmp/coverage run)
        elif [[ "${{ inputs.build_type }}" == "release" ]]; then
--- a/.github/workflows/approved-for-ci-run.yml
+++ b/.github/workflows/approved-for-ci-run.yml
@@ -0,0 +1,55 @@
+name: Handle `approved-for-ci-run` label
+# This workflow helps to run CI pipeline for PRs made by external contributors (from forks).
+
+on:
+  pull_request:
+    types:
+      # Default types that triggers a workflow ([1]):
+      # - [1] https://docs.github.com/en/actions/using-workflows/events-that-trigger-workflows#pull_request
+      - opened
+      - synchronize
+      - reopened
+      # Types that we wand to handle in addition to keep labels tidy:
+      - closed
+      # Actual magic happens here:
+      - labeled
+
+env:
+  GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+  PR_NUMBER: ${{ github.event.pull_request.number }}
+
+jobs:
+  remove-label:
+    # Remove `approved-for-ci-run` label if the workflow is triggered by changes in a PR.
+    # The PR should be reviewed and labelled manually again.
+
+    runs-on: [ ubuntu-latest ]
+
+    if: |
+      contains(fromJSON('["opened", "synchronize", "reopened", "closed"]'), github.event.action) &&
+      contains(github.event.pull_request.labels.*.name, 'approved-for-ci-run')
+
+    steps:
+      - run: gh pr --repo "${GITHUB_REPOSITORY}" edit "${PR_NUMBER}" --remove-label "approved-for-ci-run"
+
+  create-branch:
+    # Create a local branch for an `approved-for-ci-run` labelled PR to run CI pipeline in it.
+
+    runs-on: [ ubuntu-latest ]
+
+    if: |
+      github.event.action == 'labeled' &&
+      contains(github.event.pull_request.labels.*.name, 'approved-for-ci-run')
+
+    steps:
+      - run: gh pr --repo "${GITHUB_REPOSITORY}" edit "${PR_NUMBER}" --remove-label "approved-for-ci-run"
+
+      - uses: actions/checkout@v3
+        with:
+          ref: main
+
+      - run: gh pr checkout "${PR_NUMBER}"
+
+      - run: git checkout -b "ci-run/pr-${PR_NUMBER}"
+
+      - run: git push --force origin "ci-run/pr-${PR_NUMBER}"
--- a/.github/workflows/build_and_test.yml
+++ b/.github/workflows/build_and_test.yml
@@ -5,6 +5,7 @@ on:
    branches:
      - main
      - release
+      - ci-run/pr-*
  pull_request:

 defaults:
@@ -127,6 +128,11 @@ jobs:
      - name: Run cargo clippy (release)
        run: cargo hack --feature-powerset clippy --release $CLIPPY_COMMON_ARGS

+      - name: Check documentation generation
+        run: cargo doc --workspace --no-deps --document-private-items
+        env:
+            RUSTDOCFLAGS: "-Dwarnings -Arustdoc::private_intra_doc_links"
+
      # Use `${{ !cancelled() }}` to run quck tests after the longer clippy run
      - name: Check formatting
        if: ${{ !cancelled() }}
@@ -155,7 +161,7 @@ jobs:
        build_type: [ debug, release ]
    env:
      BUILD_TYPE: ${{ matrix.build_type }}
-      GIT_VERSION: ${{ github.sha }}
+      GIT_VERSION: ${{ github.event.pull_request.head.sha || github.sha }}

    steps:
      - name: Fix git ownership
@@ -174,6 +180,27 @@ jobs:
          submodules: true
          fetch-depth: 1

+      - name: Check Postgres submodules revision
+        shell: bash -euo pipefail {0}
+        run: |
+          # This is a temporary solution to ensure that the Postgres submodules revision is correct (i.e. the updated intentionally).
+          # Eventually it will be replaced by a regression test https://github.com/neondatabase/neon/pull/4603
+
+          FAILED=false
+          for postgres in postgres-v14 postgres-v15; do
+            expected=$(cat vendor/revisions.json | jq --raw-output '."'"${postgres}"'"')
+            actual=$(git rev-parse "HEAD:vendor/${postgres}")
+            if [ "${expected}" != "${actual}" ]; then
+              echo >&2 "Expected ${postgres} rev to be at '${expected}', but it is at '${actual}'"
+              FAILED=true
+            fi
+          done
+
+          if [ "${FAILED}" = "true" ]; then
+            echo >&2 "Please update vendors/revisions.json if these changes are intentional"
+            exit 1
+          fi
+
      - name: Set pg 14 revision for caching
        id: pg_v14_rev
        run: echo pg_rev=$(git rev-parse HEAD:vendor/postgres-v14) >> $GITHUB_OUTPUT
@@ -369,13 +396,11 @@ jobs:
    strategy:
      fail-fast: false
      matrix:
+        pytest_split_group: [ 1, 2, 3, 4 ]
        build_type: [ release ]
    steps:
      - name: Checkout
        uses: actions/checkout@v3
-        with:
-          submodules: true
-          fetch-depth: 1

      - name: Pytest benchmarks
        uses: ./.github/actions/run-python-test-set
@@ -384,9 +409,11 @@ jobs:
          test_selection: performance
          run_in_parallel: false
          save_perf_report: ${{ github.ref_name == 'main' }}
+          extra_params: --splits ${{ strategy.job-total }} --group ${{ matrix.pytest_split_group }}
        env:
          VIP_VAP_ACCESS_TOKEN: "${{ secrets.VIP_VAP_ACCESS_TOKEN }}"
          PERF_TEST_RESULT_CONNSTR: "${{ secrets.PERF_TEST_RESULT_CONNSTR }}"
+          TEST_RESULT_CONNSTR: "${{ secrets.REGRESS_TEST_RESULT_CONNSTR }}"
      # XXX: no coverage data handling here, since benchmarks are run on release builds,
      # while coverage is currently collected for the debug ones

@@ -614,7 +641,7 @@ jobs:
          /kaniko/executor --reproducible --snapshot-mode=redo --skip-unused-stages --cache=true
                           --cache-repo 369495373322.dkr.ecr.eu-central-1.amazonaws.com/cache
                           --context .
-                           --build-arg GIT_VERSION=${{ github.sha }}
+                           --build-arg GIT_VERSION=${{ github.event.pull_request.head.sha || github.sha }}
                           --build-arg REPOSITORY=369495373322.dkr.ecr.eu-central-1.amazonaws.com
                           --destination 369495373322.dkr.ecr.eu-central-1.amazonaws.com/neon:${{needs.tag.outputs.build-tag}}
                           --destination neondatabase/neon:${{needs.tag.outputs.build-tag}}
@@ -658,7 +685,7 @@ jobs:
          /kaniko/executor --reproducible --snapshot-mode=redo --skip-unused-stages --cache=true
                           --cache-repo 369495373322.dkr.ecr.eu-central-1.amazonaws.com/cache
                           --context .
-                           --build-arg GIT_VERSION=${{ github.sha }}
+                           --build-arg GIT_VERSION=${{ github.event.pull_request.head.sha || github.sha }}
                           --build-arg BUILD_TAG=${{needs.tag.outputs.build-tag}}
                           --build-arg REPOSITORY=369495373322.dkr.ecr.eu-central-1.amazonaws.com
                           --dockerfile Dockerfile.compute-tools
@@ -715,7 +742,7 @@ jobs:
          /kaniko/executor --reproducible --snapshot-mode=redo --skip-unused-stages --cache=true
                           --cache-repo 369495373322.dkr.ecr.eu-central-1.amazonaws.com/cache
                           --context .
-                           --build-arg GIT_VERSION=${{ github.sha }}
+                           --build-arg GIT_VERSION=${{ github.event.pull_request.head.sha || github.sha }}
                           --build-arg PG_VERSION=${{ matrix.version }}
                           --build-arg BUILD_TAG=${{needs.tag.outputs.build-tag}}
                           --build-arg REPOSITORY=369495373322.dkr.ecr.eu-central-1.amazonaws.com
@@ -742,7 +769,7 @@ jobs:
          /kaniko/executor --reproducible --snapshot-mode=redo --skip-unused-stages --cache=true \
                           --cache-repo 369495373322.dkr.ecr.eu-central-1.amazonaws.com/cache \
                           --context . \
-                           --build-arg GIT_VERSION=${{ github.sha }} \
+                           --build-arg GIT_VERSION=${{ github.event.pull_request.head.sha || github.sha }} \
                           --build-arg PG_VERSION=${{ matrix.version }} \
                           --build-arg BUILD_TAG=${{needs.tag.outputs.build-tag}} \
                           --build-arg REPOSITORY=369495373322.dkr.ecr.eu-central-1.amazonaws.com \
@@ -767,7 +794,7 @@ jobs:
      run:
        shell: sh -eu {0}
    env:
-      VM_BUILDER_VERSION: v0.11.1
+      VM_BUILDER_VERSION: v0.13.1

    steps:
      - name: Checkout
@@ -980,6 +1007,8 @@ jobs:
          done

      - name: Upload postgres-extensions to S3
+        # TODO: Reenable step after switching to the new extensions format (tar-gzipped + index.json)
+        if: false
        run: |
          for BUCKET in $(echo ${S3_BUCKETS}); do
            aws s3 cp --recursive --only-show-errors ./extensions-to-upload s3://${BUCKET}/${{ needs.tag.outputs.build-tag }}/${{ matrix.version }}
--- a/.github/workflows/neon_extra_builds.yml
+++ b/.github/workflows/neon_extra_builds.yml
@@ -3,7 +3,8 @@ name: Check neon with extra platform builds
 on:
  push:
    branches:
-    - main
+      - main
+      - ci-run/pr-*
  pull_request:

 defaults:
--- a/Cargo.lock
+++ b/Cargo.lock
@@ -158,6 +158,19 @@ dependencies = [
 "syn 1.0.109",
 ]

+[[package]]
+name = "async-compression"
+version = "0.4.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "5b0122885821398cc923ece939e24d1056a2384ee719432397fa9db87230ff11"
+dependencies = [
+ "flate2",
+ "futures-core",
+ "memchr",
+ "pin-project-lite",
+ "tokio",
+]
+
 [[package]]
 name = "async-stream"
 version = "0.3.5"
@@ -593,7 +606,7 @@ dependencies = [
 "cc",
 "cfg-if",
 "libc",
- "miniz_oxide",
+ "miniz_oxide 0.6.2",
 "object",
 "rustc-demangle",
 ]
@@ -882,9 +895,11 @@ name = "compute_tools"
 version = "0.1.0"
 dependencies = [
 "anyhow",
+ "async-compression",
 "chrono",
 "clap",
 "compute_api",
+ "flate2",
 "futures",
 "hyper",
 "notify",
@@ -1367,6 +1382,16 @@ version = "0.4.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "0ce7134b9999ecaf8bcd65542e436736ef32ddca1b3e06094cb6ec5755203b80"

+[[package]]
+name = "flate2"
+version = "1.0.26"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "3b9429470923de8e8cbd4d2dc513535400b4b3fef0319fb5c4e1f520a7bef743"
+dependencies = [
+ "crc32fast",
+ "miniz_oxide 0.7.1",
+]
+
 [[package]]
 name = "fnv"
 version = "1.0.7"
@@ -2151,6 +2176,15 @@ dependencies = [
 "adler",
 ]

+[[package]]
+name = "miniz_oxide"
+version = "0.7.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "e7810e0be55b428ada41041c41f32c9f1a42817901b4ccf45fa3d4b6561e74c7"
+dependencies = [
+ "adler",
+]
+
 [[package]]
 name = "mio"
 version = "0.8.6"
@@ -2345,9 +2379,9 @@ dependencies = [

 [[package]]
 name = "opentelemetry"
-version = "0.18.0"
+version = "0.19.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "69d6c3d7288a106c0a363e4b0e8d308058d56902adefb16f4936f417ffef086e"
+checksum = "5f4b8347cc26099d3aeee044065ecc3ae11469796b4d65d065a23a584ed92a6f"
 dependencies = [
 "opentelemetry_api",
 "opentelemetry_sdk",
@@ -2355,9 +2389,9 @@ dependencies = [

 [[package]]
 name = "opentelemetry-http"
-version = "0.7.0"
+version = "0.8.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "1edc79add46364183ece1a4542592ca593e6421c60807232f5b8f7a31703825d"
+checksum = "a819b71d6530c4297b49b3cae2939ab3a8cc1b9f382826a1bc29dd0ca3864906"
 dependencies = [
 "async-trait",
 "bytes",
@@ -2368,9 +2402,9 @@ dependencies = [

 [[package]]
 name = "opentelemetry-otlp"
-version = "0.11.0"
+version = "0.12.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d1c928609d087790fc936a1067bdc310ae702bdf3b090c3f281b713622c8bbde"
+checksum = "8af72d59a4484654ea8eb183fea5ae4eb6a41d7ac3e3bae5f4d2a282a3a7d3ca"
 dependencies = [
 "async-trait",
 "futures",
@@ -2386,48 +2420,47 @@ dependencies = [

 [[package]]
 name = "opentelemetry-proto"
-version = "0.1.0"
+version = "0.2.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d61a2f56df5574508dd86aaca016c917489e589ece4141df1b5e349af8d66c28"
+checksum = "045f8eea8c0fa19f7d48e7bc3128a39c2e5c533d5c61298c548dfefc1064474c"
 dependencies = [
 "futures",
 "futures-util",
 "opentelemetry",
 "prost",
 "tonic 0.8.3",
- "tonic-build 0.8.4",
 ]

 [[package]]
 name = "opentelemetry-semantic-conventions"
-version = "0.10.0"
+version = "0.11.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "9b02e0230abb0ab6636d18e2ba8fa02903ea63772281340ccac18e0af3ec9eeb"
+checksum = "24e33428e6bf08c6f7fcea4ddb8e358fab0fe48ab877a87c70c6ebe20f673ce5"
 dependencies = [
 "opentelemetry",
 ]

 [[package]]
 name = "opentelemetry_api"
-version = "0.18.0"
+version = "0.19.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "c24f96e21e7acc813c7a8394ee94978929db2bcc46cf6b5014fc612bf7760c22"
+checksum = "ed41783a5bf567688eb38372f2b7a8530f5a607a4b49d38dd7573236c23ca7e2"
 dependencies = [
 "fnv",
 "futures-channel",
 "futures-util",
 "indexmap",
- "js-sys",
 "once_cell",
 "pin-project-lite",
 "thiserror",
+ "urlencoding",
 ]

 [[package]]
 name = "opentelemetry_sdk"
-version = "0.18.0"
+version = "0.19.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "1ca41c4933371b61c2a2f214bf16931499af4ec90543604ec828f7a625c09113"
+checksum = "8b3a2a91fdbfdd4d212c0dcc2ab540de2c2bcbbd90be17de7a7daf8822d010c1"
 dependencies = [
 "async-trait",
 "crossbeam-channel",
@@ -2473,6 +2506,7 @@ dependencies = [
 "pageserver",
 "postgres_ffi",
 "svg_fmt",
+ "tokio",
 "utils",
 "workspace_hack",
 ]
@@ -2482,6 +2516,7 @@ name = "pageserver"
 version = "0.1.0"
 dependencies = [
 "anyhow",
+ "async-compression",
 "async-stream",
 "async-trait",
 "byteorder",
@@ -2498,6 +2533,7 @@ dependencies = [
 "enum-map",
 "enumset",
 "fail",
+ "flate2",
 "futures",
 "git-version",
 "hex",
@@ -2509,6 +2545,7 @@ dependencies = [
 "metrics",
 "nix",
 "num-traits",
+ "num_cpus",
 "once_cell",
 "pageserver_api",
 "pin-project-lite",
@@ -2901,9 +2938,9 @@ checksum = "dc375e1527247fe1a97d8b7156678dfe7c1af2fc075c9a4db3690ecd2a148068"

 [[package]]
 name = "proc-macro2"
-version = "1.0.58"
+version = "1.0.64"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "fa1fb82fc0c281dd9671101b66b771ebbe1eaf967b96ac8740dcba4b70005ca8"
+checksum = "78803b62cbf1f46fde80d7c0e803111524b9877184cfe7c3033659490ac7a7da"
 dependencies = [
 "unicode-ident",
 ]
@@ -3292,9 +3329,9 @@ dependencies = [

 [[package]]
 name = "reqwest-tracing"
-version = "0.4.4"
+version = "0.4.5"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "783e8130d2427ddd7897dd3f814d4a3aea31b05deb42a4fdf8c18258fe5aefd1"
+checksum = "1b97ad83c2fc18113346b7158d79732242002427c30f620fa817c1f32901e0a8"
 dependencies = [
 "anyhow",
 "async-trait",
@@ -3819,7 +3856,8 @@ dependencies = [
 [[package]]
 name = "sharded-slab"
 version = "0.1.4"
-source = "git+https://github.com/neondatabase/sharded-slab.git?rev=98d16753ab01c61f0a028de44167307a00efea00#98d16753ab01c61f0a028de44167307a00efea00"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "900fba806f70c630b0a382d0d825e17a0f19fcd059a2ade1ff237bcddf446b31"
 dependencies = [
 "lazy_static",
 ]
@@ -3962,7 +4000,7 @@ dependencies = [
 "tokio",
 "tokio-stream",
 "tonic 0.9.2",
- "tonic-build 0.9.2",
+ "tonic-build",
 "tracing",
 "utils",
 "workspace_hack",
@@ -4063,7 +4101,7 @@ checksum = "4b55807c0344e1e6c04d7c965f5289c39a8d94ae23ed5c0b57aabac549f871c6"
 dependencies = [
 "filetime",
 "libc",
- "xattr",
+ "xattr 0.2.3",
 ]

 [[package]]
@@ -4344,16 +4382,17 @@ dependencies = [

 [[package]]
 name = "tokio-tar"
-version = "0.3.0"
-source = "git+https://github.com/neondatabase/tokio-tar.git?rev=404df61437de0feef49ba2ccdbdd94eb8ad6e142#404df61437de0feef49ba2ccdbdd94eb8ad6e142"
+version = "0.3.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "9d5714c010ca3e5c27114c1cdeb9d14641ace49874aa5626d7149e47aedace75"
 dependencies = [
 "filetime",
 "futures-core",
 "libc",
- "redox_syscall 0.2.16",
+ "redox_syscall 0.3.5",
 "tokio",
 "tokio-stream",
- "xattr",
+ "xattr 1.0.0",
 ]

 [[package]]
@@ -4480,19 +4519,6 @@ dependencies = [
 "tracing",
 ]

-[[package]]
-name = "tonic-build"
-version = "0.8.4"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "5bf5e9b9c0f7e0a7c027dcfaba7b2c60816c7049171f679d99ee2ff65d0de8c4"
-dependencies = [
- "prettyplease 0.1.25",
- "proc-macro2",
- "prost-build",
- "quote",
- "syn 1.0.109",
-]
-
 [[package]]
 name = "tonic-build"
 version = "0.9.2"
@@ -4616,9 +4642,9 @@ dependencies = [

 [[package]]
 name = "tracing-opentelemetry"
-version = "0.18.0"
+version = "0.19.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "21ebb87a95ea13271332df069020513ab70bdb5637ca42d6e492dc3bbbad48de"
+checksum = "00a39dcf9bfc1742fa4d6215253b33a6e474be78275884c216fc2a06267b3600"
 dependencies = [
 "once_cell",
 "opentelemetry",
@@ -4817,6 +4843,7 @@ dependencies = [
 "byteorder",
 "bytes",
 "chrono",
+ "const_format",
 "criterion",
 "futures",
 "heapless",
@@ -4842,6 +4869,7 @@ dependencies = [
 "tempfile",
 "thiserror",
 "tokio",
+ "tokio-stream",
 "tracing",
 "tracing-error",
 "tracing-subscriber",
@@ -5339,6 +5367,15 @@ dependencies = [
 "libc",
 ]

+[[package]]
+name = "xattr"
+version = "1.0.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ea263437ca03c1522846a4ddafbca2542d0ad5ed9b784909d4b27b76f62bc34a"
+dependencies = [
+ "libc",
+]
+
 [[package]]
 name = "xmlparser"
 version = "0.13.5"
--- a/Cargo.toml
+++ b/Cargo.toml
@@ -32,6 +32,8 @@ license = "Apache-2.0"
 ## All dependency versions, used in the project
 [workspace.dependencies]
 anyhow = { version = "1.0", features = ["backtrace"] }
+async-compression = { version = "0.4.0", features = ["tokio", "gzip"] }
+flate2 = "1.0.26"
 async-stream = "0.3"
 async-trait = "0.1"
 aws-config = { version = "0.55", default-features = false, features=["rustls"] }
@@ -82,9 +84,9 @@ notify = "5.0.0"
 num_cpus = "1.15"
 num-traits = "0.2.15"
 once_cell = "1.13"
-opentelemetry = "0.18.0"
-opentelemetry-otlp = { version = "0.11.0", default_features=false, features = ["http-proto", "trace", "http", "reqwest-client"] }
-opentelemetry-semantic-conventions = "0.10.0"
+opentelemetry = "0.19.0"
+opentelemetry-otlp = { version = "0.12.0", default_features=false, features = ["http-proto", "trace", "http", "reqwest-client"] }
+opentelemetry-semantic-conventions = "0.11.0"
 parking_lot = "0.12"
 pbkdf2 = "0.12.1"
 pin-project-lite = "0.2"
@@ -93,7 +95,7 @@ prost = "0.11"
 rand = "0.8"
 regex = "1.4"
 reqwest = { version = "0.11", default-features = false, features = ["rustls-tls"] }
-reqwest-tracing = { version = "0.4.0", features = ["opentelemetry_0_18"] }
+reqwest-tracing = { version = "0.4.0", features = ["opentelemetry_0_19"] }
 reqwest-middleware = "0.2.0"
 reqwest-retry = "0.2.2"
 routerify = "3"
@@ -122,13 +124,14 @@ tokio-io-timeout = "1.2.0"
 tokio-postgres-rustls = "0.9.0"
 tokio-rustls = "0.23"
 tokio-stream = "0.1"
+tokio-tar = "0.3"
 tokio-util = { version = "0.7", features = ["io"] }
 toml = "0.7"
 toml_edit = "0.19"
 tonic = {version = "0.9", features = ["tls", "tls-roots"]}
 tracing = "0.1"
 tracing-error = "0.2.0"
-tracing-opentelemetry = "0.18.0"
+tracing-opentelemetry = "0.19.0"
 tracing-subscriber = { version = "0.3", default_features = false, features = ["smallvec", "fmt", "tracing-log", "std", "env-filter"] }
 url = "2.2"
 uuid = { version = "1.2", features = ["v4", "serde"] }
@@ -146,7 +149,6 @@ postgres-native-tls = { git = "https://github.com/neondatabase/rust-postgres.git
 postgres-protocol = { git = "https://github.com/neondatabase/rust-postgres.git", rev="1aaedab101b23f7612042850d8f2036810fa7c7f" }
 postgres-types = { git = "https://github.com/neondatabase/rust-postgres.git", rev="1aaedab101b23f7612042850d8f2036810fa7c7f" }
 tokio-postgres = { git = "https://github.com/neondatabase/rust-postgres.git", rev="1aaedab101b23f7612042850d8f2036810fa7c7f" }
-tokio-tar = { git = "https://github.com/neondatabase/tokio-tar.git", rev="404df61437de0feef49ba2ccdbdd94eb8ad6e142" }

 ## Other git libraries
 heapless = { default-features=false, features=[], git = "https://github.com/japaric/heapless.git", rev = "644653bf3b831c6bb4963be2de24804acf5e5001" } # upstream release pending
@@ -183,11 +185,6 @@ tonic-build = "0.9"
 # TODO: we should probably fork `tokio-postgres-rustls` instead.
 tokio-postgres = { git = "https://github.com/neondatabase/rust-postgres.git", rev="1aaedab101b23f7612042850d8f2036810fa7c7f" }

-# Changes the MAX_THREADS limit from 4096 to 32768.
-# This is a temporary workaround for using tracing from many threads in safekeepers code,
-# until async safekeepers patch is merged to the main.
-sharded-slab = { git = "https://github.com/neondatabase/sharded-slab.git", rev="98d16753ab01c61f0a028de44167307a00efea00" }
-
 ################# Binary contents sections

 [profile.release]
--- a/Dockerfile.compute-node
+++ b/Dockerfile.compute-node
@@ -132,10 +132,20 @@ RUN wget https://github.com/plv8/plv8/archive/refs/tags/v3.1.5.tar.gz -O plv8.ta
 FROM build-deps AS h3-pg-build
 COPY --from=pg-build /usr/local/pgsql/ /usr/local/pgsql/

-# packaged cmake is too old
-RUN wget https://github.com/Kitware/CMake/releases/download/v3.24.2/cmake-3.24.2-linux-x86_64.sh \
+RUN case "$(uname -m)" in \
+      "x86_64") \
+        export CMAKE_CHECKSUM=739d372726cb23129d57a539ce1432453448816e345e1545f6127296926b6754 \
+        ;; \
+      "aarch64") \
+        export CMAKE_CHECKSUM=281b42627c9a1beed03e29706574d04c6c53fae4994472e90985ef018dd29c02 \
+        ;; \
+      *) \
+        echo "Unsupported architecture '$(uname -m)'. Supported are x86_64 and aarch64" && exit 1 \
+        ;; \
+    esac && \
+    wget https://github.com/Kitware/CMake/releases/download/v3.24.2/cmake-3.24.2-linux-$(uname -m).sh \
      -q -O /tmp/cmake-install.sh \
-      && echo "739d372726cb23129d57a539ce1432453448816e345e1545f6127296926b6754 /tmp/cmake-install.sh" | sha256sum --check \
+      && echo "${CMAKE_CHECKSUM} /tmp/cmake-install.sh" | sha256sum --check \
      && chmod u+x /tmp/cmake-install.sh \
      && /tmp/cmake-install.sh --skip-license --prefix=/usr/local/ \
      && rm /tmp/cmake-install.sh
@@ -525,10 +535,10 @@ FROM build-deps AS pg-embedding-pg-build
 COPY --from=pg-build /usr/local/pgsql/ /usr/local/pgsql/

 ENV PATH "/usr/local/pgsql/bin/:$PATH"
-# 2465f831ea1f8d49c1d74f8959adb7fc277d70cd made on 05/07/2023
+# eeb3ba7c3a60c95b2604dd543c64b2f1bb4a3703 made on 15/07/2023
 # There is no release tag yet
-RUN wget https://github.com/neondatabase/pg_embedding/archive/2465f831ea1f8d49c1d74f8959adb7fc277d70cd.tar.gz -O pg_embedding.tar.gz && \
-    echo "047af2b1f664a1e6e37867bd4eeaf5934fa27d6ba3d6c4461efa388ddf7cd1d5 pg_embedding.tar.gz" | sha256sum --check && \
+RUN wget https://github.com/neondatabase/pg_embedding/archive/eeb3ba7c3a60c95b2604dd543c64b2f1bb4a3703.tar.gz -O pg_embedding.tar.gz && \
+    echo "030846df723652f99a8689ce63b66fa0c23477a7fd723533ab8a6b28ab70730f pg_embedding.tar.gz" | sha256sum --check && \
    mkdir pg_embedding-src && cd pg_embedding-src && tar xvzf ../pg_embedding.tar.gz --strip-components=1 -C . && \
    make -j $(getconf _NPROCESSORS_ONLN) && \
    make -j $(getconf _NPROCESSORS_ONLN) install && \
--- a/compute_tools/Cargo.toml
+++ b/compute_tools/Cargo.toml
@@ -6,8 +6,10 @@ license.workspace = true

 [dependencies]
 anyhow.workspace = true
+async-compression.workspace = true
 chrono.workspace = true
 clap.workspace = true
+flate2.workspace = true
 futures.workspace = true
 hyper = { workspace = true, features = ["full"] }
 notify.workspace = true
--- a/compute_tools/src/bin/compute_ctl.rs
+++ b/compute_tools/src/bin/compute_ctl.rs
@@ -223,9 +223,8 @@ fn main() -> Result<()> {
    drop(state);

    // Launch remaining service threads
-    let _monitor_handle = launch_monitor(&compute).expect("cannot launch compute monitor thread");
-    let _configurator_handle =
-        launch_configurator(&compute).expect("cannot launch configurator thread");
+    let _monitor_handle = launch_monitor(&compute);
+    let _configurator_handle = launch_configurator(&compute);

    // Start Postgres
    let mut delay_exit = false;
--- a/compute_tools/src/compute.rs
+++ b/compute_tools/src/compute.rs
@@ -1,4 +1,5 @@
 use std::fs;
+use std::io::BufRead;
 use std::os::unix::fs::PermissionsExt;
 use std::path::Path;
 use std::process::{Command, Stdio};
@@ -15,6 +16,7 @@ use utils::lsn::Lsn;

 use compute_api::responses::{ComputeMetrics, ComputeStatus};
 use compute_api::spec::{ComputeMode, ComputeSpec};
+use utils::measured_stream::MeasuredReader;

 use crate::config;
 use crate::pg_helpers::*;
@@ -140,14 +142,14 @@ fn create_neon_superuser(spec: &ComputeSpec, client: &mut Client) -> Result<()>
        .cluster
        .roles
        .iter()
-        .map(|r| format!("'{}'", escape_literal(&r.name)))
+        .map(|r| escape_literal(&r.name))
        .collect::<Vec<_>>();

    let dbs = spec
        .cluster
        .databases
        .iter()
-        .map(|db| format!("'{}'", escape_literal(&db.name)))
+        .map(|db| escape_literal(&db.name))
        .collect::<Vec<_>>();

    let roles_decl = if roles.is_empty() {
@@ -253,20 +255,52 @@ impl ComputeNode {

        let mut client = config.connect(NoTls)?;
        let basebackup_cmd = match lsn {
-            Lsn(0) => format!("basebackup {} {}", spec.tenant_id, spec.timeline_id), // First start of the compute
-            _ => format!("basebackup {} {} {}", spec.tenant_id, spec.timeline_id, lsn),
+            // HACK We don't use compression on first start (Lsn(0)) because there's no API for it
+            Lsn(0) => format!("basebackup {} {}", spec.tenant_id, spec.timeline_id),
+            _ => format!(
+                "basebackup {} {} {} --gzip",
+                spec.tenant_id, spec.timeline_id, lsn
+            ),
        };
+
        let copyreader = client.copy_out(basebackup_cmd.as_str())?;
+        let mut measured_reader = MeasuredReader::new(copyreader);
+
+        // Check the magic number to see if it's a gzip or not. Even though
+        // we might explicitly ask for gzip, an old pageserver with no implementation
+        // of gzip compression might send us uncompressed data. After some time
+        // passes we can assume all pageservers know how to compress and we can
+        // delete this check.
+        //
+        // If the data is not gzip, it will be tar. It will not be mistakenly
+        // recognized as gzip because tar starts with an ascii encoding of a filename,
+        // and 0x1f and 0x8b are unlikely first characters for any filename. Moreover,
+        // we send the "global" directory first from the pageserver, so it definitely
+        // won't be recognized as gzip.
+        let mut bufreader = std::io::BufReader::new(&mut measured_reader);
+        let gzip = {
+            let peek = bufreader.fill_buf().unwrap();
+            peek[0] == 0x1f && peek[1] == 0x8b
+        };

        // Read the archive directly from the `CopyOutReader`
        //
        // Set `ignore_zeros` so that unpack() reads all the Copy data and
        // doesn't stop at the end-of-archive marker. Otherwise, if the server
        // sends an Error after finishing the tarball, we will not notice it.
-        let mut ar = tar::Archive::new(copyreader);
-        ar.set_ignore_zeros(true);
-        ar.unpack(&self.pgdata)?;
+        if gzip {
+            let mut ar = tar::Archive::new(flate2::read::GzDecoder::new(&mut bufreader));
+            ar.set_ignore_zeros(true);
+            ar.unpack(&self.pgdata)?;
+        } else {
+            let mut ar = tar::Archive::new(&mut bufreader);
+            ar.set_ignore_zeros(true);
+            ar.unpack(&self.pgdata)?;
+        };

+        // Report metrics
+        self.state.lock().unwrap().metrics.basebackup_bytes =
+            measured_reader.get_byte_count() as u64;
        self.state.lock().unwrap().metrics.basebackup_ms = Utc::now()
            .signed_duration_since(start_time)
            .to_std()
@@ -549,6 +583,13 @@ impl ComputeNode {
            pspec.spec.cluster.cluster_id.as_deref().unwrap_or("None")
        );

+        // Log metrics so that we can search for slow operations in logs
+        let metrics = {
+            let state = self.state.lock().unwrap();
+            state.metrics.clone()
+        };
+        info!(?metrics, "compute start finished");
+
        Ok(pg)
    }

--- a/compute_tools/src/config.rs
+++ b/compute_tools/src/config.rs
@@ -47,30 +47,22 @@ pub fn write_postgres_conf(path: &Path, spec: &ComputeSpec) -> Result<()> {
    // Add options for connecting to storage
    writeln!(file, "# Neon storage settings")?;
    if let Some(s) = &spec.pageserver_connstring {
-        writeln!(
-            file,
-            "neon.pageserver_connstring='{}'",
-            escape_conf_value(s)
-        )?;
+        writeln!(file, "neon.pageserver_connstring={}", escape_conf_value(s))?;
    }
    if !spec.safekeeper_connstrings.is_empty() {
        writeln!(
            file,
-            "neon.safekeepers='{}'",
+            "neon.safekeepers={}",
            escape_conf_value(&spec.safekeeper_connstrings.join(","))
        )?;
    }
    if let Some(s) = &spec.tenant_id {
-        writeln!(
-            file,
-            "neon.tenant_id='{}'",
-            escape_conf_value(&s.to_string())
-        )?;
+        writeln!(file, "neon.tenant_id={}", escape_conf_value(&s.to_string()))?;
    }
    if let Some(s) = &spec.timeline_id {
        writeln!(
            file,
-            "neon.timeline_id='{}'",
+            "neon.timeline_id={}",
            escape_conf_value(&s.to_string())
        )?;
    }
--- a/compute_tools/src/configurator.rs
+++ b/compute_tools/src/configurator.rs
@@ -1,7 +1,6 @@
 use std::sync::Arc;
 use std::thread;

-use anyhow::Result;
 use tracing::{error, info, instrument};

 use compute_api::responses::ComputeStatus;
@@ -42,13 +41,14 @@ fn configurator_main_loop(compute: &Arc<ComputeNode>) {
    }
 }

-pub fn launch_configurator(compute: &Arc<ComputeNode>) -> Result<thread::JoinHandle<()>> {
+pub fn launch_configurator(compute: &Arc<ComputeNode>) -> thread::JoinHandle<()> {
    let compute = Arc::clone(compute);

-    Ok(thread::Builder::new()
+    thread::Builder::new()
        .name("compute-configurator".into())
        .spawn(move || {
            configurator_main_loop(&compute);
            info!("configurator thread is exited");
-        })?)
+        })
+        .expect("cannot launch configurator thread")
 }
--- a/compute_tools/src/monitor.rs
+++ b/compute_tools/src/monitor.rs
@@ -1,7 +1,6 @@
 use std::sync::Arc;
 use std::{thread, time};

-use anyhow::Result;
 use chrono::{DateTime, Utc};
 use postgres::{Client, NoTls};
 use tracing::{debug, info};
@@ -105,10 +104,11 @@ fn watch_compute_activity(compute: &ComputeNode) {
 }

 /// Launch a separate compute monitor thread and return its `JoinHandle`.
-pub fn launch_monitor(state: &Arc<ComputeNode>) -> Result<thread::JoinHandle<()>> {
+pub fn launch_monitor(state: &Arc<ComputeNode>) -> thread::JoinHandle<()> {
    let state = Arc::clone(state);

-    Ok(thread::Builder::new()
+    thread::Builder::new()
        .name("compute-monitor".into())
-        .spawn(move || watch_compute_activity(&state))?)
+        .spawn(move || watch_compute_activity(&state))
+        .expect("cannot launch compute monitor thread")
 }
--- a/compute_tools/src/pg_helpers.rs
+++ b/compute_tools/src/pg_helpers.rs
@@ -16,15 +16,26 @@ use compute_api::spec::{Database, GenericOption, GenericOptions, PgIdent, Role};

 const POSTGRES_WAIT_TIMEOUT: Duration = Duration::from_millis(60 * 1000); // milliseconds

-/// Escape a string for including it in a SQL literal
+/// Escape a string for including it in a SQL literal. Wrapping the result
+/// with `E'{}'` or `'{}'` is not required, as it returns a ready-to-use
+/// SQL string literal, e.g. `'db'''` or `E'db\\'`.
+/// See <https://github.com/postgres/postgres/blob/da98d005cdbcd45af563d0c4ac86d0e9772cd15f/src/backend/utils/adt/quote.c#L47>
+/// for the original implementation.
 pub fn escape_literal(s: &str) -> String {
-    s.replace('\'', "''").replace('\\', "\\\\")
+    let res = s.replace('\'', "''").replace('\\', "\\\\");
+
+    if res.contains('\\') {
+        format!("E'{}'", res)
+    } else {
+        format!("'{}'", res)
+    }
 }

-/// Escape a string so that it can be used in postgresql.conf.
-/// Same as escape_literal, currently.
+/// Escape a string so that it can be used in postgresql.conf. Wrapping the result
+/// with `'{}'` is not required, as it returns a ready-to-use config string.
 pub fn escape_conf_value(s: &str) -> String {
-    s.replace('\'', "''").replace('\\', "\\\\")
+    let res = s.replace('\'', "''").replace('\\', "\\\\");
+    format!("'{}'", res)
 }

 trait GenericOptionExt {
@@ -37,7 +48,7 @@ impl GenericOptionExt for GenericOption {
    fn to_pg_option(&self) -> String {
        if let Some(val) = &self.value {
            match self.vartype.as_ref() {
-                "string" => format!("{} '{}'", self.name, escape_literal(val)),
+                "string" => format!("{} {}", self.name, escape_literal(val)),
                _ => format!("{} {}", self.name, val),
            }
        } else {
@@ -49,7 +60,7 @@ impl GenericOptionExt for GenericOption {
    fn to_pg_setting(&self) -> String {
        if let Some(val) = &self.value {
            match self.vartype.as_ref() {
-                "string" => format!("{} = '{}'", self.name, escape_conf_value(val)),
+                "string" => format!("{} = {}", self.name, escape_conf_value(val)),
                _ => format!("{} = {}", self.name, val),
            }
        } else {
--- a/compute_tools/src/spec.rs
+++ b/compute_tools/src/spec.rs
@@ -397,10 +397,44 @@ pub fn handle_databases(spec: &ComputeSpec, client: &mut Client) -> Result<()> {
                // We do not check either DB exists or not,
                // Postgres will take care of it for us
                "delete_db" => {
-                    let query: String = format!("DROP DATABASE IF EXISTS {}", &op.name.pg_quote());
+                    // In Postgres we can't drop a database if it is a template.
+                    // So we need to unset the template flag first, but it could
+                    // be a retry, so we could've already dropped the database.
+                    // Check that database exists first to make it idempotent.
+                    let unset_template_query: String = format!(
+                        "
+                        DO $$
+                        BEGIN
+                            IF EXISTS(
+                                SELECT 1
+                                FROM pg_catalog.pg_database
+                                WHERE datname = {}
+                            )
+                            THEN
+                            ALTER DATABASE {} is_template false;
+                            END IF;
+                        END
+                        $$;",
+                        escape_literal(&op.name),
+                        &op.name.pg_quote()
+                    );
+                    // Use FORCE to drop database even if there are active connections.
+                    // We run this from `cloud_admin`, so it should have enough privileges.
+                    // NB: there could be other db states, which prevent us from dropping
+                    // the database. For example, if db is used by any active subscription
+                    // or replication slot.
+                    // TODO: deal with it once we allow logical replication. Proper fix should
+                    // involve returning an error code to the control plane, so it could
+                    // figure out that this is a non-retryable error, return it to the user
+                    // and fail operation permanently.
+                    let drop_db_query: String = format!(
+                        "DROP DATABASE IF EXISTS {} WITH (FORCE)",
+                        &op.name.pg_quote()
+                    );

                    warn!("deleting database '{}'", &op.name);
-                    client.execute(query.as_str(), &[])?;
+                    client.execute(unset_template_query.as_str(), &[])?;
+                    client.execute(drop_db_query.as_str(), &[])?;
                }
                "rename_db" => {
                    let new_name = op.new_name.as_ref().unwrap();
--- a/compute_tools/tests/pg_helpers_tests.rs
+++ b/compute_tools/tests/pg_helpers_tests.rs
@@ -89,4 +89,12 @@ test.escaping = 'here''s a backslash \\ and a quote '' and a double-quote " hoor
        assert_eq!(none_generic_options.find("missed_value"), None);
        assert_eq!(none_generic_options.find("invalid_value"), None);
    }
+
+    #[test]
+    fn test_escape_literal() {
+        assert_eq!(escape_literal("test"), "'test'");
+        assert_eq!(escape_literal("test'"), "'test'''");
+        assert_eq!(escape_literal("test\\'"), "E'test\\\\'''");
+        assert_eq!(escape_literal("test\\'\\'"), "E'test\\\\''\\\\'''");
+    }
 }
--- a/control_plane/src/background_process.rs
+++ b/control_plane/src/background_process.rs
@@ -10,7 +10,7 @@
 //! (non-Neon binaries don't necessarily follow our pidfile conventions).
 //! The pid stored in the file is later used to stop the service.
 //!
-//! See [`lock_file`] module for more info.
+//! See the [`lock_file`](utils::lock_file) module for more info.

 use std::ffi::OsStr;
 use std::io::Write;
--- a/control_plane/src/broker.rs
+++ b/control_plane/src/broker.rs
@@ -2,8 +2,9 @@
 //!
 //! In the local test environment, the data for each safekeeper is stored in
 //!
+//! ```text
 //!   .neon/safekeepers/<safekeeper id>
-//!
+//! ```
 use anyhow::Context;

 use std::path::PathBuf;
--- a/control_plane/src/endpoint.rs
+++ b/control_plane/src/endpoint.rs
@@ -2,7 +2,9 @@
 //!
 //! In the local test environment, the data for each endpoint is stored in
 //!
+//! ```text
 //!   .neon/endpoints/<endpoint id>
+//! ```
 //!
 //! Some basic information about the endpoint, like the tenant and timeline IDs,
 //! are stored in the `endpoint.json` file. The `endpoint.json` file is created
@@ -22,7 +24,7 @@
 //!
 //! Directory contents:
 //!
-//! ```ignore
+//! ```text
 //! .neon/endpoints/main/
 //!     compute.log               - log output of `compute_ctl` and `postgres`
 //!     endpoint.json             - serialized `EndpointConf` struct
@@ -287,7 +289,7 @@ impl Endpoint {
                        .env
                        .safekeepers
                        .iter()
-                        .map(|sk| format!("localhost:{}", sk.pg_port))
+                        .map(|sk| format!("localhost:{}", sk.get_compute_port()))
                        .collect::<Vec<String>>()
                        .join(",");
                    conf.append("neon.safekeepers", &safekeepers);
@@ -316,7 +318,7 @@ impl Endpoint {
                    .env
                    .safekeepers
                    .iter()
-                    .map(|x| x.pg_port.to_string())
+                    .map(|x| x.get_compute_port().to_string())
                    .collect::<Vec<_>>()
                    .join(",");
                let sk_hosts = vec!["localhost"; self.env.safekeepers.len()].join(",");
@@ -461,7 +463,7 @@ impl Endpoint {
                    .iter()
                    .find(|node| node.id == sk_id)
                    .ok_or_else(|| anyhow!("safekeeper {sk_id} does not exist"))?;
-                safekeeper_connstrings.push(format!("127.0.0.1:{}", sk.pg_port));
+                safekeeper_connstrings.push(format!("127.0.0.1:{}", sk.get_compute_port()));
            }
        }

--- a/control_plane/src/local_env.rs
+++ b/control_plane/src/local_env.rs
@@ -137,6 +137,7 @@ impl Default for PageServerConf {
 pub struct SafekeeperConf {
    pub id: NodeId,
    pub pg_port: u16,
+    pub pg_tenant_only_port: Option<u16>,
    pub http_port: u16,
    pub sync: bool,
    pub remote_storage: Option<String>,
@@ -149,6 +150,7 @@ impl Default for SafekeeperConf {
        Self {
            id: NodeId(0),
            pg_port: 0,
+            pg_tenant_only_port: None,
            http_port: 0,
            sync: true,
            remote_storage: None,
@@ -158,6 +160,14 @@ impl Default for SafekeeperConf {
    }
 }

+impl SafekeeperConf {
+    /// Compute is served by port on which only tenant scoped tokens allowed, if
+    /// it is configured.
+    pub fn get_compute_port(&self) -> u16 {
+        self.pg_tenant_only_port.unwrap_or(self.pg_port)
+    }
+}
+
 impl LocalEnv {
    pub fn pg_distrib_dir_raw(&self) -> PathBuf {
        self.pg_distrib_dir.clone()
--- a/control_plane/src/safekeeper.rs
+++ b/control_plane/src/safekeeper.rs
@@ -2,8 +2,9 @@
 //!
 //! In the local test environment, the data for each safekeeper is stored in
 //!
+//! ```text
 //!   .neon/safekeepers/<safekeeper id>
-//!
+//! ```
 use std::io::Write;
 use std::path::PathBuf;
 use std::process::Child;
@@ -119,45 +120,55 @@ impl SafekeeperNode {
        let availability_zone = format!("sk-{}", id_string);

        let mut args = vec![
-            "-D",
-            datadir.to_str().with_context(|| {
-                format!("Datadir path {datadir:?} cannot be represented as a unicode string")
-            })?,
-            "--id",
-            &id_string,
-            "--listen-pg",
-            &listen_pg,
-            "--listen-http",
-            &listen_http,
-            "--availability-zone",
-            &availability_zone,
+            "-D".to_owned(),
+            datadir
+                .to_str()
+                .with_context(|| {
+                    format!("Datadir path {datadir:?} cannot be represented as a unicode string")
+                })?
+                .to_owned(),
+            "--id".to_owned(),
+            id_string,
+            "--listen-pg".to_owned(),
+            listen_pg,
+            "--listen-http".to_owned(),
+            listen_http,
+            "--availability-zone".to_owned(),
+            availability_zone,
        ];
+        if let Some(pg_tenant_only_port) = self.conf.pg_tenant_only_port {
+            let listen_pg_tenant_only = format!("127.0.0.1:{}", pg_tenant_only_port);
+            args.extend(["--listen-pg-tenant-only".to_owned(), listen_pg_tenant_only]);
+        }
        if !self.conf.sync {
-            args.push("--no-sync");
+            args.push("--no-sync".to_owned());
        }

        let broker_endpoint = format!("{}", self.env.broker.client_url());
-        args.extend(["--broker-endpoint", &broker_endpoint]);
+        args.extend(["--broker-endpoint".to_owned(), broker_endpoint]);

        let mut backup_threads = String::new();
        if let Some(threads) = self.conf.backup_threads {
            backup_threads = threads.to_string();
-            args.extend(["--backup-threads", &backup_threads]);
+            args.extend(["--backup-threads".to_owned(), backup_threads]);
        } else {
            drop(backup_threads);
        }

        if let Some(ref remote_storage) = self.conf.remote_storage {
-            args.extend(["--remote-storage", remote_storage]);
+            args.extend(["--remote-storage".to_owned(), remote_storage.clone()]);
        }

        let key_path = self.env.base_data_dir.join("auth_public_key.pem");
        if self.conf.auth_enabled {
            args.extend([
-                "--auth-validation-public-key-path",
-                key_path.to_str().with_context(|| {
-                    format!("Key path {key_path:?} cannot be represented as a unicode string")
-                })?,
+                "--auth-validation-public-key-path".to_owned(),
+                key_path
+                    .to_str()
+                    .with_context(|| {
+                        format!("Key path {key_path:?} cannot be represented as a unicode string")
+                    })?
+                    .to_owned(),
            ]);
        }

--- a/docs/pageserver-thread-mgmt.md
+++ b/docs/pageserver-thread-mgmt.md
@@ -30,8 +30,8 @@ or similar, to wake up on shutdown.

 In async Rust, futures can be "cancelled" at any await point, by
 dropping the Future. For example, `tokio::select!` returns as soon as
-one of the Futures returns, and drops the others. `tokio::timeout!` is
-another example. In the Rust ecosystem, some functions are
+one of the Futures returns, and drops the others. `tokio::time::timeout`
+is another example. In the Rust ecosystem, some functions are
 cancellation-safe, meaning they can be safely dropped without
 side-effects, while others are not. See documentation of
 `tokio::select!` for examples.
@@ -42,9 +42,9 @@ function that you call cannot be assumed to be async
 cancellation-safe, and must be polled to completion.

 The downside of non-cancellation safe code is that you have to be very
-careful when using `tokio::select!`, `tokio::timeout!`, and other such
-functions that can cause a Future to be dropped. They can only be used
-with functions that are explicitly documented to be cancellation-safe,
+careful when using `tokio::select!`, `tokio::time::timeout`, and other
+such functions that can cause a Future to be dropped. They can only be
+used with functions that are explicitly documented to be cancellation-safe,
 or you need to spawn a separate task to shield from the cancellation.

 At the entry points to the code, we also take care to poll futures to
--- a/docs/rfcs/024-user-mgmt.md
+++ b/docs/rfcs/024-user-mgmt.md
@@ -0,0 +1,84 @@
+# Postgres user and database management
+
+(This supersedes the previous proposal that looked too complicated and desynchronization-prone)
+
+We've accumulated a bunch of problems with our approach to role and database management, namely:
+
+1. we don't allow role and database creation from Postgres, and users are complaining about that
+2. fine-grained role management is not possible both from Postgres and console
+
+Right now, we do store users and databases both in console and Postgres, and there are two main reasons for
+that:
+
+* we want to be able to authenticate users in proxy against the console without Postgres' involvement. Otherwise,
+malicious brute force attempts will wake up Postgres (expensive) and may exhaust the Postgres connections limit (deny of service).
+* it is handy when we can render console UI without waking up compute (e.g., show database list)
+
+This RFC doesn't talk about giving root access to the database, which is blocked by a secure runtime setup.
+
+## Overview
+
+* Add Postgres extension that sends an HTTP request each time transaction that modifies users/databases is about to commit.
+* Add user management API to internal console API. Also, the console should put a JWT token into the compute so that it can access management API.
+
+## Postgres behavior
+
+The default user role (@username) should have `CREATE ROLE`, `CREATE DB`, and `BYPASSRLS` privileges. We expose the Postgres port
+to the open internet, so we need to check password strength. Now console generates strong passwords, so there is no risk of having dumb passwords. With user-provided passwords, such risks exist.
+
+Since we store passwords in the console we should also send unencrypted password when role is created/changed. Hence communication with the console must be encrypted. Postgres also supports creating roles using hashes, in that case, we will not be able to get a raw password. So I can see the following options here:
+  * roles created via SQL will *not* have raw passwords in the console
+  * roles created via SQL will have raw passwords in the console, except ones that were created using hashes
+
+I'm leaning towards the second option here as it is a bit more consistent one -- if raw password storage is enabled then we store passwords in all cases where we can store them.
+
+To send data about roles and databases from Postgres to the console we can create the following Postgres extension:
+
+  * Intercept role/database changes in `ProcessUtility_hook`. Here we have access to the query statement with the raw password. The hook handler itself should not dial the console immediately and rather stash info in some hashmap for later use.
+  * When the transaction is about to commit we execute collected role modifications (all as one -- console should either accept all or reject all, and hence API shouldn't be REST-like). If the console request fails we can roll back the transaction. This way if the transaction is committed we know for sure that console has this information. We can use `XACT_EVENT_PRE_COMMIT` and `XACT_EVENT_PARALLEL_PRE_COMMIT` for that.
+  * Extension should be mindful of the fact that it is possible to create and delete roles within the transaction.
+  * We also need to track who is database owner, some coding around may be needed to get the current user when the database is created.
+
+## Console user management API
+
+The current public API has REST API for role management. We need to have some analog for the internal API (called mgmt API in the console code). But unlike public API here we want to have an atomic way to create several roles/databases (in cases when several roles were created in the same transaction). So something like that may work:
+
+```
+curl -X PATCH /api/v1/roles_and_databases -d '
+[
+    {"op":"create", "type":"role", "name": "kurt", "password":"lYgT3BlbkFJ2vBZrqv"},
+    {"op":"drop", "type":"role", "name": "trout"},
+    {"op":"alter", "type":"role", "name": "kilgore", "password":"3BlbkFJ2vB"},
+    {"op":"create", "type":"database", "name": "db2", "owner": "eliot"},
+]
+'
+```
+
+Makes sense not to error out on duplicated create/delete operations (see failure modes)
+
+## Managing users from the console
+
+Now console puts a spec file with the list of databases/roles and delta operations in all the compute pods. `compute_ctl` then picks up that file and stubbornly executes deltas and checks data in the spec file is the same as in the Postgres. This way if the user creates a role in the UI we restart compute with a new spec file and during the start databases/roles are created. So if Postgres send an HTTP call each time role is created we need to break recursion in that case. We can do that based on application_name or some GUC or user (local == no HTTP hook).
+
+Generally, we have several options when we are creating users via console:
+
+1. restart compute with a new spec file, execute local SQL command; cut recursion in the extension
+2. "push" spec files into running compute, execute local SQL command; cut recursion in the extension
+3. "push" spec files into running compute, execute local SQL command; let extension create those roles in the console
+4. avoid managing roles via spec files, send SQL commands to compute; let extension create those roles in the console
+
+The last option is the most straightforward one, but with the raw password storage opt-out, we will not have the password to establish an SQL connection. Also, we need a spec for provisioning purposes and to address potential desync (but that is quite unlikely). So I think the easiest approach would be:
+
+1. keep role management like it is now and cut the recursion in the extension when SQL is executed by compute_ctl
+2. add "push" endpoint to the compute_ctl to avoid compute restart during the `apply_config` operation -- that can be done as a follow up to avoid increasing scope too much
+
+## Failure modes
+
+* during role creation via SQL role was created in the console but the connection was dropped before Postgres got acknowledgment or some error happened after acknowledgment (out of disk space, deadlock, etc):
+
+  in that case, Postgres won't have a role that exists in the console. Compute restart will heal it (due to the spec file). Also if the console allows repeated creation/deletion user can repeat the transaction.
+
+
+# Scalability
+
+On my laptop, I can create 4200 roles per second. That corresponds to 363 million roles per day. Since each role creation ends up in the console database we can add some limit to the number of roles (could be reasonably big to not run into it often -- like 1k or 10k).
--- a/docs/tools.md
+++ b/docs/tools.md
@@ -0,0 +1,22 @@
+# Useful development tools
+
+This readme contains some hints on how to set up some optional development tools.
+
+## ccls
+
+[ccls](https://github.com/MaskRay/ccls) is a c/c++ language server. It requires some setup
+to work well. There are different ways to do it but here's what works for me:
+1. Make a common parent directory for all your common neon projects. (for example, `~/src/neondatabase/`)
+2. Go to `vendor/postgres-v15`
+3. Run `make clean && ./configure`
+4. Install [bear](https://github.com/rizsotto/Bear), and run `bear -- make -j4`
+5. Copy the generated `compile_commands.json` to `~/src/neondatabase` (or equivalent)
+6. Run `touch ~/src/neondatabase/.ccls-root` this will make the `compile_commands.json` file discoverable in all subdirectories
+
+With this setup you will get decent lsp mileage inside the postgres repo, and also any postgres extensions that you put in `~/src/neondatabase/`, like `pg_embedding`, or inside `~/src/neondatabase/neon/pgxn` as well.
+
+Some additional tips for various IDEs:
+
+### Emacs
+
+To improve performance: `(setq lsp-lens-enable nil)`
--- a/libs/compute_api/src/responses.rs
+++ b/libs/compute_api/src/responses.rs
@@ -71,6 +71,7 @@ pub struct ComputeMetrics {
    pub wait_for_spec_ms: u64,
    pub sync_safekeepers_ms: u64,
    pub basebackup_ms: u64,
+    pub basebackup_bytes: u64,
    pub start_postgres_ms: u64,
    pub config_ms: u64,
    pub total_startup_ms: u64,
--- a/libs/metrics/src/lib.rs
+++ b/libs/metrics/src/lib.rs
@@ -6,6 +6,7 @@ use once_cell::sync::Lazy;
 use prometheus::core::{AtomicU64, Collector, GenericGauge, GenericGaugeVec};
 pub use prometheus::opts;
 pub use prometheus::register;
+pub use prometheus::Error;
 pub use prometheus::{core, default_registry, proto};
 pub use prometheus::{exponential_buckets, linear_buckets};
 pub use prometheus::{register_counter_vec, Counter, CounterVec};
--- a/libs/metrics/src/metric_vec_duration.rs
+++ b/libs/metrics/src/metric_vec_duration.rs
@@ -1,4 +1,4 @@
-//! Helpers for observing duration on HistogramVec / CounterVec / GaugeVec / MetricVec<T>.
+//! Helpers for observing duration on `HistogramVec` / `CounterVec` / `GaugeVec` / `MetricVec<T>`.

 use std::{future::Future, time::Instant};

--- a/libs/pageserver_api/src/models.rs
+++ b/libs/pageserver_api/src/models.rs
@@ -9,6 +9,7 @@ use serde::{Deserialize, Serialize};
 use serde_with::{serde_as, DisplayFromStr};
 use strum_macros;
 use utils::{
+    completion,
    history_buffer::HistoryBufferWithDropCounter,
    id::{NodeId, TenantId, TimelineId},
    lsn::Lsn,
@@ -76,7 +77,12 @@ pub enum TenantState {
    /// system is being shut down.
    ///
    /// Transitions out of this state are possible through `set_broken()`.
-    Stopping,
+    Stopping {
+        // Because of https://github.com/serde-rs/serde/issues/2105 this has to be a named field,
+        // otherwise it will not be skipped during deserialization
+        #[serde(skip)]
+        progress: completion::Barrier,
+    },
    /// The tenant is recognized by the pageserver, but can no longer be used for
    /// any operations.
    ///
@@ -118,7 +124,7 @@ impl TenantState {
            // Why is Stopping a Maybe case? Because, during pageserver shutdown,
            // we set the Stopping state irrespective of whether the tenant
            // has finished attaching or not.
-            Self::Stopping => Maybe,
+            Self::Stopping { .. } => Maybe,
        }
    }

@@ -411,12 +417,16 @@ pub struct LayerResidenceEvent {
    pub reason: LayerResidenceEventReason,
 }

-/// The reason for recording a given [`ResidenceEvent`].
+/// The reason for recording a given [`LayerResidenceEvent`].
 #[derive(Debug, Clone, Copy, Serialize, Deserialize)]
 pub enum LayerResidenceEventReason {
    /// The layer map is being populated, e.g. during timeline load or attach.
    /// This includes [`RemoteLayer`] objects created in [`reconcile_with_remote`].
    /// We need to record such events because there is no persistent storage for the events.
+    ///
+    // https://github.com/rust-lang/rust/issues/74481
+    /// [`RemoteLayer`]: ../../tenant/storage_layer/struct.RemoteLayer.html
+    /// [`reconcile_with_remote`]: ../../tenant/struct.Timeline.html#method.reconcile_with_remote
    LayerLoad,
    /// We just created the layer (e.g., freeze_and_flush or compaction).
    /// Such layers are always [`LayerResidenceStatus::Resident`].
@@ -924,7 +934,13 @@ mod tests {
                "Activating",
            ),
            (line!(), TenantState::Active, "Active"),
-            (line!(), TenantState::Stopping, "Stopping"),
+            (
+                line!(),
+                TenantState::Stopping {
+                    progress: utils::completion::Barrier::default(),
+                },
+                "Stopping",
+            ),
            (
                line!(),
                TenantState::Broken {
--- a/libs/pageserver_api/src/reltag.rs
+++ b/libs/pageserver_api/src/reltag.rs
@@ -60,8 +60,9 @@ impl Ord for RelTag {

 /// Display RelTag in the same format that's used in most PostgreSQL debug messages:
 ///
+/// ```text
 /// <spcnode>/<dbnode>/<relnode>[_fsm|_vm|_init]
-///
+/// ```
 impl fmt::Display for RelTag {
    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
        if let Some(forkname) = forknumber_to_name(self.forknum) {
--- a/libs/postgres_ffi/src/nonrelfile_utils.rs
+++ b/libs/postgres_ffi/src/nonrelfile_utils.rs
@@ -57,9 +57,9 @@ pub fn slru_may_delete_clogsegment(segpage: u32, cutoff_page: u32) -> bool {
 // Multixact utils

 pub fn mx_offset_to_flags_offset(xid: MultiXactId) -> usize {
-    ((xid / pg_constants::MULTIXACT_MEMBERS_PER_MEMBERGROUP as u32) as u16
-        % pg_constants::MULTIXACT_MEMBERGROUPS_PER_PAGE
-        * pg_constants::MULTIXACT_MEMBERGROUP_SIZE) as usize
+    ((xid / pg_constants::MULTIXACT_MEMBERS_PER_MEMBERGROUP as u32)
+        % pg_constants::MULTIXACT_MEMBERGROUPS_PER_PAGE as u32
+        * pg_constants::MULTIXACT_MEMBERGROUP_SIZE as u32) as usize
 }

 pub fn mx_offset_to_flags_bitshift(xid: MultiXactId) -> u16 {
@@ -81,3 +81,41 @@ fn mx_offset_to_member_page(xid: u32) -> u32 {
 pub fn mx_offset_to_member_segment(xid: u32) -> i32 {
    (mx_offset_to_member_page(xid) / pg_constants::SLRU_PAGES_PER_SEGMENT) as i32
 }
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn test_multixid_calc() {
+        // Check that the mx_offset_* functions produce the same values as the
+        // corresponding PostgreSQL C macros (MXOffsetTo*). These test values
+        // were generated by calling the PostgreSQL macros with a little C
+        // program.
+        assert_eq!(mx_offset_to_member_segment(0), 0);
+        assert_eq!(mx_offset_to_member_page(0), 0);
+        assert_eq!(mx_offset_to_flags_offset(0), 0);
+        assert_eq!(mx_offset_to_flags_bitshift(0), 0);
+        assert_eq!(mx_offset_to_member_offset(0), 4);
+        assert_eq!(mx_offset_to_member_segment(1), 0);
+        assert_eq!(mx_offset_to_member_page(1), 0);
+        assert_eq!(mx_offset_to_flags_offset(1), 0);
+        assert_eq!(mx_offset_to_flags_bitshift(1), 8);
+        assert_eq!(mx_offset_to_member_offset(1), 8);
+        assert_eq!(mx_offset_to_member_segment(123456789), 2358);
+        assert_eq!(mx_offset_to_member_page(123456789), 75462);
+        assert_eq!(mx_offset_to_flags_offset(123456789), 4780);
+        assert_eq!(mx_offset_to_flags_bitshift(123456789), 8);
+        assert_eq!(mx_offset_to_member_offset(123456789), 4788);
+        assert_eq!(mx_offset_to_member_segment(u32::MAX - 1), 82040);
+        assert_eq!(mx_offset_to_member_page(u32::MAX - 1), 2625285);
+        assert_eq!(mx_offset_to_flags_offset(u32::MAX - 1), 5160);
+        assert_eq!(mx_offset_to_flags_bitshift(u32::MAX - 1), 16);
+        assert_eq!(mx_offset_to_member_offset(u32::MAX - 1), 5172);
+        assert_eq!(mx_offset_to_member_segment(u32::MAX), 82040);
+        assert_eq!(mx_offset_to_member_page(u32::MAX), 2625285);
+        assert_eq!(mx_offset_to_flags_offset(u32::MAX), 5160);
+        assert_eq!(mx_offset_to_flags_bitshift(u32::MAX), 24);
+        assert_eq!(mx_offset_to_member_offset(u32::MAX), 5176);
+    }
+}
--- a/libs/postgres_ffi/src/relfile_utils.rs
+++ b/libs/postgres_ffi/src/relfile_utils.rs
@@ -49,14 +49,16 @@ pub fn forknumber_to_name(forknum: u8) -> Option<&'static str> {
    }
 }

-///
 /// Parse a filename of a relation file. Returns (relfilenode, forknum, segno) tuple.
 ///
 /// Formats:
+///
+/// ```text
 /// <oid>
 /// <oid>_<fork name>
 /// <oid>.<segment number>
 /// <oid>_<fork name>.<segment number>
+/// ```
 ///
 /// See functions relpath() and _mdfd_segpath() in PostgreSQL sources.
 ///
--- a/libs/pq_proto/src/framed.rs
+++ b/libs/pq_proto/src/framed.rs
@@ -5,11 +5,11 @@
 //! It is similar to what tokio_util::codec::Framed with appropriate codec
 //! provides, but `FramedReader` and `FramedWriter` read/write parts can be used
 //! separately without using split from futures::stream::StreamExt (which
-//! allocates box[1] in polling internally). tokio::io::split is used for splitting
+//! allocates a [Box] in polling internally). tokio::io::split is used for splitting
 //! instead. Plus we customize error messages more than a single type for all io
 //! calls.
 //!
-//! [1] https://docs.rs/futures-util/0.3.26/src/futures_util/lock/bilock.rs.html#107
+//! [Box]: https://docs.rs/futures-util/0.3.26/src/futures_util/lock/bilock.rs.html#107
 use bytes::{Buf, BytesMut};
 use std::{
    future::Future,
@@ -117,7 +117,7 @@ impl<S: AsyncWrite + Unpin> Framed<S> {
 impl<S: AsyncRead + AsyncWrite + Unpin> Framed<S> {
    /// Split into owned read and write parts. Beware of potential issues with
    /// using halves in different tasks on TLS stream:
-    /// https://github.com/tokio-rs/tls/issues/40
+    /// <https://github.com/tokio-rs/tls/issues/40>
    pub fn split(self) -> (FramedReader<S>, FramedWriter<S>) {
        let (read_half, write_half) = tokio::io::split(self.stream);
        let reader = FramedReader {
--- a/libs/pq_proto/src/lib.rs
+++ b/libs/pq_proto/src/lib.rs
@@ -934,6 +934,15 @@ impl<'a> BeMessage<'a> {
    }
 }

+fn terminate_code(code: &[u8; 5]) -> [u8; 6] {
+    let mut terminated = [0; 6];
+    for (i, &elem) in code.iter().enumerate() {
+        terminated[i] = elem;
+    }
+
+    terminated
+}
+
 #[cfg(test)]
 mod tests {
    use super::*;
@@ -965,12 +974,3 @@ mod tests {
        assert_eq!(split_options(&params), ["foo bar", " \\", "baz ", "lol"]);
    }
 }
-
-fn terminate_code(code: &[u8; 5]) -> [u8; 6] {
-    let mut terminated = [0; 6];
-    for (i, &elem) in code.iter().enumerate() {
-        terminated[i] = elem;
-    }
-
-    terminated
-}
--- a/libs/remote_storage/src/lib.rs
+++ b/libs/remote_storage/src/lib.rs
@@ -34,12 +34,12 @@ pub const DEFAULT_REMOTE_STORAGE_MAX_CONCURRENT_SYNCS: usize = 50;
 pub const DEFAULT_REMOTE_STORAGE_MAX_SYNC_ERRORS: u32 = 10;
 /// Currently, sync happens with AWS S3, that has two limits on requests per second:
 /// ~200 RPS for IAM services
-/// https://docs.aws.amazon.com/AmazonRDS/latest/AuroraUserGuide/UsingWithRDS.IAMDBAuth.html
+/// <https://docs.aws.amazon.com/AmazonRDS/latest/AuroraUserGuide/UsingWithRDS.IAMDBAuth.html>
 /// ~3500 PUT/COPY/POST/DELETE or 5500 GET/HEAD S3 requests
-/// https://aws.amazon.com/premiumsupport/knowledge-center/s3-request-limit-avoid-throttling/
+/// <https://aws.amazon.com/premiumsupport/knowledge-center/s3-request-limit-avoid-throttling/>
 pub const DEFAULT_REMOTE_STORAGE_S3_CONCURRENCY_LIMIT: usize = 100;
 /// No limits on the client side, which currenltly means 1000 for AWS S3.
-/// https://docs.aws.amazon.com/AmazonS3/latest/API/API_ListObjectsV2.html#API_ListObjectsV2_RequestSyntax
+/// <https://docs.aws.amazon.com/AmazonS3/latest/API/API_ListObjectsV2.html#API_ListObjectsV2_RequestSyntax>
 pub const DEFAULT_MAX_KEYS_PER_LIST_RESPONSE: Option<i32> = None;

 const REMOTE_STORAGE_PREFIX_SEPARATOR: char = '/';
@@ -50,6 +50,12 @@ const REMOTE_STORAGE_PREFIX_SEPARATOR: char = '/';
 #[derive(Debug, Clone, PartialEq, Eq, PartialOrd, Ord, Hash)]
 pub struct RemotePath(PathBuf);

+impl std::fmt::Display for RemotePath {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        write!(f, "{}", self.0.display())
+    }
+}
+
 impl RemotePath {
    pub fn new(relative_path: &Path) -> anyhow::Result<Self> {
        anyhow::ensure!(
--- a/libs/remote_storage/src/local_fs.rs
+++ b/libs/remote_storage/src/local_fs.rs
@@ -7,6 +7,7 @@
 use std::{
    borrow::Cow,
    future::Future,
+    io::ErrorKind,
    path::{Path, PathBuf},
    pin::Pin,
 };
@@ -150,10 +151,7 @@ impl RemoteStorage for LocalFs {
        let mut files = vec![];
        let mut directory_queue = vec![full_path.clone()];

-        while !directory_queue.is_empty() {
-            let cur_folder = directory_queue
-                .pop()
-                .expect("queue cannot be empty: we just checked");
+        while let Some(cur_folder) = directory_queue.pop() {
            let mut entries = fs::read_dir(cur_folder.clone()).await?;
            while let Some(entry) = entries.next_entry().await? {
                let file_name: PathBuf = entry.file_name().into();
@@ -343,18 +341,14 @@ impl RemoteStorage for LocalFs {

    async fn delete(&self, path: &RemotePath) -> anyhow::Result<()> {
        let file_path = path.with_base(&self.storage_root);
-        if !file_path.exists() {
+        match fs::remove_file(&file_path).await {
+            Ok(()) => Ok(()),
+            // The file doesn't exist. This shouldn't yield an error to mirror S3's behaviour.
            // See https://docs.aws.amazon.com/AmazonS3/latest/API/API_DeleteObject.html
            // > If there isn't a null version, Amazon S3 does not remove any objects but will still respond that the command was successful.
-            return Ok(());
+            Err(e) if e.kind() == ErrorKind::NotFound => Ok(()),
+            Err(e) => Err(anyhow::anyhow!(e)),
        }
-
-        if !file_path.is_file() {
-            anyhow::bail!("{file_path:?} is not a file");
-        }
-        Ok(fs::remove_file(file_path)
-            .await
-            .map_err(|e| anyhow::anyhow!(e))?)
    }

    async fn delete_objects<'a>(&self, paths: &'a [RemotePath]) -> anyhow::Result<()> {
--- a/libs/tenant_size_model/src/calculation.rs
+++ b/libs/tenant_size_model/src/calculation.rs
@@ -21,7 +21,7 @@ use crate::{SegmentMethod, SegmentSizeResult, SizeResult, StorageModel};
 // 2. D+C+a+b
 // 3. D+A+B

-/// [`Segment`] which has had it's size calculated.
+/// `Segment` which has had its size calculated.
 #[derive(Clone, Debug)]
 struct SegmentSize {
    method: SegmentMethod,
--- a/libs/tracing-utils/src/http.rs
+++ b/libs/tracing-utils/src/http.rs
@@ -33,7 +33,7 @@ pub enum OtelName<'a> {
 /// directly into HTTP servers. However, I couldn't find one for Hyper,
 /// so I had to write our own. OpenTelemetry website has a registry of
 /// instrumentation libraries at:
-/// https://opentelemetry.io/registry/?language=rust&component=instrumentation
+/// <https://opentelemetry.io/registry/?language=rust&component=instrumentation>
 /// If a Hyper crate appears, consider switching to that.
 pub async fn tracing_handler<F, R>(
    req: Request<Body>,
--- a/libs/utils/Cargo.toml
+++ b/libs/utils/Cargo.toml
@@ -40,6 +40,12 @@ pq_proto.workspace = true
 metrics.workspace = true
 workspace_hack.workspace = true

+const_format.workspace = true
+
+# to use tokio channels as streams, this is faster to compile than async_stream
+# why is it only here? no other crate should use it, streams are rarely needed.
+tokio-stream = { version = "0.1.14" }
+
 [dev-dependencies]
 byteorder.workspace = true
 bytes.workspace = true
--- a/libs/utils/src/auth.rs
+++ b/libs/utils/src/auth.rs
@@ -16,7 +16,7 @@ use crate::id::TenantId;
 /// Algorithm to use. We require EdDSA.
 const STORAGE_TOKEN_ALGORITHM: Algorithm = Algorithm::EdDSA;

-#[derive(Debug, Serialize, Deserialize, Clone, PartialEq)]
+#[derive(Debug, Serialize, Deserialize, Clone, Copy, PartialEq)]
 #[serde(rename_all = "lowercase")]
 pub enum Scope {
    // Provides access to all data for a specific tenant (specified in `struct Claims` below)
--- a/libs/utils/src/completion.rs
+++ b/libs/utils/src/completion.rs
@@ -12,6 +12,13 @@ pub struct Completion(mpsc::Sender<()>);
 #[derive(Clone)]
 pub struct Barrier(Arc<Mutex<mpsc::Receiver<()>>>);

+impl Default for Barrier {
+    fn default() -> Self {
+        let (_, rx) = channel();
+        rx
+    }
+}
+
 impl Barrier {
    pub async fn wait(self) {
        self.0.lock().await.recv().await;
@@ -24,6 +31,15 @@ impl Barrier {
    }
 }

+impl PartialEq for Barrier {
+    fn eq(&self, other: &Self) -> bool {
+        // we don't use dyn so this is good
+        Arc::ptr_eq(&self.0, &other.0)
+    }
+}
+
+impl Eq for Barrier {}
+
 /// Create new Guard and Barrier pair.
 pub fn channel() -> (Completion, Barrier) {
    let (tx, rx) = mpsc::channel::<()>(1);
--- a/libs/utils/src/error.rs
+++ b/libs/utils/src/error.rs
@@ -0,0 +1,111 @@
+/// Create a reporter for an error that outputs similar to [`anyhow::Error`] with Display with alternative setting.
+///
+/// It can be used with `anyhow::Error` as well.
+///
+/// Why would one use this instead of converting to `anyhow::Error` on the spot? Because
+/// anyhow::Error would also capture a stacktrace on the spot, which you would later discard after
+/// formatting.
+///
+/// ## Usage
+///
+/// ```rust
+/// #[derive(Debug, thiserror::Error)]
+/// enum MyCoolError {
+///   #[error("should never happen")]
+///   Bad(#[source] std::io::Error),
+/// }
+///
+/// # fn failing_call() -> Result<(), MyCoolError> { Err(MyCoolError::Bad(std::io::ErrorKind::PermissionDenied.into())) }
+///
+/// # fn main() {
+/// use utils::error::report_compact_sources;
+///
+/// if let Err(e) = failing_call() {
+///     let e = report_compact_sources(&e);
+///     assert_eq!(format!("{e}"), "should never happen: permission denied");
+/// }
+/// # }
+/// ```
+///
+/// ## TODO
+///
+/// When we are able to describe return position impl trait in traits, this should of course be an
+/// extension trait. Until then avoid boxing with this more ackward interface.
+pub fn report_compact_sources<E: std::error::Error>(e: &E) -> impl std::fmt::Display + '_ {
+    struct AnyhowDisplayAlternateAlike<'a, E>(&'a E);
+
+    impl<E: std::error::Error> std::fmt::Display for AnyhowDisplayAlternateAlike<'_, E> {
+        fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+            write!(f, "{}", self.0)?;
+
+            // why is E a generic parameter here? hope that rustc will see through a default
+            // Error::source implementation and leave the following out if there cannot be any
+            // sources:
+            Sources(self.0.source()).try_for_each(|src| write!(f, ": {}", src))
+        }
+    }
+
+    struct Sources<'a>(Option<&'a (dyn std::error::Error + 'static)>);
+
+    impl<'a> Iterator for Sources<'a> {
+        type Item = &'a (dyn std::error::Error + 'static);
+
+        fn next(&mut self) -> Option<Self::Item> {
+            let rem = self.0;
+
+            let next = self.0.and_then(|x| x.source());
+            self.0 = next;
+            rem
+        }
+    }
+
+    AnyhowDisplayAlternateAlike(e)
+}
+
+#[cfg(test)]
+mod tests {
+    use super::report_compact_sources;
+
+    #[test]
+    fn report_compact_sources_examples() {
+        use std::fmt::Write;
+
+        #[derive(Debug, thiserror::Error)]
+        enum EvictionError {
+            #[error("cannot evict a remote layer")]
+            CannotEvictRemoteLayer,
+            #[error("stat failed")]
+            StatFailed(#[source] std::io::Error),
+            #[error("layer was no longer part of LayerMap")]
+            LayerNotFound(#[source] anyhow::Error),
+        }
+
+        let examples = [
+            (
+                line!(),
+                EvictionError::CannotEvictRemoteLayer,
+                "cannot evict a remote layer",
+            ),
+            (
+                line!(),
+                EvictionError::StatFailed(std::io::ErrorKind::PermissionDenied.into()),
+                "stat failed: permission denied",
+            ),
+            (
+                line!(),
+                EvictionError::LayerNotFound(anyhow::anyhow!("foobar")),
+                "layer was no longer part of LayerMap: foobar",
+            ),
+        ];
+
+        let mut s = String::new();
+
+        for (line, example, expected) in examples {
+            s.clear();
+
+            write!(s, "{}", report_compact_sources(&example)).expect("string grows");
+
+            assert_eq!(s, expected, "example on line {line}");
+        }
+    }
+}
--- a/libs/utils/src/http/endpoint.rs
+++ b/libs/utils/src/http/endpoint.rs
@@ -9,7 +9,6 @@ use metrics::{register_int_counter, Encoder, IntCounter, TextEncoder};
 use once_cell::sync::Lazy;
 use routerify::ext::RequestExt;
 use routerify::{Middleware, RequestInfo, Router, RouterBuilder};
-use tokio::task::JoinError;
 use tracing::{self, debug, info, info_span, warn, Instrument};

 use std::future::Future;
@@ -148,26 +147,140 @@ impl Drop for RequestCancelled {
 }

 async fn prometheus_metrics_handler(_req: Request<Body>) -> Result<Response<Body>, ApiError> {
+    use bytes::{Bytes, BytesMut};
+    use std::io::Write as _;
+    use tokio::sync::mpsc;
+    use tokio_stream::wrappers::ReceiverStream;
+
    SERVE_METRICS_COUNT.inc();

-    let mut buffer = vec![];
-    let encoder = TextEncoder::new();
+    /// An [`std::io::Write`] implementation on top of a channel sending [`bytes::Bytes`] chunks.
+    struct ChannelWriter {
+        buffer: BytesMut,
+        tx: mpsc::Sender<std::io::Result<Bytes>>,
+        written: usize,
+    }

-    let metrics = tokio::task::spawn_blocking(move || {
-        // Currently we take a lot of mutexes while collecting metrics, so it's
-        // better to spawn a blocking task to avoid blocking the event loop.
-        metrics::gather()
-    })
-    .await
-    .map_err(|e: JoinError| ApiError::InternalServerError(e.into()))?;
-    encoder.encode(&metrics, &mut buffer).unwrap();
+    impl ChannelWriter {
+        fn new(buf_len: usize, tx: mpsc::Sender<std::io::Result<Bytes>>) -> Self {
+            assert_ne!(buf_len, 0);
+            ChannelWriter {
+                // split about half off the buffer from the start, because we flush depending on
+                // capacity. first flush will come sooner than without this, but now resizes will
+                // have better chance of picking up the "other" half. not guaranteed of course.
+                buffer: BytesMut::with_capacity(buf_len).split_off(buf_len / 2),
+                tx,
+                written: 0,
+            }
+        }
+
+        fn flush0(&mut self) -> std::io::Result<usize> {
+            let n = self.buffer.len();
+            if n == 0 {
+                return Ok(0);
+            }
+
+            tracing::trace!(n, "flushing");
+            let ready = self.buffer.split().freeze();
+
+            // not ideal to call from blocking code to block_on, but we are sure that this
+            // operation does not spawn_blocking other tasks
+            let res: Result<(), ()> = tokio::runtime::Handle::current().block_on(async {
+                self.tx.send(Ok(ready)).await.map_err(|_| ())?;
+
+                // throttle sending to allow reuse of our buffer in `write`.
+                self.tx.reserve().await.map_err(|_| ())?;
+
+                // now the response task has picked up the buffer and hopefully started
+                // sending it to the client.
+                Ok(())
+            });
+            if res.is_err() {
+                return Err(std::io::ErrorKind::BrokenPipe.into());
+            }
+            self.written += n;
+            Ok(n)
+        }
+
+        fn flushed_bytes(&self) -> usize {
+            self.written
+        }
+    }
+
+    impl std::io::Write for ChannelWriter {
+        fn write(&mut self, mut buf: &[u8]) -> std::io::Result<usize> {
+            let remaining = self.buffer.capacity() - self.buffer.len();
+
+            let out_of_space = remaining < buf.len();
+
+            let original_len = buf.len();
+
+            if out_of_space {
+                let can_still_fit = buf.len() - remaining;
+                self.buffer.extend_from_slice(&buf[..can_still_fit]);
+                buf = &buf[can_still_fit..];
+                self.flush0()?;
+            }
+
+            // assume that this will often under normal operation just move the pointer back to the
+            // beginning of allocation, because previous split off parts are already sent and
+            // dropped.
+            self.buffer.extend_from_slice(buf);
+            Ok(original_len)
+        }
+
+        fn flush(&mut self) -> std::io::Result<()> {
+            self.flush0().map(|_| ())
+        }
+    }
+
+    let started_at = std::time::Instant::now();
+
+    let (tx, rx) = mpsc::channel(1);
+
+    let body = Body::wrap_stream(ReceiverStream::new(rx));
+
+    let mut writer = ChannelWriter::new(128 * 1024, tx);
+
+    let encoder = TextEncoder::new();

    let response = Response::builder()
        .status(200)
        .header(CONTENT_TYPE, encoder.format_type())
-        .body(Body::from(buffer))
+        .body(body)
        .unwrap();

+    let span = info_span!("blocking");
+    tokio::task::spawn_blocking(move || {
+        let _span = span.entered();
+        let metrics = metrics::gather();
+        let res = encoder
+            .encode(&metrics, &mut writer)
+            .and_then(|_| writer.flush().map_err(|e| e.into()));
+
+        match res {
+            Ok(()) => {
+                tracing::info!(
+                    bytes = writer.flushed_bytes(),
+                    elapsed_ms = started_at.elapsed().as_millis(),
+                    "responded /metrics"
+                );
+            }
+            Err(e) => {
+                tracing::warn!("failed to write out /metrics response: {e:#}");
+                // semantics of this error are quite... unclear. we want to error the stream out to
+                // abort the response to somehow notify the client that we failed.
+                //
+                // though, most likely the reason for failure is that the receiver is already gone.
+                drop(
+                    writer
+                        .tx
+                        .blocking_send(Err(std::io::ErrorKind::BrokenPipe.into())),
+                );
+            }
+        }
+    });
+
    Ok(response)
 }

--- a/libs/utils/src/http/json.rs
+++ b/libs/utils/src/http/json.rs
@@ -14,7 +14,7 @@ pub async fn json_request<T: for<'de> Deserialize<'de>>(
        .map_err(ApiError::BadRequest)
 }

-/// Will be removed as part of https://github.com/neondatabase/neon/issues/4282
+/// Will be removed as part of <https://github.com/neondatabase/neon/issues/4282>
 pub async fn json_request_or_empty_body<T: for<'de> Deserialize<'de>>(
    request: &mut Request<Body>,
 ) -> Result<Option<T>, ApiError> {
--- a/libs/utils/src/lib.rs
+++ b/libs/utils/src/lib.rs
@@ -63,6 +63,9 @@ pub mod rate_limit;
 /// Simple once-barrier and a guard which keeps barrier awaiting.
 pub mod completion;

+/// Reporting utilities
+pub mod error;
+
 mod failpoint_macro_helpers {

    /// use with fail::cfg("$name", "return(2000)")
@@ -109,10 +112,16 @@ pub use failpoint_macro_helpers::failpoint_sleep_helper;
 /// * building in docker (either in CI or locally)
 ///
 /// One thing to note is that .git is not available in docker (and it is bad to include it there).
-/// So everything becides docker build is covered by git_version crate, and docker uses a `GIT_VERSION` argument to get the value required.
-/// It takes variable from build process env and puts it to the rustc env. And then we can retrieve it here by using env! macro.
-/// Git version received from environment variable used as a fallback in git_version invocation.
-/// And to avoid running buildscript every recompilation, we use rerun-if-env-changed option.
+/// When building locally, the `git_version` is used to query .git. When building on CI and docker,
+/// we don't build the actual PR branch commits, but always a "phantom" would be merge commit to
+/// the target branch -- the actual PR commit from which we build from is supplied as GIT_VERSION
+/// environment variable.
+///
+/// We ended up with this compromise between phantom would be merge commits vs. pull request branch
+/// heads due to old logs becoming more reliable (github could gc the phantom merge commit
+/// anytime) in #4641.
+///
+/// To avoid running buildscript every recompilation, we use rerun-if-env-changed option.
 /// So the build script will be run only when GIT_VERSION envvar has changed.
 ///
 /// Why not to use buildscript to get git commit sha directly without procmacro from different crate?
@@ -124,25 +133,36 @@ pub use failpoint_macro_helpers::failpoint_sleep_helper;
 /// Note that with git_version prefix is `git:` and in case of git version from env its `git-env:`.
 ///
 /// #############################################################################################
-/// TODO this macro is not the way the library is intended to be used, see https://github.com/neondatabase/neon/issues/1565 for details.
-/// We use `cachepot` to reduce our current CI build times: https://github.com/neondatabase/cloud/pull/1033#issuecomment-1100935036
+/// TODO this macro is not the way the library is intended to be used, see <https://github.com/neondatabase/neon/issues/1565> for details.
+/// We use `cachepot` to reduce our current CI build times: <https://github.com/neondatabase/cloud/pull/1033#issuecomment-1100935036>
 /// Yet, it seems to ignore the GIT_VERSION env variable, passed to Docker build, even with build.rs that contains
 /// `println!("cargo:rerun-if-env-changed=GIT_VERSION");` code for cachepot cache invalidation.
 /// The problem needs further investigation and regular `const` declaration instead of a macro.
 #[macro_export]
 macro_rules! project_git_version {
    ($const_identifier:ident) => {
-        const $const_identifier: &str = git_version::git_version!(
-            prefix = "git:",
-            fallback = concat!(
-                "git-env:",
-                env!("GIT_VERSION", "Missing GIT_VERSION envvar")
-            ),
-            args = ["--abbrev=40", "--always", "--dirty=-modified"] // always use full sha
-        );
+        // this should try GIT_VERSION first only then git_version::git_version!
+        const $const_identifier: &::core::primitive::str = {
+            const __COMMIT_FROM_GIT: &::core::primitive::str = git_version::git_version! {
+                prefix = "",
+                fallback = "unknown",
+                args = ["--abbrev=40", "--always", "--dirty=-modified"] // always use full sha
+            };
+
+            const __ARG: &[&::core::primitive::str; 2] = &match ::core::option_env!("GIT_VERSION") {
+                ::core::option::Option::Some(x) => ["git-env:", x],
+                ::core::option::Option::None => ["git:", __COMMIT_FROM_GIT],
+            };
+
+            $crate::__const_format::concatcp!(__ARG[0], __ARG[1])
+        };
    };
 }

+/// Re-export for `project_git_version` macro
+#[doc(hidden)]
+pub use const_format as __const_format;
+
 /// Same as `assert!`, but evaluated during compilation and gets optimized out in runtime.
 #[macro_export]
 macro_rules! const_assert {
--- a/libs/utils/src/lock_file.rs
+++ b/libs/utils/src/lock_file.rs
@@ -1,9 +1,10 @@
 //! A module to create and read lock files.
 //!
 //! File locking is done using [`fcntl::flock`] exclusive locks.
-//! The only consumer of this module is currently [`pid_file`].
-//! See the module-level comment there for potential pitfalls
-//! with lock files that are used to store PIDs (pidfiles).
+//! The only consumer of this module is currently
+//! [`pid_file`](crate::pid_file). See the module-level comment
+//! there for potential pitfalls with lock files that are used
+//! to store PIDs (pidfiles).

 use std::{
    fs,
@@ -81,7 +82,7 @@ pub fn create_exclusive(lock_file_path: &Path) -> anyhow::Result<UnwrittenLockFi
 }

 /// Returned by [`read_and_hold_lock_file`].
-/// Check out the [`pid_file`] module for what the variants mean
+/// Check out the [`pid_file`](crate::pid_file) module for what the variants mean
 /// and potential caveats if the lock files that are used to store PIDs.
 pub enum LockFileRead {
    /// No file exists at the given path.
--- a/libs/utils/src/logging.rs
+++ b/libs/utils/src/logging.rs
@@ -112,7 +112,7 @@ pub fn init(
 ///
 /// When the return value is dropped, the hook is reverted to std default hook (prints to stderr).
 /// If the assumptions about the initialization order are not held, use
-/// [`TracingPanicHookGuard::disarm`] but keep in mind, if tracing is stopped, then panics will be
+/// [`TracingPanicHookGuard::forget`] but keep in mind, if tracing is stopped, then panics will be
 /// lost.
 #[must_use]
 pub fn replace_panic_hook_with_tracing_panic_hook() -> TracingPanicHookGuard {
--- a/libs/utils/src/measured_stream.rs
+++ b/libs/utils/src/measured_stream.rs
@@ -1,4 +1,5 @@
 use pin_project_lite::pin_project;
+use std::io::Read;
 use std::pin::Pin;
 use std::{io, task};
 use tokio::io::{AsyncRead, AsyncWrite, ReadBuf};
@@ -75,3 +76,34 @@ impl<S: AsyncWrite + Unpin, R, W: FnMut(usize)> AsyncWrite for MeasuredStream<S,
        self.project().stream.poll_shutdown(context)
    }
 }
+
+/// Wrapper for a reader that counts bytes read.
+///
+/// Similar to MeasuredStream but it's one way and it's sync
+pub struct MeasuredReader<R: Read> {
+    inner: R,
+    byte_count: usize,
+}
+
+impl<R: Read> MeasuredReader<R> {
+    pub fn new(reader: R) -> Self {
+        Self {
+            inner: reader,
+            byte_count: 0,
+        }
+    }
+
+    pub fn get_byte_count(&self) -> usize {
+        self.byte_count
+    }
+}
+
+impl<R: Read> Read for MeasuredReader<R> {
+    fn read(&mut self, buf: &mut [u8]) -> std::io::Result<usize> {
+        let result = self.inner.read(buf);
+        if let Ok(n_bytes) = result {
+            self.byte_count += n_bytes
+        }
+        result
+    }
+}
--- a/libs/utils/src/seqwait.rs
+++ b/libs/utils/src/seqwait.rs
@@ -23,9 +23,9 @@ pub enum SeqWaitError {

 /// Monotonically increasing value
 ///
-/// It is handy to store some other fields under the same mutex in SeqWait<S>
+/// It is handy to store some other fields under the same mutex in `SeqWait<S>`
 /// (e.g. store prev_record_lsn). So we allow SeqWait to be parametrized with
-/// any type that can expose counter. <V> is the type of exposed counter.
+/// any type that can expose counter. `V` is the type of exposed counter.
 pub trait MonotonicCounter<V> {
    /// Bump counter value and check that it goes forward
    /// N.B.: new_val is an actual new value, not a difference.
@@ -90,7 +90,7 @@ impl<T: Ord> Eq for Waiter<T> {}
 /// [`wait_for`]: SeqWait::wait_for
 /// [`advance`]: SeqWait::advance
 ///
-/// <S> means Storage, <V> is type of counter that this storage exposes.
+/// `S` means Storage, `V` is type of counter that this storage exposes.
 ///
 pub struct SeqWait<S, V>
 where
--- a/libs/utils/src/tracing_span_assert.rs
+++ b/libs/utils/src/tracing_span_assert.rs
@@ -1,8 +1,15 @@
 //! Assert that the current [`tracing::Span`] has a given set of fields.
 //!
+//! Can only produce meaningful positive results when tracing has been configured as in example.
+//! Absence of `tracing_error::ErrorLayer` is not detected yet.
+//!
+//! `#[cfg(test)]` code will get a pass when using the `check_fields_present` macro in case tracing
+//! is completly unconfigured.
+//!
 //! # Usage
 //!
-//! ```
+//! ```rust
+//! # fn main() {
 //! use tracing_subscriber::prelude::*;
 //! let registry = tracing_subscriber::registry()
 //!    .with(tracing_error::ErrorLayer::default());
@@ -20,23 +27,18 @@
 //!
 //! use utils::tracing_span_assert::{check_fields_present, MultiNameExtractor};
 //! let extractor = MultiNameExtractor::new("TestExtractor", ["test", "test_id"]);
-//! match check_fields_present([&extractor]) {
-//!    Ok(()) => {},
-//!    Err(missing) => {
-//!        panic!("Missing fields: {:?}", missing.into_iter().map(|f| f.name() ).collect::<Vec<_>>());
-//!    }
+//! if let Err(missing) = check_fields_present!([&extractor]) {
+//!    // if you copypaste this to a custom assert method, remember to add #[track_caller]
+//!    // to get the "user" code location for the panic.
+//!    panic!("Missing fields: {missing:?}");
 //! }
+//! # }
 //! ```
 //!
-//! Recommended reading: https://docs.rs/tracing-subscriber/0.3.16/tracing_subscriber/layer/index.html#per-layer-filtering
+//! Recommended reading: <https://docs.rs/tracing-subscriber/0.3.16/tracing_subscriber/layer/index.html#per-layer-filtering>
 //!

-use std::{
-    collections::HashSet,
-    fmt::{self},
-    hash::{Hash, Hasher},
-};
-
+#[derive(Debug)]
 pub enum ExtractionResult {
    Present,
    Absent,
@@ -71,51 +73,103 @@ impl<const L: usize> Extractor for MultiNameExtractor<L> {
    }
 }

-struct MemoryIdentity<'a>(&'a dyn Extractor);
-
-impl<'a> MemoryIdentity<'a> {
-    fn as_ptr(&self) -> *const () {
-        self.0 as *const _ as *const ()
-    }
-}
-impl<'a> PartialEq for MemoryIdentity<'a> {
-    fn eq(&self, other: &Self) -> bool {
-        self.as_ptr() == other.as_ptr()
-    }
-}
-impl<'a> Eq for MemoryIdentity<'a> {}
-impl<'a> Hash for MemoryIdentity<'a> {
-    fn hash<H: Hasher>(&self, state: &mut H) {
-        self.as_ptr().hash(state);
-    }
-}
-impl<'a> fmt::Debug for MemoryIdentity<'a> {
-    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> std::fmt::Result {
-        write!(f, "{:p}: {}", self.as_ptr(), self.0.name())
-    }
-}
-
-/// The extractor names passed as keys to [`new`].
-pub fn check_fields_present<const L: usize>(
+/// Checks that the given extractors are satisfied with the current span hierarchy.
+///
+/// This should not be called directly, but used through [`check_fields_present`] which allows
+/// `Summary::Unconfigured` only when the calling crate is being `#[cfg(test)]` as a conservative default.
+#[doc(hidden)]
+pub fn check_fields_present0<const L: usize>(
    must_be_present: [&dyn Extractor; L],
-) -> Result<(), Vec<&dyn Extractor>> {
-    let mut missing: HashSet<MemoryIdentity> =
-        HashSet::from_iter(must_be_present.into_iter().map(|r| MemoryIdentity(r)));
+) -> Result<Summary, Vec<&dyn Extractor>> {
+    let mut missing = must_be_present.into_iter().collect::<Vec<_>>();
    let trace = tracing_error::SpanTrace::capture();
    trace.with_spans(|md, _formatted_fields| {
-        missing.retain(|extractor| match extractor.0.extract(md.fields()) {
+        // when trying to understand the inner workings of how does the matching work, note that
+        // this closure might be called zero times if the span is disabled. normally it is called
+        // once per span hierarchy level.
+        missing.retain(|extractor| match extractor.extract(md.fields()) {
            ExtractionResult::Present => false,
            ExtractionResult::Absent => true,
        });
-        !missing.is_empty() // continue walking up until we've found all missing
+
+        // continue walking up until we've found all missing
+        !missing.is_empty()
    });
    if missing.is_empty() {
-        Ok(())
+        Ok(Summary::FoundEverything)
+    } else if !tracing_subscriber_configured() {
+        Ok(Summary::Unconfigured)
    } else {
-        Err(missing.into_iter().map(|mi| mi.0).collect())
+        // we can still hit here if a tracing subscriber has been configured but the ErrorLayer is
+        // missing, which can be annoying. for this case, we could probably use
+        // SpanTrace::status().
+        //
+        // another way to end up here is with RUST_LOG=pageserver=off while configuring the
+        // logging, though I guess in that case the SpanTrace::status() == EMPTY would be valid.
+        // this case is covered by test `not_found_if_tracing_error_subscriber_has_wrong_filter`.
+        Err(missing)
    }
 }

+/// Checks that the given extractors are satisfied with the current span hierarchy.
+///
+/// The macro is the preferred way of checking if fields exist while passing checks if a test does
+/// not have tracing configured.
+///
+/// Why mangled name? Because #[macro_export] will expose it at utils::__check_fields_present.
+/// However we can game a module namespaced macro for `use` purposes by re-exporting the
+/// #[macro_export] exported name with an alias (below).
+#[doc(hidden)]
+#[macro_export]
+macro_rules! __check_fields_present {
+    ($extractors:expr) => {{
+        {
+            use $crate::tracing_span_assert::{check_fields_present0, Summary::*, Extractor};
+
+            match check_fields_present0($extractors) {
+                Ok(FoundEverything) => Ok(()),
+                Ok(Unconfigured) if cfg!(test) => {
+                    // allow unconfigured in tests
+                    Ok(())
+                },
+                Ok(Unconfigured) => {
+                    panic!("utils::tracing_span_assert: outside of #[cfg(test)] expected tracing to be configured with tracing_error::ErrorLayer")
+                },
+                Err(missing) => Err(missing)
+            }
+        }
+    }}
+}
+
+pub use crate::__check_fields_present as check_fields_present;
+
+/// Explanation for why the check was deemed ok.
+///
+/// Mainly useful for testing, or configuring per-crate behaviour as in with
+/// [`check_fields_present`].
+#[derive(Debug)]
+pub enum Summary {
+    /// All extractors were found.
+    ///
+    /// Should only happen when tracing is properly configured.
+    FoundEverything,
+
+    /// Tracing has not been configured at all. This is ok for tests running without tracing set
+    /// up.
+    Unconfigured,
+}
+
+fn tracing_subscriber_configured() -> bool {
+    let mut noop_configured = false;
+    tracing::dispatcher::get_default(|d| {
+        // it is possible that this closure will not be invoked, but the current implementation
+        // always invokes it
+        noop_configured = d.is::<tracing::subscriber::NoSubscriber>();
+    });
+
+    !noop_configured
+}
+
 #[cfg(test)]
 mod tests {

@@ -123,6 +177,36 @@ mod tests {

    use super::*;

+    use std::{
+        collections::HashSet,
+        fmt::{self},
+        hash::{Hash, Hasher},
+    };
+
+    struct MemoryIdentity<'a>(&'a dyn Extractor);
+
+    impl<'a> MemoryIdentity<'a> {
+        fn as_ptr(&self) -> *const () {
+            self.0 as *const _ as *const ()
+        }
+    }
+    impl<'a> PartialEq for MemoryIdentity<'a> {
+        fn eq(&self, other: &Self) -> bool {
+            self.as_ptr() == other.as_ptr()
+        }
+    }
+    impl<'a> Eq for MemoryIdentity<'a> {}
+    impl<'a> Hash for MemoryIdentity<'a> {
+        fn hash<H: Hasher>(&self, state: &mut H) {
+            self.as_ptr().hash(state);
+        }
+    }
+    impl<'a> fmt::Debug for MemoryIdentity<'a> {
+        fn fmt(&self, f: &mut fmt::Formatter<'_>) -> std::fmt::Result {
+            write!(f, "{:p}: {}", self.as_ptr(), self.0.name())
+        }
+    }
+
    struct Setup {
        _current_thread_subscriber_guard: tracing::subscriber::DefaultGuard,
        tenant_extractor: MultiNameExtractor<2>,
@@ -159,7 +243,8 @@ mod tests {
        let setup = setup_current_thread();
        let span = tracing::info_span!("root", tenant_id = "tenant-1", timeline_id = "timeline-1");
        let _guard = span.enter();
-        check_fields_present([&setup.tenant_extractor, &setup.timeline_extractor]).unwrap();
+        let res = check_fields_present0([&setup.tenant_extractor, &setup.timeline_extractor]);
+        assert!(matches!(res, Ok(Summary::FoundEverything)), "{res:?}");
    }

    #[test]
@@ -167,8 +252,8 @@ mod tests {
        let setup = setup_current_thread();
        let span = tracing::info_span!("root", timeline_id = "timeline-1");
        let _guard = span.enter();
-        let missing =
-            check_fields_present([&setup.tenant_extractor, &setup.timeline_extractor]).unwrap_err();
+        let missing = check_fields_present0([&setup.tenant_extractor, &setup.timeline_extractor])
+            .unwrap_err();
        assert_missing(missing, vec![&setup.tenant_extractor]);
    }

@@ -185,7 +270,8 @@ mod tests {
        let span = tracing::info_span!("grandchild", timeline_id = "timeline-1");
        let _guard = span.enter();

-        check_fields_present([&setup.tenant_extractor, &setup.timeline_extractor]).unwrap();
+        let res = check_fields_present0([&setup.tenant_extractor, &setup.timeline_extractor]);
+        assert!(matches!(res, Ok(Summary::FoundEverything)), "{res:?}");
    }

    #[test]
@@ -198,7 +284,7 @@ mod tests {
        let span = tracing::info_span!("child", timeline_id = "timeline-1");
        let _guard = span.enter();

-        let missing = check_fields_present([&setup.tenant_extractor]).unwrap_err();
+        let missing = check_fields_present0([&setup.tenant_extractor]).unwrap_err();
        assert_missing(missing, vec![&setup.tenant_extractor]);
    }

@@ -207,7 +293,8 @@ mod tests {
        let setup = setup_current_thread();
        let span = tracing::info_span!("root", tenant_id = "tenant-1", timeline_id = "timeline-1");
        let _guard = span.enter();
-        check_fields_present([&setup.tenant_extractor]).unwrap();
+        let res = check_fields_present0([&setup.tenant_extractor]);
+        assert!(matches!(res, Ok(Summary::FoundEverything)), "{res:?}");
    }

    #[test]
@@ -223,7 +310,8 @@ mod tests {
        let span = tracing::info_span!("grandchild", timeline_id = "timeline-1");
        let _guard = span.enter();

-        check_fields_present([&setup.tenant_extractor]).unwrap();
+        let res = check_fields_present0([&setup.tenant_extractor]);
+        assert!(matches!(res, Ok(Summary::FoundEverything)), "{res:?}");
    }

    #[test]
@@ -231,7 +319,7 @@ mod tests {
        let setup = setup_current_thread();
        let span = tracing::info_span!("root", timeline_id = "timeline-1");
        let _guard = span.enter();
-        let missing = check_fields_present([&setup.tenant_extractor]).unwrap_err();
+        let missing = check_fields_present0([&setup.tenant_extractor]).unwrap_err();
        assert_missing(missing, vec![&setup.tenant_extractor]);
    }

@@ -245,43 +333,107 @@ mod tests {
        let span = tracing::info_span!("child", timeline_id = "timeline-1");
        let _guard = span.enter();

-        let missing = check_fields_present([&setup.tenant_extractor]).unwrap_err();
+        let missing = check_fields_present0([&setup.tenant_extractor]).unwrap_err();
        assert_missing(missing, vec![&setup.tenant_extractor]);
    }

    #[test]
-    fn tracing_error_subscriber_not_set_up() {
+    fn tracing_error_subscriber_not_set_up_straight_line() {
        // no setup
-
        let span = tracing::info_span!("foo", e = "some value");
        let _guard = span.enter();

        let extractor = MultiNameExtractor::new("E", ["e"]);
-        let missing = check_fields_present([&extractor]).unwrap_err();
-        assert_missing(missing, vec![&extractor]);
+        let res = check_fields_present0([&extractor]);
+        assert!(matches!(res, Ok(Summary::Unconfigured)), "{res:?}");
+
+        // similarly for a not found key
+        let extractor = MultiNameExtractor::new("F", ["foobar"]);
+        let res = check_fields_present0([&extractor]);
+        assert!(matches!(res, Ok(Summary::Unconfigured)), "{res:?}");
    }

    #[test]
-    #[should_panic]
-    fn panics_if_tracing_error_subscriber_has_wrong_filter() {
+    fn tracing_error_subscriber_not_set_up_with_instrument() {
+        // no setup
+
+        // demo a case where span entering is used to establish a parent child connection, but
+        // when we re-enter the subspan SpanTrace::with_spans iterates over nothing.
+        let span = tracing::info_span!("foo", e = "some value");
+        let _guard = span.enter();
+
+        let subspan = tracing::info_span!("bar", f = "foobar");
+        drop(_guard);
+
+        // normally this would work, but without any tracing-subscriber configured, both
+        // check_field_present find nothing
+        let _guard = subspan.enter();
+        let extractors: [&dyn Extractor; 2] = [
+            &MultiNameExtractor::new("E", ["e"]),
+            &MultiNameExtractor::new("F", ["f"]),
+        ];
+
+        let res = check_fields_present0(extractors);
+        assert!(matches!(res, Ok(Summary::Unconfigured)), "{res:?}");
+
+        // similarly for a not found key
+        let extractor = MultiNameExtractor::new("G", ["g"]);
+        let res = check_fields_present0([&extractor]);
+        assert!(matches!(res, Ok(Summary::Unconfigured)), "{res:?}");
+    }
+
+    #[test]
+    fn tracing_subscriber_configured() {
+        // this will fail if any utils::logging::init callers appear, but let's hope they do not
+        // appear.
+        assert!(!super::tracing_subscriber_configured());
+
+        let _g = setup_current_thread();
+
+        assert!(super::tracing_subscriber_configured());
+    }
+
+    #[test]
+    fn not_found_when_disabled_by_filter() {
        let r = tracing_subscriber::registry().with({
-            tracing_error::ErrorLayer::default().with_filter(
-                tracing_subscriber::filter::dynamic_filter_fn(|md, _| {
-                    if md.is_span() && *md.level() == tracing::Level::INFO {
-                        return false;
-                    }
-                    true
-                }),
-            )
+            tracing_error::ErrorLayer::default().with_filter(tracing_subscriber::filter::filter_fn(
+                |md| !(md.is_span() && *md.level() == tracing::Level::INFO),
+            ))
        });

        let _guard = tracing::subscriber::set_default(r);

+        // this test is a rather tricky one, it has a number of possible outcomes depending on the
+        // execution order when executed with other tests even if no test sets the global default
+        // subscriber.
+
        let span = tracing::info_span!("foo", e = "some value");
        let _guard = span.enter();

-        let extractor = MultiNameExtractor::new("E", ["e"]);
-        let missing = check_fields_present([&extractor]).unwrap_err();
-        assert_missing(missing, vec![&extractor]);
+        let extractors: [&dyn Extractor; 1] = [&MultiNameExtractor::new("E", ["e"])];
+
+        if span.is_disabled() {
+            // the tests are running single threaded, or we got lucky and no other tests subscriber
+            // was got to register their per-CALLSITE::META interest between `set_default` and
+            // creation of the span, thus the filter got to apply and registered interest of Never,
+            // so the span was never created.
+            //
+            // as the span is disabled, no keys were recorded to it, leading check_fields_present0
+            // to find an error.
+
+            let missing = check_fields_present0(extractors).unwrap_err();
+            assert_missing(missing, vec![extractors[0]]);
+        } else {
+            // when the span is enabled, it is because some other test is running at the same time,
+            // and that tests registry has filters which are interested in our above span.
+            //
+            // because the span is now enabled, all keys will be found for it. the
+            // tracing_error::SpanTrace does not consider layer filters during the span hierarchy
+            // walk (SpanTrace::with_spans), nor is the SpanTrace::status a reliable indicator in
+            // this test-induced issue.
+
+            let res = check_fields_present0(extractors);
+            assert!(matches!(res, Ok(Summary::FoundEverything)), "{res:?}");
+        }
    }
 }
--- a/pageserver/Cargo.toml
+++ b/pageserver/Cargo.toml
@@ -12,6 +12,7 @@ testing = ["fail/failpoints"]

 [dependencies]
 anyhow.workspace = true
+async-compression.workspace = true
 async-stream.workspace = true
 async-trait.workspace = true
 byteorder.workspace = true
@@ -24,6 +25,7 @@ consumption_metrics.workspace = true
 crc32c.workspace = true
 crossbeam-utils.workspace = true
 either.workspace = true
+flate2.workspace = true
 fail.workspace = true
 futures.workspace = true
 git-version.workspace = true
@@ -33,6 +35,8 @@ humantime-serde.workspace = true
 hyper.workspace = true
 itertools.workspace = true
 nix.workspace = true
+# hack to get the number of worker threads tokio uses
+num_cpus = { version = "1.15" }
 num-traits.workspace = true
 once_cell.workspace = true
 pin-project-lite.workspace = true
@@ -80,6 +84,7 @@ strum_macros.workspace = true
 criterion.workspace = true
 hex-literal.workspace = true
 tempfile.workspace = true
+tokio = { workspace = true, features = ["process", "sync", "fs", "rt", "io-util", "time", "test-util"] }

 [[bench]]
 name = "bench_layer_map"
--- a/pageserver/benches/bench_layer_map.rs
+++ b/pageserver/benches/bench_layer_map.rs
@@ -1,8 +1,8 @@
 use pageserver::keyspace::{KeyPartitioning, KeySpace};
 use pageserver::repository::Key;
 use pageserver::tenant::layer_map::LayerMap;
-use pageserver::tenant::storage_layer::{tests::LayerDescriptor, Layer, LayerFileName};
-use pageserver::tenant::storage_layer::{PersistentLayer, PersistentLayerDesc};
+use pageserver::tenant::storage_layer::LayerFileName;
+use pageserver::tenant::storage_layer::PersistentLayerDesc;
 use rand::prelude::{SeedableRng, SliceRandom, StdRng};
 use std::cmp::{max, min};
 use std::fs::File;
@@ -28,13 +28,13 @@ fn build_layer_map(filename_dump: PathBuf) -> LayerMap {
    for fname in filenames {
        let fname = fname.unwrap();
        let fname = LayerFileName::from_str(&fname).unwrap();
-        let layer = LayerDescriptor::from(fname);
+        let layer = PersistentLayerDesc::from(fname);

        let lsn_range = layer.get_lsn_range();
        min_lsn = min(min_lsn, lsn_range.start);
        max_lsn = max(max_lsn, Lsn(lsn_range.end.0 - 1));

-        updates.insert_historic(layer.layer_desc().clone());
+        updates.insert_historic(layer);
    }

    println!("min: {min_lsn}, max: {max_lsn}");
@@ -210,15 +210,15 @@ fn bench_sequential(c: &mut Criterion) {
    for i in 0..100_000 {
        let i32 = (i as u32) % 100;
        let zero = Key::from_hex("000000000000000000000000000000000000").unwrap();
-        let layer = LayerDescriptor::from(PersistentLayerDesc::new_img(
+        let layer = PersistentLayerDesc::new_img(
            TenantId::generate(),
            TimelineId::generate(),
            zero.add(10 * i32)..zero.add(10 * i32 + 1),
            Lsn(i),
            false,
            0,
-        ));
-        updates.insert_historic(layer.layer_desc().clone());
+        );
+        updates.insert_historic(layer);
    }
    updates.flush();
    println!("Finished layer map init in {:?}", now.elapsed());
--- a/pageserver/ctl/Cargo.toml
+++ b/pageserver/ctl/Cargo.toml
@@ -13,6 +13,7 @@ clap = { workspace = true, features = ["string"] }
 git-version.workspace = true
 pageserver = { path = ".." }
 postgres_ffi.workspace = true
+tokio.workspace = true
 utils.workspace = true
 svg_fmt.workspace = true
 workspace_hack.workspace = true
--- a/pageserver/ctl/src/draw_timeline_dir.rs
+++ b/pageserver/ctl/src/draw_timeline_dir.rs
@@ -7,10 +7,10 @@
 //! - The y axis represents LSN, growing upwards.
 //!
 //! Coordinates in both axis are compressed for better readability.
-//! (see https://medium.com/algorithms-digest/coordinate-compression-2fff95326fb)
+//! (see <https://medium.com/algorithms-digest/coordinate-compression-2fff95326fb>)
 //!
 //! Example use:
-//! ```
+//! ```bash
 //! $ ls test_output/test_pgbench\[neon-45-684\]/repo/tenants/$TENANT/timelines/$TIMELINE | \
 //! $   grep "__" | cargo run --release --bin pagectl draw-timeline-dir > out.svg
 //! $ firefox out.svg
@@ -20,7 +20,7 @@
 //! or from pageserver log files.
 //!
 //! TODO Consider shipping this as a grafana panel plugin:
-//!      https://grafana.com/tutorials/build-a-panel-plugin/
+//!      <https://grafana.com/tutorials/build-a-panel-plugin/>
 use anyhow::Result;
 use pageserver::repository::Key;
 use std::cmp::Ordering;
@@ -117,7 +117,8 @@ pub fn main() -> Result<()> {

        let mut lsn_diff = (lsn_end - lsn_start) as f32;
        let mut fill = Fill::None;
-        let mut margin = 0.05 * lsn_diff; // Height-dependent margin to disambiguate overlapping deltas
+        let mut ymargin = 0.05 * lsn_diff; // Height-dependent margin to disambiguate overlapping deltas
+        let xmargin = 0.05; // Height-dependent margin to disambiguate overlapping deltas
        let mut lsn_offset = 0.0;

        // Fill in and thicken rectangle if it's an
@@ -128,7 +129,7 @@ pub fn main() -> Result<()> {
                num_images += 1;
                lsn_diff = 0.3;
                lsn_offset = -lsn_diff / 2.0;
-                margin = 0.05;
+                ymargin = 0.05;
                fill = Fill::Color(rgb(0, 0, 0));
            }
            Ordering::Greater => panic!("Invalid lsn range {}-{}", lsn_start, lsn_end),
@@ -137,10 +138,10 @@ pub fn main() -> Result<()> {
        println!(
            "    {}",
            rectangle(
-                key_start as f32 + stretch * margin,
-                stretch * (lsn_max as f32 - (lsn_end as f32 - margin - lsn_offset)),
-                key_diff as f32 - stretch * 2.0 * margin,
-                stretch * (lsn_diff - 2.0 * margin)
+                key_start as f32 + stretch * xmargin,
+                stretch * (lsn_max as f32 - (lsn_end as f32 - ymargin - lsn_offset)),
+                key_diff as f32 - stretch * 2.0 * xmargin,
+                stretch * (lsn_diff - 2.0 * ymargin)
            )
            .fill(fill)
            .stroke(Stroke::Color(rgb(0, 0, 0), 0.1))
--- a/pageserver/ctl/src/layer_map_analyzer.rs
+++ b/pageserver/ctl/src/layer_map_analyzer.rs
@@ -95,7 +95,7 @@ pub(crate) fn parse_filename(name: &str) -> Option<LayerFile> {
 }

 // Finds the max_holes largest holes, ignoring any that are smaller than MIN_HOLE_LENGTH"
-fn get_holes(path: &Path, max_holes: usize) -> Result<Vec<Hole>> {
+async fn get_holes(path: &Path, max_holes: usize) -> Result<Vec<Hole>> {
    let file = FileBlockReader::new(VirtualFile::open(path)?);
    let summary_blk = file.read_blk(0)?;
    let actual_summary = Summary::des_prefix(summary_blk.as_ref())?;
@@ -129,7 +129,7 @@ fn get_holes(path: &Path, max_holes: usize) -> Result<Vec<Hole>> {
    Ok(holes)
 }

-pub(crate) fn main(cmd: &AnalyzeLayerMapCmd) -> Result<()> {
+pub(crate) async fn main(cmd: &AnalyzeLayerMapCmd) -> Result<()> {
    let storage_path = &cmd.path;
    let max_holes = cmd.max_holes.unwrap_or(DEFAULT_MAX_HOLES);

@@ -160,7 +160,7 @@ pub(crate) fn main(cmd: &AnalyzeLayerMapCmd) -> Result<()> {
                    parse_filename(&layer.file_name().into_string().unwrap())
                {
                    if layer_file.is_delta {
-                        layer_file.holes = get_holes(&layer.path(), max_holes)?;
+                        layer_file.holes = get_holes(&layer.path(), max_holes).await?;
                        n_deltas += 1;
                    }
                    layers.push(layer_file);
--- a/pageserver/ctl/src/layers.rs
+++ b/pageserver/ctl/src/layers.rs
@@ -43,8 +43,7 @@ pub(crate) enum LayerCmd {
    },
 }

-fn read_delta_file(path: impl AsRef<Path>) -> Result<()> {
-    use pageserver::tenant::blob_io::BlobCursor;
+async fn read_delta_file(path: impl AsRef<Path>) -> Result<()> {
    use pageserver::tenant::block_io::BlockReader;

    let path = path.as_ref();
@@ -78,7 +77,7 @@ fn read_delta_file(path: impl AsRef<Path>) -> Result<()> {
    Ok(())
 }

-pub(crate) fn main(cmd: &LayerCmd) -> Result<()> {
+pub(crate) async fn main(cmd: &LayerCmd) -> Result<()> {
    match cmd {
        LayerCmd::List { path } => {
            for tenant in fs::read_dir(path.join("tenants"))? {
@@ -153,7 +152,7 @@ pub(crate) fn main(cmd: &LayerCmd) -> Result<()> {
                        );

                        if layer_file.is_delta {
-                            read_delta_file(layer.path())?;
+                            read_delta_file(layer.path()).await?;
                        } else {
                            anyhow::bail!("not supported yet :(");
                        }
--- a/pageserver/ctl/src/main.rs
+++ b/pageserver/ctl/src/main.rs
@@ -72,12 +72,13 @@ struct AnalyzeLayerMapCmd {
    max_holes: Option<usize>,
 }

-fn main() -> anyhow::Result<()> {
+#[tokio::main]
+async fn main() -> anyhow::Result<()> {
    let cli = CliOpts::parse();

    match cli.command {
        Commands::Layer(cmd) => {
-            layers::main(&cmd)?;
+            layers::main(&cmd).await?;
        }
        Commands::Metadata(cmd) => {
            handle_metadata(&cmd)?;
@@ -86,7 +87,7 @@ fn main() -> anyhow::Result<()> {
            draw_timeline_dir::main()?;
        }
        Commands::AnalyzeLayerMap(cmd) => {
-            layer_map_analyzer::main(&cmd)?;
+            layer_map_analyzer::main(&cmd).await?;
        }
        Commands::PrintLayerFile(cmd) => {
            if let Err(e) = read_pg_control_file(&cmd.path) {
@@ -94,7 +95,7 @@ fn main() -> anyhow::Result<()> {
                    "Failed to read input file as a pg control one: {e:#}\n\
                    Attempting to read it as layer file"
                );
-                print_layerfile(&cmd.path)?;
+                print_layerfile(&cmd.path).await?;
            }
        }
    };
@@ -113,12 +114,12 @@ fn read_pg_control_file(control_file_path: &Path) -> anyhow::Result<()> {
    Ok(())
 }

-fn print_layerfile(path: &Path) -> anyhow::Result<()> {
+async fn print_layerfile(path: &Path) -> anyhow::Result<()> {
    // Basic initialization of things that don't change after startup
    virtual_file::init(10);
    page_cache::init(100);
    let ctx = RequestContext::new(TaskKind::DebugTool, DownloadBehavior::Error);
-    dump_layerfile_from_path(path, true, &ctx)
+    dump_layerfile_from_path(path, true, &ctx).await
 }

 fn handle_metadata(
--- a/pageserver/src/basebackup.rs
+++ b/pageserver/src/basebackup.rs
@@ -19,12 +19,6 @@ use tokio::io;
 use tokio::io::AsyncWrite;
 use tracing::*;

-/// NB: This relies on a modified version of tokio_tar that does *not* write the
-/// end-of-archive marker (1024 zero bytes), when the Builder struct is dropped
-/// without explicitly calling 'finish' or 'into_inner'!
-///
-/// See https://github.com/neondatabase/tokio-tar/pull/1
-///
 use tokio_tar::{Builder, EntryType, Header};

 use crate::context::RequestContext;
--- a/pageserver/src/bin/pageserver.rs
+++ b/pageserver/src/bin/pageserver.rs
@@ -396,8 +396,8 @@ fn start_pageserver(

            let guard = scopeguard::guard_on_success((), |_| tracing::info!("Cancelled before initial logical sizes completed"));

-            let init_sizes_done = tokio::select! {
-                _ = &mut init_sizes_done => {
+            let init_sizes_done = match tokio::time::timeout(timeout, &mut init_sizes_done).await {
+                Ok(_) => {
                    let now = std::time::Instant::now();
                    tracing::info!(
                        from_init_done_millis = (now - init_done).as_millis(),
@@ -406,7 +406,7 @@ fn start_pageserver(
                    );
                    None
                }
-                _ = tokio::time::sleep(timeout) => {
+                Err(_) => {
                    tracing::info!(
                        timeout_millis = timeout.as_millis(),
                        "Initial logical size timeout elapsed; starting background jobs"
--- a/pageserver/src/config.rs
+++ b/pageserver/src/config.rs
@@ -171,11 +171,13 @@ pub struct PageServerConf {

    pub log_format: LogFormat,

-    /// Number of concurrent [`Tenant::gather_size_inputs`] allowed.
+    /// Number of concurrent [`Tenant::gather_size_inputs`](crate::tenant::Tenant::gather_size_inputs) allowed.
    pub concurrent_tenant_size_logical_size_queries: ConfigurableSemaphore,
    /// Limit of concurrent [`Tenant::gather_size_inputs`] issued by module `eviction_task`.
    /// The number of permits is the same as `concurrent_tenant_size_logical_size_queries`.
    /// See the comment in `eviction_task` for details.
+    ///
+    /// [`Tenant::gather_size_inputs`]: crate::tenant::Tenant::gather_size_inputs
    pub eviction_task_immitated_concurrent_logical_size_queries: ConfigurableSemaphore,

    // How often to collect metrics and send them to the metrics endpoint.
@@ -570,21 +572,21 @@ impl PageServerConf {
            .join(TENANT_ATTACHING_MARKER_FILENAME)
    }

-    pub fn tenant_ignore_mark_file_path(&self, tenant_id: TenantId) -> PathBuf {
-        self.tenant_path(&tenant_id).join(IGNORED_TENANT_FILE_NAME)
+    pub fn tenant_ignore_mark_file_path(&self, tenant_id: &TenantId) -> PathBuf {
+        self.tenant_path(tenant_id).join(IGNORED_TENANT_FILE_NAME)
    }

    /// Points to a place in pageserver's local directory,
    /// where certain tenant's tenantconf file should be located.
-    pub fn tenant_config_path(&self, tenant_id: TenantId) -> PathBuf {
-        self.tenant_path(&tenant_id).join(TENANT_CONFIG_NAME)
+    pub fn tenant_config_path(&self, tenant_id: &TenantId) -> PathBuf {
+        self.tenant_path(tenant_id).join(TENANT_CONFIG_NAME)
    }

    pub fn timelines_path(&self, tenant_id: &TenantId) -> PathBuf {
        self.tenant_path(tenant_id).join(TIMELINES_SEGMENT_NAME)
    }

-    pub fn timeline_path(&self, timeline_id: &TimelineId, tenant_id: &TenantId) -> PathBuf {
+    pub fn timeline_path(&self, tenant_id: &TenantId, timeline_id: &TimelineId) -> PathBuf {
        self.timelines_path(tenant_id).join(timeline_id.to_string())
    }

@@ -594,7 +596,7 @@ impl PageServerConf {
        timeline_id: TimelineId,
    ) -> PathBuf {
        path_with_suffix_extension(
-            self.timeline_path(&timeline_id, &tenant_id),
+            self.timeline_path(&tenant_id, &timeline_id),
            TIMELINE_UNINIT_MARK_SUFFIX,
        )
    }
@@ -617,8 +619,8 @@ impl PageServerConf {

    /// Points to a place in pageserver's local directory,
    /// where certain timeline's metadata file should be located.
-    pub fn metadata_path(&self, timeline_id: TimelineId, tenant_id: TenantId) -> PathBuf {
-        self.timeline_path(&timeline_id, &tenant_id)
+    pub fn metadata_path(&self, tenant_id: &TenantId, timeline_id: &TimelineId) -> PathBuf {
+        self.timeline_path(tenant_id, timeline_id)
            .join(METADATA_FILE_NAME)
    }

@@ -993,6 +995,8 @@ impl ConfigurableSemaphore {
    /// Require a non-zero initial permits, because using permits == 0 is a crude way to disable a
    /// feature such as [`Tenant::gather_size_inputs`]. Otherwise any semaphore using future will
    /// behave like [`futures::future::pending`], just waiting until new permits are added.
+    ///
+    /// [`Tenant::gather_size_inputs`]: crate::tenant::Tenant::gather_size_inputs
    pub fn new(initial_permits: NonZeroUsize) -> Self {
        ConfigurableSemaphore {
            initial_permits,
--- a/pageserver/src/context.rs
+++ b/pageserver/src/context.rs
@@ -179,6 +179,9 @@ impl RequestContext {
    /// a context and you are unwilling to change all callers to provide one.
    ///
    /// Before we add cancellation, we should get rid of this method.
+    ///
+    /// [`attached_child`]: Self::attached_child
+    /// [`detached_child`]: Self::detached_child
    pub fn todo_child(task_kind: TaskKind, download_behavior: DownloadBehavior) -> Self {
        Self::new(task_kind, download_behavior)
    }
--- a/pageserver/src/disk_usage_eviction_task.rs
+++ b/pageserver/src/disk_usage_eviction_task.rs
@@ -60,7 +60,7 @@ use utils::serde_percent::Percent;
 use crate::{
    config::PageServerConf,
    task_mgr::{self, TaskKind, BACKGROUND_RUNTIME},
-    tenant::{self, storage_layer::PersistentLayer, Timeline},
+    tenant::{self, storage_layer::PersistentLayer, timeline::EvictionError, Timeline},
 };

 #[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
@@ -166,11 +166,11 @@ async fn disk_usage_eviction_task(
        .await;

        let sleep_until = start + task_config.period;
-        tokio::select! {
-            _ = tokio::time::sleep_until(sleep_until) => {},
-            _ = cancel.cancelled() => {
-                break
-            }
+        if tokio::time::timeout_at(sleep_until, cancel.cancelled())
+            .await
+            .is_ok()
+        {
+            break;
        }
    }
 }
@@ -305,7 +305,7 @@ pub async fn disk_usage_eviction_task_iteration_impl<U: Usage>(
    let now = SystemTime::now();
    for (i, (partition, candidate)) in candidates.iter().enumerate() {
        debug!(
-            "cand {}/{}: size={}, no_access_for={}us, parition={:?}, tenant={} timeline={} layer={}",
+            "cand {}/{}: size={}, no_access_for={}us, partition={:?}, {}/{}/{}",
            i + 1,
            candidates.len(),
            candidate.layer.file_size(),
@@ -390,13 +390,22 @@ pub async fn disk_usage_eviction_task_iteration_impl<U: Usage>(
                    assert_eq!(results.len(), batch.len());
                    for (result, layer) in results.into_iter().zip(batch.iter()) {
                        match result {
-                            Some(Ok(true)) => {
+                            Some(Ok(())) => {
                                usage_assumed.add_available_bytes(layer.file_size());
                            }
-                            Some(Ok(false)) => {
-                                // this is:
-                                // - Replacement::{NotFound, Unexpected}
-                                // - it cannot be is_remote_layer, filtered already
+                            Some(Err(EvictionError::CannotEvictRemoteLayer)) => {
+                                unreachable!("get_local_layers_for_disk_usage_eviction finds only local layers")
+                            }
+                            Some(Err(EvictionError::FileNotFound)) => {
+                                evictions_failed.file_sizes += layer.file_size();
+                                evictions_failed.count += 1;
+                            }
+                            Some(Err(
+                                e @ EvictionError::LayerNotFound(_)
+                                | e @ EvictionError::StatFailed(_),
+                            )) => {
+                                let e = utils::error::report_compact_sources(&e);
+                                warn!(%layer, "failed to evict layer: {e}");
                                evictions_failed.file_sizes += layer.file_size();
                                evictions_failed.count += 1;
                            }
@@ -404,10 +413,6 @@ pub async fn disk_usage_eviction_task_iteration_impl<U: Usage>(
                                assert!(cancel.is_cancelled());
                                return;
                            }
-                            Some(Err(e)) => {
-                                // we really shouldn't be getting this, precondition failure
-                                error!("failed to evict layer: {:#}", e);
-                            }
                        }
                    }
                }
--- a/pageserver/src/http/routes.rs
+++ b/pageserver/src/http/routes.rs
@@ -346,7 +346,7 @@ async fn timeline_create_handler(
            Err(tenant::CreateTimelineError::Other(err)) => Err(ApiError::InternalServerError(err)),
        }
    }
-    .instrument(info_span!("timeline_create", tenant = %tenant_id, timeline_id = %new_timeline_id, lsn=?request_data.ancestor_start_lsn, pg_version=?request_data.pg_version))
+    .instrument(info_span!("timeline_create", %tenant_id, timeline_id = %new_timeline_id, lsn=?request_data.ancestor_start_lsn, pg_version=?request_data.pg_version))
    .await
 }

@@ -381,7 +381,7 @@ async fn timeline_list_handler(
        }
        Ok::<Vec<TimelineInfo>, ApiError>(response_data)
    }
-    .instrument(info_span!("timeline_list", tenant = %tenant_id))
+    .instrument(info_span!("timeline_list", %tenant_id))
    .await?;

    json_response(StatusCode::OK, response_data)
@@ -418,7 +418,7 @@ async fn timeline_detail_handler(

        Ok::<_, ApiError>(timeline_info)
    }
-    .instrument(info_span!("timeline_detail", tenant = %tenant_id, timeline = %timeline_id))
+    .instrument(info_span!("timeline_detail", %tenant_id, %timeline_id))
    .await?;

    json_response(StatusCode::OK, timeline_info)
@@ -479,7 +479,7 @@ async fn tenant_attach_handler(
            remote_storage.clone(),
            &ctx,
        )
-        .instrument(info_span!("tenant_attach", tenant = %tenant_id))
+        .instrument(info_span!("tenant_attach", %tenant_id))
        .await?;
    } else {
        return Err(ApiError::BadRequest(anyhow!(
@@ -501,7 +501,7 @@ async fn timeline_delete_handler(
    let ctx = RequestContext::new(TaskKind::MgmtRequest, DownloadBehavior::Warn);

    mgr::delete_timeline(tenant_id, timeline_id, &ctx)
-        .instrument(info_span!("timeline_delete", tenant = %tenant_id, timeline = %timeline_id))
+        .instrument(info_span!("timeline_delete", %tenant_id, %timeline_id))
        .await?;

    // FIXME: needs to be an error for console to retry it. Ideally Accepted should be used and retried until 404.
@@ -519,7 +519,7 @@ async fn tenant_detach_handler(
    let state = get_state(&request);
    let conf = state.conf;
    mgr::detach_tenant(conf, tenant_id, detach_ignored.unwrap_or(false))
-        .instrument(info_span!("tenant_detach", tenant = %tenant_id))
+        .instrument(info_span!("tenant_detach", %tenant_id))
        .await?;

    json_response(StatusCode::OK, ())
@@ -542,7 +542,7 @@ async fn tenant_load_handler(
        state.remote_storage.clone(),
        &ctx,
    )
-    .instrument(info_span!("load", tenant = %tenant_id))
+    .instrument(info_span!("load", %tenant_id))
    .await?;

    json_response(StatusCode::ACCEPTED, ())
@@ -558,7 +558,7 @@ async fn tenant_ignore_handler(
    let state = get_state(&request);
    let conf = state.conf;
    mgr::ignore_tenant(conf, tenant_id)
-        .instrument(info_span!("ignore_tenant", tenant = %tenant_id))
+        .instrument(info_span!("ignore_tenant", %tenant_id))
        .await?;

    json_response(StatusCode::OK, ())
@@ -611,7 +611,7 @@ async fn tenant_status(
            attachment_status: state.attachment_status(),
        })
    }
-    .instrument(info_span!("tenant_status_handler", tenant = %tenant_id))
+    .instrument(info_span!("tenant_status_handler", %tenant_id))
    .await?;

    json_response(StatusCode::OK, tenant_info)
@@ -850,7 +850,7 @@ async fn tenant_create_handler(
        state.remote_storage.clone(),
        &ctx,
    )
-    .instrument(info_span!("tenant_create", tenant = ?target_tenant_id))
+    .instrument(info_span!("tenant_create", tenant_id = %target_tenant_id))
    .await?;

    // We created the tenant. Existing API semantics are that the tenant
@@ -912,7 +912,7 @@ async fn update_tenant_config_handler(

    let state = get_state(&request);
    mgr::set_new_tenant_config(state.conf, tenant_conf, tenant_id)
-        .instrument(info_span!("tenant_config", tenant = ?tenant_id))
+        .instrument(info_span!("tenant_config", %tenant_id))
        .await?;

    json_response(StatusCode::OK, ())
@@ -994,31 +994,29 @@ async fn timeline_gc_handler(
 // Run compaction immediately on given timeline.
 async fn timeline_compact_handler(
    request: Request<Body>,
-    _cancel: CancellationToken,
+    cancel: CancellationToken,
 ) -> Result<Response<Body>, ApiError> {
    let tenant_id: TenantId = parse_request_param(&request, "tenant_id")?;
    let timeline_id: TimelineId = parse_request_param(&request, "timeline_id")?;
    check_permission(&request, Some(tenant_id))?;

-    let ctx = RequestContext::new(TaskKind::MgmtRequest, DownloadBehavior::Download);
-    let result_receiver = mgr::immediate_compact(tenant_id, timeline_id, &ctx)
-        .await
-        .context("spawn compaction task")
-        .map_err(ApiError::InternalServerError)?;
-
-    let result: anyhow::Result<()> = result_receiver
-        .await
-        .context("receive compaction result")
-        .map_err(ApiError::InternalServerError)?;
-    result.map_err(ApiError::InternalServerError)?;
-
-    json_response(StatusCode::OK, ())
+    async {
+        let ctx = RequestContext::new(TaskKind::MgmtRequest, DownloadBehavior::Download);
+        let timeline = active_timeline_of_active_tenant(tenant_id, timeline_id).await?;
+        timeline
+            .compact(&cancel, &ctx)
+            .await
+            .map_err(ApiError::InternalServerError)?;
+        json_response(StatusCode::OK, ())
+    }
+    .instrument(info_span!("manual_compaction", %tenant_id, %timeline_id))
+    .await
 }

 // Run checkpoint immediately on given timeline.
 async fn timeline_checkpoint_handler(
    request: Request<Body>,
-    _cancel: CancellationToken,
+    cancel: CancellationToken,
 ) -> Result<Response<Body>, ApiError> {
    let tenant_id: TenantId = parse_request_param(&request, "tenant_id")?;
    let timeline_id: TimelineId = parse_request_param(&request, "timeline_id")?;
@@ -1031,13 +1029,13 @@ async fn timeline_checkpoint_handler(
            .await
            .map_err(ApiError::InternalServerError)?;
        timeline
-            .compact(&ctx)
+            .compact(&cancel, &ctx)
            .await
            .map_err(ApiError::InternalServerError)?;

        json_response(StatusCode::OK, ())
    }
-    .instrument(info_span!("manual_checkpoint", tenant_id = %tenant_id, timeline_id = %timeline_id))
+    .instrument(info_span!("manual_checkpoint", %tenant_id, %timeline_id))
    .await
 }

@@ -1143,7 +1141,7 @@ async fn disk_usage_eviction_run(
    let Some(storage) = state.remote_storage.clone() else {
        return Err(ApiError::InternalServerError(anyhow::anyhow!(
            "remote storage not configured, cannot run eviction iteration"
-        )))
+        )));
    };

    let state = state.disk_usage_eviction_state.clone();
--- a/pageserver/src/metrics.rs
+++ b/pageserver/src/metrics.rs
@@ -6,7 +6,6 @@ use metrics::{
    IntCounterVec, IntGauge, IntGaugeVec, UIntGauge, UIntGaugeVec,
 };
 use once_cell::sync::Lazy;
-use pageserver_api::models::TenantState;
 use strum::VariantNames;
 use strum_macros::{EnumVariantNames, IntoStaticStr};
 use utils::id::{TenantId, TimelineId};
@@ -84,11 +83,10 @@ pub static STORAGE_TIME_GLOBAL: Lazy<HistogramVec> = Lazy::new(|| {
    .expect("failed to define a metric")
 });

-static READ_NUM_FS_LAYERS: Lazy<HistogramVec> = Lazy::new(|| {
-    register_histogram_vec!(
+pub(crate) static READ_NUM_FS_LAYERS: Lazy<Histogram> = Lazy::new(|| {
+    register_histogram!(
        "pageserver_read_num_fs_layers",
        "Number of persistent layers accessed for processing a read request, including those in the cache",
-        &["tenant_id", "timeline_id"],
        vec![1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 10.0, 20.0, 50.0, 100.0],
    )
    .expect("failed to define a metric")
@@ -112,11 +110,10 @@ pub static MATERIALIZED_PAGE_CACHE_HIT_DIRECT: Lazy<IntCounter> = Lazy::new(|| {
    .expect("failed to define a metric")
 });

-static GET_RECONSTRUCT_DATA_TIME: Lazy<HistogramVec> = Lazy::new(|| {
-    register_histogram_vec!(
+pub(crate) static GET_RECONSTRUCT_DATA_TIME: Lazy<Histogram> = Lazy::new(|| {
+    register_histogram!(
        "pageserver_getpage_get_reconstruct_data_seconds",
        "Time spent in get_reconstruct_value_data",
-        &["tenant_id", "timeline_id"],
        CRITICAL_OP_BUCKETS.into(),
    )
    .expect("failed to define a metric")
@@ -246,11 +243,10 @@ pub static PAGE_CACHE_SIZE: Lazy<PageCacheSizeMetrics> = Lazy::new(|| PageCacheS
    },
 });

-static WAIT_LSN_TIME: Lazy<HistogramVec> = Lazy::new(|| {
-    register_histogram_vec!(
+pub(crate) static WAIT_LSN_TIME: Lazy<Histogram> = Lazy::new(|| {
+    register_histogram!(
        "pageserver_wait_lsn_seconds",
        "Time spent waiting for WAL to arrive",
-        &["tenant_id", "timeline_id"],
        CRITICAL_OP_BUCKETS.into(),
    )
    .expect("failed to define a metric")
@@ -309,11 +305,24 @@ static CURRENT_LOGICAL_SIZE: Lazy<UIntGaugeVec> = Lazy::new(|| {
    .expect("failed to define current logical size metric")
 });

-pub static TENANT_STATE_METRIC: Lazy<UIntGaugeVec> = Lazy::new(|| {
+pub(crate) static TENANT_STATE_METRIC: Lazy<UIntGaugeVec> = Lazy::new(|| {
    register_uint_gauge_vec!(
        "pageserver_tenant_states_count",
        "Count of tenants per state",
-        &["tenant_id", "state"]
+        &["state"]
+    )
+    .expect("Failed to register pageserver_tenant_states_count metric")
+});
+
+/// A set of broken tenants.
+///
+/// These are expected to be so rare that a set is fine. Set as in a new timeseries per each broken
+/// tenant.
+pub(crate) static BROKEN_TENANTS_SET: Lazy<UIntGaugeVec> = Lazy::new(|| {
+    register_uint_gauge_vec!(
+        "pageserver_broken_tenants_count",
+        "Set of broken tenants",
+        &["tenant_id"]
    )
    .expect("Failed to register pageserver_tenant_states_count metric")
 });
@@ -385,7 +394,7 @@ pub static UNEXPECTED_ONDEMAND_DOWNLOADS: Lazy<IntCounter> = Lazy::new(|| {
    .expect("failed to define a metric")
 });

-/// Each [`Timeline`]'s  [`EVICTIONS_WITH_LOW_RESIDENCE_DURATION`] metric.
+/// Each `Timeline`'s  [`EVICTIONS_WITH_LOW_RESIDENCE_DURATION`] metric.
 #[derive(Debug)]
 pub struct EvictionsWithLowResidenceDuration {
    data_source: &'static str,
@@ -499,23 +508,31 @@ const STORAGE_IO_TIME_BUCKETS: &[f64] = &[
    30.000,   // 30000 ms
 ];

-const STORAGE_IO_TIME_OPERATIONS: &[&str] = &[
-    "open", "close", "read", "write", "seek", "fsync", "gc", "metadata",
-];
-
-const STORAGE_IO_SIZE_OPERATIONS: &[&str] = &["read", "write"];
-
-pub static STORAGE_IO_TIME: Lazy<HistogramVec> = Lazy::new(|| {
+/// Tracks time taken by fs operations near VirtualFile.
+///
+/// Operations:
+/// - open ([`std::fs::OpenOptions::open`])
+/// - close (dropping [`std::fs::File`])
+/// - close-by-replace (close by replacement algorithm)
+/// - read (`read_at`)
+/// - write (`write_at`)
+/// - seek (modify internal position or file length query)
+/// - fsync ([`std::fs::File::sync_all`])
+/// - metadata ([`std::fs::File::metadata`])
+pub(crate) static STORAGE_IO_TIME: Lazy<HistogramVec> = Lazy::new(|| {
    register_histogram_vec!(
        "pageserver_io_operations_seconds",
        "Time spent in IO operations",
-        &["operation", "tenant_id", "timeline_id"],
+        &["operation"],
        STORAGE_IO_TIME_BUCKETS.into()
    )
    .expect("failed to define a metric")
 });

-pub static STORAGE_IO_SIZE: Lazy<IntGaugeVec> = Lazy::new(|| {
+const STORAGE_IO_SIZE_OPERATIONS: &[&str] = &["read", "write"];
+
+// Needed for the https://neonprod.grafana.net/d/5uK9tHL4k/picking-tenant-for-relocation?orgId=1
+pub(crate) static STORAGE_IO_SIZE: Lazy<IntGaugeVec> = Lazy::new(|| {
    register_int_gauge_vec!(
        "pageserver_io_operations_bytes_total",
        "Total amount of bytes read/written in IO operations",
@@ -541,6 +558,17 @@ pub static SMGR_QUERY_TIME: Lazy<HistogramVec> = Lazy::new(|| {
    .expect("failed to define a metric")
 });

+// keep in sync with control plane Go code so that we can validate
+// compute's basebackup_ms metric with our perspective in the context of SLI/SLO.
+static COMPUTE_STARTUP_BUCKETS: Lazy<[f64; 28]> = Lazy::new(|| {
+    // Go code uses milliseconds. Variable is called `computeStartupBuckets`
+    [
+        5, 10, 20, 30, 50, 70, 100, 120, 150, 200, 250, 300, 350, 400, 450, 500, 600, 800, 1000,
+        1500, 2000, 2500, 3000, 5000, 10000, 20000, 40000, 60000,
+    ]
+    .map(|ms| (ms as f64) / 1000.0)
+});
+
 pub struct BasebackupQueryTime(HistogramVec);
 pub static BASEBACKUP_QUERY_TIME: Lazy<BasebackupQueryTime> = Lazy::new(|| {
    BasebackupQueryTime({
@@ -548,7 +576,7 @@ pub static BASEBACKUP_QUERY_TIME: Lazy<BasebackupQueryTime> = Lazy::new(|| {
            "pageserver_basebackup_query_seconds",
            "Histogram of basebackup queries durations, by result type",
            &["result"],
-            CRITICAL_OP_BUCKETS.into(),
+            COMPUTE_STARTUP_BUCKETS.to_vec(),
        )
        .expect("failed to define a metric")
    })
@@ -594,7 +622,7 @@ static REMOTE_TIMELINE_CLIENT_CALLS_STARTED_HIST: Lazy<HistogramVec> = Lazy::new
         at a given instant. It gives you a better idea of the queue depth \
         than plotting the gauge directly, since operations may complete faster \
         than the sampling interval.",
-        &["tenant_id", "timeline_id", "file_kind", "op_kind"],
+        &["file_kind", "op_kind"],
        // The calls_unfinished gauge is an integer gauge, hence we have integer buckets.
        vec![0.0, 1.0, 2.0, 4.0, 6.0, 8.0, 10.0, 15.0, 20.0, 40.0, 60.0, 80.0, 100.0, 500.0],
    )
@@ -651,13 +679,13 @@ impl RemoteOpFileKind {
    }
 }

-pub static REMOTE_OPERATION_TIME: Lazy<HistogramVec> = Lazy::new(|| {
+pub(crate) static REMOTE_OPERATION_TIME: Lazy<HistogramVec> = Lazy::new(|| {
    register_histogram_vec!(
        "pageserver_remote_operation_seconds",
        "Time spent on remote storage operations. \
        Grouped by tenant, timeline, operation_kind and status. \
        Does not account for time spent waiting in remote timeline client's queues.",
-        &["tenant_id", "timeline_id", "file_kind", "op_kind", "status"]
+        &["file_kind", "op_kind", "status"]
    )
    .expect("failed to define a metric")
 });
@@ -818,7 +846,7 @@ pub static WAL_REDO_RECORD_COUNTER: Lazy<IntCounter> = Lazy::new(|| {
    .unwrap()
 });

-/// Similar to [`prometheus::HistogramTimer`] but does not record on drop.
+/// Similar to `prometheus::HistogramTimer` but does not record on drop.
 pub struct StorageTimeMetricsTimer {
    metrics: StorageTimeMetrics,
    start: Instant,
@@ -876,7 +904,7 @@ impl StorageTimeMetrics {

    /// Starts timing a new operation.
    ///
-    /// Note: unlike [`prometheus::HistogramTimer`] the returned timer does not record on drop.
+    /// Note: unlike `prometheus::HistogramTimer` the returned timer does not record on drop.
    pub fn start_timer(&self) -> StorageTimeMetricsTimer {
        StorageTimeMetricsTimer::new(self.clone())
    }
@@ -886,7 +914,6 @@ impl StorageTimeMetrics {
 pub struct TimelineMetrics {
    tenant_id: String,
    timeline_id: String,
-    pub get_reconstruct_data_time_histo: Histogram,
    pub flush_time_histo: StorageTimeMetrics,
    pub compact_time_histo: StorageTimeMetrics,
    pub create_images_time_histo: StorageTimeMetrics,
@@ -895,9 +922,7 @@ pub struct TimelineMetrics {
    pub load_layer_map_histo: StorageTimeMetrics,
    pub garbage_collect_histo: StorageTimeMetrics,
    pub last_record_gauge: IntGauge,
-    pub wait_lsn_time_histo: Histogram,
    pub resident_physical_size_gauge: UIntGauge,
-    pub read_num_fs_layers: Histogram,
    /// copy of LayeredTimeline.current_logical_size
    pub current_logical_size_gauge: UIntGauge,
    pub num_persistent_files_created: IntCounter,
@@ -914,9 +939,6 @@ impl TimelineMetrics {
    ) -> Self {
        let tenant_id = tenant_id.to_string();
        let timeline_id = timeline_id.to_string();
-        let get_reconstruct_data_time_histo = GET_RECONSTRUCT_DATA_TIME
-            .get_metric_with_label_values(&[&tenant_id, &timeline_id])
-            .unwrap();
        let flush_time_histo =
            StorageTimeMetrics::new(StorageTimeOperation::LayerFlush, &tenant_id, &timeline_id);
        let compact_time_histo =
@@ -937,9 +959,6 @@ impl TimelineMetrics {
        let last_record_gauge = LAST_RECORD_LSN
            .get_metric_with_label_values(&[&tenant_id, &timeline_id])
            .unwrap();
-        let wait_lsn_time_histo = WAIT_LSN_TIME
-            .get_metric_with_label_values(&[&tenant_id, &timeline_id])
-            .unwrap();
        let resident_physical_size_gauge = RESIDENT_PHYSICAL_SIZE
            .get_metric_with_label_values(&[&tenant_id, &timeline_id])
            .unwrap();
@@ -955,16 +974,12 @@ impl TimelineMetrics {
        let evictions = EVICTIONS
            .get_metric_with_label_values(&[&tenant_id, &timeline_id])
            .unwrap();
-        let read_num_fs_layers = READ_NUM_FS_LAYERS
-            .get_metric_with_label_values(&[&tenant_id, &timeline_id])
-            .unwrap();
        let evictions_with_low_residence_duration =
            evictions_with_low_residence_duration_builder.build(&tenant_id, &timeline_id);

        TimelineMetrics {
            tenant_id,
            timeline_id,
-            get_reconstruct_data_time_histo,
            flush_time_histo,
            compact_time_histo,
            create_images_time_histo,
@@ -973,7 +988,6 @@ impl TimelineMetrics {
            garbage_collect_histo,
            load_layer_map_histo,
            last_record_gauge,
-            wait_lsn_time_histo,
            resident_physical_size_gauge,
            current_logical_size_gauge,
            num_persistent_files_created,
@@ -982,7 +996,6 @@ impl TimelineMetrics {
            evictions_with_low_residence_duration: std::sync::RwLock::new(
                evictions_with_low_residence_duration,
            ),
-            read_num_fs_layers,
        }
    }
 }
@@ -991,15 +1004,12 @@ impl Drop for TimelineMetrics {
    fn drop(&mut self) {
        let tenant_id = &self.tenant_id;
        let timeline_id = &self.timeline_id;
-        let _ = GET_RECONSTRUCT_DATA_TIME.remove_label_values(&[tenant_id, timeline_id]);
        let _ = LAST_RECORD_LSN.remove_label_values(&[tenant_id, timeline_id]);
-        let _ = WAIT_LSN_TIME.remove_label_values(&[tenant_id, timeline_id]);
        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]);
        let _ = EVICTIONS.remove_label_values(&[tenant_id, timeline_id]);
-        let _ = READ_NUM_FS_LAYERS.remove_label_values(&[tenant_id, timeline_id]);

        self.evictions_with_low_residence_duration
            .write()
@@ -1011,9 +1021,6 @@ impl Drop for TimelineMetrics {
            let _ =
                STORAGE_TIME_COUNT_PER_TIMELINE.remove_label_values(&[op, tenant_id, timeline_id]);
        }
-        for op in STORAGE_IO_TIME_OPERATIONS {
-            let _ = STORAGE_IO_TIME.remove_label_values(&[op, tenant_id, timeline_id]);
-        }

        for op in STORAGE_IO_SIZE_OPERATIONS {
            let _ = STORAGE_IO_SIZE.remove_label_values(&[op, tenant_id, timeline_id]);
@@ -1028,9 +1035,7 @@ impl Drop for TimelineMetrics {
 pub fn remove_tenant_metrics(tenant_id: &TenantId) {
    let tid = tenant_id.to_string();
    let _ = TENANT_SYNTHETIC_SIZE_METRIC.remove_label_values(&[&tid]);
-    for state in TenantState::VARIANTS {
-        let _ = TENANT_STATE_METRIC.remove_label_values(&[&tid, state]);
-    }
+    // we leave the BROKEN_TENANTS_SET entry if any
 }

 use futures::Future;
@@ -1045,9 +1050,7 @@ pub struct RemoteTimelineClientMetrics {
    tenant_id: String,
    timeline_id: String,
    remote_physical_size_gauge: Mutex<Option<UIntGauge>>,
-    remote_operation_time: Mutex<HashMap<(&'static str, &'static str, &'static str), Histogram>>,
    calls_unfinished_gauge: Mutex<HashMap<(&'static str, &'static str), IntGauge>>,
-    calls_started_hist: Mutex<HashMap<(&'static str, &'static str), Histogram>>,
    bytes_started_counter: Mutex<HashMap<(&'static str, &'static str), IntCounter>>,
    bytes_finished_counter: Mutex<HashMap<(&'static str, &'static str), IntCounter>>,
 }
@@ -1057,14 +1060,13 @@ impl RemoteTimelineClientMetrics {
        RemoteTimelineClientMetrics {
            tenant_id: tenant_id.to_string(),
            timeline_id: timeline_id.to_string(),
-            remote_operation_time: Mutex::new(HashMap::default()),
            calls_unfinished_gauge: Mutex::new(HashMap::default()),
-            calls_started_hist: Mutex::new(HashMap::default()),
            bytes_started_counter: Mutex::new(HashMap::default()),
            bytes_finished_counter: Mutex::new(HashMap::default()),
            remote_physical_size_gauge: Mutex::new(None),
        }
    }
+
    pub fn remote_physical_size_gauge(&self) -> UIntGauge {
        let mut guard = self.remote_physical_size_gauge.lock().unwrap();
        guard
@@ -1078,26 +1080,17 @@ impl RemoteTimelineClientMetrics {
            })
            .clone()
    }
+
    pub fn remote_operation_time(
        &self,
        file_kind: &RemoteOpFileKind,
        op_kind: &RemoteOpKind,
        status: &'static str,
    ) -> Histogram {
-        let mut guard = self.remote_operation_time.lock().unwrap();
        let key = (file_kind.as_str(), op_kind.as_str(), status);
-        let metric = guard.entry(key).or_insert_with(move || {
-            REMOTE_OPERATION_TIME
-                .get_metric_with_label_values(&[
-                    &self.tenant_id.to_string(),
-                    &self.timeline_id.to_string(),
-                    key.0,
-                    key.1,
-                    key.2,
-                ])
-                .unwrap()
-        });
-        metric.clone()
+        REMOTE_OPERATION_TIME
+            .get_metric_with_label_values(&[key.0, key.1, key.2])
+            .unwrap()
    }

    fn calls_unfinished_gauge(
@@ -1125,19 +1118,10 @@ impl RemoteTimelineClientMetrics {
        file_kind: &RemoteOpFileKind,
        op_kind: &RemoteOpKind,
    ) -> Histogram {
-        let mut guard = self.calls_started_hist.lock().unwrap();
        let key = (file_kind.as_str(), op_kind.as_str());
-        let metric = guard.entry(key).or_insert_with(move || {
-            REMOTE_TIMELINE_CLIENT_CALLS_STARTED_HIST
-                .get_metric_with_label_values(&[
-                    &self.tenant_id.to_string(),
-                    &self.timeline_id.to_string(),
-                    key.0,
-                    key.1,
-                ])
-                .unwrap()
-        });
-        metric.clone()
+        REMOTE_TIMELINE_CLIENT_CALLS_STARTED_HIST
+            .get_metric_with_label_values(&[key.0, key.1])
+            .unwrap()
    }

    fn bytes_started_counter(
@@ -1256,7 +1240,7 @@ impl RemoteTimelineClientMetrics {
    /// Update the metrics that change when a call to the remote timeline client instance starts.
    ///
    /// Drop the returned guard object once the operation is finished to updates corresponding metrics that track completions.
-    /// Or, use [`RemoteTimelineClientCallMetricGuard::will_decrement_manually`] and [`call_end`] if that
+    /// Or, use [`RemoteTimelineClientCallMetricGuard::will_decrement_manually`] and [`call_end`](Self::call_end) if that
    /// is more suitable.
    /// Never do both.
    pub(crate) fn call_begin(
@@ -1289,7 +1273,7 @@ impl RemoteTimelineClientMetrics {

    /// Manually udpate the metrics that track completions, instead of using the guard object.
    /// Using the guard object is generally preferable.
-    /// See [`call_begin`] for more context.
+    /// See [`call_begin`](Self::call_begin) for more context.
    pub(crate) fn call_end(
        &self,
        file_kind: &RemoteOpFileKind,
@@ -1317,15 +1301,10 @@ impl Drop for RemoteTimelineClientMetrics {
            tenant_id,
            timeline_id,
            remote_physical_size_gauge,
-            remote_operation_time,
            calls_unfinished_gauge,
-            calls_started_hist,
            bytes_started_counter,
            bytes_finished_counter,
        } = self;
-        for ((a, b, c), _) in remote_operation_time.get_mut().unwrap().drain() {
-            let _ = REMOTE_OPERATION_TIME.remove_label_values(&[tenant_id, timeline_id, a, b, c]);
-        }
        for ((a, b), _) in calls_unfinished_gauge.get_mut().unwrap().drain() {
            let _ = REMOTE_TIMELINE_CLIENT_CALLS_UNFINISHED_GAUGE.remove_label_values(&[
                tenant_id,
@@ -1334,14 +1313,6 @@ impl Drop for RemoteTimelineClientMetrics {
                b,
            ]);
        }
-        for ((a, b), _) in calls_started_hist.get_mut().unwrap().drain() {
-            let _ = REMOTE_TIMELINE_CLIENT_CALLS_STARTED_HIST.remove_label_values(&[
-                tenant_id,
-                timeline_id,
-                a,
-                b,
-            ]);
-        }
        for ((a, b), _) in bytes_started_counter.get_mut().unwrap().drain() {
            let _ = REMOTE_TIMELINE_CLIENT_BYTES_STARTED_COUNTER.remove_label_values(&[
                tenant_id,
--- a/pageserver/src/page_service.rs
+++ b/pageserver/src/page_service.rs
@@ -10,6 +10,7 @@
 //

 use anyhow::Context;
+use async_compression::tokio::write::GzipEncoder;
 use bytes::Buf;
 use bytes::Bytes;
 use futures::Stream;
@@ -31,8 +32,10 @@ use std::str;
 use std::str::FromStr;
 use std::sync::Arc;
 use std::time::Duration;
+use tokio::io::AsyncWriteExt;
 use tokio::io::{AsyncRead, AsyncWrite};
 use tokio_util::io::StreamReader;
+use tracing::field;
 use tracing::*;
 use utils::id::ConnectionId;
 use utils::{
@@ -51,6 +54,7 @@ use crate::metrics::{LIVE_CONNECTIONS_COUNT, SMGR_QUERY_TIME};
 use crate::task_mgr;
 use crate::task_mgr::TaskKind;
 use crate::tenant;
+use crate::tenant::debug_assert_current_span_has_tenant_and_timeline_id;
 use crate::tenant::mgr;
 use crate::tenant::mgr::GetTenantError;
 use crate::tenant::{Tenant, Timeline};
@@ -238,6 +242,7 @@ pub async fn libpq_listener_main(
    Ok(())
 }

+#[instrument(skip_all, fields(peer_addr))]
 async fn page_service_conn_main(
    conf: &'static PageServerConf,
    broker_client: storage_broker::BrokerClientChannel,
@@ -260,6 +265,7 @@ async fn page_service_conn_main(
        .context("could not set TCP_NODELAY")?;

    let peer_addr = socket.peer_addr().context("get peer address")?;
+    tracing::Span::current().record("peer_addr", field::display(peer_addr));

    // setup read timeout of 10 minutes. the timeout is rather arbitrary for requirements:
    // - long enough for most valid compute connections
@@ -362,7 +368,7 @@ impl PageServerHandler {
        }
    }

-    #[instrument(skip(self, pgb, ctx))]
+    #[instrument(skip_all)]
    async fn handle_pagerequests<IO>(
        &self,
        pgb: &mut PostgresBackend<IO>,
@@ -373,6 +379,8 @@ impl PageServerHandler {
    where
        IO: AsyncRead + AsyncWrite + Send + Sync + Unpin,
    {
+        debug_assert_current_span_has_tenant_and_timeline_id();
+
        // NOTE: pagerequests handler exits when connection is closed,
        //       so there is no need to reset the association
        task_mgr::associate_with(Some(tenant_id), Some(timeline_id));
@@ -473,7 +481,7 @@ impl PageServerHandler {
    }

    #[allow(clippy::too_many_arguments)]
-    #[instrument(skip(self, pgb, ctx))]
+    #[instrument(skip_all, fields(%base_lsn, end_lsn=%_end_lsn, %pg_version))]
    async fn handle_import_basebackup<IO>(
        &self,
        pgb: &mut PostgresBackend<IO>,
@@ -487,6 +495,8 @@ impl PageServerHandler {
    where
        IO: AsyncRead + AsyncWrite + Send + Sync + Unpin,
    {
+        debug_assert_current_span_has_tenant_and_timeline_id();
+
        task_mgr::associate_with(Some(tenant_id), Some(timeline_id));
        // Create empty timeline
        info!("creating new timeline");
@@ -531,7 +541,7 @@ impl PageServerHandler {
        Ok(())
    }

-    #[instrument(skip(self, pgb, ctx))]
+    #[instrument(skip_all, fields(%start_lsn, %end_lsn))]
    async fn handle_import_wal<IO>(
        &self,
        pgb: &mut PostgresBackend<IO>,
@@ -544,6 +554,7 @@ impl PageServerHandler {
    where
        IO: AsyncRead + AsyncWrite + Send + Sync + Unpin,
    {
+        debug_assert_current_span_has_tenant_and_timeline_id();
        task_mgr::associate_with(Some(tenant_id), Some(timeline_id));

        let timeline = get_active_tenant_timeline(tenant_id, timeline_id, &ctx).await?;
@@ -738,7 +749,7 @@ impl PageServerHandler {
    }

    #[allow(clippy::too_many_arguments)]
-    #[instrument(skip(self, pgb, ctx))]
+    #[instrument(skip_all, fields(?lsn, ?prev_lsn, %full_backup))]
    async fn handle_basebackup_request<IO>(
        &mut self,
        pgb: &mut PostgresBackend<IO>,
@@ -747,11 +758,14 @@ impl PageServerHandler {
        lsn: Option<Lsn>,
        prev_lsn: Option<Lsn>,
        full_backup: bool,
+        gzip: bool,
        ctx: RequestContext,
    ) -> anyhow::Result<()>
    where
        IO: AsyncRead + AsyncWrite + Send + Sync + Unpin,
    {
+        debug_assert_current_span_has_tenant_and_timeline_id();
+
        let started = std::time::Instant::now();

        // check that the timeline exists
@@ -772,8 +786,9 @@ impl PageServerHandler {
        pgb.write_message_noflush(&BeMessage::CopyOutResponse)?;
        pgb.flush().await?;

-        // Send a tarball of the latest layer on the timeline
-        {
+        // Send a tarball of the latest layer on the timeline. Compress if not
+        // fullbackup. TODO Compress in that case too (tests need to be updated)
+        if full_backup {
            let mut writer = pgb.copyout_writer();
            basebackup::send_basebackup_tarball(
                &mut writer,
@@ -784,6 +799,40 @@ impl PageServerHandler {
                &ctx,
            )
            .await?;
+        } else {
+            let mut writer = pgb.copyout_writer();
+            if gzip {
+                let mut encoder = GzipEncoder::with_quality(
+                    writer,
+                    // NOTE using fast compression because it's on the critical path
+                    //      for compute startup. For an empty database, we get
+                    //      <100KB with this method. The Level::Best compression method
+                    //      gives us <20KB, but maybe we should add basebackup caching
+                    //      on compute shutdown first.
+                    async_compression::Level::Fastest,
+                );
+                basebackup::send_basebackup_tarball(
+                    &mut encoder,
+                    &timeline,
+                    lsn,
+                    prev_lsn,
+                    full_backup,
+                    &ctx,
+                )
+                .await?;
+                // shutdown the encoder to ensure the gzip footer is written
+                encoder.shutdown().await?;
+            } else {
+                basebackup::send_basebackup_tarball(
+                    &mut writer,
+                    &timeline,
+                    lsn,
+                    prev_lsn,
+                    full_backup,
+                    &ctx,
+                )
+                .await?;
+            }
        }

        pgb.write_message_noflush(&BeMessage::CopyDone)?;
@@ -862,6 +911,7 @@ where
        Ok(())
    }

+    #[instrument(skip_all, fields(tenant_id, timeline_id))]
    async fn process_query(
        &mut self,
        pgb: &mut PostgresBackend<IO>,
@@ -883,6 +933,10 @@ where
            let timeline_id = TimelineId::from_str(params[1])
                .with_context(|| format!("Failed to parse timeline id from {}", params[1]))?;

+            tracing::Span::current()
+                .record("tenant_id", field::display(tenant_id))
+                .record("timeline_id", field::display(timeline_id));
+
            self.check_permission(Some(tenant_id))?;

            self.handle_pagerequests(pgb, tenant_id, timeline_id, ctx)
@@ -902,6 +956,10 @@ where
            let timeline_id = TimelineId::from_str(params[1])
                .with_context(|| format!("Failed to parse timeline id from {}", params[1]))?;

+            tracing::Span::current()
+                .record("tenant_id", field::display(tenant_id))
+                .record("timeline_id", field::display(timeline_id));
+
            self.check_permission(Some(tenant_id))?;

            let lsn = if params.len() >= 3 {
@@ -913,6 +971,19 @@ where
                None
            };

+            let gzip = if params.len() >= 4 {
+                if params[3] == "--gzip" {
+                    true
+                } else {
+                    return Err(QueryError::Other(anyhow::anyhow!(
+                        "Parameter in position 3 unknown {}",
+                        params[3],
+                    )));
+                }
+            } else {
+                false
+            };
+
            metrics::metric_vec_duration::observe_async_block_duration_by_result(
                &*crate::metrics::BASEBACKUP_QUERY_TIME,
                async move {
@@ -923,6 +994,7 @@ where
                        lsn,
                        None,
                        false,
+                        gzip,
                        ctx,
                    )
                    .await?;
@@ -948,6 +1020,10 @@ where
            let timeline_id = TimelineId::from_str(params[1])
                .with_context(|| format!("Failed to parse timeline id from {}", params[1]))?;

+            tracing::Span::current()
+                .record("tenant_id", field::display(tenant_id))
+                .record("timeline_id", field::display(timeline_id));
+
            self.check_permission(Some(tenant_id))?;
            let timeline = get_active_tenant_timeline(tenant_id, timeline_id, &ctx).await?;

@@ -979,6 +1055,10 @@ where
            let timeline_id = TimelineId::from_str(params[1])
                .with_context(|| format!("Failed to parse timeline id from {}", params[1]))?;

+            tracing::Span::current()
+                .record("tenant_id", field::display(tenant_id))
+                .record("timeline_id", field::display(timeline_id));
+
            // The caller is responsible for providing correct lsn and prev_lsn.
            let lsn = if params.len() > 2 {
                Some(
@@ -1000,8 +1080,17 @@ where
            self.check_permission(Some(tenant_id))?;

            // Check that the timeline exists
-            self.handle_basebackup_request(pgb, tenant_id, timeline_id, lsn, prev_lsn, true, ctx)
-                .await?;
+            self.handle_basebackup_request(
+                pgb,
+                tenant_id,
+                timeline_id,
+                lsn,
+                prev_lsn,
+                true,
+                false,
+                ctx,
+            )
+            .await?;
            pgb.write_message_noflush(&BeMessage::CommandComplete(b"SELECT 1"))?;
        } else if query_string.starts_with("import basebackup ") {
            // Import the `base` section (everything but the wal) of a basebackup.
@@ -1033,6 +1122,10 @@ where
            let pg_version = u32::from_str(params[4])
                .with_context(|| format!("Failed to parse pg_version from {}", params[4]))?;

+            tracing::Span::current()
+                .record("tenant_id", field::display(tenant_id))
+                .record("timeline_id", field::display(timeline_id));
+
            self.check_permission(Some(tenant_id))?;

            match self
@@ -1077,6 +1170,10 @@ where
            let end_lsn = Lsn::from_str(params[3])
                .with_context(|| format!("Failed to parse Lsn from {}", params[3]))?;

+            tracing::Span::current()
+                .record("tenant_id", field::display(tenant_id))
+                .record("timeline_id", field::display(timeline_id));
+
            self.check_permission(Some(tenant_id))?;

            match self
@@ -1108,6 +1205,8 @@ where
            let tenant_id = TenantId::from_str(params[0])
                .with_context(|| format!("Failed to parse tenant id from {}", params[0]))?;

+            tracing::Span::current().record("tenant_id", field::display(tenant_id));
+
            self.check_permission(Some(tenant_id))?;

            let tenant = get_active_tenant_with_timeout(tenant_id, &ctx).await?;
--- a/pageserver/src/pgdatadir_mapping.rs
+++ b/pageserver/src/pgdatadir_mapping.rs
@@ -1131,7 +1131,7 @@ impl<'a> DatadirModification<'a> {
    /// context, breaking the atomicity is OK. If the import is interrupted, the
    /// whole import fails and the timeline will be deleted anyway.
    /// (Or to be precise, it will be left behind for debugging purposes and
-    /// ignored, see https://github.com/neondatabase/neon/pull/1809)
+    /// ignored, see <https://github.com/neondatabase/neon/pull/1809>)
    ///
    /// Note: A consequence of flushing the pending operations is that they
    /// won't be visible to subsequent operations until `commit`. The function
--- a/pageserver/src/task_mgr.rs
+++ b/pageserver/src/task_mgr.rs
@@ -130,11 +130,25 @@ pub static WALRECEIVER_RUNTIME: Lazy<Runtime> = Lazy::new(|| {
 pub static BACKGROUND_RUNTIME: Lazy<Runtime> = Lazy::new(|| {
    tokio::runtime::Builder::new_multi_thread()
        .thread_name("background op worker")
+        // if you change the number of worker threads please change the constant below
        .enable_all()
        .build()
        .expect("Failed to create background op runtime")
 });

+pub(crate) static BACKGROUND_RUNTIME_WORKER_THREADS: Lazy<usize> = Lazy::new(|| {
+    // force init and thus panics
+    let _ = BACKGROUND_RUNTIME.handle();
+    // replicates tokio-1.28.1::loom::sys::num_cpus which is not available publicly
+    // tokio would had already panicked for parsing errors or NotUnicode
+    //
+    // this will be wrong if any of the runtimes gets their worker threads configured to something
+    // else, but that has not been needed in a long time.
+    std::env::var("TOKIO_WORKER_THREADS")
+        .map(|s| s.parse::<usize>().unwrap())
+        .unwrap_or_else(|_e| usize::max(1, num_cpus::get()))
+});
+
 #[derive(Debug, Clone, Copy)]
 pub struct PageserverTaskId(u64);

@@ -205,7 +219,7 @@ pub enum TaskKind {
    ///
    /// Walreceiver uses its own abstraction called `TaskHandle` to represent the activity of establishing and handling a connection.
    /// That abstraction doesn't use `task_mgr`.
-    /// The [`WalReceiverManager`] task ensures that this `TaskHandle` task does not outlive the [`WalReceiverManager`] task.
+    /// The `WalReceiverManager` task ensures that this `TaskHandle` task does not outlive the `WalReceiverManager` task.
    /// For the `RequestContext` that we hand to the TaskHandle, we use the [`WalReceiverConnectionHandler`] task kind.
    ///
    /// Once the connection is established, the `TaskHandle` task creates a
@@ -213,16 +227,21 @@ pub enum TaskKind {
    /// the `Connection` object.
    /// A `CancellationToken` created by the `TaskHandle` task ensures
    /// that the [`WalReceiverConnectionPoller`] task will cancel soon after as the `TaskHandle` is dropped.
+    ///
+    /// [`WalReceiverConnectionHandler`]: Self::WalReceiverConnectionHandler
+    /// [`WalReceiverConnectionPoller`]: Self::WalReceiverConnectionPoller
    WalReceiverManager,

-    /// The `TaskHandle` task that executes [`walreceiver_connection::handle_walreceiver_connection`].
+    /// The `TaskHandle` task that executes `handle_walreceiver_connection`.
    /// Not a `task_mgr` task, but we use this `TaskKind` for its `RequestContext`.
    /// See the comment on [`WalReceiverManager`].
+    ///
+    /// [`WalReceiverManager`]: Self::WalReceiverManager
    WalReceiverConnectionHandler,

    /// The task that polls the `tokio-postgres::Connection` object.
-    /// Spawned by task [`WalReceiverConnectionHandler`].
-    /// See the comment on [`WalReceiverManager`].
+    /// Spawned by task [`WalReceiverConnectionHandler`](Self::WalReceiverConnectionHandler).
+    /// See the comment on [`WalReceiverManager`](Self::WalReceiverManager).
    WalReceiverConnectionPoller,

    // Garbage collection worker. One per tenant
@@ -506,17 +525,13 @@ pub async fn shutdown_tasks(
                    warn!(name = task.name, tenant_id = ?tenant_id, timeline_id = ?timeline_id, kind = ?task_kind, "stopping left-over");
                }
            }
-            let join_handle = tokio::select! {
-                biased;
-                _ = &mut join_handle => { None },
-                _ = tokio::time::sleep(std::time::Duration::from_secs(1)) => {
-                    // allow some time to elapse before logging to cut down the number of log
-                    // lines.
-                    info!("waiting for {} to shut down", task.name);
-                    Some(join_handle)
-                }
-            };
-            if let Some(join_handle) = join_handle {
+            if tokio::time::timeout(std::time::Duration::from_secs(1), &mut join_handle)
+                .await
+                .is_err()
+            {
+                // allow some time to elapse before logging to cut down the number of log
+                // lines.
+                info!("waiting for {} to shut down", task.name);
                // we never handled this return value, but:
                // - we don't deschedule which would lead to is_cancelled
                // - panics are already logged (is_panicked)
@@ -544,7 +559,7 @@ pub fn current_task_id() -> Option<PageserverTaskId> {
 pub async fn shutdown_watcher() {
    let token = SHUTDOWN_TOKEN
        .try_with(|t| t.clone())
-        .expect("shutdown_requested() called in an unexpected task or thread");
+        .expect("shutdown_watcher() called in an unexpected task or thread");

    token.cancelled().await;
 }
--- a/pageserver/src/tenant.rs
+++ b/pageserver/src/tenant.rs
@@ -20,6 +20,7 @@ use storage_broker::BrokerClientChannel;
 use tokio::sync::watch;
 use tokio::sync::OwnedMutexGuard;
 use tokio::task::JoinSet;
+use tokio_util::sync::CancellationToken;
 use tracing::*;
 use utils::completion;
 use utils::crashsafe::path_with_suffix_extension;
@@ -84,6 +85,25 @@ use utils::{
    lsn::{Lsn, RecordLsn},
 };

+/// Declare a failpoint that can use the `pause` failpoint action.
+/// We don't want to block the executor thread, hence, spawn_blocking + await.
+macro_rules! pausable_failpoint {
+    ($name:literal) => {
+        if cfg!(feature = "testing") {
+            tokio::task::spawn_blocking({
+                let current = tracing::Span::current();
+                move || {
+                    let _entered = current.entered();
+                    tracing::info!("at failpoint {}", $name);
+                    fail::fail_point!($name);
+                }
+            })
+            .await
+            .expect("spawn_blocking");
+        }
+    };
+}
+
 pub mod blob_io;
 pub mod block_io;
 pub mod disk_btree;
@@ -102,7 +122,7 @@ pub mod mgr;
 pub mod tasks;
 pub mod upload_queue;

-mod timeline;
+pub(crate) mod timeline;

 pub mod size;

@@ -114,7 +134,7 @@ pub use timeline::{
 // re-export this function so that page_cache.rs can use it.
 pub use crate::tenant::ephemeral_file::writeback as writeback_ephemeral_file;

-// re-export for use in storage_sync.rs
+// re-export for use in remote_timeline_client.rs
 pub use crate::tenant::metadata::save_metadata;

 // re-export for use in walreceiver
@@ -262,7 +282,7 @@ pub enum DeleteTimelineError {
 }

 pub enum SetStoppingError {
-    AlreadyStopping,
+    AlreadyStopping(completion::Barrier),
    Broken,
 }

@@ -299,10 +319,6 @@ impl std::fmt::Display for WaitToBecomeActiveError {
    }
 }

-pub(crate) enum ShutdownError {
-    AlreadyStopping,
-}
-
 struct DeletionGuard(OwnedMutexGuard<bool>);

 impl DeletionGuard {
@@ -410,7 +426,7 @@ impl Tenant {
                    .layers
                    .read()
                    .await
-                    .0
+                    .layer_map()
                    .iter_historic_layers()
                    .next()
                    .is_some(),
@@ -421,8 +437,8 @@ impl Tenant {
        if !picked_local {
            save_metadata(
                self.conf,
-                timeline_id,
-                tenant_id,
+                &tenant_id,
+                &timeline_id,
                up_to_date_metadata,
                first_save,
            )
@@ -451,7 +467,7 @@ impl Tenant {
    ) -> anyhow::Result<Arc<Tenant>> {
        // TODO dedup with spawn_load
        let tenant_conf =
-            Self::load_tenant_config(conf, tenant_id).context("load tenant config")?;
+            Self::load_tenant_config(conf, &tenant_id).context("load tenant config")?;

        let wal_redo_manager = Arc::new(PostgresRedoManager::new(conf, tenant_id));
        let tenant = Arc::new(Tenant::new(
@@ -560,7 +576,7 @@ impl Tenant {
                .map(move |res| {
                    res.with_context(|| format!("download index part for timeline {timeline_id}"))
                })
-                .instrument(info_span!("download_index_part", timeline=%timeline_id)),
+                .instrument(info_span!("download_index_part", %timeline_id)),
            );
        }
        // Wait for all the download tasks to complete & collect results.
@@ -646,7 +662,7 @@ impl Tenant {
        span::debug_assert_current_span_has_tenant_id();

        info!("downloading index file for timeline {}", timeline_id);
-        tokio::fs::create_dir_all(self.conf.timeline_path(&timeline_id, &self.tenant_id))
+        tokio::fs::create_dir_all(self.conf.timeline_path(&self.tenant_id, &timeline_id))
            .await
            .context("Failed to create new timeline directory")?;

@@ -724,7 +740,7 @@ impl Tenant {
    ) -> Arc<Tenant> {
        span::debug_assert_current_span_has_tenant_id();

-        let tenant_conf = match Self::load_tenant_config(conf, tenant_id) {
+        let tenant_conf = match Self::load_tenant_config(conf, &tenant_id) {
            Ok(conf) => conf,
            Err(e) => {
                error!("load tenant config failed: {:?}", e);
@@ -835,7 +851,7 @@ impl Tenant {
                            timeline_uninit_mark_file.display()
                        )
                    })?;
-                let timeline_dir = self.conf.timeline_path(&timeline_id, &self.tenant_id);
+                let timeline_dir = self.conf.timeline_path(&self.tenant_id, &timeline_id);
                if let Err(e) =
                    remove_timeline_and_uninit_mark(&timeline_dir, timeline_uninit_mark_file)
                {
@@ -880,7 +896,7 @@ impl Tenant {
                if let Ok(timeline_id) =
                    file_name.to_str().unwrap_or_default().parse::<TimelineId>()
                {
-                    let metadata = load_metadata(self.conf, timeline_id, self.tenant_id)
+                    let metadata = load_metadata(self.conf, &self.tenant_id, &timeline_id)
                        .context("failed to load metadata")?;
                    timelines_to_load.insert(timeline_id, metadata);
                } else {
@@ -1153,7 +1169,7 @@ impl Tenant {
        )
    }

-    /// Helper for unit tests to create an emtpy timeline.
+    /// Helper for unit tests to create an empty timeline.
    ///
    /// The timeline is has state value `Active` but its background loops are not running.
    // This makes the various functions which anyhow::ensure! for Active state work in tests.
@@ -1320,7 +1336,11 @@ impl Tenant {
    /// This function is periodically called by compactor task.
    /// Also it can be explicitly requested per timeline through page server
    /// api's 'compact' command.
-    pub async fn compaction_iteration(&self, ctx: &RequestContext) -> anyhow::Result<()> {
+    pub async fn compaction_iteration(
+        &self,
+        cancel: &CancellationToken,
+        ctx: &RequestContext,
+    ) -> anyhow::Result<()> {
        anyhow::ensure!(
            self.is_active(),
            "Cannot run compaction iteration on inactive tenant"
@@ -1348,8 +1368,8 @@ impl Tenant {

        for (timeline_id, timeline) in &timelines_to_compact {
            timeline
-                .compact(ctx)
-                .instrument(info_span!("compact_timeline", timeline = %timeline_id))
+                .compact(cancel, ctx)
+                .instrument(info_span!("compact_timeline", %timeline_id))
                .await?;
        }

@@ -1440,12 +1460,12 @@ impl Tenant {
            let layer_removal_guard = timeline.layer_removal_cs.lock().await;
            info!("got layer_removal_cs.lock(), deleting layer files");

-            // NB: storage_sync upload tasks that reference these layers have been cancelled
+            // NB: remote_timeline_client upload tasks that reference these layers have been cancelled
            //     by the caller.

            let local_timeline_directory = self
                .conf
-                .timeline_path(&timeline.timeline_id, &self.tenant_id);
+                .timeline_path(&self.tenant_id, &timeline.timeline_id);

            fail::fail_point!("timeline-delete-before-rm", |_| {
                Err(anyhow::anyhow!("failpoint: timeline-delete-before-rm"))?
@@ -1498,20 +1518,7 @@ impl Tenant {
            remote_client.delete_all().await.context("delete_all")?
        };

-        // Have a failpoint that can use the `pause` failpoint action.
-        // We don't want to block the executor thread, hence, spawn_blocking + await.
-        if cfg!(feature = "testing") {
-            tokio::task::spawn_blocking({
-                let current = tracing::Span::current();
-                move || {
-                    let _entered = current.entered();
-                    tracing::info!("at failpoint in_progress_delete");
-                    fail::fail_point!("in_progress_delete");
-                }
-            })
-            .await
-            .expect("spawn_blocking");
-        }
+        pausable_failpoint!("in_progress_delete");

        {
            // Remove the timeline from the map.
@@ -1715,7 +1722,7 @@ impl Tenant {
        self.state.send_modify(|current_state| {
            use pageserver_api::models::ActivatingFrom;
            match &*current_state {
-                TenantState::Activating(_) | TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping => {
+                TenantState::Activating(_) | TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => {
                    panic!("caller is responsible for calling activate() only on Loading / Attaching tenants, got {state:?}", state = current_state);
                }
                TenantState::Loading => {
@@ -1779,7 +1786,16 @@ impl Tenant {
    /// - detach + ignore (freeze_and_flush == false)
    ///
    /// This will attempt to shutdown even if tenant is broken.
-    pub(crate) async fn shutdown(&self, freeze_and_flush: bool) -> Result<(), ShutdownError> {
+    ///
+    /// `shutdown_progress` is a [`completion::Barrier`] for the shutdown initiated by this call.
+    /// If the tenant is already shutting down, we return a clone of the first shutdown call's
+    /// `Barrier` as an `Err`. This not-first caller can use the returned barrier to join with
+    /// the ongoing shutdown.
+    async fn shutdown(
+        &self,
+        shutdown_progress: completion::Barrier,
+        freeze_and_flush: bool,
+    ) -> Result<(), completion::Barrier> {
        span::debug_assert_current_span_has_tenant_id();
        // Set tenant (and its timlines) to Stoppping state.
        //
@@ -1798,12 +1814,16 @@ impl Tenant {
        // But the tenant background loops are joined-on in our caller.
        // It's mesed up.
        // we just ignore the failure to stop
-        match self.set_stopping().await {
+
+        match self.set_stopping(shutdown_progress).await {
            Ok(()) => {}
            Err(SetStoppingError::Broken) => {
                // assume that this is acceptable
            }
-            Err(SetStoppingError::AlreadyStopping) => return Err(ShutdownError::AlreadyStopping),
+            Err(SetStoppingError::AlreadyStopping(other)) => {
+                // give caller the option to wait for this this shutdown
+                return Err(other);
+            }
        };

        if freeze_and_flush {
@@ -1835,7 +1855,7 @@ impl Tenant {
    /// This function waits for the tenant to become active if it isn't already, before transitioning it into Stopping state.
    ///
    /// This function is not cancel-safe!
-    async fn set_stopping(&self) -> Result<(), SetStoppingError> {
+    async fn set_stopping(&self, progress: completion::Barrier) -> Result<(), SetStoppingError> {
        let mut rx = self.state.subscribe();

        // cannot stop before we're done activating, so wait out until we're done activating
@@ -1847,7 +1867,7 @@ impl Tenant {
                );
                false
            }
-            TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping {} => true,
+            TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => true,
        })
        .await
        .expect("cannot drop self.state while on a &self method");
@@ -1862,7 +1882,7 @@ impl Tenant {
                // FIXME: due to time-of-check vs time-of-use issues, it can happen that new timelines
                // are created after the transition to Stopping. That's harmless, as the Timelines
                // won't be accessible to anyone afterwards, because the Tenant is in Stopping state.
-                *current_state = TenantState::Stopping;
+                *current_state = TenantState::Stopping { progress };
                // Continue stopping outside the closure. We need to grab timelines.lock()
                // and we plan to turn it into a tokio::sync::Mutex in a future patch.
                true
@@ -1874,9 +1894,9 @@ impl Tenant {
                err = Some(SetStoppingError::Broken);
                false
            }
-            TenantState::Stopping => {
+            TenantState::Stopping { progress } => {
                info!("Tenant is already in Stopping state");
-                err = Some(SetStoppingError::AlreadyStopping);
+                err = Some(SetStoppingError::AlreadyStopping(progress.clone()));
                false
            }
        });
@@ -1920,7 +1940,7 @@ impl Tenant {
                );
                false
            }
-            TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping {} => true,
+            TenantState::Active | TenantState::Broken { .. } | TenantState::Stopping { .. } => true,
        })
        .await
        .expect("cannot drop self.state while on a &self method");
@@ -1943,7 +1963,7 @@ impl Tenant {
                    warn!("Tenant is already in Broken state");
                }
                // This is the only "expected" path, any other path is a bug.
-                TenantState::Stopping => {
+                TenantState::Stopping { .. } => {
                    warn!(
                        "Marking Stopping tenant as Broken state, reason: {}",
                        reason
@@ -1976,7 +1996,7 @@ impl Tenant {
                TenantState::Active { .. } => {
                    return Ok(());
                }
-                TenantState::Broken { .. } | TenantState::Stopping => {
+                TenantState::Broken { .. } | TenantState::Stopping { .. } => {
                    // There's no chance the tenant can transition back into ::Active
                    return Err(WaitToBecomeActiveError::WillNotBecomeActive {
                        tenant_id: self.tenant_id,
@@ -2179,28 +2199,53 @@ impl Tenant {
        let (state, mut rx) = watch::channel(state);

        tokio::spawn(async move {
-            let mut current_state: &'static str = From::from(&*rx.borrow_and_update());
            let tid = tenant_id.to_string();
-            TENANT_STATE_METRIC
-                .with_label_values(&[&tid, current_state])
-                .inc();
-            loop {
-                match rx.changed().await {
-                    Ok(()) => {
-                        let new_state: &'static str = From::from(&*rx.borrow_and_update());
-                        TENANT_STATE_METRIC
-                            .with_label_values(&[&tid, current_state])
-                            .dec();
-                        TENANT_STATE_METRIC
-                            .with_label_values(&[&tid, new_state])
-                            .inc();

-                        current_state = new_state;
-                    }
-                    Err(_sender_dropped_error) => {
-                        info!("Tenant dropped the state updates sender, quitting waiting for tenant state change");
-                        return;
-                    }
+            fn inspect_state(state: &TenantState) -> ([&'static str; 1], bool) {
+                ([state.into()], matches!(state, TenantState::Broken { .. }))
+            }
+
+            let mut tuple = inspect_state(&rx.borrow_and_update());
+
+            let is_broken = tuple.1;
+            let mut counted_broken = if !is_broken {
+                // the tenant might be ignored and reloaded, so first remove any previous set
+                // element. it most likely has already been scraped, as these are manual operations
+                // right now. most likely we will add it back very soon.
+                drop(crate::metrics::BROKEN_TENANTS_SET.remove_label_values(&[&tid]));
+                false
+            } else {
+                // add the id to the set right away, there should not be any updates on the channel
+                // after
+                crate::metrics::BROKEN_TENANTS_SET
+                    .with_label_values(&[&tid])
+                    .set(1);
+                true
+            };
+
+            loop {
+                let labels = &tuple.0;
+                let current = TENANT_STATE_METRIC.with_label_values(labels);
+                current.inc();
+
+                if rx.changed().await.is_err() {
+                    // tenant has been dropped; decrement the counter because a tenant with that
+                    // state is no longer in tenant map, but allow any broken set item to exist
+                    // still.
+                    current.dec();
+                    break;
+                }
+
+                current.dec();
+                tuple = inspect_state(&rx.borrow_and_update());
+
+                let is_broken = tuple.1;
+                if is_broken && !counted_broken {
+                    counted_broken = true;
+                    // insert the tenant_id (back) into the set
+                    crate::metrics::BROKEN_TENANTS_SET
+                        .with_label_values(&[&tid])
+                        .inc();
                }
            }
        });
@@ -2226,7 +2271,7 @@ impl Tenant {
    /// Locate and load config
    pub(super) fn load_tenant_config(
        conf: &'static PageServerConf,
-        tenant_id: TenantId,
+        tenant_id: &TenantId,
    ) -> anyhow::Result<TenantConfOpt> {
        let target_config_path = conf.tenant_config_path(tenant_id);
        let target_config_display = target_config_path.display();
@@ -2813,7 +2858,7 @@ impl Tenant {
        timeline_struct.init_empty_layer_map(start_lsn);

        if let Err(e) =
-            self.create_timeline_files(&uninit_mark.timeline_path, new_timeline_id, new_metadata)
+            self.create_timeline_files(&uninit_mark.timeline_path, &new_timeline_id, new_metadata)
        {
            error!("Failed to create initial files for timeline {tenant_id}/{new_timeline_id}, cleaning up: {e:?}");
            cleanup_timeline_directory(uninit_mark);
@@ -2832,7 +2877,7 @@ impl Tenant {
    fn create_timeline_files(
        &self,
        timeline_path: &Path,
-        new_timeline_id: TimelineId,
+        new_timeline_id: &TimelineId,
        new_metadata: &TimelineMetadata,
    ) -> anyhow::Result<()> {
        crashsafe::create_dir(timeline_path).context("Failed to create timeline directory")?;
@@ -2843,8 +2888,8 @@ impl Tenant {

        save_metadata(
            self.conf,
+            &self.tenant_id,
            new_timeline_id,
-            self.tenant_id,
            new_metadata,
            true,
        )
@@ -2867,7 +2912,7 @@ impl Tenant {
            timelines.get(&timeline_id).is_none(),
            "Timeline {tenant_id}/{timeline_id} already exists in pageserver's memory"
        );
-        let timeline_path = self.conf.timeline_path(&timeline_id, &tenant_id);
+        let timeline_path = self.conf.timeline_path(&tenant_id, &timeline_id);
        anyhow::ensure!(
            !timeline_path.exists(),
            "Timeline {} already exists, cannot create its uninit mark file",
@@ -2998,10 +3043,10 @@ pub(crate) enum CreateTenantFilesMode {
 pub(crate) fn create_tenant_files(
    conf: &'static PageServerConf,
    tenant_conf: TenantConfOpt,
-    tenant_id: TenantId,
+    tenant_id: &TenantId,
    mode: CreateTenantFilesMode,
 ) -> anyhow::Result<PathBuf> {
-    let target_tenant_directory = conf.tenant_path(&tenant_id);
+    let target_tenant_directory = conf.tenant_path(tenant_id);
    anyhow::ensure!(
        !target_tenant_directory
            .try_exists()
@@ -3052,7 +3097,7 @@ pub(crate) fn create_tenant_files(
 fn try_create_target_tenant_dir(
    conf: &'static PageServerConf,
    tenant_conf: TenantConfOpt,
-    tenant_id: TenantId,
+    tenant_id: &TenantId,
    mode: CreateTenantFilesMode,
    temporary_tenant_dir: &Path,
    target_tenant_directory: &Path,
@@ -3076,7 +3121,7 @@ fn try_create_target_tenant_dir(
    }

    let temporary_tenant_timelines_dir = rebase_directory(
-        &conf.timelines_path(&tenant_id),
+        &conf.timelines_path(tenant_id),
        target_tenant_directory,
        temporary_tenant_dir,
    )
@@ -3088,7 +3133,7 @@ fn try_create_target_tenant_dir(
    )
    .with_context(|| format!("resolve tenant {tenant_id} temporary config path"))?;

-    Tenant::persist_tenant_config(&tenant_id, &temporary_tenant_config_path, tenant_conf, true)?;
+    Tenant::persist_tenant_config(tenant_id, &temporary_tenant_config_path, tenant_conf, true)?;

    crashsafe::create_dir(&temporary_tenant_timelines_dir).with_context(|| {
        format!(
@@ -3195,7 +3240,7 @@ impl Drop for Tenant {
    }
 }
 /// Dump contents of a layer file to stdout.
-pub fn dump_layerfile_from_path(
+pub async fn dump_layerfile_from_path(
    path: &Path,
    verbose: bool,
    ctx: &RequestContext,
@@ -3209,8 +3254,16 @@ pub fn dump_layerfile_from_path(
    file.read_exact_at(&mut header_buf, 0)?;

    match u16::from_be_bytes(header_buf) {
-        crate::IMAGE_FILE_MAGIC => ImageLayer::new_for_path(path, file)?.dump(verbose, ctx)?,
-        crate::DELTA_FILE_MAGIC => DeltaLayer::new_for_path(path, file)?.dump(verbose, ctx)?,
+        crate::IMAGE_FILE_MAGIC => {
+            ImageLayer::new_for_path(path, file)?
+                .dump(verbose, ctx)
+                .await?
+        }
+        crate::DELTA_FILE_MAGIC => {
+            DeltaLayer::new_for_path(path, file)?
+                .dump(verbose, ctx)
+                .await?
+        }
        magic => bail!("unrecognized magic identifier: {:?}", magic),
    }

@@ -3344,14 +3397,18 @@ pub mod harness {
        pub async fn load(&self) -> (Arc<Tenant>, RequestContext) {
            let ctx = RequestContext::new(TaskKind::UnitTest, DownloadBehavior::Error);
            (
-                self.try_load(&ctx)
+                self.try_load(&ctx, None)
                    .await
                    .expect("failed to load test tenant"),
                ctx,
            )
        }

-        pub async fn try_load(&self, ctx: &RequestContext) -> anyhow::Result<Arc<Tenant>> {
+        pub async fn try_load(
+            &self,
+            ctx: &RequestContext,
+            remote_storage: Option<remote_storage::GenericRemoteStorage>,
+        ) -> anyhow::Result<Arc<Tenant>> {
            let walredo_mgr = Arc::new(TestRedoManager);

            let tenant = Arc::new(Tenant::new(
@@ -3360,7 +3417,7 @@ pub mod harness {
                TenantConfOpt::from(self.tenant_conf),
                walredo_mgr,
                self.tenant_id,
-                None,
+                remote_storage,
            ));
            tenant
                .load(None, ctx)
@@ -3376,7 +3433,7 @@ pub mod harness {
        }

        pub fn timeline_path(&self, timeline_id: &TimelineId) -> PathBuf {
-            self.conf.timeline_path(timeline_id, &self.tenant_id)
+            self.conf.timeline_path(&self.tenant_id, timeline_id)
        }
    }

@@ -3422,6 +3479,7 @@ mod tests {
    use hex_literal::hex;
    use once_cell::sync::Lazy;
    use rand::{thread_rng, Rng};
+    use tokio_util::sync::CancellationToken;

    static TEST_KEY: Lazy<Key> =
        Lazy::new(|| Key::from_slice(&hex!("112222222233333333444444445500000001")));
@@ -3898,7 +3956,11 @@ mod tests {
        metadata_bytes[8] ^= 1;
        std::fs::write(metadata_path, metadata_bytes)?;

-        let err = harness.try_load(&ctx).await.err().expect("should fail");
+        let err = harness
+            .try_load(&ctx, None)
+            .await
+            .err()
+            .expect("should fail");
        // get all the stack with all .context, not tonly the last one
        let message = format!("{err:#}");
        let expected = "Failed to parse metadata bytes from path";
@@ -3939,7 +4001,7 @@ mod tests {
        drop(writer);

        tline.freeze_and_flush().await?;
-        tline.compact(&ctx).await?;
+        tline.compact(&CancellationToken::new(), &ctx).await?;

        let writer = tline.writer().await;
        writer
@@ -3949,7 +4011,7 @@ mod tests {
        drop(writer);

        tline.freeze_and_flush().await?;
-        tline.compact(&ctx).await?;
+        tline.compact(&CancellationToken::new(), &ctx).await?;

        let writer = tline.writer().await;
        writer
@@ -3959,7 +4021,7 @@ mod tests {
        drop(writer);

        tline.freeze_and_flush().await?;
-        tline.compact(&ctx).await?;
+        tline.compact(&CancellationToken::new(), &ctx).await?;

        let writer = tline.writer().await;
        writer
@@ -3969,7 +4031,7 @@ mod tests {
        drop(writer);

        tline.freeze_and_flush().await?;
-        tline.compact(&ctx).await?;
+        tline.compact(&CancellationToken::new(), &ctx).await?;

        assert_eq!(
            tline.get(*TEST_KEY, Lsn(0x10), &ctx).await?,
@@ -4038,7 +4100,7 @@ mod tests {
                .update_gc_info(Vec::new(), cutoff, Duration::ZERO, &ctx)
                .await?;
            tline.freeze_and_flush().await?;
-            tline.compact(&ctx).await?;
+            tline.compact(&CancellationToken::new(), &ctx).await?;
            tline.gc().await?;
        }

@@ -4115,7 +4177,7 @@ mod tests {
                .update_gc_info(Vec::new(), cutoff, Duration::ZERO, &ctx)
                .await?;
            tline.freeze_and_flush().await?;
-            tline.compact(&ctx).await?;
+            tline.compact(&CancellationToken::new(), &ctx).await?;
            tline.gc().await?;
        }

@@ -4203,7 +4265,7 @@ mod tests {
                .update_gc_info(Vec::new(), cutoff, Duration::ZERO, &ctx)
                .await?;
            tline.freeze_and_flush().await?;
-            tline.compact(&ctx).await?;
+            tline.compact(&CancellationToken::new(), &ctx).await?;
            tline.gc().await?;
        }

@@ -4329,13 +4391,13 @@ mod tests {
        // assert freeze_and_flush exercised the initdb optimization
        {
            let state = tline.flush_loop_state.lock().unwrap();
-            let
-                timeline::FlushLoopState::Running {
-                    expect_initdb_optimization,
-                    initdb_optimization_count,
-                } = *state else {
-                    panic!("unexpected state: {:?}", *state);
-                };
+            let timeline::FlushLoopState::Running {
+                expect_initdb_optimization,
+                initdb_optimization_count,
+            } = *state
+            else {
+                panic!("unexpected state: {:?}", *state);
+            };
            assert!(expect_initdb_optimization);
            assert!(initdb_optimization_count > 0);
        }
@@ -4370,7 +4432,7 @@ mod tests {

        assert!(!harness
            .conf
-            .timeline_path(&TIMELINE_ID, &tenant.tenant_id)
+            .timeline_path(&tenant.tenant_id, &TIMELINE_ID)
            .exists());

        assert!(!harness
--- a/pageserver/src/tenant/blob_io.rs
+++ b/pageserver/src/tenant/blob_io.rs
@@ -16,29 +16,19 @@ use crate::tenant::block_io::{BlockCursor, BlockReader};
 use std::cmp::min;
 use std::io::{Error, ErrorKind};

-/// For reading
-pub trait BlobCursor {
+impl<R> BlockCursor<R>
+where
+    R: BlockReader,
+{
    /// Read a blob into a new buffer.
-    fn read_blob(&mut self, offset: u64) -> Result<Vec<u8>, std::io::Error> {
+    pub fn read_blob(&mut self, offset: u64) -> Result<Vec<u8>, std::io::Error> {
        let mut buf = Vec::new();
        self.read_blob_into_buf(offset, &mut buf)?;
        Ok(buf)
    }
-
    /// Read blob into the given buffer. Any previous contents in the buffer
    /// are overwritten.
-    fn read_blob_into_buf(
-        &mut self,
-        offset: u64,
-        dstbuf: &mut Vec<u8>,
-    ) -> Result<(), std::io::Error>;
-}
-
-impl<R> BlobCursor for BlockCursor<R>
-where
-    R: BlockReader,
-{
-    fn read_blob_into_buf(
+    pub fn read_blob_into_buf(
        &mut self,
        offset: u64,
        dstbuf: &mut Vec<u8>,
--- a/pageserver/src/tenant/disk_btree.rs
+++ b/pageserver/src/tenant/disk_btree.rs
@@ -442,7 +442,7 @@ where
    writer: W,

    ///
-    /// stack[0] is the current root page, stack.last() is the leaf.
+    /// `stack[0]` is the current root page, `stack.last()` is the leaf.
    ///
    /// We maintain the length of the stack to be always greater than zero.
    /// Two exceptions are:
--- a/pageserver/src/tenant/ephemeral_file.rs
+++ b/pageserver/src/tenant/ephemeral_file.rs
@@ -55,7 +55,7 @@ impl EphemeralFile {
        l.next_file_id += 1;

        let filename = conf
-            .timeline_path(&timeline_id, &tenant_id)
+            .timeline_path(&tenant_id, &timeline_id)
            .join(PathBuf::from(format!("ephemeral-{}", file_id)));

        let file = VirtualFile::open_with_options(
@@ -328,7 +328,7 @@ fn to_io_error(e: anyhow::Error, context: &str) -> io::Error {
 #[cfg(test)]
 mod tests {
    use super::*;
-    use crate::tenant::blob_io::{BlobCursor, BlobWriter};
+    use crate::tenant::blob_io::BlobWriter;
    use crate::tenant::block_io::BlockCursor;
    use rand::{seq::SliceRandom, thread_rng, RngCore};
    use std::fs;
@@ -346,7 +346,7 @@ mod tests {

        let tenant_id = TenantId::from_str("11000000000000000000000000000000").unwrap();
        let timeline_id = TimelineId::from_str("22000000000000000000000000000000").unwrap();
-        fs::create_dir_all(conf.timeline_path(&timeline_id, &tenant_id))?;
+        fs::create_dir_all(conf.timeline_path(&tenant_id, &timeline_id))?;

        Ok((conf, tenant_id, timeline_id))
    }
--- a/pageserver/src/tenant/layer_map.rs
+++ b/pageserver/src/tenant/layer_map.rs
@@ -16,7 +16,7 @@
 //! Other read methods are less critical but still impact performance of background tasks.
 //!
 //! This data structure relies on a persistent/immutable binary search tree. See the
-//! following lecture for an introduction https://www.youtube.com/watch?v=WqCWghETNDc&t=581s
+//! following lecture for an introduction <https://www.youtube.com/watch?v=WqCWghETNDc&t=581s>
 //! Summary: A persistent/immutable BST (and persistent data structures in general) allows
 //! you to modify the tree in such a way that each modification creates a new "version"
 //! of the tree. When you modify it, you get a new version, but all previous versions are
@@ -40,7 +40,7 @@
 //! afterwards. We can add layers as long as they have larger LSNs than any previous layer in
 //! the map, but if we need to remove a layer, or insert anything with an older LSN, we need
 //! to throw away most of the persistent BST and build a new one, starting from the oldest
-//! LSN. See `LayerMap::flush_updates()`.
+//! LSN. See [`LayerMap::flush_updates()`].
 //!

 mod historic_layer_coverage;
@@ -60,7 +60,6 @@ use utils::lsn::Lsn;
 use historic_layer_coverage::BufferedHistoricLayerCoverage;
 pub use historic_layer_coverage::LayerKey;

-use super::storage_layer::range_eq;
 use super::storage_layer::PersistentLayerDesc;

 ///
@@ -365,7 +364,7 @@ impl LayerMap {
    }

    pub fn is_l0(layer: &PersistentLayerDesc) -> bool {
-        range_eq(&layer.get_key_range(), &(Key::MIN..Key::MAX))
+        layer.get_key_range() == (Key::MIN..Key::MAX)
    }

    /// This function determines which layers are counted in `count_deltas`:
@@ -397,7 +396,7 @@ impl LayerMap {
        }

        // Case 2
-        if range_eq(partition_range, &(Key::MIN..Key::MAX)) {
+        if partition_range == &(Key::MIN..Key::MAX) {
            return true;
        }

@@ -627,17 +626,17 @@ impl LayerMap {

    /// debugging function to print out the contents of the layer map
    #[allow(unused)]
-    pub fn dump(&self, verbose: bool, ctx: &RequestContext) -> Result<()> {
+    pub async fn dump(&self, verbose: bool, ctx: &RequestContext) -> Result<()> {
        println!("Begin dump LayerMap");

        println!("open_layer:");
        if let Some(open_layer) = &self.open_layer {
-            open_layer.dump(verbose, ctx)?;
+            open_layer.dump(verbose, ctx).await?;
        }

        println!("frozen_layers:");
        for frozen_layer in self.frozen_layers.iter() {
-            frozen_layer.dump(verbose, ctx)?;
+            frozen_layer.dump(verbose, ctx).await?;
        }

        println!("historic_layers:");
@@ -652,19 +651,35 @@ impl LayerMap {
 #[cfg(test)]
 mod tests {
    use super::LayerMap;
-    use crate::tenant::storage_layer::{tests::LayerDescriptor, LayerFileName};
+    use crate::tenant::storage_layer::LayerFileName;
    use std::str::FromStr;
    use std::sync::Arc;

    mod l0_delta_layers_updated {

        use crate::tenant::{
-            storage_layer::{PersistentLayer, PersistentLayerDesc},
-            timeline::LayerFileManager,
+            storage_layer::{AsLayerDesc, PersistentLayerDesc},
+            timeline::layer_manager::LayerFileManager,
        };

        use super::*;

+        struct LayerObject(PersistentLayerDesc);
+
+        impl AsLayerDesc for LayerObject {
+            fn layer_desc(&self) -> &PersistentLayerDesc {
+                &self.0
+            }
+        }
+
+        impl LayerObject {
+            fn new(desc: PersistentLayerDesc) -> Self {
+                LayerObject(desc)
+            }
+        }
+
+        type TestLayerFileManager = LayerFileManager<LayerObject>;
+
        #[test]
        fn for_full_range_delta() {
            // l0_delta_layers are used by compaction, and should observe all buffered updates
@@ -701,18 +716,18 @@ mod tests {

            let layer = "000000000000000000000000000000000000-FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF__0000000053423C21-0000000053424D69";
            let layer = LayerFileName::from_str(layer).unwrap();
-            let layer = LayerDescriptor::from(layer);
+            let layer = PersistentLayerDesc::from(layer);

            // same skeletan construction; see scenario below
-            let not_found = Arc::new(layer.clone());
-            let new_version = Arc::new(layer);
+            let not_found = Arc::new(LayerObject::new(layer.clone()));
+            let new_version = Arc::new(LayerObject::new(layer));

            // after the immutable storage state refactor, the replace operation
            // will not use layer map any more. We keep it here for consistency in test cases
            // and can remove it in the future.
            let _map = LayerMap::default();

-            let mut mapping = LayerFileManager::new();
+            let mut mapping = TestLayerFileManager::new();

            mapping
                .replace_and_verify(not_found, new_version)
@@ -721,10 +736,10 @@ mod tests {

        fn l0_delta_layers_updated_scenario(layer_name: &str, expected_l0: bool) {
            let name = LayerFileName::from_str(layer_name).unwrap();
-            let skeleton = LayerDescriptor::from(name);
+            let skeleton = PersistentLayerDesc::from(name);

-            let remote = Arc::new(skeleton.clone());
-            let downloaded = Arc::new(skeleton);
+            let remote = Arc::new(LayerObject::new(skeleton.clone()));
+            let downloaded = Arc::new(LayerObject::new(skeleton));

            let mut map = LayerMap::default();
            let mut mapping = LayerFileManager::new();
--- a/pageserver/src/tenant/layer_map/historic_layer_coverage.rs
+++ b/pageserver/src/tenant/layer_map/historic_layer_coverage.rs
@@ -122,8 +122,7 @@ impl<Value: Clone> HistoricLayerCoverage<Value> {
        self.head = self
            .historic
            .iter()
-            .rev()
-            .next()
+            .next_back()
            .map(|(_, v)| v.clone())
            .unwrap_or_default();
    }
@@ -412,7 +411,7 @@ fn test_persistent_overlapping() {
 /// still be more critical.
 ///
 /// See this for more on persistent and retroactive techniques:
-/// https://www.youtube.com/watch?v=WqCWghETNDc&t=581s
+/// <https://www.youtube.com/watch?v=WqCWghETNDc&t=581s>
 pub struct BufferedHistoricLayerCoverage<Value> {
    /// A persistent layer map that we rebuild when we need to retroactively update
    historic_coverage: HistoricLayerCoverage<Value>,
--- a/pageserver/src/tenant/layer_map/layer_coverage.rs
+++ b/pageserver/src/tenant/layer_map/layer_coverage.rs
@@ -2,7 +2,7 @@ use std::ops::Range;

 // NOTE the `im` crate has 20x more downloads and also has
 // persistent/immutable BTree. But it's bugged so rpds is a
-// better choice https://github.com/neondatabase/neon/issues/3395
+// better choice <https://github.com/neondatabase/neon/issues/3395>
 use rpds::RedBlackTreeMapSync;

 /// Data structure that can efficiently:
@@ -11,7 +11,7 @@ use rpds::RedBlackTreeMapSync;
 /// - insert layers in non-decreasing lsn.start order
 ///
 /// For a detailed explanation and justification of this approach, see:
-/// https://neon.tech/blog/persistent-structures-in-neons-wal-indexing
+/// <https://neon.tech/blog/persistent-structures-in-neons-wal-indexing>
 ///
 /// NOTE The struct is parameterized over Value for easier
 ///      testing, but in practice it's some sort of layer.
@@ -113,8 +113,7 @@ impl<Value: Clone> LayerCoverage<Value> {
    pub fn query(&self, key: i128) -> Option<Value> {
        self.nodes
            .range(..=key)
-            .rev()
-            .next()?
+            .next_back()?
            .1
            .as_ref()
            .map(|(_, v)| v.clone())
--- a/pageserver/src/tenant/manifest.rs
+++ b/pageserver/src/tenant/manifest.rs
@@ -24,7 +24,7 @@
 //! Currently, this is not used in the system. Future refactors will ensure
 //! the storage state will be recorded in this file, and the system can be
 //! recovered from this file. This is tracked in
-//! https://github.com/neondatabase/neon/issues/4418
+//! <https://github.com/neondatabase/neon/issues/4418>

 use std::io::{self, Read, Write};

--- a/pageserver/src/tenant/metadata.rs
+++ b/pageserver/src/tenant/metadata.rs
@@ -1,10 +1,12 @@
 //! Every image of a certain timeline from [`crate::tenant::Tenant`]
 //! has a metadata that needs to be stored persistently.
 //!
-//! Later, the file gets is used in [`crate::remote_storage::storage_sync`] as a part of
+//! Later, the file gets used in [`remote_timeline_client`] as a part of
 //! external storage import and export operations.
 //!
 //! The module contains all structs and related helper methods related to timeline metadata.
+//!
+//! [`remote_timeline_client`]: super::remote_timeline_client

 use std::fs::{File, OpenOptions};
 use std::io::Write;
@@ -232,13 +234,13 @@ impl TimelineMetadata {
 /// Save timeline metadata to file
 pub fn save_metadata(
    conf: &'static PageServerConf,
-    timeline_id: TimelineId,
-    tenant_id: TenantId,
+    tenant_id: &TenantId,
+    timeline_id: &TimelineId,
    data: &TimelineMetadata,
    first_save: bool,
 ) -> anyhow::Result<()> {
    let _enter = info_span!("saving metadata").entered();
-    let path = conf.metadata_path(timeline_id, tenant_id);
+    let path = conf.metadata_path(tenant_id, timeline_id);
    // use OpenOptions to ensure file presence is consistent with first_save
    let mut file = VirtualFile::open_with_options(
        &path,
@@ -267,10 +269,10 @@ pub fn save_metadata(

 pub fn load_metadata(
    conf: &'static PageServerConf,
-    timeline_id: TimelineId,
-    tenant_id: TenantId,
+    tenant_id: &TenantId,
+    timeline_id: &TimelineId,
 ) -> anyhow::Result<TimelineMetadata> {
-    let metadata_path = conf.metadata_path(timeline_id, tenant_id);
+    let metadata_path = conf.metadata_path(tenant_id, timeline_id);
    let metadata_bytes = std::fs::read(&metadata_path).with_context(|| {
        format!(
            "Failed to read metadata bytes from path {}",
--- a/pageserver/src/tenant/mgr.rs
+++ b/pageserver/src/tenant/mgr.rs
@@ -184,9 +184,9 @@ pub fn schedule_local_tenant_processing(
            format!("Could not parse tenant id out of the tenant dir name in path {tenant_path:?}")
        })?;

-    let tenant_ignore_mark = conf.tenant_ignore_mark_file_path(tenant_id);
+    let tenant_ignore_mark = conf.tenant_ignore_mark_file_path(&tenant_id);
    anyhow::ensure!(
-        !conf.tenant_ignore_mark_file_path(tenant_id).exists(),
+        !conf.tenant_ignore_mark_file_path(&tenant_id).exists(),
        "Cannot load tenant, ignore mark found at {tenant_ignore_mark:?}"
    );

@@ -233,11 +233,17 @@ pub fn schedule_local_tenant_processing(
 /// That could be easily misinterpreted by control plane, the consumer of the
 /// management API. For example, it could attach the tenant on a different pageserver.
 /// We would then be in split-brain once this pageserver restarts.
-#[instrument]
+#[instrument(skip_all)]
 pub async fn shutdown_all_tenants() {
+    shutdown_all_tenants0(&TENANTS).await
+}
+
+async fn shutdown_all_tenants0(tenants: &tokio::sync::RwLock<TenantsMap>) {
+    use utils::completion;
+
    // Prevent new tenants from being created.
    let tenants_to_shut_down = {
-        let mut m = TENANTS.write().await;
+        let mut m = tenants.write().await;
        match &mut *m {
            TenantsMap::Initializing => {
                *m = TenantsMap::ShuttingDown(HashMap::default());
@@ -262,14 +268,41 @@ pub async fn shutdown_all_tenants() {
    for (tenant_id, tenant) in tenants_to_shut_down {
        join_set.spawn(
            async move {
-                let freeze_and_flush = true;
+                // ordering shouldn't matter for this, either we store true right away or never
+                let ordering = std::sync::atomic::Ordering::Relaxed;
+                let joined_other = std::sync::atomic::AtomicBool::new(false);

-                match tenant.shutdown(freeze_and_flush).await {
-                    Ok(()) => debug!("tenant successfully stopped"),
-                    Err(super::ShutdownError::AlreadyStopping) => {
-                        warn!("tenant was already shutting down")
+                let mut shutdown = std::pin::pin!(async {
+                    let freeze_and_flush = true;
+
+                    let res = {
+                        let (_guard, shutdown_progress) = completion::channel();
+                        tenant.shutdown(shutdown_progress, freeze_and_flush).await
+                    };
+
+                    if let Err(other_progress) = res {
+                        // join the another shutdown in progress
+                        joined_other.store(true, ordering);
+                        other_progress.wait().await;
                    }
-                }
+                });
+
+                // in practice we might not have a lot time to go, since systemd is going to
+                // SIGKILL us at 10s, but we can try. delete tenant might take a while, so put out
+                // a warning.
+                let warning = std::time::Duration::from_secs(5);
+                let mut warning = std::pin::pin!(tokio::time::sleep(warning));
+
+                tokio::select! {
+                    _ = &mut shutdown => {},
+                    _ = &mut warning => {
+                        let joined_other = joined_other.load(ordering);
+                        warn!(%joined_other, "waiting for the shutdown to complete");
+                        shutdown.await;
+                    }
+                };
+
+                debug!("tenant successfully stopped");
            }
            .instrument(info_span!("shutdown", %tenant_id)),
        );
@@ -310,7 +343,7 @@ pub async fn create_tenant(
        // We're holding the tenants lock in write mode while doing local IO.
        // If this section ever becomes contentious, introduce a new `TenantState::Creating`
        // and do the work in that state.
-        let tenant_directory = super::create_tenant_files(conf, tenant_conf, tenant_id, CreateTenantFilesMode::Create)?;
+        let tenant_directory = super::create_tenant_files(conf, tenant_conf, &tenant_id, CreateTenantFilesMode::Create)?;
        // TODO: tenant directory remains on disk if we bail out from here on.
        //       See https://github.com/neondatabase/neon/issues/4233

@@ -344,14 +377,9 @@ pub async fn set_new_tenant_config(
    info!("configuring tenant {tenant_id}");
    let tenant = get_tenant(tenant_id, true).await?;

-    let tenant_config_path = conf.tenant_config_path(tenant_id);
-    Tenant::persist_tenant_config(
-        &tenant.tenant_id(),
-        &tenant_config_path,
-        new_tenant_conf,
-        false,
-    )
-    .map_err(SetNewTenantConfigError::Persist)?;
+    let tenant_config_path = conf.tenant_config_path(&tenant_id);
+    Tenant::persist_tenant_config(&tenant_id, &tenant_config_path, new_tenant_conf, false)
+        .map_err(SetNewTenantConfigError::Persist)?;
    tenant.set_new_tenant_config(new_tenant_conf);
    Ok(())
 }
@@ -418,6 +446,15 @@ pub async fn detach_tenant(
    conf: &'static PageServerConf,
    tenant_id: TenantId,
    detach_ignored: bool,
+) -> Result<(), TenantStateError> {
+    detach_tenant0(conf, &TENANTS, tenant_id, detach_ignored).await
+}
+
+async fn detach_tenant0(
+    conf: &'static PageServerConf,
+    tenants: &tokio::sync::RwLock<TenantsMap>,
+    tenant_id: TenantId,
+    detach_ignored: bool,
 ) -> Result<(), TenantStateError> {
    let local_files_cleanup_operation = |tenant_id_to_clean| async move {
        let local_tenant_directory = conf.tenant_path(&tenant_id_to_clean);
@@ -430,12 +467,13 @@ pub async fn detach_tenant(
    };

    let removal_result =
-        remove_tenant_from_memory(tenant_id, local_files_cleanup_operation(tenant_id)).await;
+        remove_tenant_from_memory(tenants, tenant_id, local_files_cleanup_operation(tenant_id))
+            .await;

    // Ignored tenants are not present in memory and will bail the removal from memory operation.
    // Before returning the error, check for ignored tenant removal case — we only need to clean its local files then.
    if detach_ignored && matches!(removal_result, Err(TenantStateError::NotFound(_))) {
-        let tenant_ignore_mark = conf.tenant_ignore_mark_file_path(tenant_id);
+        let tenant_ignore_mark = conf.tenant_ignore_mark_file_path(&tenant_id);
        if tenant_ignore_mark.exists() {
            info!("Detaching an ignored tenant");
            local_files_cleanup_operation(tenant_id)
@@ -457,7 +495,7 @@ pub async fn load_tenant(
 ) -> Result<(), TenantMapInsertError> {
    tenant_map_insert(tenant_id, || {
        let tenant_path = conf.tenant_path(&tenant_id);
-        let tenant_ignore_mark = conf.tenant_ignore_mark_file_path(tenant_id);
+        let tenant_ignore_mark = conf.tenant_ignore_mark_file_path(&tenant_id);
        if tenant_ignore_mark.exists() {
            std::fs::remove_file(&tenant_ignore_mark)
                .with_context(|| format!("Failed to remove tenant ignore mark {tenant_ignore_mark:?} during tenant loading"))?;
@@ -477,8 +515,16 @@ pub async fn ignore_tenant(
    conf: &'static PageServerConf,
    tenant_id: TenantId,
 ) -> Result<(), TenantStateError> {
-    remove_tenant_from_memory(tenant_id, async {
-        let ignore_mark_file = conf.tenant_ignore_mark_file_path(tenant_id);
+    ignore_tenant0(conf, &TENANTS, tenant_id).await
+}
+
+async fn ignore_tenant0(
+    conf: &'static PageServerConf,
+    tenants: &tokio::sync::RwLock<TenantsMap>,
+    tenant_id: TenantId,
+) -> Result<(), TenantStateError> {
+    remove_tenant_from_memory(tenants, tenant_id, async {
+        let ignore_mark_file = conf.tenant_ignore_mark_file_path(&tenant_id);
        fs::File::create(&ignore_mark_file)
            .await
            .context("Failed to create ignore mark file")
@@ -525,7 +571,7 @@ pub async fn attach_tenant(
    ctx: &RequestContext,
 ) -> Result<(), TenantMapInsertError> {
    tenant_map_insert(tenant_id, || {
-        let tenant_dir = create_tenant_files(conf, tenant_conf, tenant_id, CreateTenantFilesMode::Attach)?;
+        let tenant_dir = create_tenant_files(conf, tenant_conf, &tenant_id, CreateTenantFilesMode::Attach)?;
        // TODO: tenant directory remains on disk if we bail out from here on.
        //       See https://github.com/neondatabase/neon/issues/4233

@@ -602,18 +648,21 @@ where
 /// If the cleanup fails, tenant will stay in memory in [`TenantState::Broken`] state, and another removal
 /// operation would be needed to remove it.
 async fn remove_tenant_from_memory<V, F>(
+    tenants: &tokio::sync::RwLock<TenantsMap>,
    tenant_id: TenantId,
    tenant_cleanup: F,
 ) -> Result<V, TenantStateError>
 where
    F: std::future::Future<Output = anyhow::Result<V>>,
 {
+    use utils::completion;
+
    // It's important to keep the tenant in memory after the final cleanup, to avoid cleanup races.
    // The exclusive lock here ensures we don't miss the tenant state updates before trying another removal.
    // tenant-wde cleanup operations may take some time (removing the entire tenant directory), we want to
    // avoid holding the lock for the entire process.
    let tenant = {
-        TENANTS
+        tenants
            .write()
            .await
            .get(&tenant_id)
@@ -621,14 +670,20 @@ where
            .ok_or(TenantStateError::NotFound(tenant_id))?
    };

+    // allow pageserver shutdown to await for our completion
+    let (_guard, progress) = completion::channel();
+
+    // whenever we remove a tenant from memory, we don't want to flush and wait for upload
    let freeze_and_flush = false;

    // shutdown is sure to transition tenant to stopping, and wait for all tasks to complete, so
    // that we can continue safely to cleanup.
-    match tenant.shutdown(freeze_and_flush).await {
+    match tenant.shutdown(progress, freeze_and_flush).await {
        Ok(()) => {}
-        Err(super::ShutdownError::AlreadyStopping) => {
-            return Err(TenantStateError::IsStopping(tenant_id))
+        Err(_other) => {
+            // if pageserver shutdown or other detach/ignore is already ongoing, we don't want to
+            // wait for it but return an error right away because these are distinct requests.
+            return Err(TenantStateError::IsStopping(tenant_id));
        }
    }

@@ -637,14 +692,14 @@ where
        .with_context(|| format!("Failed to run cleanup for tenant {tenant_id}"))
    {
        Ok(hook_value) => {
-            let mut tenants_accessor = TENANTS.write().await;
+            let mut tenants_accessor = tenants.write().await;
            if tenants_accessor.remove(&tenant_id).is_none() {
                warn!("Tenant {tenant_id} got removed from memory before operation finished");
            }
            Ok(hook_value)
        }
        Err(e) => {
-            let tenants_accessor = TENANTS.read().await;
+            let tenants_accessor = tenants.read().await;
            match tenants_accessor.get(&tenant_id) {
                Some(tenant) => {
                    tenant.set_broken(e.to_string()).await;
@@ -695,7 +750,7 @@ pub async fn immediate_gc(
            fail::fail_point!("immediate_gc_task_pre");
            let result = tenant
                .gc_iteration(Some(timeline_id), gc_horizon, pitr, &ctx)
-                .instrument(info_span!("manual_gc", tenant = %tenant_id, timeline = %timeline_id))
+                .instrument(info_span!("manual_gc", %tenant_id, %timeline_id))
                .await;
                // FIXME: `gc_iteration` can return an error for multiple reasons; we should handle it
                // better once the types support it.
@@ -713,53 +768,108 @@ pub async fn immediate_gc(
    Ok(wait_task_done)
 }

-pub async fn immediate_compact(
-    tenant_id: TenantId,
-    timeline_id: TimelineId,
-    ctx: &RequestContext,
-) -> Result<tokio::sync::oneshot::Receiver<anyhow::Result<()>>, ApiError> {
-    let guard = TENANTS.read().await;
+#[cfg(test)]
+mod tests {
+    use std::collections::HashMap;
+    use std::sync::Arc;
+    use tracing::{info_span, Instrument};

-    let tenant = guard
-        .get(&tenant_id)
-        .map(Arc::clone)
-        .with_context(|| format!("tenant {tenant_id}"))
-        .map_err(|e| ApiError::NotFound(e.into()))?;
+    use super::{super::harness::TenantHarness, TenantsMap};

-    let timeline = tenant
-        .get_timeline(timeline_id, true)
-        .map_err(|e| ApiError::NotFound(e.into()))?;
+    #[tokio::test(start_paused = true)]
+    async fn shutdown_joins_remove_tenant_from_memory() {
+        // the test is a bit ugly with the lockstep together with spawned tasks. the aim is to make
+        // sure `shutdown_all_tenants0` per-tenant processing joins in any active
+        // remove_tenant_from_memory calls, which is enforced by making the operation last until
+        // we've ran `shutdown_all_tenants0` for a long time.

-    // Run in task_mgr to avoid race with tenant_detach operation
-    let ctx = ctx.detached_child(TaskKind::Compaction, DownloadBehavior::Download);
-    let (task_done, wait_task_done) = tokio::sync::oneshot::channel();
-    task_mgr::spawn(
-        &tokio::runtime::Handle::current(),
-        TaskKind::Compaction,
-        Some(tenant_id),
-        Some(timeline_id),
-        &format!(
-            "timeline_compact_handler compaction run for tenant {tenant_id} timeline {timeline_id}"
-        ),
-        false,
-        async move {
-            let result = timeline
-                .compact(&ctx)
-                .instrument(
-                    info_span!("manual_compact", tenant = %tenant_id, timeline = %timeline_id),
-                )
-                .await;
+        let (t, _ctx) = TenantHarness::create("shutdown_joins_detach")
+            .unwrap()
+            .load()
+            .await;

-            match task_done.send(result) {
-                Ok(_) => (),
-                Err(result) => error!("failed to send compaction result: {result:?}"),
-            }
-            Ok(())
-        },
-    );
+        // harness loads it to active, which is forced and nothing is running on the tenant

-    // drop the guard until after we've spawned the task so that timeline shutdown will wait for the task
-    drop(guard);
+        let id = t.tenant_id();

-    Ok(wait_task_done)
+        // tenant harness configures the logging and we cannot escape it
+        let _e = info_span!("testing", tenant_id = %id).entered();
+
+        let tenants = HashMap::from([(id, t.clone())]);
+        let tenants = Arc::new(tokio::sync::RwLock::new(TenantsMap::Open(tenants)));
+
+        let (until_cleanup_completed, can_complete_cleanup) = utils::completion::channel();
+        let (until_cleanup_started, cleanup_started) = utils::completion::channel();
+
+        // start a "detaching operation", which will take a while, until can_complete_cleanup
+        let cleanup_task = {
+            let jh = tokio::spawn({
+                let tenants = tenants.clone();
+                async move {
+                    let cleanup = async move {
+                        drop(until_cleanup_started);
+                        can_complete_cleanup.wait().await;
+                        anyhow::Ok(())
+                    };
+                    super::remove_tenant_from_memory(&tenants, id, cleanup).await
+                }
+                .instrument(info_span!("foobar", tenant_id = %id))
+            });
+
+            // now the long cleanup should be in place, with the stopping state
+            cleanup_started.wait().await;
+            jh
+        };
+
+        let mut cleanup_progress = std::pin::pin!(t
+            .shutdown(utils::completion::Barrier::default(), false)
+            .await
+            .unwrap_err()
+            .wait());
+
+        let mut shutdown_task = {
+            let (until_shutdown_started, shutdown_started) = utils::completion::channel();
+
+            let shutdown_task = tokio::spawn(async move {
+                drop(until_shutdown_started);
+                super::shutdown_all_tenants0(&tenants).await;
+            });
+
+            shutdown_started.wait().await;
+            shutdown_task
+        };
+
+        // if the joining in is removed from shutdown_all_tenants0, the shutdown_task should always
+        // get to complete within timeout and fail the test. it is expected to continue awaiting
+        // until completion or SIGKILL during normal shutdown.
+        //
+        // the timeout is long to cover anything that shutdown_task could be doing, but it is
+        // handled instantly because we use tokio's time pausing in this test. 100s is much more than
+        // what we get from systemd on shutdown (10s).
+        let long_time = std::time::Duration::from_secs(100);
+        tokio::select! {
+            _ = &mut shutdown_task => unreachable!("shutdown must continue, until_cleanup_completed is not dropped"),
+            _ = &mut cleanup_progress => unreachable!("cleanup progress must continue, until_cleanup_completed is not dropped"),
+            _ = tokio::time::sleep(long_time) => {},
+        }
+
+        // allow the remove_tenant_from_memory and thus eventually the shutdown to continue
+        drop(until_cleanup_completed);
+
+        let (je, ()) = tokio::join!(shutdown_task, cleanup_progress);
+        je.expect("Tenant::shutdown shutdown not have panicked");
+        cleanup_task
+            .await
+            .expect("no panicking")
+            .expect("remove_tenant_from_memory failed");
+
+        futures::future::poll_immediate(
+            t.shutdown(utils::completion::Barrier::default(), false)
+                .await
+                .unwrap_err()
+                .wait(),
+        )
+        .await
+        .expect("the stopping progress must still be complete");
+    }
 }
--- a/pageserver/src/tenant/remote_timeline_client.rs
+++ b/pageserver/src/tenant/remote_timeline_client.rs
@@ -135,7 +135,7 @@
 //! - Initiate upload queue with that [`IndexPart`].
 //! - Reschedule all lost operations by comparing the local filesystem state
 //!   and remote state as per [`IndexPart`]. This is done in
-//!   [`Timeline::timeline_init_and_sync`] and [`Timeline::reconcile_with_remote`].
+//!   [`Tenant::timeline_init_and_sync`] and [`Timeline::reconcile_with_remote`].
 //!
 //! Note that if we crash during file deletion between the index update
 //! that removes the file from the list of files, and deleting the remote file,
@@ -163,8 +163,8 @@
 //!   - download their remote [`IndexPart`]s
 //!   - create `Timeline` struct and a `RemoteTimelineClient`
 //!   - initialize the client's upload queue with its `IndexPart`
-//!   - create [`RemoteLayer`] instances for layers that are referenced by `IndexPart`
-//!     but not present locally
+//!   - create [`RemoteLayer`](super::storage_layer::RemoteLayer) instances
+//!     for layers that are referenced by `IndexPart` but not present locally
 //!   - schedule uploads for layers that are only present locally.
 //!   - if the remote `IndexPart`'s metadata was newer than the metadata in
 //!     the local filesystem, write the remote metadata to the local filesystem
@@ -198,6 +198,8 @@
 //! in remote storage.
 //! But note that we don't test any of this right now.
 //!
+//! [`Tenant::timeline_init_and_sync`]: super::Tenant::timeline_init_and_sync
+//! [`Timeline::reconcile_with_remote`]: super::Timeline::reconcile_with_remote

 mod delete;
 mod download;
@@ -442,8 +444,8 @@ impl RemoteTimelineClient {
        let index_part = download::download_index_part(
            self.conf,
            &self.storage_impl,
-            self.tenant_id,
-            self.timeline_id,
+            &self.tenant_id,
+            &self.timeline_id,
        )
        .measure_remote_op(
            self.tenant_id,
@@ -748,25 +750,13 @@ impl RemoteTimelineClient {
            stopped.deleted_at = SetDeletedFlagProgress::NotRunning;
        });

-        // Have a failpoint that can use the `pause` failpoint action.
-        // We don't want to block the executor thread, hence, spawn_blocking + await.
-        if cfg!(feature = "testing") {
-            tokio::task::spawn_blocking({
-                let current = tracing::Span::current();
-                move || {
-                    let _entered = current.entered();
-                    tracing::info!("at failpoint persist_deleted_index_part");
-                    fail::fail_point!("persist_deleted_index_part");
-                }
-            })
-            .await
-            .expect("spawn_blocking");
-        }
+        pausable_failpoint!("persist_deleted_index_part");
+
        upload::upload_index_part(
            self.conf,
            &self.storage_impl,
-            self.tenant_id,
-            self.timeline_id,
+            &self.tenant_id,
+            &self.timeline_id,
            &index_part_with_deleted_at,
        )
        .await?;
@@ -841,7 +831,7 @@ impl RemoteTimelineClient {

        // Do not delete index part yet, it is needed for possible retry. If we remove it first
        // and retry will arrive to different pageserver there wont be any traces of it on remote storage
-        let timeline_path = self.conf.timeline_path(&self.timeline_id, &self.tenant_id);
+        let timeline_path = self.conf.timeline_path(&self.tenant_id, &self.timeline_id);
        let timeline_storage_path = self.conf.remote_path(&timeline_path)?;

        let remaining = self
@@ -852,14 +842,16 @@ impl RemoteTimelineClient {
        let remaining: Vec<RemotePath> = remaining
            .into_iter()
            .filter(|p| p.object_name() != Some(IndexPart::FILE_NAME))
+            .inspect(|path| {
+                if let Some(name) = path.object_name() {
+                    info!(%name, "deleting a file not referenced from index_part.json");
+                } else {
+                    warn!(%path, "deleting a nameless or non-utf8 object not referenced from index_part.json");
+                }
+            })
            .collect();

        if !remaining.is_empty() {
-            warn!(
-                "Found {} files not bound to index_file.json, proceeding with their deletion",
-                remaining.len()
-            );
-            warn!("About to remove {} files", remaining.len());
            self.storage_impl.delete_objects(&remaining).await?;
        }

@@ -868,7 +860,7 @@ impl RemoteTimelineClient {
        debug!("deleting index part");
        self.storage_impl.delete(&index_file_path).await?;

-        info!(deletions_queued, "done deleting, including index_part.json");
+        info!(prefix=%timeline_storage_path, referenced=deletions_queued, not_referenced=%remaining.len(), "done deleting in timeline prefix, including index_part.json");

        Ok(())
    }
@@ -933,11 +925,11 @@ impl RemoteTimelineClient {

            // Assign unique ID to this task
            upload_queue.task_counter += 1;
-            let task_id = upload_queue.task_counter;
+            let upload_task_id = upload_queue.task_counter;

            // Add it to the in-progress map
            let task = Arc::new(UploadTask {
-                task_id,
+                task_id: upload_task_id,
                op: next_op,
                retries: AtomicU32::new(0),
            });
@@ -947,6 +939,8 @@ impl RemoteTimelineClient {

            // Spawn task to perform the task
            let self_rc = Arc::clone(self);
+            let tenant_id = self.tenant_id;
+            let timeline_id = self.timeline_id;
            task_mgr::spawn(
                self.runtime.handle(),
                TaskKind::RemoteUploadTask,
@@ -958,7 +952,7 @@ impl RemoteTimelineClient {
                    self_rc.perform_upload_task(task).await;
                    Ok(())
                }
-                .instrument(info_span!(parent: None, "remote_upload", tenant = %self.tenant_id, timeline = %self.timeline_id, upload_task_id = %task_id)),
+                .instrument(info_span!(parent: None, "remote_upload", %tenant_id, %timeline_id, %upload_task_id)),
            );

            // Loop back to process next task
@@ -1003,7 +997,7 @@ impl RemoteTimelineClient {
                UploadOp::UploadLayer(ref layer_file_name, ref layer_metadata) => {
                    let path = &self
                        .conf
-                        .timeline_path(&self.timeline_id, &self.tenant_id)
+                        .timeline_path(&self.tenant_id, &self.timeline_id)
                        .join(layer_file_name.file_name());
                    upload::upload_timeline_layer(
                        self.conf,
@@ -1024,8 +1018,8 @@ impl RemoteTimelineClient {
                    let res = upload::upload_index_part(
                        self.conf,
                        &self.storage_impl,
-                        self.tenant_id,
-                        self.timeline_id,
+                        &self.tenant_id,
+                        &self.timeline_id,
                        index_part,
                    )
                    .measure_remote_op(
@@ -1044,7 +1038,7 @@ impl RemoteTimelineClient {
                UploadOp::Delete(delete) => {
                    let path = &self
                        .conf
-                        .timeline_path(&self.timeline_id, &self.tenant_id)
+                        .timeline_path(&self.tenant_id, &self.timeline_id)
                        .join(delete.layer_file_name.file_name());
                    delete::delete_layer(self.conf, &self.storage_impl, path)
                        .measure_remote_op(
--- a/pageserver/src/tenant/remote_timeline_client/delete.rs
+++ b/pageserver/src/tenant/remote_timeline_client/delete.rs
@@ -19,9 +19,10 @@ pub(super) async fn delete_layer<'a>(

    let path_to_delete = conf.remote_path(local_layer_path)?;

-    // XXX: If the deletion fails because the object already didn't exist,
-    // it would be good to just issue a warning but consider it success.
-    // https://github.com/neondatabase/neon/issues/2934
+    // We don't want to print an error if the delete failed if the file has
+    // already been deleted. Thankfully, in this situation S3 already
+    // does not yield an error. While OS-provided local file system APIs do yield
+    // errors, we avoid them in the `LocalFs` wrapper.
    storage.delete(&path_to_delete).await.with_context(|| {
        format!("Failed to delete remote layer from storage at {path_to_delete:?}")
    })
--- a/pageserver/src/tenant/remote_timeline_client/download.rs
+++ b/pageserver/src/tenant/remote_timeline_client/download.rs
@@ -46,7 +46,7 @@ pub async fn download_layer_file<'a>(
 ) -> Result<u64, DownloadError> {
    debug_assert_current_span_has_tenant_and_timeline_id();

-    let timeline_path = conf.timeline_path(&timeline_id, &tenant_id);
+    let timeline_path = conf.timeline_path(&tenant_id, &timeline_id);

    let local_path = timeline_path.join(layer_file_name.file_name());

@@ -229,11 +229,11 @@ pub async fn list_remote_timelines<'a>(
 pub(super) async fn download_index_part(
    conf: &'static PageServerConf,
    storage: &GenericRemoteStorage,
-    tenant_id: TenantId,
-    timeline_id: TimelineId,
+    tenant_id: &TenantId,
+    timeline_id: &TimelineId,
 ) -> Result<IndexPart, DownloadError> {
    let index_part_path = conf
-        .metadata_path(timeline_id, tenant_id)
+        .metadata_path(tenant_id, timeline_id)
        .with_file_name(IndexPart::FILE_NAME);
    let part_storage_path = conf
        .remote_path(&index_part_path)
--- a/pageserver/src/tenant/remote_timeline_client/upload.rs
+++ b/pageserver/src/tenant/remote_timeline_client/upload.rs
@@ -2,7 +2,7 @@

 use anyhow::{bail, Context};
 use fail::fail_point;
-use std::path::Path;
+use std::{io::ErrorKind, path::Path};
 use tokio::fs;

 use crate::{config::PageServerConf, tenant::remote_timeline_client::index::IndexPart};
@@ -11,12 +11,14 @@ use utils::id::{TenantId, TimelineId};

 use super::index::LayerFileMetadata;

+use tracing::info;
+
 /// Serializes and uploads the given index part data to the remote storage.
 pub(super) async fn upload_index_part<'a>(
    conf: &'static PageServerConf,
    storage: &'a GenericRemoteStorage,
-    tenant_id: TenantId,
-    timeline_id: TimelineId,
+    tenant_id: &TenantId,
+    timeline_id: &TimelineId,
    index_part: &'a IndexPart,
 ) -> anyhow::Result<()> {
    tracing::trace!("uploading new index part");
@@ -31,7 +33,7 @@ pub(super) async fn upload_index_part<'a>(
    let index_part_bytes = tokio::io::BufReader::new(std::io::Cursor::new(index_part_bytes));

    let index_part_path = conf
-        .metadata_path(timeline_id, tenant_id)
+        .metadata_path(tenant_id, timeline_id)
        .with_file_name(IndexPart::FILE_NAME);
    let storage_path = conf.remote_path(&index_part_path)?;

@@ -56,9 +58,21 @@ pub(super) async fn upload_timeline_layer<'a>(
    });
    let storage_path = conf.remote_path(source_path)?;

-    let source_file = fs::File::open(&source_path)
-        .await
-        .with_context(|| format!("Failed to open a source file for layer {source_path:?}"))?;
+    let source_file_res = fs::File::open(&source_path).await;
+    let source_file = match source_file_res {
+        Ok(source_file) => source_file,
+        Err(e) if e.kind() == ErrorKind::NotFound => {
+            // If we encounter this arm, it wasn't intended, but it's also not
+            // a big problem, if it's because the file was deleted before an
+            // upload. However, a nonexistent file can also be indicative of
+            // something worse, like when a file is scheduled for upload before
+            // it has been written to disk yet.
+            info!(path = %source_path.display(), "File to upload doesn't exist. Likely the file has been deleted and an upload is not required any more.");
+            return Ok(());
+        }
+        Err(e) => Err(e)
+            .with_context(|| format!("Failed to open a source file for layer {source_path:?}"))?,
+    };

    let fs_size = source_file
        .metadata()
--- a/pageserver/src/tenant/size.rs
+++ b/pageserver/src/tenant/size.rs
@@ -110,11 +110,11 @@ pub struct TimelineInputs {
 ///
 /// Tenant size does not consider the latest state, but only the state until next_gc_cutoff, which
 /// is updated on-demand, during the start of this calculation and separate from the
-/// [`Timeline::latest_gc_cutoff`].
+/// [`TimelineInputs::latest_gc_cutoff`].
 ///
 /// For timelines in general:
 ///
-/// ```ignore
+/// ```text
 /// 0-----|---------|----|------------| · · · · · |·> lsn
 ///   initdb_lsn  branchpoints*  next_gc_cutoff  latest
 /// ```
--- a/pageserver/src/tenant/span.rs
+++ b/pageserver/src/tenant/span.rs
@@ -5,16 +5,13 @@ use utils::tracing_span_assert::{check_fields_present, MultiNameExtractor};
 pub(crate) fn debug_assert_current_span_has_tenant_id() {}

 #[cfg(debug_assertions)]
-pub(crate) static TENANT_ID_EXTRACTOR: once_cell::sync::Lazy<MultiNameExtractor<2>> =
-    once_cell::sync::Lazy::new(|| MultiNameExtractor::new("TenantId", ["tenant_id", "tenant"]));
+pub(crate) static TENANT_ID_EXTRACTOR: once_cell::sync::Lazy<MultiNameExtractor<1>> =
+    once_cell::sync::Lazy::new(|| MultiNameExtractor::new("TenantId", ["tenant_id"]));

 #[cfg(debug_assertions)]
 #[track_caller]
 pub(crate) fn debug_assert_current_span_has_tenant_id() {
-    if let Err(missing) = check_fields_present([&*TENANT_ID_EXTRACTOR]) {
-        panic!(
-            "missing extractors: {:?}",
-            missing.into_iter().map(|e| e.name()).collect::<Vec<_>>()
-        )
+    if let Err(missing) = check_fields_present!([&*TENANT_ID_EXTRACTOR]) {
+        panic!("missing extractors: {missing:?}")
    }
 }
--- a/pageserver/src/tenant/storage_layer.rs
+++ b/pageserver/src/tenant/storage_layer.rs
@@ -41,7 +41,7 @@ pub use inmemory_layer::InMemoryLayer;
 pub use layer_desc::{PersistentLayerDesc, PersistentLayerKey};
 pub use remote_layer::RemoteLayer;

-use super::layer_map::BatchedUpdates;
+use super::timeline::layer_manager::LayerManager;

 pub fn range_overlaps<T>(a: &Range<T>, b: &Range<T>) -> bool
 where
@@ -54,13 +54,6 @@ where
    }
 }

-pub fn range_eq<T>(a: &Range<T>, b: &Range<T>) -> bool
-where
-    T: PartialEq<T>,
-{
-    a.start == b.start && a.end == b.end
-}
-
 /// Struct used to communicate across calls to 'get_value_reconstruct_data'.
 ///
 /// Before first call, you can fill in 'page_img' if you have an older cached
@@ -169,6 +162,9 @@ impl LayerAccessStats {
    /// The caller is responsible for recording a residence event
    /// using [`record_residence_event`] before calling `latest_activity`.
    /// If they don't, [`latest_activity`] will return `None`.
+    ///
+    /// [`record_residence_event`]: Self::record_residence_event
+    /// [`latest_activity`]: Self::latest_activity
    pub(crate) fn empty_will_record_residence_event_later() -> Self {
        LayerAccessStats(Mutex::default())
    }
@@ -176,8 +172,11 @@ impl LayerAccessStats {
    /// Create an empty stats object and record a [`LayerLoad`] event with the given residence status.
    ///
    /// See [`record_residence_event`] for why you need to do this while holding the layer map lock.
+    ///
+    /// [`LayerLoad`]: LayerResidenceEventReason::LayerLoad
+    /// [`record_residence_event`]: Self::record_residence_event
    pub(crate) fn for_loading_layer(
-        layer_map_lock_held_witness: &BatchedUpdates<'_>,
+        layer_map_lock_held_witness: &LayerManager,
        status: LayerResidenceStatus,
    ) -> Self {
        let new = LayerAccessStats(Mutex::new(LayerAccessStatsLocked::default()));
@@ -194,9 +193,11 @@ impl LayerAccessStats {
    /// The `new_status` is not recorded in `self`.
    ///
    /// See [`record_residence_event`] for why you need to do this while holding the layer map lock.
+    ///
+    /// [`record_residence_event`]: Self::record_residence_event
    pub(crate) fn clone_for_residence_change(
        &self,
-        layer_map_lock_held_witness: &BatchedUpdates<'_>,
+        layer_map_lock_held_witness: &LayerManager,
        new_status: LayerResidenceStatus,
    ) -> LayerAccessStats {
        let clone = {
@@ -228,7 +229,7 @@ impl LayerAccessStats {
    ///
    pub(crate) fn record_residence_event(
        &self,
-        _layer_map_lock_held_witness: &BatchedUpdates<'_>,
+        _layer_map_lock_held_witness: &LayerManager,
        status: LayerResidenceStatus,
        reason: LayerResidenceEventReason,
    ) {
@@ -301,11 +302,13 @@ impl LayerAccessStats {
    /// implementation error. This function logs a rate-limited warning in that case.
    ///
    /// TODO: use type system to avoid the need for `fallback`.
-    /// The approach in https://github.com/neondatabase/neon/pull/3775
+    /// The approach in <https://github.com/neondatabase/neon/pull/3775>
    /// could be used to enforce that a residence event is recorded
    /// before a layer is added to the layer map. We could also have
    /// a layer wrapper type that holds the LayerAccessStats, and ensure
    /// that that type can only be produced by inserting into the layer map.
+    ///
+    /// [`record_residence_event`]: Self::record_residence_event
    pub(crate) fn latest_activity(&self) -> Option<SystemTime> {
        let locked = self.0.lock().unwrap();
        let inner = &locked.for_eviction_policy;
@@ -330,12 +333,13 @@ impl LayerAccessStats {
 }

 /// Supertrait of the [`Layer`] trait that captures the bare minimum interface
-/// required by [`LayerMap`].
+/// required by [`LayerMap`](super::layer_map::LayerMap).
 ///
 /// All layers should implement a minimal `std::fmt::Debug` without tenant or
 /// timeline names, because those are known in the context of which the layers
 /// are used in (timeline).
-pub trait Layer: std::fmt::Debug + std::fmt::Display + Send + Sync {
+#[async_trait::async_trait]
+pub trait Layer: std::fmt::Debug + std::fmt::Display + Send + Sync + 'static {
    /// Range of keys that this layer covers
    fn get_key_range(&self) -> Range<Key>;

@@ -365,7 +369,7 @@ pub trait Layer: std::fmt::Debug + std::fmt::Display + Send + Sync {
    /// is available. If this returns ValueReconstructResult::Continue, look up
    /// the predecessor layer and call again with the same 'reconstruct_data' to
    /// collect more data.
-    fn get_value_reconstruct_data(
+    async fn get_value_reconstruct_data(
        &self,
        key: Key,
        lsn_range: Range<Lsn>,
@@ -374,15 +378,21 @@ pub trait Layer: std::fmt::Debug + std::fmt::Display + Send + Sync {
    ) -> Result<ValueReconstructResult>;

    /// Dump summary of the contents of the layer to stdout
-    fn dump(&self, verbose: bool, ctx: &RequestContext) -> Result<()>;
+    async fn dump(&self, verbose: bool, ctx: &RequestContext) -> Result<()>;
 }

-/// Returned by [`Layer::iter`]
+/// Returned by [`PersistentLayer::iter`]
 pub type LayerIter<'i> = Box<dyn Iterator<Item = Result<(Key, Lsn, Value)>> + 'i + Send>;

-/// Returned by [`Layer::key_iter`]
+/// Returned by [`PersistentLayer::key_iter`]
 pub type LayerKeyIter<'i> = Box<dyn Iterator<Item = (Key, Lsn, u64)> + 'i + Send>;

+/// Get a layer descriptor from a layer.
+pub trait AsLayerDesc {
+    /// Get the layer descriptor.
+    fn layer_desc(&self) -> &PersistentLayerDesc;
+}
+
 /// A Layer contains all data in a "rectangle" consisting of a range of keys and
 /// range of LSNs.
 ///
@@ -396,10 +406,8 @@ pub type LayerKeyIter<'i> = Box<dyn Iterator<Item = (Key, Lsn, u64)> + 'i + Send
 /// A delta layer contains all modifications within a range of LSNs and keys.
 /// An image layer is a snapshot of all the data in a key-range, at a single
 /// LSN.
-pub trait PersistentLayer: Layer {
-    /// Get the layer descriptor.
-    fn layer_desc(&self) -> &PersistentLayerDesc;
-
+pub trait PersistentLayer: Layer + AsLayerDesc {
+    /// Identify the tenant this layer belongs to
    fn get_tenant_id(&self) -> TenantId {
        self.layer_desc().tenant_id
    }
@@ -435,6 +443,10 @@ pub trait PersistentLayer: Layer {
        None
    }

+    fn downcast_delta_layer(self: Arc<Self>) -> Option<std::sync::Arc<DeltaLayer>> {
+        None
+    }
+
    fn is_remote_layer(&self) -> bool {
        false
    }
@@ -465,119 +477,32 @@ pub fn downcast_remote_layer(
 pub mod tests {
    use super::*;

-    /// Holds metadata about a layer without any content. Used mostly for testing.
-    ///
-    /// To use filenames as fixtures, parse them as [`LayerFileName`] then convert from that to a
-    /// LayerDescriptor.
-    #[derive(Clone, Debug)]
-    pub struct LayerDescriptor {
-        base: PersistentLayerDesc,
-    }
-
-    impl From<PersistentLayerDesc> for LayerDescriptor {
-        fn from(base: PersistentLayerDesc) -> Self {
-            Self { base }
-        }
-    }
-
-    impl Layer for LayerDescriptor {
-        fn get_value_reconstruct_data(
-            &self,
-            _key: Key,
-            _lsn_range: Range<Lsn>,
-            _reconstruct_data: &mut ValueReconstructState,
-            _ctx: &RequestContext,
-        ) -> Result<ValueReconstructResult> {
-            todo!("This method shouldn't be part of the Layer trait")
-        }
-
-        fn dump(&self, _verbose: bool, _ctx: &RequestContext) -> Result<()> {
-            todo!()
-        }
-
-        /// Boilerplate to implement the Layer trait, always use layer_desc for persistent layers.
-        fn get_key_range(&self) -> Range<Key> {
-            self.layer_desc().key_range.clone()
-        }
-
-        /// Boilerplate to implement the Layer trait, always use layer_desc for persistent layers.
-        fn get_lsn_range(&self) -> Range<Lsn> {
-            self.layer_desc().lsn_range.clone()
-        }
-
-        /// Boilerplate to implement the Layer trait, always use layer_desc for persistent layers.
-        fn is_incremental(&self) -> bool {
-            self.layer_desc().is_incremental
-        }
-    }
-
-    /// Boilerplate to implement the Layer trait, always use layer_desc for persistent layers.
-    impl std::fmt::Display for LayerDescriptor {
-        fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
-            write!(f, "{}", self.layer_desc().short_id())
-        }
-    }
-
-    impl PersistentLayer for LayerDescriptor {
-        fn layer_desc(&self) -> &PersistentLayerDesc {
-            &self.base
-        }
-
-        fn local_path(&self) -> Option<PathBuf> {
-            unimplemented!()
-        }
-
-        fn iter(&self, _: &RequestContext) -> Result<LayerIter<'_>> {
-            unimplemented!()
-        }
-
-        fn key_iter(&self, _: &RequestContext) -> Result<LayerKeyIter<'_>> {
-            unimplemented!()
-        }
-
-        fn delete_resident_layer_file(&self) -> Result<()> {
-            unimplemented!()
-        }
-
-        fn info(&self, _: LayerAccessStatsReset) -> HistoricLayerInfo {
-            unimplemented!()
-        }
-
-        fn access_stats(&self) -> &LayerAccessStats {
-            unimplemented!()
-        }
-    }
-
-    impl From<DeltaFileName> for LayerDescriptor {
+    impl From<DeltaFileName> for PersistentLayerDesc {
        fn from(value: DeltaFileName) -> Self {
-            LayerDescriptor {
-                base: PersistentLayerDesc::new_delta(
-                    TenantId::from_array([0; 16]),
-                    TimelineId::from_array([0; 16]),
-                    value.key_range,
-                    value.lsn_range,
-                    233,
-                ),
-            }
+            PersistentLayerDesc::new_delta(
+                TenantId::from_array([0; 16]),
+                TimelineId::from_array([0; 16]),
+                value.key_range,
+                value.lsn_range,
+                233,
+            )
        }
    }

-    impl From<ImageFileName> for LayerDescriptor {
+    impl From<ImageFileName> for PersistentLayerDesc {
        fn from(value: ImageFileName) -> Self {
-            LayerDescriptor {
-                base: PersistentLayerDesc::new_img(
-                    TenantId::from_array([0; 16]),
-                    TimelineId::from_array([0; 16]),
-                    value.key_range,
-                    value.lsn,
-                    false,
-                    233,
-                ),
-            }
+            PersistentLayerDesc::new_img(
+                TenantId::from_array([0; 16]),
+                TimelineId::from_array([0; 16]),
+                value.key_range,
+                value.lsn,
+                false,
+                233,
+            )
        }
    }

-    impl From<LayerFileName> for LayerDescriptor {
+    impl From<LayerFileName> for PersistentLayerDesc {
        fn from(value: LayerFileName) -> Self {
            match value {
                LayerFileName::Delta(d) => Self::from(d),
--- a/pageserver/src/tenant/storage_layer/delta_layer.rs
+++ b/pageserver/src/tenant/storage_layer/delta_layer.rs
@@ -7,14 +7,18 @@
 //! must be page images or WAL records with the 'will_init' flag set, so that
 //! they can be replayed without referring to an older page version.
 //!
-//! The delta files are stored in timelines/<timeline_id> directory.  Currently,
+//! The delta files are stored in `timelines/<timeline_id>` directory.  Currently,
 //! there are no subdirectories, and each delta file is named like this:
 //!
-//!    <key start>-<key end>__<start LSN>-<end LSN
+//! ```text
+//!    <key start>-<key end>__<start LSN>-<end LSN>
+//! ```
 //!
 //! For example:
 //!
+//! ```text
 //!    000000067F000032BE0000400000000020B6-000000067F000032BE0000400000000030B6__000000578C6B29-0000000057A50051
+//! ```
 //!
 //! Every delta file consists of three parts: "summary", "index", and
 //! "values". The summary is a fixed size header at the beginning of the file,
@@ -27,7 +31,7 @@ use crate::config::PageServerConf;
 use crate::context::RequestContext;
 use crate::page_cache::{PageReadGuard, PAGE_SZ};
 use crate::repository::{Key, Value, KEY_SIZE};
-use crate::tenant::blob_io::{BlobCursor, BlobWriter, WriteBlobWriter};
+use crate::tenant::blob_io::{BlobWriter, WriteBlobWriter};
 use crate::tenant::block_io::{BlockBuf, BlockCursor, BlockReader, FileBlockReader};
 use crate::tenant::disk_btree::{DiskBtreeBuilder, DiskBtreeReader, VisitDirection};
 use crate::tenant::storage_layer::{
@@ -47,6 +51,7 @@ use std::io::{Seek, SeekFrom};
 use std::ops::Range;
 use std::os::unix::fs::FileExt;
 use std::path::{Path, PathBuf};
+use std::sync::Arc;
 use tracing::*;

 use utils::{
@@ -56,8 +61,8 @@ use utils::{
 };

 use super::{
-    DeltaFileName, Layer, LayerAccessStats, LayerAccessStatsReset, LayerIter, LayerKeyIter,
-    PathOrConf, PersistentLayerDesc,
+    AsLayerDesc, DeltaFileName, Layer, LayerAccessStats, LayerAccessStatsReset, LayerIter,
+    LayerKeyIter, PathOrConf, PersistentLayerDesc,
 };

 ///
@@ -218,9 +223,10 @@ impl std::fmt::Debug for DeltaLayerInner {
    }
 }

+#[async_trait::async_trait]
 impl Layer for DeltaLayer {
    /// debugging function to print out the contents of the layer
-    fn dump(&self, verbose: bool, ctx: &RequestContext) -> Result<()> {
+    async fn dump(&self, verbose: bool, ctx: &RequestContext) -> Result<()> {
        println!(
            "----- delta layer for ten {} tli {} keys {}-{} lsn {}-{} size {} ----",
            self.desc.tenant_id,
@@ -295,7 +301,7 @@ impl Layer for DeltaLayer {
        Ok(())
    }

-    fn get_value_reconstruct_data(
+    async fn get_value_reconstruct_data(
        &self,
        key: Key,
        lsn_range: Range<Lsn>,
@@ -403,10 +409,16 @@ impl std::fmt::Display for DeltaLayer {
    }
 }

-impl PersistentLayer for DeltaLayer {
+impl AsLayerDesc for DeltaLayer {
    fn layer_desc(&self) -> &PersistentLayerDesc {
        &self.desc
    }
+}
+
+impl PersistentLayer for DeltaLayer {
+    fn downcast_delta_layer(self: Arc<Self>) -> Option<std::sync::Arc<DeltaLayer>> {
+        Some(self)
+    }

    fn local_path(&self) -> Option<PathBuf> {
        Some(self.path())
@@ -459,22 +471,22 @@ impl PersistentLayer for DeltaLayer {
 impl DeltaLayer {
    fn path_for(
        path_or_conf: &PathOrConf,
-        timeline_id: TimelineId,
-        tenant_id: TenantId,
+        tenant_id: &TenantId,
+        timeline_id: &TimelineId,
        fname: &DeltaFileName,
    ) -> PathBuf {
        match path_or_conf {
            PathOrConf::Path(path) => path.clone(),
            PathOrConf::Conf(conf) => conf
-                .timeline_path(&timeline_id, &tenant_id)
+                .timeline_path(tenant_id, timeline_id)
                .join(fname.to_string()),
        }
    }

    fn temp_path_for(
        conf: &PageServerConf,
-        timeline_id: TimelineId,
-        tenant_id: TenantId,
+        tenant_id: &TenantId,
+        timeline_id: &TimelineId,
        key_start: Key,
        lsn_range: &Range<Lsn>,
    ) -> PathBuf {
@@ -484,7 +496,7 @@ impl DeltaLayer {
            .map(char::from)
            .collect();

-        conf.timeline_path(&timeline_id, &tenant_id).join(format!(
+        conf.timeline_path(tenant_id, timeline_id).join(format!(
            "{}-XXX__{:016X}-{:016X}.{}.{}",
            key_start,
            u64::from(lsn_range.start),
@@ -606,8 +618,8 @@ impl DeltaLayer {
    pub fn path(&self) -> PathBuf {
        Self::path_for(
            &self.path_or_conf,
-            self.desc.timeline_id,
-            self.desc.tenant_id,
+            &self.desc.tenant_id,
+            &self.desc.timeline_id,
            &self.layer_name(),
        )
    }
@@ -655,7 +667,7 @@ impl DeltaLayerWriterInner {
        //
        // Note: This overwrites any existing file. There shouldn't be any.
        // FIXME: throw an error instead?
-        let path = DeltaLayer::temp_path_for(conf, timeline_id, tenant_id, key_start, &lsn_range);
+        let path = DeltaLayer::temp_path_for(conf, &tenant_id, &timeline_id, key_start, &lsn_range);

        let mut file = VirtualFile::create(&path)?;
        // make room for the header block
@@ -770,8 +782,8 @@ impl DeltaLayerWriterInner {
        // FIXME: throw an error instead?
        let final_path = DeltaLayer::path_for(
            &PathOrConf::Conf(self.conf),
-            self.timeline_id,
-            self.tenant_id,
+            &self.tenant_id,
+            &self.timeline_id,
            &DeltaFileName {
                key_range: self.key_start..key_end,
                lsn_range: self.lsn_range,
@@ -798,7 +810,7 @@ impl DeltaLayerWriterInner {
 ///
 /// # Note
 ///
-/// As described in https://github.com/neondatabase/neon/issues/2650, it's
+/// As described in <https://github.com/neondatabase/neon/issues/2650>, it's
 /// possible for the writer to drop before `finish` is actually called. So this
 /// could lead to odd temporary files in the directory, exhausting file system.
 /// This structure wraps `DeltaLayerWriterInner` and also contains `Drop`
--- a/pageserver/src/tenant/storage_layer/filename.rs
+++ b/pageserver/src/tenant/storage_layer/filename.rs
@@ -57,8 +57,9 @@ impl Ord for DeltaFileName {

 /// Represents the filename of a DeltaLayer
 ///
+/// ```text
 ///    <key start>-<key end>__<LSN start>-<LSN end>
-///
+/// ```
 impl DeltaFileName {
    ///
    /// Parse a string as a delta file name. Returns None if the filename does not
@@ -162,7 +163,9 @@ impl ImageFileName {
 ///
 /// Represents the filename of an ImageLayer
 ///
+/// ```text
 ///    <key start>-<key end>__<LSN>
+/// ```
 impl ImageFileName {
    ///
    /// Parse a string as an image file name. Returns None if the filename does not
--- a/pageserver/src/tenant/storage_layer/image_layer.rs
+++ b/pageserver/src/tenant/storage_layer/image_layer.rs
@@ -7,11 +7,15 @@
 //! timelines/<timeline_id> directory.  Currently, there are no
 //! subdirectories, and each image layer file is named like this:
 //!
+//! ```text
 //!    <key start>-<key end>__<LSN>
+//! ```
 //!
 //! For example:
 //!
+//! ```text
 //!    000000067F000032BE0000400000000070B6-000000067F000032BE0000400000000080B6__00000000346BC568
+//! ```
 //!
 //! Every image layer file consists of three parts: "summary",
 //! "index", and "values".  The summary is a fixed size header at the
@@ -23,7 +27,7 @@ use crate::config::PageServerConf;
 use crate::context::RequestContext;
 use crate::page_cache::PAGE_SZ;
 use crate::repository::{Key, KEY_SIZE};
-use crate::tenant::blob_io::{BlobCursor, BlobWriter, WriteBlobWriter};
+use crate::tenant::blob_io::{BlobWriter, WriteBlobWriter};
 use crate::tenant::block_io::{BlockBuf, BlockReader, FileBlockReader};
 use crate::tenant::disk_btree::{DiskBtreeBuilder, DiskBtreeReader, VisitDirection};
 use crate::tenant::storage_layer::{
@@ -53,7 +57,9 @@ use utils::{
 };

 use super::filename::ImageFileName;
-use super::{Layer, LayerAccessStatsReset, LayerIter, PathOrConf, PersistentLayerDesc};
+use super::{
+    AsLayerDesc, Layer, LayerAccessStatsReset, LayerIter, PathOrConf, PersistentLayerDesc,
+};

 ///
 /// Header stored in the beginning of the file
@@ -149,9 +155,10 @@ impl std::fmt::Debug for ImageLayerInner {
    }
 }

+#[async_trait::async_trait]
 impl Layer for ImageLayer {
    /// debugging function to print out the contents of the layer
-    fn dump(&self, verbose: bool, ctx: &RequestContext) -> Result<()> {
+    async fn dump(&self, verbose: bool, ctx: &RequestContext) -> Result<()> {
        println!(
            "----- image layer for ten {} tli {} key {}-{} at {} is_incremental {} size {} ----",
            self.desc.tenant_id,
@@ -183,7 +190,7 @@ impl Layer for ImageLayer {
    }

    /// Look up given page in the file
-    fn get_value_reconstruct_data(
+    async fn get_value_reconstruct_data(
        &self,
        key: Key,
        lsn_range: Range<Lsn>,
@@ -241,11 +248,13 @@ impl std::fmt::Display for ImageLayer {
    }
 }

-impl PersistentLayer for ImageLayer {
+impl AsLayerDesc for ImageLayer {
    fn layer_desc(&self) -> &PersistentLayerDesc {
        &self.desc
    }
+}

+impl PersistentLayer for ImageLayer {
    fn local_path(&self) -> Option<PathBuf> {
        Some(self.path())
    }
@@ -288,7 +297,7 @@ impl ImageLayer {
        match path_or_conf {
            PathOrConf::Path(path) => path.to_path_buf(),
            PathOrConf::Conf(conf) => conf
-                .timeline_path(&timeline_id, &tenant_id)
+                .timeline_path(&tenant_id, &timeline_id)
                .join(fname.to_string()),
        }
    }
@@ -305,7 +314,7 @@ impl ImageLayer {
            .map(char::from)
            .collect();

-        conf.timeline_path(&timeline_id, &tenant_id)
+        conf.timeline_path(&tenant_id, &timeline_id)
            .join(format!("{fname}.{rand_string}.{TEMP_FILE_SUFFIX}"))
    }

@@ -656,7 +665,7 @@ impl ImageLayerWriterInner {
 ///
 /// # Note
 ///
-/// As described in https://github.com/neondatabase/neon/issues/2650, it's
+/// As described in <https://github.com/neondatabase/neon/issues/2650>, it's
 /// possible for the writer to drop before `finish` is actually called. So this
 /// could lead to odd temporary files in the directory, exhausting file system.
 /// This structure wraps `ImageLayerWriterInner` and also contains `Drop`
--- a/pageserver/src/tenant/storage_layer/inmemory_layer.rs
+++ b/pageserver/src/tenant/storage_layer/inmemory_layer.rs
@@ -7,7 +7,7 @@
 use crate::config::PageServerConf;
 use crate::context::RequestContext;
 use crate::repository::{Key, Value};
-use crate::tenant::blob_io::{BlobCursor, BlobWriter};
+use crate::tenant::blob_io::BlobWriter;
 use crate::tenant::block_io::BlockReader;
 use crate::tenant::ephemeral_file::EphemeralFile;
 use crate::tenant::storage_layer::{ValueReconstructResult, ValueReconstructState};
@@ -110,6 +110,7 @@ impl InMemoryLayer {
    }
 }

+#[async_trait::async_trait]
 impl Layer for InMemoryLayer {
    fn get_key_range(&self) -> Range<Key> {
        Key::MIN..Key::MAX
@@ -132,7 +133,7 @@ impl Layer for InMemoryLayer {
    }

    /// debugging function to print out the contents of the layer
-    fn dump(&self, verbose: bool, _ctx: &RequestContext) -> Result<()> {
+    async fn dump(&self, verbose: bool, _ctx: &RequestContext) -> Result<()> {
        let inner = self.inner.read().unwrap();

        let end_str = inner
@@ -183,7 +184,7 @@ impl Layer for InMemoryLayer {
    }

    /// Look up given value in the layer.
-    fn get_value_reconstruct_data(
+    async fn get_value_reconstruct_data(
        &self,
        key: Key,
        lsn_range: Range<Lsn>,
--- a/pageserver/src/tenant/storage_layer/remote_layer.rs
+++ b/pageserver/src/tenant/storage_layer/remote_layer.rs
@@ -4,9 +4,9 @@
 use crate::config::PageServerConf;
 use crate::context::RequestContext;
 use crate::repository::Key;
-use crate::tenant::layer_map::BatchedUpdates;
 use crate::tenant::remote_timeline_client::index::LayerFileMetadata;
 use crate::tenant::storage_layer::{Layer, ValueReconstructResult, ValueReconstructState};
+use crate::tenant::timeline::layer_manager::LayerManager;
 use anyhow::{bail, Result};
 use pageserver_api::models::HistoricLayerInfo;
 use std::ops::Range;
@@ -20,12 +20,12 @@ use utils::{

 use super::filename::{DeltaFileName, ImageFileName};
 use super::{
-    DeltaLayer, ImageLayer, LayerAccessStats, LayerAccessStatsReset, LayerIter, LayerKeyIter,
-    LayerResidenceStatus, PersistentLayer, PersistentLayerDesc,
+    AsLayerDesc, DeltaLayer, ImageLayer, LayerAccessStats, LayerAccessStatsReset, LayerIter,
+    LayerKeyIter, LayerResidenceStatus, PersistentLayer, PersistentLayerDesc,
 };

 /// RemoteLayer is a not yet downloaded [`ImageLayer`] or
-/// [`crate::storage_layer::DeltaLayer`].
+/// [`DeltaLayer`](super::DeltaLayer).
 ///
 /// RemoteLayer might be downloaded on-demand during operations which are
 /// allowed download remote layers and during which, it gets replaced with a
@@ -50,6 +50,8 @@ pub struct RemoteLayer {
    /// It is very unlikely to accumulate these in the Timeline's LayerMap, but having this avoids
    /// a possible fast loop between `Timeline::get_reconstruct_data` and
    /// `Timeline::download_remote_layer`, which also logs.
+    ///
+    /// [`ongoing_download`]: Self::ongoing_download
    pub(crate) download_replacement_failure: std::sync::atomic::AtomicBool,
 }

@@ -63,8 +65,9 @@ impl std::fmt::Debug for RemoteLayer {
    }
 }

+#[async_trait::async_trait]
 impl Layer for RemoteLayer {
-    fn get_value_reconstruct_data(
+    async fn get_value_reconstruct_data(
        &self,
        _key: Key,
        _lsn_range: Range<Lsn>,
@@ -75,7 +78,7 @@ impl Layer for RemoteLayer {
    }

    /// debugging function to print out the contents of the layer
-    fn dump(&self, _verbose: bool, _ctx: &RequestContext) -> Result<()> {
+    async fn dump(&self, _verbose: bool, _ctx: &RequestContext) -> Result<()> {
        println!(
            "----- remote layer for ten {} tli {} keys {}-{} lsn {}-{} is_delta {} is_incremental {} size {} ----",
            self.desc.tenant_id,
@@ -115,11 +118,13 @@ impl std::fmt::Display for RemoteLayer {
    }
 }

-impl PersistentLayer for RemoteLayer {
+impl AsLayerDesc for RemoteLayer {
    fn layer_desc(&self) -> &PersistentLayerDesc {
        &self.desc
    }
+}

+impl PersistentLayer for RemoteLayer {
    fn local_path(&self) -> Option<PathBuf> {
        None
    }
@@ -222,7 +227,7 @@ impl RemoteLayer {
    /// Create a Layer struct representing this layer, after it has been downloaded.
    pub fn create_downloaded_layer(
        &self,
-        layer_map_lock_held_witness: &BatchedUpdates<'_>,
+        layer_map_lock_held_witness: &LayerManager,
        conf: &'static PageServerConf,
        file_size: u64,
    ) -> Arc<dyn PersistentLayer> {
--- a/pageserver/src/tenant/tasks.rs
+++ b/pageserver/src/tenant/tasks.rs
@@ -111,7 +111,7 @@ async fn compaction_loop(tenant: Arc<Tenant>, cancel: CancellationToken) {
                Duration::from_secs(10)
            } else {
                // Run compaction
-                if let Err(e) = tenant.compaction_iteration(&ctx).await {
+                if let Err(e) = tenant.compaction_iteration(&cancel, &ctx).await {
                    error!("Compaction failed, retrying in {:?}: {e:?}", wait_duration);
                    wait_duration
                } else {
@@ -122,12 +122,12 @@ async fn compaction_loop(tenant: Arc<Tenant>, cancel: CancellationToken) {
            warn_when_period_overrun(started_at.elapsed(), period, "compaction");

            // Sleep
-            tokio::select! {
-                _ = cancel.cancelled() => {
-                    info!("received cancellation request during idling");
-                    break;
-                },
-                _ = tokio::time::sleep(sleep_duration) => {},
+            if tokio::time::timeout(sleep_duration, cancel.cancelled())
+                .await
+                .is_ok()
+            {
+                info!("received cancellation request during idling");
+                break;
            }
        }
    }
@@ -196,12 +196,12 @@ async fn gc_loop(tenant: Arc<Tenant>, cancel: CancellationToken) {
            warn_when_period_overrun(started_at.elapsed(), period, "gc");

            // Sleep
-            tokio::select! {
-                _ = cancel.cancelled() => {
-                    info!("received cancellation request during idling");
-                    break;
-                },
-                _ = tokio::time::sleep(sleep_duration) => {},
+            if tokio::time::timeout(sleep_duration, cancel.cancelled())
+                .await
+                .is_ok()
+            {
+                info!("received cancellation request during idling");
+                break;
            }
        }
    }
@@ -263,9 +263,9 @@ pub(crate) async fn random_init_delay(
        rng.gen_range(Duration::ZERO..=period)
    };

-    tokio::select! {
-        _ = cancel.cancelled() => Err(Cancelled),
-        _ = tokio::time::sleep(d) => Ok(()),
+    match tokio::time::timeout(d, cancel.cancelled()).await {
+        Ok(_) => Err(Cancelled),
+        Err(_) => Ok(()),
    }
 }

--- a/pageserver/src/tenant/timeline.rs
+++ b/pageserver/src/tenant/timeline.rs
--- a/pageserver/src/tenant/timeline/eviction_task.rs
+++ b/pageserver/src/tenant/timeline/eviction_task.rs
@@ -30,6 +30,7 @@ use crate::{
    tenant::{
        config::{EvictionPolicy, EvictionPolicyLayerAccessThreshold},
        storage_layer::PersistentLayer,
+        timeline::EvictionError,
        LogicalSizeCalculationCause, Tenant,
    },
 };
@@ -100,11 +101,11 @@ impl Timeline {
            match cf {
                ControlFlow::Break(()) => break,
                ControlFlow::Continue(sleep_until) => {
-                    tokio::select! {
-                        _ = cancel.cancelled() => {
-                            break;
-                        }
-                        _ = tokio::time::sleep_until(sleep_until) => { }
+                    if tokio::time::timeout_at(sleep_until, cancel.cancelled())
+                        .await
+                        .is_ok()
+                    {
+                        break;
                    }
                }
            }
@@ -198,10 +199,10 @@ impl Timeline {
        // So, we just need to deal with this.
        let candidates: Vec<Arc<dyn PersistentLayer>> = {
            let guard = self.layers.read().await;
-            let (layers, mapping) = &*guard;
+            let layers = guard.layer_map();
            let mut candidates = Vec::new();
            for hist_layer in layers.iter_historic_layers() {
-                let hist_layer = mapping.get_from_desc(&hist_layer);
+                let hist_layer = guard.get_from_desc(&hist_layer);
                if hist_layer.is_remote_layer() {
                    continue;
                }
@@ -270,20 +271,22 @@ impl Timeline {
                None => {
                    stats.skipped_for_shutdown += 1;
                }
-                Some(Ok(true)) => {
-                    debug!("evicted layer {l:?}");
+                Some(Ok(())) => {
                    stats.evicted += 1;
                }
-                Some(Ok(false)) => {
-                    debug!("layer is not evictable: {l:?}");
+                Some(Err(EvictionError::CannotEvictRemoteLayer)) => {
                    stats.not_evictable += 1;
                }
-                Some(Err(e)) => {
-                    // This variant is the case where an unexpected error happened during eviction.
-                    // Expected errors that result in non-eviction are `Some(Ok(false))`.
-                    // So, dump Debug here to gather as much info as possible in this rare case.
-                    warn!("failed to evict layer {l:?}: {e:?}");
-                    stats.errors += 1;
+                Some(Err(EvictionError::FileNotFound)) => {
+                    // compaction/gc removed the file while we were waiting on layer_removal_cs
+                    stats.not_evictable += 1;
+                }
+                Some(Err(
+                    e @ EvictionError::LayerNotFound(_) | e @ EvictionError::StatFailed(_),
+                )) => {
+                    let e = utils::error::report_compact_sources(&e);
+                    warn!(layer = %l, "failed to evict layer: {e}");
+                    stats.not_evictable += 1;
                }
            }
        }
--- a/pageserver/src/tenant/timeline/layer_manager.rs
+++ b/pageserver/src/tenant/timeline/layer_manager.rs
@@ -0,0 +1,391 @@
+use anyhow::{bail, ensure, Context, Result};
+use std::{collections::HashMap, sync::Arc};
+use tracing::trace;
+use utils::{
+    id::{TenantId, TimelineId},
+    lsn::{AtomicLsn, Lsn},
+};
+
+use crate::{
+    config::PageServerConf,
+    metrics::TimelineMetrics,
+    tenant::{
+        layer_map::{BatchedUpdates, LayerMap},
+        storage_layer::{
+            AsLayerDesc, DeltaLayer, ImageLayer, InMemoryLayer, Layer, PersistentLayer,
+            PersistentLayerDesc, PersistentLayerKey, RemoteLayer,
+        },
+        timeline::compare_arced_layers,
+    },
+};
+
+/// Provides semantic APIs to manipulate the layer map.
+pub struct LayerManager {
+    layer_map: LayerMap,
+    layer_fmgr: LayerFileManager,
+}
+
+/// After GC, the layer map changes will not be applied immediately. Users should manually apply the changes after
+/// scheduling deletes in remote client.
+pub struct ApplyGcResultGuard<'a>(BatchedUpdates<'a>);
+
+impl ApplyGcResultGuard<'_> {
+    pub fn flush(self) {
+        self.0.flush();
+    }
+}
+
+impl LayerManager {
+    pub fn create() -> Self {
+        Self {
+            layer_map: LayerMap::default(),
+            layer_fmgr: LayerFileManager::new(),
+        }
+    }
+
+    pub fn get_from_desc(&self, desc: &PersistentLayerDesc) -> Arc<dyn PersistentLayer> {
+        self.layer_fmgr.get_from_desc(desc)
+    }
+
+    /// Get an immutable reference to the layer map.
+    ///
+    /// We expect users only to be able to get an immutable layer map. If users want to make modifications,
+    /// they should use the below semantic APIs. This design makes us step closer to immutable storage state.
+    pub fn layer_map(&self) -> &LayerMap {
+        &self.layer_map
+    }
+
+    /// Get a mutable reference to the layer map. This function will be removed once `flush_frozen_layer`
+    /// gets a refactor.
+    pub fn layer_map_mut(&mut self) -> &mut LayerMap {
+        &mut self.layer_map
+    }
+
+    /// Replace layers in the layer file manager, used in evictions and layer downloads.
+    pub fn replace_and_verify(
+        &mut self,
+        expected: Arc<dyn PersistentLayer>,
+        new: Arc<dyn PersistentLayer>,
+    ) -> Result<()> {
+        self.layer_fmgr.replace_and_verify(expected, new)
+    }
+
+    /// Called from `load_layer_map`. Initialize the layer manager with:
+    /// 1. all on-disk layers
+    /// 2. next open layer (with disk disk_consistent_lsn LSN)
+    pub fn initialize_local_layers(
+        &mut self,
+        on_disk_layers: Vec<Arc<dyn PersistentLayer>>,
+        next_open_layer_at: Lsn,
+    ) {
+        let mut updates = self.layer_map.batch_update();
+        for layer in on_disk_layers {
+            Self::insert_historic_layer(layer, &mut updates, &mut self.layer_fmgr);
+        }
+        updates.flush();
+        self.layer_map.next_open_layer_at = Some(next_open_layer_at);
+    }
+
+    /// Initialize when creating a new timeline, called in `init_empty_layer_map`.
+    pub fn initialize_empty(&mut self, next_open_layer_at: Lsn) {
+        self.layer_map.next_open_layer_at = Some(next_open_layer_at);
+    }
+
+    pub fn initialize_remote_layers(
+        &mut self,
+        corrupted_local_layers: Vec<Arc<dyn PersistentLayer>>,
+        remote_layers: Vec<Arc<RemoteLayer>>,
+    ) {
+        let mut updates = self.layer_map.batch_update();
+        for layer in corrupted_local_layers {
+            Self::remove_historic_layer(layer, &mut updates, &mut self.layer_fmgr);
+        }
+        for layer in remote_layers {
+            Self::insert_historic_layer(layer, &mut updates, &mut self.layer_fmgr);
+        }
+        updates.flush();
+    }
+
+    /// Open a new writable layer to append data if there is no open layer, otherwise return the current open layer,
+    /// called within `get_layer_for_write`.
+    pub fn get_layer_for_write(
+        &mut self,
+        lsn: Lsn,
+        last_record_lsn: Lsn,
+        conf: &'static PageServerConf,
+        timeline_id: TimelineId,
+        tenant_id: TenantId,
+    ) -> Result<Arc<InMemoryLayer>> {
+        ensure!(lsn.is_aligned());
+
+        ensure!(
+            lsn > last_record_lsn,
+            "cannot modify relation after advancing last_record_lsn (incoming_lsn={}, last_record_lsn={})\n{}",
+            lsn,
+            last_record_lsn,
+            std::backtrace::Backtrace::force_capture(),
+        );
+
+        // Do we have a layer open for writing already?
+        let layer = if let Some(open_layer) = &self.layer_map.open_layer {
+            if open_layer.get_lsn_range().start > lsn {
+                bail!(
+                    "unexpected open layer in the future: open layers starts at {}, write lsn {}",
+                    open_layer.get_lsn_range().start,
+                    lsn
+                );
+            }
+
+            Arc::clone(open_layer)
+        } else {
+            // No writeable layer yet. Create one.
+            let start_lsn = self
+                .layer_map
+                .next_open_layer_at
+                .context("No next open layer found")?;
+
+            trace!(
+                "creating in-memory layer at {}/{} for record at {}",
+                timeline_id,
+                start_lsn,
+                lsn
+            );
+
+            let new_layer = InMemoryLayer::create(conf, timeline_id, tenant_id, start_lsn)?;
+            let layer = Arc::new(new_layer);
+
+            self.layer_map.open_layer = Some(layer.clone());
+            self.layer_map.next_open_layer_at = None;
+
+            layer
+        };
+
+        Ok(layer)
+    }
+
+    /// Called from `freeze_inmem_layer`, returns true if successfully frozen.
+    pub fn try_freeze_in_memory_layer(
+        &mut self,
+        Lsn(last_record_lsn): Lsn,
+        last_freeze_at: &AtomicLsn,
+    ) {
+        let end_lsn = Lsn(last_record_lsn + 1);
+
+        if let Some(open_layer) = &self.layer_map.open_layer {
+            let open_layer_rc = Arc::clone(open_layer);
+            // Does this layer need freezing?
+            open_layer.freeze(end_lsn);
+
+            // The layer is no longer open, update the layer map to reflect this.
+            // We will replace it with on-disk historics below.
+            self.layer_map.frozen_layers.push_back(open_layer_rc);
+            self.layer_map.open_layer = None;
+            self.layer_map.next_open_layer_at = Some(end_lsn);
+            last_freeze_at.store(end_lsn);
+        }
+    }
+
+    /// Add image layers to the layer map, called from `create_image_layers`.
+    pub fn track_new_image_layers(&mut self, image_layers: Vec<ImageLayer>) {
+        let mut updates = self.layer_map.batch_update();
+        for layer in image_layers {
+            Self::insert_historic_layer(Arc::new(layer), &mut updates, &mut self.layer_fmgr);
+        }
+        updates.flush();
+    }
+
+    /// Flush a frozen layer and add the written delta layer to the layer map.
+    pub fn finish_flush_l0_layer(
+        &mut self,
+        delta_layer: Option<DeltaLayer>,
+        frozen_layer_for_check: &Arc<InMemoryLayer>,
+    ) {
+        let l = self.layer_map.frozen_layers.pop_front();
+        let mut updates = self.layer_map.batch_update();
+
+        // Only one thread may call this function at a time (for this
+        // timeline). If two threads tried to flush the same frozen
+        // layer to disk at the same time, that would not work.
+        assert!(compare_arced_layers(&l.unwrap(), frozen_layer_for_check));
+
+        if let Some(delta_layer) = delta_layer {
+            Self::insert_historic_layer(Arc::new(delta_layer), &mut updates, &mut self.layer_fmgr);
+        }
+        updates.flush();
+    }
+
+    /// Called when compaction is completed.
+    pub fn finish_compact_l0(
+        &mut self,
+        layer_removal_cs: Arc<tokio::sync::OwnedMutexGuard<()>>,
+        compact_from: Vec<Arc<dyn PersistentLayer>>,
+        compact_to: Vec<Arc<dyn PersistentLayer>>,
+        metrics: &TimelineMetrics,
+    ) -> Result<()> {
+        let mut updates = self.layer_map.batch_update();
+        for l in compact_to {
+            Self::insert_historic_layer(l, &mut updates, &mut self.layer_fmgr);
+        }
+        for l in compact_from {
+            // NB: the layer file identified by descriptor `l` is guaranteed to be present
+            // in the LayerFileManager because compaction kept holding `layer_removal_cs` the entire
+            // time, even though we dropped `Timeline::layers` inbetween.
+            Self::delete_historic_layer(
+                layer_removal_cs.clone(),
+                l,
+                &mut updates,
+                metrics,
+                &mut self.layer_fmgr,
+            )?;
+        }
+        updates.flush();
+        Ok(())
+    }
+
+    /// Called when garbage collect the timeline. Returns a guard that will apply the updates to the layer map.
+    pub fn finish_gc_timeline(
+        &mut self,
+        layer_removal_cs: Arc<tokio::sync::OwnedMutexGuard<()>>,
+        gc_layers: Vec<Arc<dyn PersistentLayer>>,
+        metrics: &TimelineMetrics,
+    ) -> Result<ApplyGcResultGuard> {
+        let mut updates = self.layer_map.batch_update();
+        for doomed_layer in gc_layers {
+            Self::delete_historic_layer(
+                layer_removal_cs.clone(),
+                doomed_layer,
+                &mut updates,
+                metrics,
+                &mut self.layer_fmgr,
+            )?; // FIXME: schedule succeeded deletions in timeline.rs `gc_timeline` instead of in batch?
+        }
+        Ok(ApplyGcResultGuard(updates))
+    }
+
+    /// Helper function to insert a layer into the layer map and file manager.
+    fn insert_historic_layer(
+        layer: Arc<dyn PersistentLayer>,
+        updates: &mut BatchedUpdates<'_>,
+        mapping: &mut LayerFileManager,
+    ) {
+        updates.insert_historic(layer.layer_desc().clone());
+        mapping.insert(layer);
+    }
+
+    /// Helper function to remove a layer into the layer map and file manager
+    fn remove_historic_layer(
+        layer: Arc<dyn PersistentLayer>,
+        updates: &mut BatchedUpdates<'_>,
+        mapping: &mut LayerFileManager,
+    ) {
+        updates.remove_historic(layer.layer_desc().clone());
+        mapping.remove(layer);
+    }
+
+    /// Removes the layer from local FS (if present) and from memory.
+    /// Remote storage is not affected by this operation.
+    fn delete_historic_layer(
+        // we cannot remove layers otherwise, since gc and compaction will race
+        _layer_removal_cs: Arc<tokio::sync::OwnedMutexGuard<()>>,
+        layer: Arc<dyn PersistentLayer>,
+        updates: &mut BatchedUpdates<'_>,
+        metrics: &TimelineMetrics,
+        mapping: &mut LayerFileManager,
+    ) -> anyhow::Result<()> {
+        if !layer.is_remote_layer() {
+            layer.delete_resident_layer_file()?;
+            let layer_file_size = layer.file_size();
+            metrics.resident_physical_size_gauge.sub(layer_file_size);
+        }
+
+        // TODO Removing from the bottom of the layer map is expensive.
+        //      Maybe instead discard all layer map historic versions that
+        //      won't be needed for page reconstruction for this timeline,
+        //      and mark what we can't delete yet as deleted from the layer
+        //      map index without actually rebuilding the index.
+        updates.remove_historic(layer.layer_desc().clone());
+        mapping.remove(layer);
+
+        Ok(())
+    }
+
+    pub(crate) fn contains(&self, layer: &Arc<dyn PersistentLayer>) -> bool {
+        self.layer_fmgr.contains(layer)
+    }
+}
+
+pub struct LayerFileManager<T: AsLayerDesc + ?Sized = dyn PersistentLayer>(
+    HashMap<PersistentLayerKey, Arc<T>>,
+);
+
+impl<T: AsLayerDesc + ?Sized> LayerFileManager<T> {
+    fn get_from_desc(&self, desc: &PersistentLayerDesc) -> Arc<T> {
+        // The assumption for the `expect()` is that all code maintains the following invariant:
+        // A layer's descriptor is present in the LayerMap => the LayerFileManager contains a layer for the descriptor.
+        self.0
+            .get(&desc.key())
+            .with_context(|| format!("get layer from desc: {}", desc.filename()))
+            .expect("not found")
+            .clone()
+    }
+
+    pub(crate) fn insert(&mut self, layer: Arc<T>) {
+        let present = self.0.insert(layer.layer_desc().key(), layer.clone());
+        if present.is_some() && cfg!(debug_assertions) {
+            panic!("overwriting a layer: {:?}", layer.layer_desc())
+        }
+    }
+
+    pub(crate) fn contains(&self, layer: &Arc<T>) -> bool {
+        self.0.contains_key(&layer.layer_desc().key())
+    }
+
+    pub(crate) fn new() -> Self {
+        Self(HashMap::new())
+    }
+
+    pub(crate) fn remove(&mut self, layer: Arc<T>) {
+        let present = self.0.remove(&layer.layer_desc().key());
+        if present.is_none() && cfg!(debug_assertions) {
+            panic!(
+                "removing layer that is not present in layer mapping: {:?}",
+                layer.layer_desc()
+            )
+        }
+    }
+
+    pub(crate) fn replace_and_verify(&mut self, expected: Arc<T>, new: Arc<T>) -> Result<()> {
+        let key = expected.layer_desc().key();
+        let other = new.layer_desc().key();
+
+        let expected_l0 = LayerMap::is_l0(expected.layer_desc());
+        let new_l0 = LayerMap::is_l0(new.layer_desc());
+
+        fail::fail_point!("layermap-replace-notfound", |_| anyhow::bail!(
+            "layermap-replace-notfound"
+        ));
+
+        anyhow::ensure!(
+            key == other,
+            "expected and new layer have different keys: {key:?} != {other:?}"
+        );
+
+        anyhow::ensure!(
+            expected_l0 == new_l0,
+            "one layer is l0 while the other is not: {expected_l0} != {new_l0}"
+        );
+
+        if let Some(layer) = self.0.get_mut(&key) {
+            anyhow::ensure!(
+                compare_arced_layers(&expected, layer),
+                "another layer was found instead of expected, expected={expected:?}, new={new:?}",
+                expected = Arc::as_ptr(&expected),
+                new = Arc::as_ptr(layer),
+            );
+            *layer = new;
+            Ok(())
+        } else {
+            anyhow::bail!("layer was not found");
+        }
+    }
+}
--- a/pageserver/src/tenant/timeline/span.rs
+++ b/pageserver/src/tenant/timeline/span.rs
@@ -7,19 +7,14 @@ pub(crate) fn debug_assert_current_span_has_tenant_and_timeline_id() {}
 #[cfg(debug_assertions)]
 #[track_caller]
 pub(crate) fn debug_assert_current_span_has_tenant_and_timeline_id() {
-    static TIMELINE_ID_EXTRACTOR: once_cell::sync::Lazy<MultiNameExtractor<2>> =
-        once_cell::sync::Lazy::new(|| {
-            MultiNameExtractor::new("TimelineId", ["timeline_id", "timeline"])
-        });
+    static TIMELINE_ID_EXTRACTOR: once_cell::sync::Lazy<MultiNameExtractor<1>> =
+        once_cell::sync::Lazy::new(|| MultiNameExtractor::new("TimelineId", ["timeline_id"]));

    let fields: [&dyn Extractor; 2] = [
        &*crate::tenant::span::TENANT_ID_EXTRACTOR,
        &*TIMELINE_ID_EXTRACTOR,
    ];
-    if let Err(missing) = check_fields_present(fields) {
-        panic!(
-            "missing extractors: {:?}",
-            missing.into_iter().map(|e| e.name()).collect::<Vec<_>>()
-        )
+    if let Err(missing) = check_fields_present!(fields) {
+        panic!("missing extractors: {missing:?}")
    }
 }
--- a/pageserver/src/tenant/timeline/uninit.rs
+++ b/pageserver/src/tenant/timeline/uninit.rs
@@ -132,7 +132,7 @@ impl<'t> UninitializedTimeline<'t> {
 impl Drop for UninitializedTimeline<'_> {
    fn drop(&mut self) {
        if let Some((_, uninit_mark)) = self.raw_timeline.take() {
-            let _entered = info_span!("drop_uninitialized_timeline", tenant = %self.owning_tenant.tenant_id, timeline = %self.timeline_id).entered();
+            let _entered = info_span!("drop_uninitialized_timeline", tenant_id = %self.owning_tenant.tenant_id, timeline_id = %self.timeline_id).entered();
            error!("Timeline got dropped without initializing, cleaning its files");
            cleanup_timeline_directory(uninit_mark);
        }
--- a/pageserver/src/tenant/timeline/walreceiver/connection_manager.rs
+++ b/pageserver/src/tenant/timeline/walreceiver/connection_manager.rs
@@ -6,7 +6,7 @@
 //! Current connection state is tracked too, to ensure it's not getting stale.
 //!
 //! After every connection or storage broker update fetched, the state gets updated correspondingly and rechecked for the new conneciton leader,
-//! then a [re]connection happens, if necessary.
+//! then a (re)connection happens, if necessary.
 //! Only WAL streaming task expects to be finished, other loops (storage broker, connection management) never exit unless cancelled explicitly via the dedicated channel.

 use std::{collections::HashMap, num::NonZeroU64, ops::ControlFlow, sync::Arc, time::Duration};
@@ -266,7 +266,7 @@ pub struct ConnectionManagerStatus {
 impl ConnectionManagerStatus {
    /// Generates a string, describing current connection status in a form, suitable for logging.
    pub fn to_human_readable_string(&self) -> String {
-        let mut resulting_string = "WalReceiver status".to_string();
+        let mut resulting_string = String::new();
        match &self.existing_connection {
            Some(connection) => {
                if connection.has_processed_wal {
--- a/pageserver/src/virtual_file.rs
+++ b/pageserver/src/virtual_file.rs
@@ -149,12 +149,10 @@ impl OpenFiles {
        // old file.
        //
        if let Some(old_file) = slot_guard.file.take() {
-            // We do not have information about tenant_id/timeline_id of evicted file.
-            // It is possible to store path together with file or use filepath crate,
-            // but as far as close() is not expected to be fast, it is not so critical to gather
-            // precise per-tenant statistic here.
+            // the normal path of dropping VirtualFile uses "close", use "close-by-replace" here to
+            // distinguish the two.
            STORAGE_IO_TIME
-                .with_label_values(&["close", "-", "-"])
+                .with_label_values(&["close-by-replace"])
                .observe_closure_duration(|| drop(old_file));
        }

@@ -208,7 +206,7 @@ impl VirtualFile {
        }
        let (handle, mut slot_guard) = get_open_files().find_victim_slot();
        let file = STORAGE_IO_TIME
-            .with_label_values(&["open", &tenant_id, &timeline_id])
+            .with_label_values(&["open"])
            .observe_closure_duration(|| open_options.open(path))?;

        // Strip all options other than read and write.
@@ -271,7 +269,7 @@ impl VirtualFile {
                            // Found a cached file descriptor.
                            slot.recently_used.store(true, Ordering::Relaxed);
                            return Ok(STORAGE_IO_TIME
-                                .with_label_values(&[op, &self.tenant_id, &self.timeline_id])
+                                .with_label_values(&[op])
                                .observe_closure_duration(|| func(file)));
                        }
                    }
@@ -298,12 +296,12 @@ impl VirtualFile {

        // Open the physical file
        let file = STORAGE_IO_TIME
-            .with_label_values(&["open", &self.tenant_id, &self.timeline_id])
+            .with_label_values(&["open"])
            .observe_closure_duration(|| self.open_options.open(&self.path))?;

        // Perform the requested operation on it
        let result = STORAGE_IO_TIME
-            .with_label_values(&[op, &self.tenant_id, &self.timeline_id])
+            .with_label_values(&[op])
            .observe_closure_duration(|| func(&file));

        // Store the File in the slot and update the handle in the VirtualFile
@@ -333,13 +331,11 @@ impl Drop for VirtualFile {
        let mut slot_guard = slot.inner.write().unwrap();
        if slot_guard.tag == handle.tag {
            slot.recently_used.store(false, Ordering::Relaxed);
-            // Unlike files evicted by replacement algorithm, here
-            // we group close time by tenant_id/timeline_id.
-            // At allows to compare number/time of "normal" file closes
-            // with file eviction.
+            // there is also operation "close-by-replace" for closes done on eviction for
+            // comparison.
            STORAGE_IO_TIME
-                .with_label_values(&["close", &self.tenant_id, &self.timeline_id])
-                .observe_closure_duration(|| slot_guard.file.take());
+                .with_label_values(&["close"])
+                .observe_closure_duration(|| drop(slot_guard.file.take()));
        }
    }
 }
--- a/pageserver/src/walredo.rs
+++ b/pageserver/src/walredo.rs
@@ -175,8 +175,8 @@ impl WalRedoManager for PostgresRedoManager {
        let mut img = base_img.map(|p| p.1);
        let mut batch_neon = can_apply_in_neon(&records[0].1);
        let mut batch_start = 0;
-        for i in 1..records.len() {
-            let rec_neon = can_apply_in_neon(&records[i].1);
+        for (i, record) in records.iter().enumerate().skip(1) {
+            let rec_neon = can_apply_in_neon(&record.1);

            if rec_neon != batch_neon {
                let result = if batch_neon {
@@ -685,7 +685,7 @@ impl PostgresRedoManager {
            // as close-on-exec by default, but that's not enough, since we use
            // libraries that directly call libc open without setting that flag.
            .close_fds()
-            .spawn_no_leak_child()
+            .spawn_no_leak_child(self.tenant_id)
            .map_err(|e| {
                Error::new(
                    e.kind(),
@@ -989,6 +989,7 @@ impl PostgresRedoManager {
 /// Wrapper type around `std::process::Child` which guarantees that the child
 /// will be killed and waited-for by this process before being dropped.
 struct NoLeakChild {
+    tenant_id: TenantId,
    child: Option<Child>,
 }

@@ -1007,9 +1008,12 @@ impl DerefMut for NoLeakChild {
 }

 impl NoLeakChild {
-    fn spawn(command: &mut Command) -> io::Result<Self> {
+    fn spawn(tenant_id: TenantId, command: &mut Command) -> io::Result<Self> {
        let child = command.spawn()?;
-        Ok(NoLeakChild { child: Some(child) })
+        Ok(NoLeakChild {
+            tenant_id,
+            child: Some(child),
+        })
    }

    fn kill_and_wait(mut self) {
@@ -1056,11 +1060,16 @@ impl Drop for NoLeakChild {
            Some(child) => child,
            None => return,
        };
+        let tenant_id = self.tenant_id;
        // Offload the kill+wait of the child process into the background.
        // If someone stops the runtime, we'll leak the child process.
        // We can ignore that case because we only stop the runtime on pageserver exit.
        BACKGROUND_RUNTIME.spawn(async move {
            tokio::task::spawn_blocking(move || {
+                // Intentionally don't inherit the tracing context from whoever is dropping us.
+                // This thread here is going to outlive of our dropper.
+                let span = tracing::info_span!("walredo", %tenant_id);
+                let _entered = span.enter();
                Self::kill_and_wait_impl(child);
            })
            .await
@@ -1069,12 +1078,12 @@ impl Drop for NoLeakChild {
 }

 trait NoLeakChildCommandExt {
-    fn spawn_no_leak_child(&mut self) -> io::Result<NoLeakChild>;
+    fn spawn_no_leak_child(&mut self, tenant_id: TenantId) -> io::Result<NoLeakChild>;
 }

 impl NoLeakChildCommandExt for Command {
-    fn spawn_no_leak_child(&mut self) -> io::Result<NoLeakChild> {
-        NoLeakChild::spawn(self)
+    fn spawn_no_leak_child(&mut self, tenant_id: TenantId) -> io::Result<NoLeakChild> {
+        NoLeakChild::spawn(tenant_id, self)
    }
 }

--- a/pgxn/hnsw/hnsw.control
+++ b/pgxn/hnsw/hnsw.control
@@ -1,4 +1,4 @@
-comment = 'hnsw index'
+comment = '** Deprecated ** Please use pg_embedding instead'
 default_version = '0.1.0'
 module_pathname = '$libdir/hnsw'
 relocatable = true
--- a/Show More
+++ b/Show More